Mitigating filter bubbles: Diverse and explainable recommender systems

Abstract

In recent years, the surge in online content has necessitated the development of intelligent recommender systems capable of offering personalized suggestions to users. However, these systems often encapsulate users within a “filter bubble”, limiting their exposure to a narrow range of content. This study introduces a novel approach to address this issue by integrating a novel diversity module into a knowledge graph-based explainable recommender system. Utilizing the Movie Lens 1M dataset, this research pioneers in fostering a more nuanced and transparent user experience, thereby enhancing user trust and broadening the spectrum of recommendations. Looking ahead, we aim to further refine this system by incorporating an explicit feedback loop and leveraging Natural Language Processing (NLP) techniques to provide users with insightful explanations of recommendations, including a comprehensive analysis of filter bubbles. This initiative marks a significant stride towards creating a more inclusive and informed recommendation landscape, promising users not only a wider array of content but also a deeper understanding of the recommendation mechanisms at play.

Keywords

Recommender system explainable recommendations filter bubble knowledge graph diversity

1 Introduction

In the contemporary digital era, recommender systems have emerged as a pivotal resource, assisting individuals in navigating through a plethora of options across various categories. These sophisticated systems have carved a niche for themselves by offering personalized filtering services, thereby facilitating tailored suggestions for users [4]. A recommender system is a software system that uses algorithms to analyse data and make personalized recommendations to users [5].

Serving as automated platforms, they meticulously sift through a vast array of online content, encompassing movies, products, music, news articles, and much more, to curate a selection aligned with individual preferences and tastes. The escalating advancements in crafting systems capable of delivering personalized insights have propelled the field of recommender systems to the forefront of research and innovation in recent times.

In the evolving landscape of technology, recommender systems have emerged as a focal point of research, continually presenting a myriad of challenges that researchers strive to overcome [6, 7]. A significant hurdle in this domain is the pervasive issue of the “Filter Bubble” in recommender systems. This term encapsulates the scenario where users find themselves ensnared in a cycle of recommendations that predominantly align with their existing preferences and inclinations, thereby curtailing their access to a broader spectrum of information and diverse content. Essentially, the system crafts a “bubble” around them, narrowing down their exposure to a limited array of items and perspectives.

The phenomenon known as the “filter bubble” first came to light in 2009, during a period when search engines started tailoring search outcomes for individuals based on a variety of factors including their previous online activities, preferences, and search histories. This personalization strategy, while enhancing the relevance of search results, inadvertently fostered an environment where users could be exposed to a limited scope of information, often aligned with their existing beliefs and preferences. Consequently, when different users input the same search query, they might be presented with divergent results, potentially fostering isolated information spheres, and limiting the diversity of perspectives encountered. This scenario raises concerns about the creation of informational silos, where individuals are potentially shielded from a broader array of viewpoints and opinions, thus hindering a more comprehensive understanding of various topics [8].

In an endeavor to counteract the effects of the filter bubble in recommender systems, this study proposes the utilization of an explainable recommender system. We have leveraged an explainable recommender system called CAFE (Coarse to Fine Neural Symbolic Network) [1]. In this work we have incorporated a novel diversity method into our explainable recommender system to mitigate the filter bubble issue. This initiative not only seeks to broaden the horizons for users but also aspires to foster a more inclusive and expansive recommendation environment.

In recent times, Explainable Recommender Systems have garnered significant attention, particularly in the realm of e-commerce. These sophisticated systems go beyond merely suggesting items to users; they also elucidate the underlying reasons behind each recommendation, fostering a transparent and trustful interaction between the users and the system. The construction of such recommender systems is intricately linked with the deployment of Knowledge Graphs [19], a tool that adeptly addresses both facets of the process: crafting tailored recommendations and providing insightful explanations for those suggestions.

By integrating the capabilities of knowledge graphs, recommender systems are not only able to refine the quality of recommendations but also significantly augment their explainability, offering users a deeper understanding and insight into the rationale behind each suggestion [3]. This dual functionality not only enhances user trust but also facilitates a more nuanced and enriched user experience, fostering a symbiotic relationship between the system and its users.

Remarkably, the prevalent application of explainable recommender systems is found within the e-commerce sector, where they primarily analyze datasets from platforms such as Amazon, focusing on categories like beauty products and other consumer goods. Contrary to this established trend, our study takes a divergent path. We have ventured into the realm of entertainment, crafting an explainable movie recommender system that leverages the rich data from the Movie Lens 1M dataset. This departure from the norm allows us to explore new dimensions of personalization and explanation in the context of movie recommendations. Furthermore, we have adopted the CAFE model as our foundational framework, a choice driven by its proven efficacy and adaptability. This model has been meticulously trained on the same dataset.

The explainable knowledge-based recommender system we are employing operates on the principles of a connection-based path embedding model, [2]. This sophisticated approach meticulously enumerates potential paths, facilitating a nuanced and comprehensive analysis of the relationships that exist between user-item pairs as well as item-item pairs. By doing so, it promises to offer a more granular insight into the intricate web of connections that dictate user preferences and item affiliations, paving the way for more precise and insightful recommendations.

This paper will delve deeper into our innovative approach, shedding light on the potential of a diversified, transparent, and enriched recommender system to transcend the limitations of the ‘Filter Bubble’. Through our exploration, we aim to foster a more inclusive recommendation environment, where users can benefit from a broader spectrum of information and insights, thereby enhancing the overall user experience.

2 Related work

2.1 Explainable recommender systems using knowledge graphs (KG)

Over the past several years, the research community has been fervently exploring avenues to create comprehensive explanations for the recommendations presented to users. One pioneering study in this domain is delineated in [10], where the researchers delve into the intricacies of the explanation interfaces within Automated Collaborative Filtering systems, setting a precedent in the field.

In the context of social networks, a noteworthy contribution has been made by [11], who ventured into the realm of music recommendations. This study illuminated the potential of leveraging social cues to enhance the recommendation process, demonstrating the influence of a user’s close friend’s preferences on the recommendations, with the depth of friendship being ascertained through an analysis of their interactions on platforms like Facebook.

Shifting the focus to the integration of knowledge graphs in crafting explainable recommendations, a significant stride was made by [12]. This team harnessed the potential of knowledge graph embeddings to craft recommendations that not only cater to user preferences but also provide insightful explanations. This was achieved through a meticulous process of soft matching, post the selection of recommended items, thereby adding a layer of transparency to the recommendation process.

Further advancing in this trajectory, the study documented in [3] introduced a novel approach that amalgamated reinforcement learning with reasoning over knowledge graphs, paving the way for the generation of recommendations accompanied by coherent explanations. This innovative approach has been a beacon, inspiring further research in this direction.

Adding a new dimension to this evolving field, the research presented in [13] broke new ground by developing a method that seamlessly integrates heterogeneous information networks with matrix factorization techniques. This initiative aimed to foster personalized entity-based recommendations, marking a pioneering step in utilizing paths within knowledge graphs to enhance the recommendation process. This groundbreaking approach has spurred a wave of research endeavors, encouraging scholars to explore the potential of knowledge graph pathways in refining recommendations.

2.2 Filter bubble in recommender systems

The phenomenon of filter bubbles in recommender systems has been a focal point of extensive discussions and research, polarizing opinions within the scholarly community. A segment of researchers staunchly believes that the filter bubble is not a significant concern, arguing that recommender systems actually facilitate users in uncovering items that would otherwise remain undiscovered in the absence of such recommendations [14]. This perspective is further substantiated by the findings of [15], who undertook meticulous experiments to scrutinize the effects of both implicit and explicit personalization on the diversity of content presented in Google News. Interestingly, their investigations did not corroborate the existence of a filter bubble, thereby challenging the prevalent notion surrounding it.

Contrastingly, a substantial body of research posits that the filter bubble indeed poses a significant challenge, often resulting in users being repeatedly exposed to similar items, thereby gradually alienating them from a diverse array of content [16, 17]. This viewpoint is echoed in the study conducted by [18], who embarked on an analytical journey to assess the influence of various collaborative filtering algorithms on the diversity and novelty of the recommendations generated. Through a comprehensive examination of different user configurations, they sought to gauge the impact of user preferences on the diversity and novelty of the recommendations. Their findings underscored the detrimental effects of the filter bubble on user behavior, highlighting its role in fostering a more insular user experience.

In response to this growing concern, several innovative approaches have been proposed to counteract the filter bubble phenomenon. For instance, [19] ventured to mitigate the filter bubble effect by incorporating elements of unexpectedness in their recommendation algorithms, aiming to infuse a sense of novelty and diversity in the user experience. Similarly, [20] introduced a groundbreaking recommender prototype that empowers users to actively participate in diminishing the influence of filter bubbles, fostering a more inclusive and varied content discovery process.

Furthermore, the study by [21] presented a novel strategy to manage filter bubbles effectively by leveraging Concept Activation Vectors. These vectors were utilized to address specific topical dimensions, such as political polarization, aiming to create a more balanced and diversified content landscape. By developing a state-of-the-art VAE-based recommender system, they succeeded in diversifying latent embeddings along targeted dimensions, all the while ensuring the topical relevance of the content, thus paving a new path in the ongoing efforts to curb the filter bubble effect in recommender systems.

2.3 Diversity in recommender systems

In the realm of recommender systems, the integration of diversity stands as a prevalent strategy to counteract the effects of filter bubbles, fostering a more expansive and varied range of recommendations for users [22]. The methodologies to infuse diversity into these systems can be broadly categorized into two distinct approaches: post-processing techniques and end-to-end methods.

The post-processing approach, a widely adopted strategy, primarily focuses on enhancing the diversity of recommendations during the latter stages of the recommendation process [23]. This technique is characterized by the implementation of re-ranking strategies, wherein the initially generated recommendations are subjected to further analysis and adjustments to introduce a broader spectrum of options, thereby preventing the concentration of similar items and promoting a more varied selection [24]. This method essentially serves to refine the recommendation list, ensuring that users are presented with a more diverse array of choices, which in turn helps in breaking the cycle of repetitive and confined suggestions.

Conversely, the end-to-end method adopts a more holistic approach, wherein the objectives of both accuracy and diversity are harmoniously balanced throughout the entire process, encompassing both the training and inference stages [25]. This approach is designed to inherently foster diversity from the inception of the recommendation process, ensuring that the system is attuned to the nuances of user preferences while also maintaining a broad perspective that encourages the inclusion of diverse items in the recommendations. This method, therefore, presents a comprehensive solution, seamlessly integrating diversity considerations into the core functioning of the recommender system, thereby promoting a more enriched and varied user experience.

In a notable contribution to this field, [26] embarked on a mission to address the pervasive issue of filter bubbles by conceptualizing and developing a cross-domain matrix factorization technique. This innovative approach leverages adaptive diversity regularization, coupled with social tagging mechanisms and collective matrix factorization strategies, to enhance the performance of the recommendation process. Through this method, they have managed to create a system that not only offers improved recommendation performance but also actively works to mitigate the filter bubble effect, paving the way for a more inclusive and diverse recommendation landscape.

3 Methodology

3.1 Overview

In this research, our endeavor was to amplify the capabilities of the explainable recommender model, which served as the foundational framework for our project. Our primary innovation centers around the integration of a diversity module, a pivotal step towards mitigating the filter bubble effect commonly observed in recommender systems. It is imperative to strike a harmonious balance in incorporating diversity; excessive inclusion can potentially lead to unsatisfactory recommendations, thereby diminishing user satisfaction.

The introduction of the diversity module is grounded in the principles of Shannon entropy, a strategy aimed at bolstering the diversity within recommendations without compromising on the accuracy levels. Essentially, Shannon entropy quantifies the uncertainty inherent in a random variable. In the context of our study, it serves to gauge the diversity or uncertainty of genres that a user engages with during their interactions on the platform. This segment meticulously explores the architecture of the original explainable recommender model and provides an in-depth analysis of the modifications undertaken to assimilate the diversity module, thereby enhancing its efficacy and inclusivity in generating recommendations. Figure 1 shows a detailed explanation of the model which includes Diversity.

Fig. 1

Depiction of the Enhanced Diversity based Cafe Model, a Multi-Stage KG Reasoning Approach [1]. (a) Utilizing a Knowledge Graph (KG) and initiating with a specific user, the objective is to facilitate multi-step path reasoning to generate recommendations. (b) During the initial phase, a customized user profile is developed, incorporating a novel diversity module that considers historical user interactions within the KG, aiming to broaden the spectrum of recommendations. (c) To effectively integrate the user profile in the path reasoning process, a repository of neural symbolic reasoning units, now enriched with diversity elements, is maintained to foster a more expansive recommendation environment. (d) In the subsequent phase, a layout tree is assembled utilizing the modules, which are now fine-tuned to incorporate diversity aspects based on the user profile, enhancing the recommendation quality and breadth. This structure is leveraged by the E-UCPG algorithm to create (e) a series of paths that culminate in a diverse set of recommendations, promoting a more inclusive and enriched user experience.

3.2 Explainable recommender system’s model: A brief recap

In our research, we have adopted an explainable Recommender System called “Coarse-to-Fine Neural Symbolic Reasoning (CAFE)” model [1], a pioneering framework in the domain of movie recommender systems. This model is particularly adept at harnessing the power of Knowledge Graphs (KGs), a rich repository of user-item interaction data, to craft personalized user profiles that are central to enhancing the recommendation process and also enable explainable recommendations [3, 12, 27].

The CAFE model is a beacon in the evolving landscape of movie recommender systems, offering a nuanced methodology to decipher intricate user-item interaction patterns. These patterns are meticulously mapped in KGs, which serve as fertile ground for harvesting a wealth of information about users and the movies they interact with. By analyzing these interactions, the model is capable of generating detailed user profiles that encapsulate a coarse representation of user behaviors. These profiles, therefore, become instrumental in guiding the path-finding process, paving the way for the derivation of reasoning paths that are aligned with fine-grained predictions.

To navigate through the complex structure of KGs, the model employs a path finding algorithm called Enhanced User-Centric Path generation Algorithm (E-UCPG). This algorithm is grounded in neural symbolic reasoning modules, which are proficient in identifying a series of paths across the KG swiftly and effectively. The algorithm acts as a guiding light, directing the system towards potential movies of interest based on discernible user behaviors captured from historical data.

Furthermore, the model operates on a coarse-to-fine paradigm. Initially, it focuses on crafting a user pro file that embodies prominent user behavior patterns extracted from historical interactions. This profile serves as a roadmap, offering critical insights into potential reasoning paths that are more likely to resonate with the user’s preferences. As the process progresses, the model transitions to a fine-grained modeling phase, where an advanced path inference com ponent takes charge. This component, guided by the insights gleaned from the user profile, orchestrates a multi-step path reasoning process, ensuring a direct and effective path to high-quality recommendations.

Algorithm 1 Enhanced User-Centric Path generation (E-UCPG) Algorithm
1. Input: user U, profile P.
2. Output: Diverse recommendation paths DP.
3. procedure Run
4. Generate tree PT from profile P.
5. Set root, U_ref, PathSet to root(PT), U, {{U}}.
6. Queue Q starts with CHILDREN(root).
7. while Q not EMPTY do
8. node ← Q.pop(), parent ← PARENT(node).
9. U_ref ← UpdateRef(U, parent).
10. PathSet_node ← {}.
11. for path in PathSet_parent do
12. Extend path using valid edges from G
13. Merge new paths into PathSet_node
14. Q ← Q UNION CHILDREN(node).
15. ApplyDiversity(PathSet_node)
16. return CompliePaths(PT).

The Enhanced User-Centric Path generation Algorithm is given in Algorithm 1. By embracing this sophisticated approach, the model stands poised to redefine the sphere of movie recommender systems, promising not only superior recommendations but also comprehensive explanations that illuminate the underlying rationale behind each recommendation. This transparency not only fosters a deeper understanding but also significantly enhances user trust in the recommendations presented. The E-UCPG algorithm generates personalized recommendation paths by expanding user-derived tree, ensuring diversity in the resultant paths. It concludes by compiling these diverse paths into a comprehensive recommendation set.

3.2.1 Entity embedding model

In the dynamic landscape of recommender systems, the Entity Embedding Module stands as a cornerstone, orchestrating the initialization of embeddings for a myriad of entities encapsulated within the Knowledge Graph. This module excels in analyzing and integrating various attributes intrinsically linked to each entity, thereby sculpting a dense representation that serves as the foundation of the recommendation process.

Mathematically, the initialization of the embeddings can be represented as given in Equation (1): $W = U (- E / 0.5, E / 0.5)$ (1) Where: •

W: Weight matrix of the embeddings

•

U (a, b): Uniform distribution function with range [a, b]

•

E: Embedding size

At its nucleus, the module cultivates a profound comprehension of user-item interactions by fabricating a vibrant, multidimensional space where entities are delineated as vectors. These vectors harbor the quintessential characteristics and patterns extracted from the attributes of the entities, as formulated in Equation (2): $Vi = f (Ai)$ (2) Where: •

Vi: Vector representation of entity i

•

Ai: Attribute set of entity i

•

f: Transformation function that maps attributes to the vector space

Initially, these embeddings are conceived through a uniform distribution, a tactic that guarantees a harmonized and impartial inception for the model. As the training phase unfolds, these embeddings are meticulously refined to mirror the intrinsic patterns and subtleties discerned in the data more accurately, enhancing the precision and effectiveness of the recommendation system.

Moreover, the module is fortified with tools to effortlessly incorporate diversity metrics, a maneuver pivotal in nurturing a broader and varied recommendation set. By exploiting these diversity metrics, the module adeptly maintains a fine equilibrium, ensuring the recommendations are not only precise but also span a broad array of user inclinations, thereby alleviating the prevalent filter bubble effect in recommender systems.

Through its advanced methodology in entity representation coupled with a dedication to promoting diversity, the Entity Embedding Module emerges as a beacon of innovation, steering the evolution of sophisticated and proficient recommender systems.

3.2.2 Relational module

In the intricate framework of our explainable recommender system, the Relation Module emerges as a vital component, central to the computational and predictive functionalities of the system. This module is proficient in determining the log probabilities of forthcoming output entities during the forward propagation stage, a crucial step in crafting accurate and reliable recommendations.

The Relation Module functions by amalgamating the embeddings of entities that are actively participating in a certain relation, thereby forging a fresh representation. This newly formed representation is pivotal in forecasting the tail entity within the Knowledge Graph, a procedure deeply rooted in mathematical computations and probabilistic principles.

From a mathematical standpoint, the operation can be delineated as shown in Equation (3): $Y = f (X) + Eh$ (3) Where: •

Y: Output (New Representation)

•

f(X): A function denoting the amalgamation and transformation of entity embeddings

•

Eh: Embedding of the head entity

The module is architecturally designed with a series of linear transformations and batch normalizations, harmonized with activation functions to facilitate a smooth transition and integration of data between various layers.

Through the concerted efforts of the Relation Module, our model aims to deliver recommendations that are not only precise but also grounded in mathematical rigor and computational efficiency, paving the way for a new era of intelligent and responsive recommender systems.

3.2.3 Integration of diversity module

In an endeavor to counteract the prevalent filter bubble effect observed in recommender systems and to cultivate a broader spectrum of recommendations, we have ingeniously integrated a diversity module into our explainable recommender model. This innovative module leverages the principles of Shannon entropy [32], a fundamental concept in information theory, to ascertain the diversity score for each user, grounded on their historical interactions.

Shannon entropy for diversity calculation

The diversity module operates on the premise of Shannon entropy, a mathematical concept rooted in information theory, which serves as a potent tool to gauge the diversity embedded in a user’s interaction history. This is mathematically represented by Equation (4): $H (X) = - \sum p (x_{i}) {log}_{2} p (x_{i})$ (4) Where,

H(X): Denotes the Shannon entropy, a metric that encapsulates the diversity in the user’s interaction history.

p(x_i): Represents the probability of occurrence of an item x_i in the user’s interaction history.

This entropy score emerges as a robust indicator of diversity, where elevated values signify a richer and more varied interaction history.

By harnessing the Shannon entropy, the module adeptly computes the diversity score for each user, drawing upon their interaction history. These computed scores are then seamlessly amalgamated into the symbolic network, a critical facet of our model. This integration profoundly impacts both the ranking and regularization losses during the training phase, steering the model towards generating a more diverse and inclusive set of recommendations. Through this strategic integration, the model aspires to transcend the limitations of the filter bubble effect, fostering a recommendation environment that is both diverse and reflective of a broader range of user preferences.

3.2.4 Integration into symbolic network

In our refined approach, the diversity scores, computed through the application of Shannon entropy, are seamlessly woven into the symbolic network, a vital facet of our advanced explainable recommender system. These scores play a pivotal role in modulating the regularization and ranking losses throughout the training phase, thereby steering the model towards crafting a broader spectrum of recommendations. This amalgamation is meticulously executed by merging the diversity scores with the outcomes at each juncture within the symbolic network, a strategy that significantly impacts the ensuing layers and the ultimate recommendation output.

To conclude, the incorporation of the diversity module significantly enhances our evolved model, positioning it to deliver a more expansive and inclusive range of recommendations. This effectively mitigates the filter bubble issue that is prevalent in many existing recommender systems. This novel approach retains the core tenets of our initial explainable recommender model, introducing a revolutionary mechanism that promotes greater diversity in the recommendations. Consequently, users are presented with a broader and more nuanced array of choices, fostering a more enriched user experience.

4 Experiments and results

In our study, we embarked on a comprehensive series of experiments to ascertain the efficacy of our proposed model. These experiments were meticulously designed to scrutinize the performance of the model from various angles, employing a diverse range of metrics for a more rounded analysis.

Furthermore, we engaged in a comparative study where the results yielded by our model were compared against those obtained from other baseline models, establishing a clear benchmark and highlighting the potential advancements our model brings to the table.

In addition to this, we undertook a rigorous process of hyperparameter tuning, where various configurations were tested to fine-tune the model to its optimal state. This step was crucial in ensuring that the model not only performs at its peak but also demonstrates stability and reliability across different scenarios, thereby showcasing its robustness and adaptability in delivering optimized results.

4.1 Dataset

In our research, we utilized the well-established MovieLens 1M (ML1M) dataset as the foundational data source to develop and refine our model. This dataset encapsulates a rich collection of one million evaluations, where users have rated various movies on a scale from 0 to 5. A detailed breakdown of the dataset’s characteristics can be found in Table 1, which delineates the number of users, the variety of movies available, and the total interactions recorded. Table 2 shows the stats of ML1M Knowledge graph.

Table 1
Detailed analysis of the MovieLens 1M dataset

Dataset No. of users No. of items Interactions

ML 1M 6040 3883 1000209

Dataset	No. of users	No. of items	Interactions
ML 1M	6040	3883	1000209

Table 2

Detailed Analysis of the MovieLens 1M KG

KG	No. of	No. of	No. of Relation
Entities	relations	types
ML 1M	16899	1830766	10

Simultaneously, we incorporated a Knowledge Graph (KG) –a sophisticated semantic network that graphically illustrates the intricate interconnections between various entities, which can encompass objects, events, circumstances, or abstract concepts. This graph is typically housed within a Knowledge Base (KB), where it serves as a visual representation of the complex web of relationships, thereby earning the moniker “knowledge graph”.

A knowledge graph is fundamentally constructed from three primary elements: nodes, edges, and labels. Nodes can represent a diverse array of entities, including but not limited to individuals, places, or objects. These nodes are interconnected through edges, which elucidate the nature of the relationships between them. To illustrate, consider the movie “Titanic” and the actor Leonardo DiCaprio as nodes. An edge in this context could signify the “starring” relationship where Leonardo DiCaprio played a lead role in “Titanic”, or alternatively, it could denote a “starred by” relationship, indicating that the movie “Titanic” featured Leonardo DiCaprio in a prominent role.

To adapt this structure for recommendation systems, an initial step involves associating each item in the offline dataset with its corresponding entity within a Knowledge Base. This specific knowledge graph for the MovieLens dataset was meticulously extracted and documented by [28].

4.2 Data preprocessing

In the field of data science, the initial step of preprocessing is crucial as it lays the groundwork for integrating and analyzing data from diverse sources. When it comes to augmenting the MovieLens 1M dataset with knowledge graphs (KGs), this step takes on a heightened significance. In this section, we explore the various facets of this initial phase, emphasizing the methods used to tackle common issues such as missing links in the KG and the importance of threshold discarding.

To begin with, one of the primary challenges faced is the mismatch between the number of items listed in the original dataset and those that have been successfully mapped onto the KG. This issue arises due to occasional lapses in linking products from the dataset to their respective entities in the KG, leaving a gap in information regarding certain items. To remedy this, a detailed verification process is initiated to pinpoint and eliminate unmatched products along with their related ratings from the dataset, thereby fostering a more synchronized data structure.

Additionally, the preprocessing stage incorporates the vital technique of threshold discarding. This strategy is essential when dealing with datasets that are sparse or have a scant number of ratings per user. Here, a predetermined threshold value is set, and items and users that are mentioned less than this threshold value in the ratings are removed. Utilizing various data manipulation tools, a sequence of operations is carried out to identify and discard less prevalent users and items, thus refining the dataset to its k-core and improving its density.

Furthermore, this stage requires the consistent application of changes across different data frames to ensure uniformity and integrity. This involves deleting users and items that are no longer found in the updated ratings data frame, a move that guarantees the dataset remains dense and in line with the objectives set.

As we approach the end of the preprocessing phase, a set of functions are employed to extend the removal of items to the KG, guaranteeing that the modifications are mirrored across all pertinent data structures. This step is critical in preserving the consistency and reliability of the knowledge graph, setting the stage for deeper analysis and accurate predictions.

In conclusion, the preprocessing phase outlined here acts as a solid guideline, assisting individuals in the complex task of refining and enhancing datasets integrated with knowledge graphs. Through a series of well-planned steps and approaches, it facilitates the development of a synchronized and enriched data structure, poised to facilitate insightful analyses and support informed decision-making in the sphere of movie recommendations.

4.3 Evaluation metrics overview

To accurately gauge the efficacy of our newly developed model, we have integrated a variety of metrics into our evaluation framework. These metrics, including Recall, NDCG, W-NDCG, and Coverage, serve as robust tools to scrutinize the performance of the recommendation system from different angles. Here, we delve deeper into each of these metrics, elucidating their significance and the mathematical formulations that underpin them.

Recall: Recall is a critical metric that assesses the accuracy of the recommendation system. Essentially, it quantifies the proportion of relevant items that have been successfully identified and recommended by the model. The mathematical representation of Recall is given by Equation (5): $Recall = \frac{Number of relevant items recommended}{Total number of relevant items}$ (5)

This metric serves as a reliable indicator of the model’s ability to pinpoint and suggest items that are genuinely of interest to the users, thereby reflecting the quality of the recommendations provided.

NDCG (Normalized Discounted Cumulative Gain): NDCG is a vital metric that evaluates the quality of the rankings produced by the recommendation system. It measures the relevance of the items listed at the top of the recommendations, which is particularly significant in the context of movie suggestions. The mathematical expression for NDCG is given by Equation (6): $NDCG = \frac{1}{z} \sum_{i = 1}^{n} \frac{{rel}_{i}}{{log}_{2} (i + 1)}$ (6) where: •

rel_i is the relevance of the item at position i.

•

Z is a normalization factor to ensure the NDCG score lies between 0 and 1.

•

n is the number of items in the list.

W-NDCG [34]: This metric is an extension of NDCG, where each item in the recommendation list is assigned a different weight. These weights can be determined based on various factors such as the category of the movie or user preferences. The formula for W-NDCG is given by Equation (7): $W - NDCG = \frac{1}{z} \cdot \sum_{i = 1}^{n} \frac{{rel}_{i \cdot w_{i}}}{{log}_{2} (i + 1)}$ (7) where: •

wi is the weight assigned to the item at position i

This nuanced approach allows for a more tailored evaluation, taking into account the potential varying significance of different items in the list.

Precision: Precision is another vital metric that evaluates the effectiveness of a recommender system. It measures the proportion of recommended items that are indeed relevant to the users, thereby indicating the accuracy of the suggestions made by the system. The mathematical representation of Precision is given by Equation (8): $Precision = \frac{No . of Relevant items Recommended}{Total no . of Items Recommended}$ (8)

Coverage: Coverage is a metric that assesses the diversity and inclusivity of the recommendations generated by the system. Essentially, it evaluates the system’s ability to recommend a wide array of items from the training set to the test set, promoting a diverse range of suggestions. The mathematical representation of Coverage is given by Equation (9) $\begin{matrix} Coverage = \\ \frac{Number of Unique Items Recommended}{Total number of unique Items in the Training Set} \end{matrix}$ (9)

This metric is instrumental in ensuring that the recommendation system is capable of catering to a broad spectrum of user preferences, fostering a more inclusive and comprehensive recommendation experience.

In conclusion, these metrics collectively offer a comprehensive framework for evaluating the performance of the recommendation system, facilitating a nuanced analysis that encompasses various critical aspects of the recommendation process.

4.4 Baselines

In our endeavor to validate the efficacy of our newly proposed model, we have undertaken a meticulous comparison with several contemporary baseline models that are recognized as the pinnacle of the current state of the art. These baseline models encompass the following:

1.
FM (Factorization Machine) [29]: A widely acknowledged model that utilizes factorization to predict the ratings in a recommendation system. It is known for its versatility and ability to handle a wide range of data patterns.
2.
NFM (Neural Factorization Machine) [30]: This model represents an evolution of the FM, incorporating neural networks to enhance the learning capabilities and prediction accuracy. It is adept at capturing complex patterns within the data, offering a nuanced approach to recommendation systems.
3.
Fairco [31]: Fairco stands as a revolutionary ranking algorithm that embodies the principles of fairness and equity. It is designed to facilitate balanced exposure opportunities across various item groups, thereby promoting a more inclusive and representative recommendation list. This algorithm empowers users with the ability to control and influence the ranking process, fostering a more personalized and equitable user experience.
4.
Diversity [24]: The Diversity model is renowned for its innovative approach to enhancing the diversity of recommendations. It employs a re-ranking method that aims to minimize the similarity between items within a recommendation list. By doing so, it ensures a broader and more varied range of suggestions, enhancing user satisfaction and engagement.

The FM and NFM models integrated with the principles of Fairco and Diversity are used as baselines.

Apart from this we have also compared our model with diversity-based recommender systems where diversity is calculated using Shannon Entropy [32] method.

We anticipate that this comprehensive comparison will shed light on the potential advantages of our proposed model, paving the way for a new era of intelligent and inclusive recommendation systems.
4.5 Hyperparameter settings

In the process of optimizing our model, we adhered to a specific set of hyperparameter settings to facilitate effective training. These settings are delineated as follows:

Epoch Count: We conducted the training over a span of 40 epochs, allowing the model to learn and adapt progressively through each iteration.

Checkpoint Intervals: The model’s progress was monitored at intervals of 100 steps, enabling timely checkpoints to track and evaluate the performance dynamically.

Batch Dimension: A batch size of 256 was chosen.

Learning Rate: Adopted a learning rate of 0.05, a moderate pace that ensures steady learning without overshooting the optimal solutions.

Dropout Proportion: To prevent overfitting and promote a generalized model, a dropout rate of 0.4 was implemented, thereby randomly nullifying 40% of the connections during training.

Embedding Dimensionality: The embedding layer was configured with a size of 100, offering a rich yet manageable feature space for representing the data intricacies.

Through this configuration, we aimed to create a conducive environment for the model to learn and adapt, potentially leading to superior performance and reliable predictions.

Hence, to conclude, we meticulously crafted a series of experiments to evaluate a novel recommendation model, utilizing the enriched MovieLens 1M dataset integrated with a detailed Knowledge Graph (KG). This initiative involved a critical data preprocessing phase to address the challenges of missing links in the KG and to refine the dataset through threshold discarding. We established a robust evaluation framework comprising metrics like Recall, NDCG, W-NDCG, and Coverage to assess the system’s performance comprehensively. A comparative analysis with state-of-the-art baseline models, including FM, NFM, Fairco, and Diversity, was conducted to underscore the advancements our model introduces. Furthermore, we delineated the optimal hyperparameter settings adopted during the training phase, aiming to foster efficient learning and reliable predictions, thereby paving the way for a new benchmark in recommendation systems.

4.6 Performance

We evaluated our model and all the baselines in terms of four representative top-N recommendation metrics which we have discussed earlier. These are: Recall, NDCG, W-NDCG and Coverage. The results are shown in Tables 3 and 4 respectively. Table 3 shows the performance of our model in comparison to FM baselines and similarly in Table 4 we compare our model’s performance with NFM baselines.

Table 3
Performance comparison of our diversity based explainable recommender system with FM baselines. The best results are highlighted in bold

Models Recall NDCG W-NDCG Coverage

Random 0.0029 0.0031 0.0029 0.085024

FM [29] 0.0659 0.0536 0.0485 0.0866

FM-Fairco [31] 0.0605 0.0506 0.0459 0.093799

FM-Diversity [24] 0.0531 0.0473 0.0428 0.095407

Our Model 0.07831 0.08635 0.095 0.260

Models	Recall	NDCG	W-NDCG	Coverage
Random	0.0029	0.0031	0.0029	0.085024
FM [29]	0.0659	0.0536	0.0485	0.0866
FM-Fairco [31]	0.0605	0.0506	0.0459	0.093799
FM-Diversity [24]	0.0531	0.0473	0.0428	0.095407
Our Model	0.07831	0.08635	0.095	0.260

Table 4

Performance comparison of our diversity based explainable recommender system with NFM baselines. The best results are highlighted in bold

Models	Recall	NDCG	W-NDCG	Coverage
Random	0.0029	0.0031	0.0029	0.085024
NFM [30]	0.0651	0.0556	0.0501	0.088854
NFM-Fairco [31]	0.0626	0.0516	0.0470	0.093715
NFM-Diversity [24]	0.0522	0.0481	0.0438	0.097018
Our Model	0.07831	0.08635	0.095	0.260

Upon scrutinizing Table 3, which delineates the performance juxtaposition of our diversity-based explainable recommender system against FM baselines, it is evident that our model has significantly outperformed the others across all metrics.

Notably, our model manifests a remarkable enhancement in the Recall metric, registering a score of 0.07831, which is a substantial improvement compared to the FM baseline which stands at 0.0659. This indicates a heightened ability of our model to successfully recommend items that are indeed relevant to the users.

Furthermore, our model exhibits a superior performance in terms of the NDCG and W-NDCG metrics, with scores of 0.08635 and 0.095 respectively, showcasing its proficiency in ranking the recommended items in a manner that is more aligned with user preferences. Moreover, the Coverage metric of our model, recorded at 0.260, almost doubles that of the nearest competitor, FM-Diversity at 0.095407, illustrating the model’s expansive reach and its capacity to recommend a more diverse set of items.

Transitioning to Table 4, where our model is pitted against NFM baselines, a similar trend of superiority is observed. Our model continues to lead, echoing the performance demonstrated in Table 3. It maintains a consistent Recall score of 0.07831, which again surpasses the closest competitor, NFM-Fairco, which managed a score of 0.0626. This consistency in Recall metric across different baseline comparisons accentuates the robustness of our model.

The NDCG and W-NDCG metrics further corroborate the efficacy of our model, with scores mirroring those in Table 3, thereby establishing its supremacy in delivering more personalized and relevant recommendations. The Coverage metric, standing at 0.260, reinforces the model’s capability in offering a broader and more diverse range of recommendations, fostering an environment that mitigates the filter bubble effect significantly.

As we can see from Table 5 our model performs better in comparison to normal diversity based recommender system [32]. Notably, in this case normal NDCG metric is not used for analysis instead NDCG@K metric is used. Here the value of K is 20. The formula for NDCG@K is given by: $NDCG @ K = \frac{1}{z} \sum_{i = 1}^{k} \frac{2^{re l_{i}} - 1}{{log}^{2} (i + 1)}$ (10)

Table 5

Performance comparison of our diversity based explainable recommender system with diversity based recommender system where Shannon Entropy is used to calculate diversity. The best results are shown in bold

Model	NDCG@20
[32]	0.246
Our Model	0.30789

Hence, we can clearly see that the proposed model performs significantly better in terms of NDCG@K, showcasing its proficiency in giving users better top k recommendations which are both diverse and aligned to their preferences.

Moreover Figs. 2 and 3 give a comparative analysis of the model and the baselines (FM and NFM) for all the metrics. The explainable recommender model performs significantly better in all aspects, especially in NDCG and Coverage metrics.

Fig. 2

Comparative Analysis between proposed Model and the FM baselines Based on Key Performance Metrics.

Fig. 3

Comparative Analysis between proposed Model and the NFM baselines Based on Key Performance Metrics.

In conclusion, the results unequivocally demonstrate the better performance of the diversity-based explainable recommender system in comparison to both FM and NFM baselines. The substantial improvements in all the evaluated metrics underscore the high potential of the model, paving the way for more inclusive, diverse, and user-centric recommendations.

The performance dynamics of the model varies by the parameter ‘K’, which represents the number of top recommendations considered in the evaluation metrics of precision and recall. In the context of recommender systems, ‘K’ serves as a critical parameter that influences the granularity of the recommendations. A smaller value of ‘K’ implies a focus on the topmost recommendations, thereby emphasizing higher precision, while a larger ‘K’ value broadens the scope, potentially enhancing recall by considering a more extensive set of recommendations.

Figures 4 and 5 vividly illustrate the variations in recall and precision metrics at different ‘K’ values, providing a comprehensive view of how the model’s performance fluctuates with changes in ‘K’. This analysis is pivotal in fine-tuning the model to achieve a harmonious balance between precision and recall, thereby optimizing the overall effectiveness of the recommender system.

Fig. 4

Variation Of Precision values at Different values of Top K Recommendations.

Fig. 5

Variation Of Recall values at Different values of Top K Recommendations.

We can clearly see that as the value of K is increasing the Recall as it should is increasing, and the precision is decreasing since more items are recommended the proportion of relevant items to the total number of recommended items decreases.

Figure 6 shows a graph between Recall and Precision. The Recall vs Precision graph stands as a pivotal instrument, offering a detailed insight into the efficacy of our recommender system.

Fig. 6

Recall V/S Precision Graph for different values of Top-K recommendations.

As illustrated in the graph, at a K value of 10, we observe a recall of 0.05069 and a precision of 0.10767, indicating an initial high precision but lower recall rate. As we escalate the K value to 100, the recall notably improves to 0.1128, showcasing the system’s enhanced ability to identify relevant items, albeit at a slightly reduced precision of 0.03133. This graphical representation vividly encapsulates the trade-off between recall and precision at various thresholds, assisting us in identifying a balanced point where the system can offer recommendations that are both inclusive and accurate. Through this analytical lens, we are empowered to fine-tune our model, fostering a recommendation environment that is adept at catering to diverse user preferences while maintaining a commendable level of accuracy.

5 Conclusion and future work

In this study, we have innovatively utilized the CAFE model, a knowledge graph-based explainable recommender system, enhancing it with a newly developed diversity module. This module operates based on the principles of Shannon Entropy, a strategy that has propelled us forward in breaking the monotonous cycle of repeated suggestions, paving the way for a wider range of recommendations and thereby elevating the overall user experience.

Traditionally, explainable recommender systems have been heavily reliant on e-commerce data. However, we chose to break away from the norm, steering our focus towards the entertainment sector. By harnessing the Movie Lens 1M dataset to train our model, we have ventured into a relatively unexplored domain. This shift has opened up new avenues for personalization and explanation, especially in the realm of movie recommendations. Our approach not only promises a richer and more nuanced user experience but also fosters a harmonious relationship between the system and its users, building a foundation of trust and transparency.

Looking ahead, we plan to augment our recommender system further by introducing an explicit feedback loop. This enhancement is designed to create a more agile and responsive system, adept at adjusting to the changing preferences and behaviors of users. Additionally, we are eager to explore the potential of Natural Language Processing (NLP) techniques more deeply. This exploration aims to provide users with detailed insights into the recommendations, including a comprehensive analysis of filter bubbles. These forthcoming developments are set to amplify the system’s explainability, granting users a profound understanding of the underlying processes that steer the recommendation journey.

Moreover, we are committed to devising effective strategies to elucidate the concept of filter bubbles to users, nurturing a user community that is both informed and aware. This endeavor echoes our dedication to cultivating a more inclusive and expansive recommendation landscape, where users are not mere recipients of suggestions but active contributors to the recommendation process.

In conclusion, this study serves as a guiding light in the continuous journey to enhance the functionalities of recommender systems, setting the stage for a more inclusive, open, and enriched user experience. Through our pioneering approach, we aim to overcome the constraints of current systems, fostering a recommendation environment that offers users access to a wider array of information and insights. We are optimistic that our efforts will shape a future where recommender systems are not just tools offering suggestions but platforms encouraging informed decisions and a diversity of viewpoints.

References

Yikun Xian , et al., CAFE: Coarse-to-fine neural symbolic reasoning for explainable recommendation, Proceedings of the 29th ACM International Conference on Information & Knowledge Management, 2020.

Balloccu

Boratto

Fenu

Marras

, Post Processing Recommender Systems with Knowledge Graphs for Recency, Popularity, and Diversity of Explanations, In proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (2022), 646–656.

Yikun Xian , et al., Reinforcement knowledge graph reasoning for explainable recommendation, Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval, 2019.

Steffen Rendle , Factorization Machines with LibFM, ACM Transactions on Intelligent Systems and Technology 3(3) (2012), https://doi.org/10.1145/2168752.2168771

Beel Joeran Corinna Breitinger Bela Gipp Stefan Langer , Research-paper recommender systems: A literature survey, International Journal on Digital Libraries 17(4) (2018), 305–38.

Patrik Dokoupil Long-Term Fairness for Group Recommender Systems with Large Groups.” In RecSys 2022 – Proceedings of the 16thACMConference on Recommender Systems 2022. https://doi.org/10.1145/3523227.3547424.

Aya Sayed Yassine Himeur Abdullah Alsalemi Faycal Bensaali Abbes Amira , Intelligent Edge Based Recommender System for Internet of Energy Applications, IEEE Systems Journal 16(3) (2022), https://doi.org/10.1109/JSYST.2021.3124793

Felfernig

Friedrich

Zanker

Jannach

, Constraint-Based Recommender Systems, In Recommender Systems Handbook, Second Edition, 2015. https://doi.org/10.1007/978-1-4899-7637-6_5.

Aidan Hogan Eva Blomqvist Michael Cochez Claudia d’Amato Gerard de Melo Claudio Gutierrez José Emilio Labra Gayo Sabrina Kirrane Sebastian Neumaier Axel Polleres Roberto Navigli Axel-Cyrille Ngonga Ngomo Sabbir Rashid

Anisa Rula Lukas Schmelzeisen Juan Sequeda Steffen Staab Antoine Zimmermann , Knowledge Graphs. ArXiv 2003.02320 2020 (2020).

10.

Herlocker

Konstan

Riedl

, Explaining collaborative Fltering recommendations, In CSCW (2000), 241–250.

11.

Sharma

Cosley

, Do Social Explanations Work?: Studying and Modeling the Effects of Social Explanations in Recommender Systems, In WWW 13 (2013), 1133–1144.

12.

Qingyao Ai Vahid Azizi Xu Chen Yongfeng Zhang , Learning Heterogeneous Knowledge Base Embeddings for Explainable Recommendation, Algorithms 2018 (2018).

13.

Huang Yafan , et al., Path-enhanced explainable recommendation with knowledge graphs, World Wide Web 24(5) (1769), 1789.

14.

Tien Nguyen

Pik-Mai Hui

Maxwell Harper Loren Terveen Joseph Konstan

, Exploring the Filter Bubble: The Effect of Using Recommender Systems on Content Diversity, In WWW. ACM (2014), 677–686.

15.

Haim Mario, Andreas Graefe Hans Bernd Brosius , Burst of the Filter Bubble?: Effects of Personalization on the Diversity of Google News, Digital Journalism 6(3) (2018), 330–343. https://doi.org/10.1080/21670811.2017.1338145

16.

Mahsa Badami Olfa Nasraoui Patrick Shafto , PrCP: Prerecommendation Counter-Polarization, In KDIR (2018), 280–287.

17.

Seth Flaxman Sharad Goel Justin Rao

, Filter bubbles, echo chambers, and online news consumption, Public Opinion Quarterly 80(S1 (2016), 298–320.

18.

Flore Vancompernolle Vromman François Fouss , Filter bubbles created by collaborative filtering algorithms themselves, fact or fiction? An experimental comparison, ACM International Conference Proceeding Series 2021 (2021), 153–59. https://doi.org/10.1145/3498851.3498945.

19.

Pan Li Maofei Que Zhichao Jiang Hu

Y.A.O.

Alexander Tuzhilin , PURS: Personalized Unexpected Recommender System for Improving User Satisfaction, RecSys 2020 –14th ACM Conference on Recommender Systems (2020), 279–88. https://doi.org/10.1145/3383313.3412238

20.

Wenjie Wang Fuli Feng Liqiang Nie Tat Seng Chua , User-Controllable Recommendation Against Filter Bubbles. SIGIR 2022 – Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. Vol. 1. Association for Computing Machinery, 2022, https://doi.org/10.1145/3477495.3532075

21.

Gao Zhaolin Tianshu Shen Zheda Mai Mohamed Reda Bouadjenek Isaac Waller Ashton Anderson Ron Bodkin Scott Sanner , Mitigating the Filter Bubble While Maintaining Relevance: Targeted Diversification withVAEBased Recommender Systems, SIGIR 2022 - Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. Vol. 1. Association for Computing Machinery. 2022, https://doi.org/10.1145/3477495.3531890.

22.

Peizhe Cheng Shuaiqiang Wang Jun Ma Jiankai Sun Hui Xiong , Learning to Recommend Accurate and Diverse Items, In WWW IW3C2 (2017), 183–192.

23.

Jaime Carbonell Jade Goldstein , The use of MMR, diversity-based reranking for reordering documents and producing summaries, In SIGIR. ACM 1998, 335–336.

24.

Cai-Nicolas Ziegler Sean McNee

Joseph Konstan

Georg Lausen , Improving recommendation lists through topic diversification, In WWW. ACM 2005, 22–32

25.

Laming Chen Guoxin Zhang Hanning Zhou , Fast greedy map inference for determinantal point process to improve recommendation diversity, In NeurIPS 2018, 5627–5638.

26.

Sun Jianshan Jian Song Yuanchun Jiang Yezheng Liu Jun Li , Prick the Filter Bubble: A Novel Cross Domain Recommendation Model with Adaptive Diversity Regularization, Electronic Markets 32(1) (2022), 101–21. https://doi.org/10.1007/s12525-021-00492-1

27.

Hongwei Wang Fuzheng Zhang Mengdi Zhang Jure Leskovec Miao Zhao Wenjie Li Zhongyuan Wang , Knowledge-aware graph neural networks with label smoothness regularization for recommender systems, In KDD, 2019.

28.

Yixin Cao , et al., Unifying knowledge graph learning and recommendation: Towards a better understanding of user preferences, The World Wide Web Conference (2019).

29.

Steffen Rendle , Factorization Machines. In ICDM. IEEE, 2010, 995–1000.

30.

Xiangnan He Tat-Seng Chua , Neural Factorization Machines for Sparse Predictive Analytics. In SIGIR. ACM (2017), 355–364.

31.

MMorik Ashudeep Singh Jessica Hong Thorsten Joachims , Ashudeep Singh, Jessica Hong and Thorsten Joachims, Controlling Fairness and Bias in Dynamic Learning-to-Rank, In SIGIR. ACM 2020, 429–438.

32.

Farzad Eskandanian Bamshad Mobasher Robin Burke A clustering approach for personalizing diversity in collaborative recommender systems, Proceedings of the 25th Conference on User Modeling, Adaptation and Personalization 2017 (2022).

33.

Järvelin

Kekäläinen

, IR evaluation methods for retrieving highly relevant documents. In: Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval 2000 (2022), 41–48.

34.

Yan-Martin Tamm Rinchin Damdinov Alexey Vasilev , Quality metrics in recommender systems: Do we calculate metrics consistently?, Proceedings of the 15th ACM Conference on Recommender Systems (2021).