From detection to diagnosis: Discursive preclusion and the infrastructural politics of visibility on Chinese platforms

Abstract

This study investigates how algorithmic infrastructures on Chinese social media govern visibility not through post-hoc deletion but through pre-emptive design. Focusing on Xiaohongshu during the 2025 China—U.S. trade conflict, it constructs a three-part comparative corpus—over 30,000 political comments from Xiaohongshu, 5000 non-political comments from the same platform, and 1200 Weibo comments on the same issue. A multi-stage diagnostic framework integrates weakly supervised anomaly detection and clustering, behavioral—semantic network analysis, topic modeling (Latent Dirichlet Allocation, LDA), and sentiment analysis to trace how expressive diversity is filtered across platforms. The results reveal no evidence of coordinated automation or bot activity but demonstrate a pronounced convergence in tone and theme: Xiaohongshu's political discourse is affectively positive, rhetorically moderate, and thematically compressed. In contrast, its non-political baseline displays greater affective heterogeneity, while Weibo's corpus shows a wider distribution of topics and discursive framings despite a similarly positive sentiment bias. These contrasts indicate that discursive alignment on Xiaohongshu arises not from manipulation or censorship, but from infrastructural filtration—ranking algorithms, participation thresholds, and affective heuristics that quietly define what can surface. The study conceptualizes this anticipatory logic as discursive preclusion, a form of algorithmic governance that renders dissent statistically improbable rather than overtly suppressed. Methodologically, it advances a “detection-to-diagnosis” approach that interprets silence and convergence as evidence of design-based control; conceptually, it reframes platform power as the governance of legibility and affect rather than of speech itself.

Keywords

Discursive preclusion algorithmic governance platform visibility infrastructural power affective convergence Chinese social media

Introduction

What if the most powerful operations of media governance today lie not in what is deleted, but in what is never made visible at all? In contemporary platform societies, discursive control increasingly takes the form of preclusion—infrastructural mechanisms that shape the very conditions of visibility before speech emerges (Lomborg et al., 2024). Platform architectures—ranking algorithms, interface constraints, and engagement heuristics—do not simply govern what users say; they determine what becomes sayable, surfaceable, and publicly salient (Van Dalen, 2023; Papaevangelou and Siapera, 2025). This anticipatory mode of control, often described as algorithmic gatekeeping (Barzilai-Nahon, 2008) or infrastructural power (Plantin et al., 2018), shifts the focus from managing messages to structuring expressibility itself (Gillespie, 2018; Bucher, 2018; Ananny and Crawford, 2018).

One of the most prominent concerns in this landscape has been the role of social bots—automated accounts deployed to simulate engagement, amplify preferred narratives, and manufacture consensus at scale (Ferrara et al., 2016; Woolley and Howard, 2016). Studies of electoral interference, protest discourse, and geopolitical propaganda have made bots central to discussions of computational manipulation and discursive engineering (Howard et al., 2018a; Freelon et al., 2018; Prier, 2020). But this emphasis on visible distortion invites a broader question: what if the absence of such signals is itself a form of control? Especially in tightly curated environments, the seeming lack of interference may reflect not neutrality, but a different regime of power—one that governs expression by rendering deviance statistically unlikely from the outset.

This study investigates one such case: user discourse on Xiaohongshu (Rednote), a Chinese lifestyle and commerce platform, during the 2025 China—U.S. trade conflict. To assess whether the observed discursive convergence reflects manipulation or organic consensus, we construct a three-part comparative dataset: (1) over 30,000 comments on Xiaohongshu posts related to the trade conflict, (2) a baseline sample of approximately 5000 comments from non-political Xiaohongshu content during the same period, and (3) a parallel sample of around 1200 comments from Weibo posts on the same political issue. We employ a multi-stage diagnostic pipeline—combining weakly supervised bot detection, multimodal comment clustering, sentiment-topic modeling, and semantic network comparison—to evaluate cross-platform variation in engagement patterns, emotional profiles, and coordination signatures.

Our findings reveal no coherent traces of bot-like coordination in Xiaohongshu's political corpus. Instead, we observe sentiment convergence, topic uniformity, and rhetorical moderation—patterns absent in non-political content and diverging from Weibo's framing. While we cannot rule out the possibility that such convergence partly reflects a genuinely narrow range of user sentiments, taken together these cross-platform and cross-context contrasts make a purely organic explanation insufficient and point toward an infrastructural shaping of visibility. In this sense, rather than relying on artificial amplification, Xiaohongshu is more plausibly governed through forms of pre-emptive visibility filtering—via participation thresholds, engagement metrics, and content-sorting logics that quietly condition what can circulate at scale. We treat this interpretation as diagnostic rather than determinative: discursive preclusion names the patterned co-occurrence of high engagement, weak coordination, and semantic-affective compression, rather than a directly observed filtering mechanism. By treating silence not as a null result but as an empirical outcome, this study reframes debates on algorithmic governance away from who manipulates discourse and toward how platforms engineer the improbability of dissent.

Literature review and theoretical framework

From propaganda to platform governance

Early accounts of digital manipulation centered on computational propaganda and the detection of automated agents. Research on social bots documented how automated or semi-automated accounts could simulate engagement, amplify preferred frames, and manufacture consensus at scale (Ferrara et al., 2016; Woolley and Howard, 2016; Varol et al., 2017). Empirical studies connected these practices to electoral interference, protest communication, and geopolitical influence operations (Howard et al., 2018a; Freelon et al., 2018; Prier, 2020). This body of work established an influential analytic template: manipulation is indexed by the presence of coordinated inauthentic behavior, and methodological progress is measured by ever more sensitive techniques for identifying it.

As platforms matured, however, scholars began to argue that this presence-based lens could not fully account for how influence operates in platformized media systems (Helmond and Van der Vlist, 2019; Napoli, 2019b). Classic models of gatekeeping and ideological filtering—long used to explain editorial power in mass media (Shoemaker and Vos, 2009; McQuail, 1992; Herman and Chomsky, 2021)—were reinterpreted for an algorithmic environment in which selection is dispersed across ranking functions, recommendation pipelines, and feedback signals (Napoli, 2019a; Van Dalen, 2023). Rather than discrete editorial decisions, control increasingly inheres in infrastructures that condition circulation in advance: computational processes and interface architectures that mediate not only what becomes newsworthy, but what can surface as legible expression at all (Gillespie, 2018; Bucher, 2018; Striphas, 2015).

This shift suggests a broader reorientation—from a focus on overt interventions (e.g. bot amplification, post-hoc removal) (Wright, 2022; Pedrazzi and Oehmer-Pedrazzi, 2023) to the analysis of infrastructural arrangements that structure the field of possible utterances before manipulation would even be necessary (Cohen, 2023). In other words, platform governance is not reducible to who speaks or which messages are deleted; it encompasses the technical and organizational conditions that make some expressions highly probable and others practically unobservable. The implication for research design is a parallel shift: from detection toward diagnosis—away from locating manipulators and toward mapping how infrastructures allocate visibility, legibility, and traction within the attention economy. This reorientation sets the stage for examining filtration logics and the politics of visibility, situating our later formulation of discursive preclusion as a specifically infrastructural mode of control.

Infrastructural filtration and the politics of visibility

Within-platform studies, the question of control has shifted from message management to infrastructural mediation. Algorithms, interface design, and engagement metrics collectively configure how content circulates and, crucially, what never enters circulation at all. As Gillespie (2018) and Bucher (2018) argue, these architectures enact a new form of custodianship of visibility, where filtering mechanisms determine not only what gains prominence but what remains unseen. Platforms govern through design: the logic of ranking, recommendation, and feedback loops substitutes for editorial selection, embedding normative assumptions into the infrastructure of participation.

This infrastructural turn reframes visibility as a stratified and actively managed resource. Rather than a binary of deletion or promotion, content is algorithmically sorted into gradations of perceptibility and neglect. Algorithmic systems enact forms of cultural judgment by deciding which information attains epistemic status and which recedes into background noise, making filtration an epistemological as well as technical process (Hyzen et al., 2025; Gillespie, 2020). On social media platforms, such logic materializes through ranking functions, personalized feeds, and engagement heuristics that privilege emotional coherence while suppressing discordant expression (DeVito, 2017; Napoli, 2019a; Zeng and Kaye, 2022). Visibility therefore depends not on equal access but on compatibility with the platform's normative expectations—a condition of legibility rather than a universal right.

These stratified and algorithmically mediated visibility conditions are not merely technical arrangements; they are expressions of infrastructural power. Rather than issuing commands or enforcing explicit prohibitions, infrastructural power works by configuring the material and algorithmic conditions under which action, meaning, and visibility become possible in the first place (Mann, 1984; Plantin et al., 2018; Van Dijck et al., 2018). When platforms define how content is ranked, how audiences are assembled, and how engagement is measured, infrastructural power is exercised by shaping social outcomes without appearing to intervene in them directly. Because these operations are embedded in technical systems and routine interactions, their effects are often experienced as natural or neutral, even as they systematically advantage certain forms of expression over others. In this sense, infrastructural power governs not by dictating what users must say, but by calibrating the conditions under which some voices become easier to articulate, circulate, and accumulate visibility than others.

These dynamics constitute a politics of visibility rooted in omission as much as promotion (Roberts, 2019). What a platform suppresses or fails to render legible may shape public discourse as powerfully as what it amplifies. Moreover, infrastructural governance often unfolds opaquely: filtering decisions tend to precede and exceed moderation, producing an anticipatory mode of control in which power operates through invisibility rather than enforcement (Ananny and Crawford, 2018). Users adapt to these hidden hierarchies, internalizing them through self-moderation and strategic self-presentation (Duffy and Meisner, 2023; Hallinan et al., 2025). In this sense, platform power lies less in overt censorship than in the silent calibration of expressibility—the determination of which voices, affects, and perspectives are structurally enabled to appear.

Infrastructural governance in the Chinese context

Research on Chinese digital platforms has highlighted the structural integration of regulatory imperatives into everyday infrastructural design (Xu, 2024; Wang, 2024). Rather than operating through episodic or post-hoc moderation, many systems embed control into the very processes that determine whether content becomes visible at all (Liu and Yang, 2022). This sequencing of automated and human review has been characterized as a layered form of governance, where the moment of visibility itself becomes an object of regulation (Chen and Shi, 2025). Such governance is neither purely technical nor purely administrative—it unfolds through the coordination of infrastructures, institutional mandates, and user expectations that together delimit the boundaries of expressibility (Li and Zhou, 2024).

This embeddedness of control has been theorized as part of a broader regime of infrastructural sovereignty, in which political and commercial logics are translated into algorithmic operations (Bratton, 2016; Liang and Li, 2025). Under this regime, platforms integrate compliance mechanisms into ranking systems, recommendation models, and interface cues that subtly shape how participation occurs (Ye et al., 2025; Zheng, 2024). These systems do not simply restrict expression ex post; they define, in advance, the range of affective and thematic registers that can circulate. Affective coherence and normative alignment are not merely rewarded but infrastructurally incentivized, forming part of the predictive models through which content gains traction (Zhang, 2025).

Within this framework, control operates less through direct intervention than through anticipatory calibration (Bucher et al., 2021). Visibility is managed by design: through thresholds, filters, and feedback systems that govern expression at its point of emergence (Duffy and Meisner, 2023). What is absent in such environments is not beyond regulation but the result of infrastructural selection, where some possibilities never acquire the conditions necessary for appearance. This anticipatory mode of governance clarifies how discursive uniformity can emerge even without overt censorship, setting the theoretical foundation for understanding discursive preclusion as an infrastructural logic of the sayable.

Defining discursive preclusion

Building on the preceding discussion of algorithmic filtration and infrastructural governance, this study introduces the concept of discursive preclusion to describe a mode of control that operates before visibility takes shape. Whereas censorship and moderation intervene after discourse emerges—by deleting, demoting, or demonetizing nonconforming content—preclusion functions prior to expression, delimiting what can materialize within the platform's communicative space (Shang, 2025). It marks a shift from the politics of removal to the politics of possibility: the orchestration of silence through architecture rather than authority.

Discursive preclusion differs from related notions such as shadow banning and algorithmic gatekeeping. Shadow banning restricts circulation post factum, functioning as a reactive sanction within an existing expressive field (Myers West, 2018; Hojati and Nault, 2025). Gatekeeping, whether editorial or algorithmic, presumes a discrete decision about what passes through a predefined threshold of publication (Shoemaker and Vos, 2009; Napoli, 2019b; DeVito, 2017). By contrast, preclusion governs at a deeper infrastructural level—it encodes normative assumptions into ranking algorithms, content classifiers, and engagement thresholds that collectively determine what can appear as legible discourse. What remains unseen is not merely deprioritized but rendered statistically improbable within the system's design parameters.

Analytically, discursive preclusion operates along a temporal continuum of infrastructural governance rather than within the domain of moderation. At its anticipatory end, filtration mechanisms constrain discourse at the moment of generation—through keyword gating, affective calibration, or interface-level affordances that determine what can become visible. At its retrospective end, more conventional moderation practices retract visibility after content has already circulated. Between these poles lie hybrid regimes where users internalize algorithmic expectations, pre-adjusting their tone, affect, or topic selection in anticipation of what will be rendered legible. In this sense, control unfolds as a temporally staged orchestration of visibility, transforming silence from an incidental absence into a structural condition of participation.

Through this framework, discursive preclusion redefines governance in platform environments. It redirects analytical attention from manipulation and censorship as isolated acts toward the infrastructural arrangements that make certain utterances systematically unlikely. Silence, under this model, is not a vacuum of expression but an engineered outcome. Recognizing preclusion as an anticipatory form of infrastructural power helps explain how ostensibly participatory platforms cultivate affective uniformity and rhetorical moderation through the pre-emptive calibration of visibility itself.

Methodology

Corpus construction and comparative design

This study constructs a three-part comparative dataset designed to capture both the political and non-political dynamics of discourse circulation across Chinese social media. The primary corpus comprises approximately 30,000 comments posted on Xiaohongshu during the 2025 wave of China—U.S. trade tensions. Posts were collected between March and April 2025—a period marked by renewed diplomatic friction, digital sanctions, and intensified online discussion. Keywords such as “贸易战” (trade war), “关税” (tariff), and “中美关系” (China—U.S. relations) guided data retrieval, and all comments were cleaned for duplicates and spam while preserving emojis and expressive markers that carry platform-specific affective meaning.

To contextualize these patterns, two comparative corpora were assembled. The first serves as a baseline sample, containing roughly 5000 comments from non-political Xiaohongshu posts in the same period. This dataset captures the platform's default interactional norms and affective repertoires under conditions of low political salience. The second constitutes a parallel sample of approximately 1200 Weibo comments on the same trade-related issue, providing a cross-platform reference from a system governed by a distinct moderation model. Xiaohongshu was selected as the primary research site because of its hybrid identity as both a lifestyle-oriented community and a semi-regulated social platform. This dual nature makes it analytically valuable for examining how commercial recommendation systems and regulatory constraints intersect in shaping discursive visibility.

To contextualize these patterns, two comparative corpora were assembled. The first serves as a baseline sample, containing roughly 5000 comments from non-political Xiaohongshu posts in the same period. This dataset captures the platform's default interactional norms and affective repertoires under conditions of low political salience. The second constitutes a parallel sample of approximately 1200 Weibo comments on the same trade-related issue, providing a cross-platform reference from a system governed by a distinct moderation model. The sizes of these comparative corpora are intentionally smaller than the main Xiaohongshu political dataset, as their purpose is not statistical representativeness but structural benchmarking—establishing reference distributions of topical and affective variation against which patterns in the high-salience corpus can be interpreted. Xiaohongshu was selected as the primary research site because of its hybrid identity as both a lifestyle-oriented community and a semi-regulated social platform. This dual nature makes it analytically valuable for examining how commercial recommendation systems and regulatory constraints intersect in shaping discursive visibility.

This three-part design enables both within-platform and cross-platform comparison. Within Xiaohongshu, contrasting political and non-political discourse reveals how topic sensitivity interacts with platform affordances to shape expressive behavior. Across platforms, juxtaposing Xiaohongshu and Weibo highlights how differing infrastructural regimes—pre-emptive versus reactive moderation—produce distinct patterns of emotional alignment and thematic diversity. Together, these datasets form the empirical foundation for diagnosing how visibility is structured through infrastructural rather than agentic mechanisms.

All data analyzed in this study were publicly accessible at the time of collection. User identifiers, profile links, and metadata unrelated to textual content were removed prior to analysis to ensure anonymity. Following established ethical guidelines for computational social research, the study treats social media discourse as public communication while remaining attentive to users’ expectations of contextual integrity. No attempt was made to trace or infer the identity of individual users, and all examples used in the analysis were paraphrased or aggregated to minimize re-identification risks.

Analytical framework: From detection to diagnosis

The analytical strategy proceeds from detection to diagnosis—not to identify coordinated manipulation in a narrow sense, but to examine how discursive uniformity emerges within infrastructural conditions of visibility. Instead of treating computational tools as ends in themselves, the framework mobilizes them heuristically to infer the relational dynamics of participation, alignment, and filtration. The analysis unfolds across three interrelated levels—individual, group, and structural—each corresponding to a different dimension of visibility production.

At the individual level, behavioral and linguistic indicators were assessed to identify outlier comments that deviate from typical interactional norms. Nine heuristic features—spanning structural/behavioral (comment length, exact-duplicate count, multi-reply count, URL presence, default-username pattern), linguistic (type—token ratio, punctuation richness, question markers, emotion words), and lexical signals (ad- related phrases)—served as labeling functions within a weakly supervised model (Snorkel) (Ratner et al., 2020; Yang et al., 2020). The LabelModel (Bach et al., 2017) produced probabilistic anomaly scores that yielded an initial triage (Normal, Moderately Suspicious, Highly Suspicious) using quantile-based cutoffs calibrated on the empirical score distribution. These anomalies are read not as proof of automation but as micro-level entry points into how platform affordances condition legible expression.

At the group level, clustering was applied only to the Moderately Suspicious subset to refine this initial triage. Seven standardized behavioral—linguistic features (e.g. length, likes, duplication, multi-reply, ad-flag, TTR, punctuation ratio) served as input for KMeans clustering; the optimal number of clusters was identified empirically through silhouette analysis (McQueen, 1967; Rousseeuw, 1987). For each cluster, mean anomaly scores were computed and relabeled by a quartile rule (Aggarwal, 2017): clusters in the top quartile were upgraded to Highly Suspicious, those in the bottom quartile downgraded to Normal, and the remainder stayed Moderately Suspicious. This produces a refined label set, distinguishing voluntary imitation from structurally induced alignment by identifying stylistic/affective formations that are consistently extreme versus ordinary.

At the structural level, network modeling was conducted exclusively on the refined Highly Suspicious set (users associated with comments upgraded or remaining at the high tier) to test whether coordination persists beyond individual/group anomalies. We constructed (1) a behavioral co-occurrence graph linking user pairs who repeatedly appeared under the same posts (thresholded at the 90th percentile of non-zero co-occurrences) (Newman, 2003) and (2) a semantic similarity graph linking user pairs whose SentenceTransformer (paraphrase-MiniLM-L6- v2) (Reimers and Gurevych, 2019) mean-embeddings exceeded a 90th-percentile cosine similarity. We then analyzed their intersection graph (edges present in both layers), applying Louvain community detection (on a composite edge weight combining co-occurrence counts with text-similarity strength) (Blondel et al., 2008) to evaluate cohesiveness. Finally, Pearson correlation tests (Pearson, 1895) between node-level behavioral degree and text-similarity degree assessed whether textual convergence coincides with synchronized participation. The target here is not to “catch” covert coordination, but to diagnose whether any residual clustering reflects organized manipulation or an architectural effect of visibility shaping.

Across these three levels, the analysis moves from micro-level participation to macro-level structure, shifting the emphasis from identifying suspicious actors to understanding how visibility itself becomes stratified. In doing so, the study reframes “detection” as a critical diagnostic tool for uncovering the infrastructural logic that governs what can be seen, said, and amplified.

Topic and sentiment modeling

To analyze how expression on the platform is organized semantically and affectively, the study applies Latent Dirichlet Allocation (LDA) topic modeling (Blei et al., 2003) and sentiment analysis (Pang et al., 2008) as complementary diagnostic tools. These methods do not aim to infer user intentions but to expose the structural boundaries of what becomes sayable and emotionally permissible within the platform's communicative field.

The LDA model was implemented using scikit-learn's LatentDirich- letAllocation module (Pedregosa et al., 2011) with a batch learning method and 10 training iterations. Preprocessing followed two stages: first, character-level cleaning removed digits, ASCII punctuation, and Latin letters while retaining emoji placeholders (e.g. [点赞R]) that encode platform-specific affective expressions; second, tokenization was performed using the Jieba segmenter (Sun, 2012). Tokens shorter than two characters or containing no Chinese characters were filtered out, ensuring that only linguistically meaningful terms contributed to topic inference.

A CountVectorizer was applied to construct the term-frequency matrix. To determine the optimal number of topics, coherence and perplexity metrics were computed across multiple model configurations. The highest UMass coherence (Mimno et al., 2011) indicated a compact thematic structure consistent with discursive convergence. The final model extracted four topic clusters, each represented by its top 10 keywords with the highest topic probability. These topics are not treated as discrete themes but as structural indicators of discursive concentration—how topical diversity becomes compressed under specific visibility constraints. In parallel, sentiment analysis was performed using the Chinese-language package SnowNLP (Isnowfy, 2012) to assign a sentiment score between 0 (negative) and 1 (positive) to each comment. Scores were averaged by topic and plotted as histograms and bar charts to capture the emotional bandwidth of the corpus. A narrow sentiment range—particularly when paired with low topical variation—was interpreted as evidence of affective convergence, where expressions cluster around institutionally permissible tones or moods.

Together, topic and sentiment modeling provide a meso-level diagnostic of visibility filtering. By situating linguistic and emotional regularities within the platform's infrastructural logics, this step reveals how discursive uniformity is not only algorithmically permitted but also affectively reinforced—a reflection of governance through both semantic narrowing and emotional calibration.

Baseline datasets and comparative procedures

To ensure interpretive validity, the same diagnostic framework was applied across all three corpora, positioning the non-political Xiaohongshu sample and the Weibo dataset as comparative anchors rather than control groups in a technical sense (Ragin, 2014). The non-political Xiaohongshu corpus functions as a baseline environment (Thelwall et al., 2012), reflecting the platform's ordinary affective and semantic dynamics under low-salience conditions. To construct this baseline, we sampled highly engaged non-political posts from the same time period on Xiaohongshu, focusing on platform-typical domains such as beauty, travel, and everyday lifestyle. Only posts exceeding 1000 comments were included, mirroring the selection criteria used for political posts, so that aside from topical content, other factors such as visibility, engagement intensity, and interactional scale were held as constant as possible. Replicating the same detection, clustering, and modeling procedures allows us to delineate what constitutes ordinary variation in engagement when political sensitivity is absent. Deviations observed in the trade-war corpus can therefore be read as infrastructural effects rather than behavioral noise.

The Weibo corpus, by contrast, serves as a parallel environment—a cross-platform reference that captures a distinct regime of moderation temporality (Plantin et al., 2018). Weibo typically follows a post-first moderation model, in which comments are published before being reviewed, flagged, or removed (King et al., 2017). Because our Weibo sample was collected approximately six months after the original posts were published, the dataset can be assumed to reflect discourse that had already passed through Weibo's post-hoc moderation and removal processes. Accordingly, the Weibo data were analyzed solely through topic and sentiment modeling, excluding bot detection and anomaly scoring stages. This adjustment isolates the effects of temporal sequencing in governance rather than differences in user demographics or topic composition.

The baseline and parallel corpora operationalize a comparative diagnostic logic: they reveal how discursive uniformity may emerge not from coordinated manipulation or collective consensus, but from divergent infrastructural pathways through which visibility is preconditioned, filtered, or retrospectively managed.

Findings: Diagnosing the infrastructural production of absence

Fragmented visibility and the absence of coordination

Across all three diagnostic layers—individual anomalies, group clustering, and structural networks—the analysis of the Xiaohongshu political corpus reveals not the presence of coordinated manipulation, but a landscape of fragmentation. The absence of collective coherence, as shown below, is itself indicative of how visibility is conditioned by infrastructural constraints rather than organized interference.

At the individual level, weak supervision identified anomalous traces within the Xiaohongshu political corpus, amounting to approximately 6.5% of all 30,000 comments. Nine heuristic labeling functions were combined into a probabilistic anomaly score, classifying comments as

Normal: 23,904, Moderately Suspicious: 4,543, Highly Suspicious: 1653.

These preliminary results suggested a limited number of outlier behaviors, which then became the basis for a fine-grained inspection.

To refine these categories, the “Moderately Suspicious” subset (4543 comments) was clustered using KMeans over seven behavioral—linguistic features. Based on average anomaly scores, three clusters (2, 7, 9) were upgraded to “Highly Suspicious,” while two (6, 8) were reclassified as “Normal.” The updated distribution was as follows:

Normal: 26,261, Moderately Suspicious: 1,714, Highly Suspicious: 2125. This reallocation demoted 2357 comments to “Normal” and promoted 472 to “Highly Suspicious.” The separability of clusters, visualized via principal component analysis (Hotelling, 1933), demonstrates that the seven features effectively capture structured heterogeneity rather than random noise (Figure 1).

Figure 1.

PCA-based visualization of clustered suspicious comments. PCA: principal component analysis.

Moving from individual traces to relational patterns, three network diagnostics were applied to the 2125 users associated with highly suspicious comments. The behavioral co-occurrence network (G_cooc) linked users commenting under the same post, thresholded at the 90th percentile (MIN COOC = 1). As Figure 2 illustrates, the resulting graph was extremely sparse: 415 users were isolated, most components contained fewer than 10 users, and the largest connected component—only 90 nodes—represented under 5% of all high-suspicion users. Such dispersion suggests minimal behavioral overlap, inconsistent with coordinated posting.

Figure 2.

Behavioral co-occurrence network of highly suspicious users.

To test whether semantic coordination compensated for behavioral dispersion, a textual similarity network (G_text) was constructed using transformer-based sentence embeddings to measure semantic alignment among users. User pairs exceeding cosine similarity of 0.8772 (90th percentile) were linked. The resulting network, shown in Figure 3, again exhibited fragmentation: 1365 users formed only weak links, with 30.4% (415 users) remaining isolated. The sparse dyads and triads suggest that even the most semantically similar users did not employ templated or repeated language.

Figure 3.

Textual similarity network of highly suspicious users.

When the two networks were intersected (G_both), coordination signatures nearly vanished (Figure 4). Only a few user pairs shared both behavioral and semantic overlap. Louvain community detection (Lambiotte et al., 2008) identified many micro-communities: over 70% of nodes were singletons, and 90% of the remaining clusters contained fewer than 10 users. The largest community had 90 users (less than 5% of all high-suspicion accounts), yet its average text-similarity degree stayed below 0.15. Correlation between users’ behavioral and semantic degrees yielded r = 0.0125, p = 0.6442, indicating no statistically significant relationship between the two dimensions, reflecting a lack of observable coupling at the empirical level. Together, these results demonstrate that anomalous users do not form organized clusters but remain atomized across the communicative field.

Figure 4.

Intersection network combining behavioral and semantic links.

The empirical picture that emerges is one of systemic isolation. The absence of cohesive clusters across all three modalities—behavioral, semantic, and hybrid—suggests that anomalous traces are not the footprint of coordination but the residue of infrastructural fragmentation. Participation appears de-collectivized: visibility is not eliminated through deletion but diffused across disconnected nodes and weakly articulated ties.

Viewed through the lens of discursive preclusion, this fragmentation constitutes an infrastructural outcome rather than a behavioral anomaly. Instead of suppressing individual posts, the platform's architecture statistically neutralizes the formation of density itself—ensuring that affective alignment occurs without social accumulation. The absence of coordination is thus not evidence of openness, but the expression of a controlled dispersal: a mode of governance that manages visibility not by silencing discourse, but by preventing it from cohering.

Semantic narrowing and emotional convergence

If the previous section demonstrated the absence of coordinated collectivity, the semantic and affective layers reveal a parallel pattern of convergence. Rather than a pluralistic debate, Xiaohongshu's political comment field displays a compressed range of meanings and emotions—what may be called semantic narrowing and emotional convergence.

LDA topic modeling identifies only four coherent clusters across the entire 30,100-comment corpus, suggesting a notably restricted thematic landscape. Based on UMass coherence scoring, the optimal number of topics was determined as k = 4, with a peak coherence of approximately 130.6994. Below are the 10 most probable words associated with each topic.

Topic 0: 断层 (disruption), 趋势 (trend), 世永一 (the world's first forever), 绝对 (absolute), 数据 (data), 霸权主义 (hegemonism), 入股 (buy shares), 不亏 (no loss), 同时 (simultaneously), 一辈子 (a lifetime)

Topic 1: 政治(politics), 数据(data), 国内(domestic), 外贸(foreign trade), 鸡蛋(egg), 视频(video), 确实(indeed), 以前(before), 公司 (company), 品牌 (brand)

Topic 2: 美国 (United States), 中国 (China), 关税 (tariff), 人民 (people), 特朗普 (Trump), 贸易战 (trade war), 希望 (hope), 股市 (stock market), 战争 (war), 问题 (issue)

Topic 3: 中国 (China), 美国 (United States), 支持 (support), 退税 (export rebate), 华为 (Huawei), 涨价 (price increase), 关税 (tariff),出口 (export), 手机 (mobile phone), 价格 (price)

These clusters represent four overlapping domains of political discourse: (1) anti-hegemonic narratives and perceptions of U.S. decline (Topic 0), (2) everyday economic life and consumer adjustment (Topic 1), (3) collective framings of trade rivalry and strategy (Topic 2), and (4) transactional concerns such as tax rebates and product pricing (Topic 3). While thematically distinct, their narrow scope—geopolitics, consumption, resilience, and adjustment—marks a drastic reduction in topical diversity, pointing to structural curation rather than spontaneous focus.

Figure 5 visualizes the distribution of topic assignments. Topic 0 dominates, indicating the centrality of anti-hegemonic rhetoric and the repetition of geopolitical tropes. The prevalence of such discourse—rather than pluralized or conflicting framings—suggests that visibility is unevenly distributed toward ideologically compatible expression.

Figure 5.

Distribution of comments assigned to each topic. Topic 0 (anti-hegemonic discourse) is the most prevalent, while Topic 1 (domestic economy) is least represented.

Sentiment analysis corroborates this diagnosis. Using SnowNLP, a Chinese-language classifier that outputs continuous polarity scores between 0 (negative) and 1 (positive), the overall distribution of sentiments shows a distinct bimodal pattern (Figure 6). Peaks occur near 0.5 and 1.0, corresponding to neutral and highly positive tones, while negative affect is statistically rare—an affective configuration consistent with the platform's tendency toward moderated expression and constrained polarity.

Figure 6.

Distribution of sentiment scores across all comments. Peaks at 0.5 and 1.0 indicate dominance of neutral and positive affect.

When disaggregated by topic, this affective uniformity becomes even clearer. As shown in Table 1, all topic-level averages exceed 0.57, with Topic 2—discussions of the trade war—reaching the highest mean of 0.719. Even in the most conflict-oriented discourse, sentiment remains predominantly positive or conciliatory.

Table 1.

Average sentiment scores by topic (SnowNLP).

Topic	0	1	2	3
Mean sentiment	0.575350	0.642754	0.719332	0.653168
Score

Across all topics, the prevailing affect oscillates between calm neutrality and patriotic optimism. Comments invoking trade conflict or tariff pressure frequently reframe loss and competition through motivational idioms—“中国会更强 (China will be stronger)” or “支持国产 (support domestic brands)”—emphasizing resilience and collective composure. Expressions of anger, irony, or despair are statistically marginal, often confined to isolated users rather than circulating broadly.

The overall affective landscape that emerges is one of constrained polarity. Emotional tones cluster around neutrality and mild positivity, with negative affect rarely surfacing. This narrowing of sentiment distribution suggests a platform environment that privileges moderation and filters out extreme or dissenting expressions. Such convergence marks a pronounced reduction in both semantic range and affective bandwidth within the political corpus, setting the stage for the subsequent discussion of how these dynamics are temporally structured by platform governance.

The overall affective landscape that emerges is one of constrained polarity. Emotional tones cluster around neutrality and mild positivity, with negative affect rarely surfacing. While such a narrowing of sentiment distribution could in principle reflect a genuinely homogeneous user mood, its co-occurrence with a highly fragmented and weakly coordinated interaction structure suggests that this convergence is unlikely to be purely organic. In this context, the compression of both affective and semantic range is more consistent with a platform environment that privileges moderation and conditions which expressions can gain visibility, rather than with spontaneous consensus alone. Such convergence marks a pronounced reduction in both semantic range and affective bandwidth within the political corpus, setting the stage for the subsequent discussion of how these dynamics are temporally structured by platform governance.

Discursive preclusion across platform temporalities

The comparative baselines clarify how discursive preclusion functions as an infrastructural mode of control that shapes the conditions of expressibility before content even appears. Whereas retrospective moderation, exemplified by Weibo, tends to operate after content circulation, Xiaohongshu appears to rely on a more pre-emptive structuring of visibility: expressions seem to remain most visible under low-salience conditions, while politically sensitive discourse is likely subject to filtration or downranking before reaching wide exposure. The following comparative analysis illustrates how these temporal architectures manifest empirically.

Anticipatory Preclusion: Xiaohongshu's Latent Governance applying the same anomaly detection and network diagnostic procedures as in the main corpus, no evidence of automated coordination or bot activity was identified in the non-political Xiaohongshu dataset. The intersection network (G_both) in Figure 7 reveals two densely connected clusters alongside multiple smaller ones, each representing organically formed micro-communities. Such dense modularity indicates natural social self-organization: users connect through shared affective interests—fashion, daily life, or consumption—without centralized coordination. The absence of fragmentation contrasts sharply with the main corpus, where similar structures dissolve into sparsity, suggesting that algorithmic filtration suppresses collective formation under high-salience conditions.

Figure 7.

Intersection network (G_both) for the non-political Xiaohongshu corpus. Dense clusters indicate organic social self-organization under low-salience conditions.

Topic modeling of the same corpus (Figure 8) identifies eight major clusters, dominated by lifestyle, esthetics, and consumer practices. Topic 0, which centers on everyday routines and emotional sharing, accounts for approximately 30.7% of all comments, while the remaining topics cover areas ranging from product evaluation to interpersonal empathy. This semantic diversity contrasts with the main corpus, where only four topics emerged and were heavily compressed around geopolitical and nationalist frames. The difference indicates that when ideological salience is absent, Xiaohongshu's infrastructural logic permits topical proliferation, allowing heterogeneous forms of expression to coexist.

Figure 8.

Topic distribution in the non-political Xiaohongshu corpus, showing eight distinct thematic clusters centered on lifestyle and affective exchange.

Sentiment analysis (Figure 9) further supports this openness. The distribution spans the full 0–1 range, with peaks at both extremes, suggesting a balanced mixture of positive enthusiasm and negative critique. Unlike the politically charged dataset, where emotion converged toward mild positivity, here affective heterogeneity persists: irony, frustration, and excitement coexist within a stable expressive ecology. This broad range reflects a dormant rather than absent mechanism of preclusion—an architecture capable of constraint, but inactive under low-risk conditions.

Figure 9.

Sentiment distribution (non-political Xiaohongshu). The wide affective bandwidth reflects latent rather than active preclusion.

In sum, Xiaohongshu's openness under ordinary conditions is conditional rather than inherent. Its expressive pluralism reveals not the absence of control, but its dormancy: a pre-emptive architecture awaiting activation once discursive risk arises. Preclusion thus exists not as censorship-in-action but as a potentiality embedded in infrastructural design—a silent readiness to delimit expression before it begins.

In sum, Xiaohongshu's openness under ordinary conditions appears contingent rather than inherent. In non-political contexts, the platform supports dense social clustering, wide topical dispersion, and broad affective bandwidth—patterns that stand in marked contrast to the fragmentation, semantic compression, and affective convergence observed in the political corpus. While this divergence could partly reflect differences between everyday and political talk, the magnitude and systematicity of the contrast tend to align more closely with expressive pluralism on Xiaohongshu being conditioned by contextual sensitivity rather than guaranteed by default. Preclusion thus appears not as censorship-in-action but as a latent infrastructural capacity—a design logic that can, under conditions of heightened discursive risk, narrow what becomes visible before expression has a chance to circulate.

Retrospective Moderation: Weibo as a Conventional Counterpoint Weibo, by contrast, exemplifies a traditional post-hoc model of control in which content first circulates and is later reviewed, downranked, or removed. Topic modeling (Figure 10) identifies 18 clusters, including policy debate, satire, and collective frustration—discursive forms largely absent from Xiaohongshu. Because our Weibo dataset was collected several months after the original posts, the visible corpus reflects discourse that has already passed through Weibo's retrospective moderation. Yet such post-hoc filtering rarely produces a fully homogenized field: while the most sensitive or overtly contentious content is selectively removed, substantial thematic diversity typically remains. This residual plurality stands in marked contrast to Xiaohongshu's political corpus, where both topical and affective ranges are sharply compressed despite similarly high engagement levels. The comparison suggests a difference in governance temporality: on Weibo, diversity enters first and is then partially pruned, whereas on Xiaohongshu, many forms of divergence appear less likely to surface at all.

Figure 10.

Topic distribution in the Weibo corpus. Eighteen topics reveal greater initial semantic diversity before retrospective moderation.

Sentiment analysis (Figure 11) further corroborates this temporal volatility. The distribution is skewed toward positivity, with a continuous but low-frequency tail across the neutral and mildly negative range. Rather than exhibiting clear peaks of antagonism, the data indicate a narrow affective bandwidth centered on neutral and mildly positive tones. Negativity remains statistically marginal because within a post-hoc regime it is displaced through selective removal and downranking, leaving behind a moderated residue that remains more thematically plural than affectively diverse. This configuration aligns with prior accounts of retrospective moderation or reactive governance (Zhu et al., 2012, 2013; Chen et al., 2023), where expressions of dissent tend to disappear from public view after circulation. This residual plurality stands in contrast to Xiaohongshu's political corpus, where both topical and affective ranges are sharply compressed.

Figure 11.

Sentiment distribution (Weibo). The strong positive skew and attenuated midrange reflect a narrowing of affective variation after initial circulation, exemplifying retrospective moderation.

Discussion

Reconfiguring visibility and governance

The absence of coordinated manipulation revealed in this study should not be read as neutrality. It signals a mode of governance that operates through design rather than intervention. On Xiaohongshu, discourse is shaped before it appears: ranking algorithms, comment filters, and participation thresholds quietly define what can surface and what will sink. Visibility is not a democratic given but an infrastructural privilege, allocated to expressions that are affectively stable and stylistically legible. In such an environment, silence is not an accident—it is architecture.

This pre-emptive form of control contrasts sharply with platforms such as Weibo, where moderation typically unfolds after expression. There, content circulates briefly before being removed or demoted; on Xiaohongshu, it rarely circulates at all if deemed risky. The two models embody different temporal logics of power: one reactive and event-driven, the other anticipatory and procedural. Users on the latter platform encounter a sense of calm not because conflict is absent, but because dissent has been engineered to disperse before it coheres. Consensus, in this sense, is a product of infrastructure, not ideology.

Understanding governance as an infrastructural process reframes what control looks like in contemporary media systems. Platforms do not simply moderate content—they choreograph participation. By designing the conditions of expressibility, they transform censorship from a discrete act into a continuous state of calibration. Users internalize these cues, learning through experience which tones travel and which disappear. Over time, legibility becomes self-regulation: participants adjust to the rhythm of the system long before explicit rules are invoked.

Such orchestration exemplifies a positive form of power—what Foucault (1990) described as the productivity of governance. Power here does not repress speech; it generates the boundaries within which speech can occur. Xiaohongshu's architecture produces predictability and affective coherence not by silencing users, but by scripting visibility itself. What remains unseen is not forbidden but improbable, filtered out by a system that manages expression through design. Infrastructural stability, not open debate, becomes the metric of communicative order.

From bot detection to infrastructural diagnosis

The movement from bot detection to infrastructural diagnosis marks a shift in how control is recognized and evidenced. Traditional studies of computational propaganda locate manipulation in the visible actions of coordinated or automated agents (Woolley and Howard, 2018; Bradshaw and Howard, 2017; Howard et al., 2018b). Within that framework, influence is measured by its traces—fake accounts, synchronized timing, or linguistic uniformity. Yet the empirical absence of such coordination in Xiaohongshu's political discourse reveals the limits of this approach. When power operates through architecture rather than agency, what must be uncovered is not manipulation itself but the conditions that make it unnecessary.

This diagnostic stance redefines both the object and method of inquiry. Detection assumes transparency: it seeks signs of interference that can be verified. Diagnosis, by contrast, accepts opacity as constitutive of platform power. The analytical pipeline used here—anomaly scoring, clustering, and semantic network analysis—was never meant to expose hidden actors; it serves instead to map how infrastructures distribute visibility and isolate participation. The absence of coordination thus becomes an empirical clue. It suggests that fragmentation and uniformity are not opposites but twin outcomes of design—dividing users enough to prevent accumulation while aligning them within an acceptable affective range.

Adopting diagnosis as a research logic also reshapes what counts as data. Silence, uniform tone, and null findings cease to be analytical dead ends; they become evidence of how systems structure expressibility. In environments governed by algorithmic opacity, effects are often the only observable dimension of power. By reading these effects symptomatically—through the narrowing of topics, the smoothing of sentiment, the dispersal of dissent—research can trace how governance materializes without visible enforcement. In this sense, infrastructural diagnosis is not a technique but an epistemology: it turns the study of manipulation into a study of the architectures that preclude its necessity.

This reorientation carries wider implications for platform research. As control migrates from content policy to design logic, critical scholarship must learn to interpret absence as patterned intention. What is missing from circulation can be as revealing as what remains. The analytical challenge, therefore, is to detect governance not through interventionary acts but through the silent regularities of infrastructure—through what platforms make improbable, rather than what they make visible.

Governing affect: Emotional legibility and platform incentives

If infrastructural power regulates what can appear, it also shapes how it is likely to feel. The emotional coherence observed on Xiaohongshu—an overwhelming tone of reassurance, optimism, and patriotic calm—points to how governance may extend beyond discourse into effect itself. The platform does not simply silence emotion; it appears to organize its visibility. Posts that express composure, warmth, or civic pride tend to circulate more widely in the observable corpus, while irritation, irony, or fatigue remain statistically marginal. In this sense, emotional legibility becomes a precondition for reach: to be seen, one must feel in ways that align with dominant affective registers.

This dynamic resonates with what Ahmed (2013) describes as the circulation of emotion as alignment. Feelings do not simply arise; they attach to certain forms and subjects, binding users to collective norms. On Xiaohongshu, this alignment appears to be reflected in engagement patterns: positive tones are more prevalent in highly visible comments, which in turn tends to reinforce the prominence of similar expressions. While the internal mechanics of ranking remain opaque, the observed co-occurrence of affective positivity and visibility suggests a feedback relation in which emotional coherence is more compatible with infrastructural circulation than affective friction.

For users, this form of regulation rarely presents itself as coercion. Governance is experienced more as rhythm than as rule. Through repeated exposure, participants learn which affects travel, which tones stall, and which simply fade from view. Over time, these patterned outcomes foster what Berlant (2011) terms cruel optimism: the attachment to positivity as a condition of belonging. Users come to internalize the affective grammar of the system, adjusting their voice in order to remain both intelligible and rewarded. Regulation thus becomes habitual, its effectiveness measured less by enforcement than by ease.

Seen in this light, affective convergence is unlikely to be only a byproduct of spontaneous social consensus, especially given the highly fragmented interaction structure documented earlier. Rather, it tends to align more closely with an infrastructural environment in which certain emotional tones are more legible, more circulable, and therefore more likely to be reproduced at scale. Affective convergence here refers not simply to similarity of feeling, but to a patterned compression of the affective range under conditions where visibility, engagement incentives, and interactional affordances jointly make some emotional registers easier to sustain in public view than others (Swart, 2021). In a fragmented comment ecology, such coherence need not depend on users amplifying one another; it can arise when the visible field is continuously tilted toward reassurance, warmth, and civic composure, while discordant affect remains less likely to accumulate presence or traction (Reviglio and Agosti, 2020; Metzler and Garcia, 2024). In this sense, convergence is relational not because users coordinate, but because the same infrastructural conditions repeatedly channel heterogeneous expressions into a narrower emotional horizon. Here, emotional legibility functions as an invisible infrastructure of participation, shaping what kinds of feelings can plausibly appear in public view.

Limitations and future directions

Several limitations should be acknowledged when interpreting the findings of this study. First, the analysis focuses on a single platform and a single political episode, offering interpretive depth but constraining generalizability. Xiaohongshu's hybrid identity—as a commercial lifestyle community operating within a semi-regulated media environment—shapes its infrastructural design in distinctive ways. The dynamics of preclusion identified here may therefore not apply uniformly across platforms with different ownership structures, audience compositions, or governance logics. Comparative investigations, particularly those including platforms such as Douyin, Weibo, or international analogs, would help clarify how similar architectures of anticipatory control manifest across varied socio-technical contexts.

A second limitation arises from data accessibility. This study relies entirely on front-end, publicly visible content, which allows the mapping of discursive outcomes but not the procedural mechanisms through which visibility is filtered. Proprietary algorithms, moderation protocols, and internal ranking heuristics remain opaque to external observation, making causal explanation necessarily inferential. The analytical framework here should therefore be read as diagnostic rather than demonstrative.

Future work could address this limitation through methodological triangulation—combining computational diagnostics with ethnographic fieldwork, interface analysis, or collaborative access to anonymized back-end datasets—to reveal how infrastructural preclusion operates in real time.

Finally, the analysis offers a cross-sectional snapshot of discourse during a particular moment of geopolitical tension. Yet infrastructural power is inherently dynamic: algorithmic thresholds evolve alongside commercial incentives, policy directives, and user adaptation. Longitudinal or event-based research could trace how semantic narrowing and affective convergence fluctuate as governance routines stabilize or shift. Despite these constraints, the study provides a diagnostic entry point for examining visibility under algorithmic opacity and calls for comparative research across platforms and time to capture the evolving logics of design-based control. These directions point toward a broader research agenda on how infrastructural governance evolves across platforms, media systems, and political contexts.

Conclusion

This study examined public discourse on Xiaohongshu during the 2025 China—U.S. trade conflict using a three-part comparative design: a large political corpus from Xiaohongshu, a contemporaneous non-political baseline from the same platform, and a parallel Weibo sample on the same issue. A multi-stage diagnostic pipeline—combining weak supervision anomaly scoring, clustering, and intersected behavioral—semantic network analysis—found no evidence of coordinated automation. At the same time, the Xiaohongshu political corpus displayed a compressed topical range and affective convergence, in contrast to the platform's non-political baseline (characterized by greater topical diversity, denser organic clustering, and a distinct negative-valence peak indicating affective heterogeneity) and to the Weibo sample (showing comparable emotional positivity to the main corpus but a more dispersed and less convergent topic structure). Taken together, these patterns suggest that the absence of manipulation does not imply neutrality; rather, it reflects a designed condition of visibility.

The findings indicate that Xiaohongshu's platform logic may shape expression upstream, making certain utterances statistically less likely to surface. We conceptualize this as discursive preclusion: an infrastructural mode of governance that does not delete content after the fact but conditions what can appear in the first place. Evidence for this configuration is reflected in the corpus itself—limited topic proliferation and a sentiment distribution skewed toward neutral or positive tones in the political sample. By contrast, the non-political baseline displays a wider affective bandwidth, and the Weibo sample shows a more dispersed and less convergent topic structure despite a similarly positive sentiment bias. Under this reading, convergence without coordination is not a paradox but a signature of governance by design.

Conceptually, the study reframes platform power as a shift from moderating speech to governing legibility. Methodologically, it advances a detection-to-diagnosis approach suited to opaque infrastructures: treating silence, uniformity, and fragmentation as analyzable outputs rather than null results. Empirically, the cross-platform and within-platform contrasts show how different architectures produce distinct visibility regimes without relying on automation or overt enforcement. Ultimately, this study demonstrates how design governs visibility itself—shaping not only what becomes visible, but what can become sayable.

Footnotes

ORCID iDs

Jie Xin

Lihui Zheng

Wan Ni

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Social Science Fund of China, (grant number 25BXW092).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data and code availability

All data and code underlying the findings of this study are openly available at the following GitHub repository: . com/xinjiexinjie/socialbots. The repository contains the anonymized dataset, full analysis pipeline (including anomaly scoring, clustering, and network construction), and all scripts used for figure generation and statistical modeling.

References

Aggarwal

(2017) Outlier Analysis. 2nd edition. Cham, Switzerland: Springer

Ahmed

(2013) The cultural politics of emotion. New York, NY, USA: Routledge.

Ananny

Crawford

(2018) Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media & Society 20(3): 973–989.

Bach

Ratner

, et al. (2017) Learning the structure of generative models without labeled data. In: International Conference on Machine Learning. PMLR, pp.273–282.

Barzilai-Nahon

(2008) Toward a theory of network gatekeeping: A framework for exploring information control. Journal of the American Society for Information Science and Technology 59(9): 1493–1512.

Berlant

(2011) Cruel optimism. Durham, NC, USA: Duke University Press.

Blei

Jordan

(2003) Latent Dirichlet allocation. Journal of Machine Learning Research 3(Jan): 993–1022.

Blondel

Guillaume

Lambiotte

, et al. (2008) Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment 2008(10): 10008.

Bradshaw

Howard

(2017) Troops, trolls and troublemakers: A global inventory of organized social media manipulation. In: Computational Propaganda Research Project of Oxford Internet Institute, December 2017, Oxford, UK, pp.1–37.

10.

Bratton

(2016) The stack: On software and sovereignty. Cambridge, MA, USA: MIT Press.

11.

Bucher

Schou

Waldkirch

(2021) Pacifying the algorithm–anticipatory compliance in the face of algorithmic management in the gig economy. Organization 28(1): 44–67.

12.

Bucher

(2018) If…then: Algorithmic power and politics. New York, NY, USA: Oxford University Press.

13.

Chen

Shi

(2025) Proactive governance by official administrators on Chinese social media platforms: Boundary discourse and governance legitimacy. Media, Culture & Society 48(1): 90–107.

14.

Chen

Xie

Wang

, et al. (2023) How we express ourselves freely: Censorship, self-censorship, and anti-censorship on a Chinese social media. In: International Conference on Information. Springer, pp.93–108.

15.

Cohen

(2023) Infrastructuring the digital public sphere. Yale Journal of Law & Technology 25(Spring): 1–40.

16.

DeVito

(2017) From editors to algorithms: A values-based approach to understanding story selection in the Facebook news feed. Digital Journalism 5(6): 753–773.

17.

Duffy

Meisner

(2023) Platform governance at the margins: Social media creators’ experiences with algorithmic (in) visibility. Media, Culture & Society 45(2): 285–304.

18.

Ferrara

Varol

Davis

, et al. (2016) The rise of social bots. Communications of the ACM 59(7): 96–104.

19.

Foucault

(1990) The history of sexuality: An introduction, volume 1. New York, NY, USA: Vintage.

20.

Freelon

McIlwain

Clark

(2018) Quantifying the power and consequences of social media protest. New Media & Society 20(3): 990–1011.

21.

Gillespie

(2018) Custodians of the Internet: Platforms, content moderation, and the hidden decisions that shape social media. New Haven, CT, USA: Yale University Press.

22.

Gillespie

(2020) Content moderation, AI, and the question of scale. Big Data & Society 7(2): 2053951720943234.

23.

Hallinan

Reynolds

Kuperberg

, et al. (2025) Aspirational platform governance: How creators legitimise content moderation through accusations of bias. Internet Policy Review 14(1): 1–28.

24.

Helmond

Van der Vlist

(2019) Social media and platform historiography: Challenges and opportunities. TMG–Journal for media History 22(1): 6–34.

25.

Herman

Chomsky

(2021) Manufacturing consent. In: Power and Inequality. New York, NY, USA: Routledge, 198–206.

26.

Hojati

Nault

(2025) Content moderation with shadowbanning. Information Systems Research. DOI:10.1287/isre.2024.1140.

27.

Hotelling

(1933) Analysis of a complex of statistical variables into principal components. Journal of Educational Psychology 24(6): 417.

28.

Howard

Ganesh

Liotsiou

, et al. (2018a) The IRA, social media and political polarization in the united states, 2012–2018. Report, Project on Computational Propaganda, Oxford Internet Institute, University of Oxford.

29.

Howard

Woolley

Calo

(2018b) Algorithms, bots, and political communication in the US 2016 election: The challenge of automated political communication for election law and administration. Journal of Information Technology & Politics 15(2): 81–93.

30.

Hyzen

Van den Bulck

Puppis

, et al. (2025) Epistemic welfare and algorithmic recommender systems: Overcoming the epistemic crisis in the digitalized public sphere. Communication Theory 36(1): 46–57.

31.

Isnowfy (2012) Snownlp: A python library for processing Chinese text. https://github. com/isnowfy/snownlp. Accessed: 2025-06-04.

32.

King

Pan

Roberts

(2017) How the Chinese government fabricates social media posts for strategic distraction, not engaged argument. American Political Science Review 111(3): 484–501.

33.

Lambiotte

Delvenne

Barahona

(2008) Laplacian dynamics and multiscale modular structure in networks. arXiv preprint arXiv:0812.1770.

34.

Zhou

(2024) When content moderation is not about content: How Chinese social media platforms moderate content and why it matters. New Media & Society 27(11): 6139–6155.

35.

Liang

(2025) Geopolitics in platform governance: Algorithmic sovereignty and data localization in China and the United States. Policy & Internet 17(3): e70014.

36.

Liu

Yang

(2022) “dual-track” platform governance on content: A comparative study between China and United States. Policy & Internet 14(2): 304–323.

37.

Lomborg

Sick

Flensburg

, et al. (2024) Monitoring infrastructural power: Methodological challenges in studying mobile infrastructures for datafication. Internet Policy Review 13(2): 1–28.

38.

Mann

(1984) The autonomous power of the state: Its origins, mechanisms and results. European Journal of Sociology/Archives EuropéEnnes de Sociologie 25(2): 185–213.

39.

McQuail

(1992) Media Performance: Mass Communication and the Public Interest. London, UK: Sage Publications.

40.

McQueen

(1967) Some methods of classification and analysis of multivariate observations. In: Proc. of 5th Berkeley Symposium on Math. Stat. and Prob, pp.281–297.

41.

Metzler

Garcia

(2024) Social drivers and algorithmic mechanisms on digital media. Perspectives on Psychological Science 19(5): 735–748.

42.

Mimno

Wallach

Talley

, et al. (2011) Optimizing semantic coherence in topic models. In: Proceedings of the 2011 conference on empirical methods in natural language processing, pp.262–272.

43.

Myers West

(2018) Censored, suspended, shadowbanned: User interpretations of content moderation on social media platforms. New Media & Society 20(11): 4366–4383.

44.

Napoli

(2019a) Algorithmic gatekeeping and the transformation of news organizations. In: Social media and the public interest: Media regulation in the disinformation age.New York, NY, USA: Columbia University Press, 53–79.

45.

Napoli

(2019b) Social media and the public interest: Media regulation in the disinformation age. New York, NY, USA: Columbia University Press.

46.

Newman

(2003) The structure and function of complex networks. SIAM review 45(2): 167–256.

47.

Pang

Lee

, et al. (2008) Opinion mining and sentiment analysis. Foundations and Trends® in Information Retrieval 2(1–2): 1–135.

48.

Papaevangelou

Siapera

(2025) State, platform capitalism and infrastructural power: Microsoft’s data centres in Greece 2.0. Platforms & Society 2: 29768624251323325.

49.

Pearson

(1895) Vii. Note on regression and inheritance in the case of two parents. proceedings of the Royal Society of London 58(347–352): 240–242.

50.

Pedrazzi

Oehmer-Pedrazzi

(2023) Governance of misinformation by public service bots: The potential of automated social networking site accounts operated by public service media for correcting misinformation. In: Public Service Media’s Contribution to Society. Nordicom: University of Gothenburg, 193–218.

51.

Pedregosa

Varoquaux

Gramfort

, et al. (2011) Scikit-learn: Machine learning in python. The Journal of Machine Learning Research 12: 2825–2830.

52.

Plantin

Lagoze

Edwards

, et al. (2018) Infrastructure studies meet platform studies in the age of google and Facebook. New media & Society 20(1): 293–310.

53.

Prier

(2020) Commanding the trend: Social media as information warfare. In: Information warfare in the age of cyber conflict. London, UK; New York, NY, USA: Routledge, 88–113.

54.

Ragin

(2014) The comparative method: Moving beyond qualitative and quantitative strategies. Berkeley, CA, USA: University of California Press.

55.

Ratner

Bach

Ehrenberg

, et al. (2020) Snorkel: Rapid training data creation with weak supervision. The VLDB Journal 29(2): 709–730.

56.

Reimers

Gurevych

(2019) Sentence-bert: Sentence embeddings using siamese bert- networks. arXiv preprint arXiv:1908.10084.

57.

Reviglio

Agosti

(2020) Thinking outside the black-box: The case for “algorithmic sovereignty” in social media. Social Media+ Society 6(2): 2056305120915613.

58.

Roberts

(2019) Behind the Screen: Content Moderation in the Shadows of Social Media. New Haven, CT, USA: Yale University Press.

59.

Rousseeuw

(1987) Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational And Applied Mathematics 20(1): 53–65.

60.

Shang

(2025) Shifting platform governance: Examining participatory content moderation on a Chinese platform bilibili. Information, Communication & Society. DOI:10.1080/1369118X.2025.2520004.

61.

Shoemaker

Vos

(2009) Gatekeeping theory. New York, NY, USA: Routledge.

62.

Striphas

(2015) Algorithmic culture. European Journal of Cultural Studies 18(4–5): 395–412.

63.

Sun

(2012) Jieba: Chinese text segmentation. https://github.com/fxsjy/jieba . (Accessed: 2025–06–04).

64.

Swart

(2021) Experiencing algorithms: How young people understand, feel about, and engage with algorithmic news selection on social media. Social Media+ Society 7(2): 20563051211008828.

65.

Thelwall

Buckley

Paltoglou

(2012) Sentiment strength detection for the social web. Journal of the American Society for Information Science and Technology 63(1): 163–173.

66.

Van Dalen

(2023) Algorithmic gatekeeping for professional communicators: Power, trust, and legitimacy. London, UK: Taylor & Francis.

67.

Van Dijck

Poell

De Waal

(2018) The platform society: Public values in a connective world. Oxford University Press.

68.

Varol

Ferrara

Davis

, et al. (2017) Online human-bot interactions: Detection, estimation, and characterization.

69.

Wang

(2024) Regulation of algorithmic decision-making in China: Development, problems and implications. Singapore Journal of Legal Studies 2024: 276–305.

70.

Woolley

Howard

(2016) Automation, algorithms, and politics: Political communica- tion, computational propaganda, and autonomous agents—introduction. International Journal of Communication 10: 4882–4890.

71.

Woolley

Howard

(2018) Computational propaganda: Political parties, politicians, and political manipulation on social media. New York, NY, USA: Oxford University Press.

72.

Wright

(2022) Automated platform governance through visibility and scale: On the transformational power of automoderator. Social Media+ Society 8(1): 20563051221077020.

73.

(2024) Opening the ‘black box’of algorithms: Regulation of algorithms in China. Communication Research and Practice 10(3): 288–296.

74.

Yang

Varol

Hui

, et al. (2020) Scalable and generalizable social bot detection through data selection.

75.

Huang

Krijnen

(2025) Douyin’s playful platform governance: Platform’s self- regulation and content creators’ participatory surveillance. International Journal of Cultural Studies 28(1): 80–98.

76.

Zeng

Kaye

DBV

(2022) From content moderation to visibility moderation: A case study of platform governance on tiktok. Policy & Internet 14(1): 79–95.

77.

Zhang

(2025) Affective counterpublics under constraint: Emotion, platform governance, and AI ethics discourse on Chinese social media. AI & SOCIETY 41(1): 595–609.

78.

Zheng

(2024) Unintended consequences: Data practice in the backstage of social media. The Journal of Chinese Sociology 11(1): 11.

79.

Zhu

Phipps

Pridgen

, et al. (2012) Tracking and quantifying censorship on a Chinese microblogging site. arXiv preprint arXiv:1211.6166.

80.

Zhu

Phipps

Pridgen

, et al. (2013) The velocity of censorship: High- fidelity detection of microblog post deletions. In: 22nd USENIX Security Symposium (USENIX Security 13), pp.227–240.