E-Commerce Network Search System Based on Target Webpage Positioning and Sentiment Analysis Recommendation

Abstract

With the rise of e-commerce network search systems, product search efficiency and user satisfaction have become increasingly important. To address the low accuracy of consumer sentiment analysis in existing product recommendation scenarios, a webpage localization and sentiment analysis recommendation model is proposed that combines an improved web search algorithm with a bidirectional long short-term memory network and an attention mechanism. An e-commerce network search system is then designed around this model. Experimental results show that the sentiment analysis recommendation model achieves an accuracy of 98.88% and an average mean squared error of 1.027, outperforming all comparison models. The average root-mean-square error is 0.476, recall is 98.92%, the F1 score is 97.78%, and the recognition accuracy for each of the four emotional tendencies exceeds 95%. In addition, the integrated system delivers an average search time of 87.6 ms, a central processing unit occupancy of 44.68%, a missed-search rate of 1.42%, and a user satisfaction of 99.34%, all superior to the comparison systems. The system offers a ready-to-deploy solution for sentiment-aware product search and provides a theoretical basis for future research in e-commerce search systems.

Keywords

attention mechanism bidirectional long short-term memory network document object model e-commerce search system improved page ranking algorithm sentiment analysis recommendation

Introduction

The rapid growth of e-commerce has made cross-platform price comparison essential for consumers, yet existing search systems suffer from low accuracy in sentiment analysis for product recommendations.^1–2 Many experts have conducted relevant research, but the results are not ideal. N. Modi and J. Singh used eye-tracking technology to explore the impact of e-commerce graphic layout on consumer attention, providing a basis for understanding user browsing behavior, but did not involve improving search results or page positioning accuracy.³ In response to the problem of low efficiency in detecting logical vulnerabilities in the dialogue and communication system of e-commerce websites, W. Yu et al. proposed a modeling and analysis method for multiparticipant and multisession e-commerce interaction processes based on color web, but the effect was not satisfactory.⁴ S.K. Subramaniam et al. introduced an approach grounded in visual search and real-time personalized engine to tackle the challenge of instantly matching products in current e-commerce searches. Experimental results showed that this method outperformed comparative methods in terms of recall and recall rates.⁵ In response to the problem of low accuracy in predicting consumer behavior, S. Wei et al. introduced a technique for capturing user interests utilizing a multihead attention mechanism, with experiments demonstrating its effectiveness.⁶ Y. An et al. proposed a clustering framework based on Deep Clustering Network to solve the problem of difficult manual extraction of product advantages, disadvantages, and trend information in e-commerce reviews. The experimental results showed that the framework could accurately extract consumer opinions and predict demand trends.⁷

Accurate webpage localization and sentiment analysis are critical for improving online shopping efficiency. PageRank (PR) leverages link relationships with iterative optimization to enhance search quality,⁸ while bidirectional long short-term memory (BiLSTM) captures bidirectional context for improved sequence understanding in sentiment analysis.⁹ Many experts have conducted relevant research. For instance, V. Srinivas and P. Gowda proposed an improved PR algorithm to improve search engine accuracy. The performance of this algorithm was compared and analyzed, and the results showed that it was superior to the comparison algorithm.¹⁰ In response to the problem of low efficiency in user retrieval of scientific papers, A. Lubis and E. Prasiwenningrum proposed a PR-based visual search system. Comparative experimental outcomes revealed that the system had good search performance.¹¹ S. Murugaiyan and S.R. Uyyala proposed a hybrid model based on BiLSTM and deep convolutional neural network to accurately recognize human emotions. The model was tested on four benchmark datasets, and the results showed that the emotion classification accuracy of the model was above 90%.¹² In response to the problem of ignoring contextual information in emotion recognition, M. Wankhade et al. proposed a model that combined BiLSTM and transformer bidirectional encoder representation. Experimental results showed that this model performed better than the comparison model.¹³ Q. Zhao et al. proposed a sentiment analysis framework that combines specific parts of speech and BiLSTM to address the issue of neglecting word orientation in sentiment analysis of online e-commerce comment information. The framework was trained on three pre trained corpora, and the results showed that it had good robustness.¹⁴

Existing methods treat web page positioning and sentiment classification as separate tasks, leading to suboptimal performance: positioning algorithms ignore user emotions, while sentiment analysis disregards page relevance. This study proposes an integrated approach: (1) an improved PR method combining Shark-Search and term frequency–inverse document frequency (TF-IDF) for precise target page positioning, and (2) a BiLSTM-attention-based sentiment analysis model for recommendation. Finally, these components are integrated into an e-commerce search system to enhance purchasing efficiency and satisfaction. The novelty lies in the fusion of PR-Shark-Search for positioning and BiLSTM-attention for sentiment analysis.

Methods and Materials

Design of target webpage localization method for improved PR by integrating Shark-Search and TF-IDF

Current search systems suffer from low accuracy in sentiment analysis for product recommendations. This study addresses this by integrating an improved PR algorithm with TF-IDF and Shark-Search for target webpage positioning and a BiLSTM-attention-based sentiment analysis model. These components form an e-commerce search system designed to enhance recommendation accuracy. PR is a ranking algorithm based on the link relationship between webpages, which adjusts the page ranking results to improve the ranking of more important webpages in search results, thus improving the pertinence and caliber of search outcomes.^15–16 However, the PR algorithm suffers from irrelevant page links in thematic results.¹⁷ Shark-Search, a heuristic algorithm for focused crawlers, offers rapid response to webpage updates and efficient thematic relevance assessment, effectively compensating for PR’s deficiencies. TF-IDF is a statistical method widely used in information retrieval and text mining, characterized by strong anti-interference capabilities and simple computation.¹⁸ This study combines PR, Shark-Search, and TF-IDF to construct an improved target webpage positioning method. The workflow of this method is illustrated in Figure 1.

FIG. 1.

Detailed workflow of the improved PR algorithm combined with Shark-Search for target webpage positioning. PR, PageRank.

Figure 1a shows the process begins with initializing PR values and Shark-Search scores. PR values are calculated using the PR algorithm, while Shark-Search scores are computed based on webpage content and link relationships. The algorithm then calculates topic relevance scores and compares them against a threshold; pages exceeding the threshold are added to the topic queue. Convergence is checked next. If satisfied, the topic webpage queue is output; otherwise, iteration continues until the maximum count is reached. Figure 1b illustrates that the algorithm first employs TF-IDF to extract product feature words from webpage text and constructs a feature lexicon,¹⁹ then calculates feature word weights using equation (1).

ω_{i, j} = a f_{i, j} \times i d f_{i, j} = \frac{m_{i, j}}{\sum_{k} m_{k, j}} \times \log \frac{| D |}{| {j : a_{i} \in d_{j} |}

(1)

In equation (1), $a_{i}$ is the feature word, $d_{j}$ is the document, $a f_{i, j}$ is the word frequency of $a_{i}$ in $d_{j}$ , $ω_{i, j}$ is the weight of $a_{i}$ in the document, $m_{k, j}$ is the number of words $a_{i}$ appears in $d_{j}$ , $| D |$ is the total number of documents in the dataset, and $| {j : a_{i} \in d_{j} |$ is the number of documents containing $a_{i}$ . $i d f_{i, j}$ is the inverse document frequency. Second, the weight is normalized by equation (2).

ω_{i, j} = \frac{a f_{i, j} \times \log (| D | / | {j : a_{i} \in d_{j} | + 0.01)}{\sqrt{{\sum_{i = 1}^{m} [a f_{i, j} \times \log (| D | / | {j : a_{i} \in d_{j} | + 0.01)]}^{2}}}

(2)

Next, the Skip Gram model is used to generate a vector representation in a high-dimensional space for each feature word. The cosine similarity algorithm is used to calculate the semantic similarity of all feature words and topic words. Then, the Shark-Search algorithm and PR algorithm are combined to evaluate the topic relevance $Poten_Sco (child_url)$ of text and link dimensions in webpages, as shown in equation (3).

P oten_Sco (child_url) = E \times Inher (child_url) + F \times Nei (child_url) + G \times P R (child_url)

(3)

In equation (3), $E$ , $F$ , and $G$ are integers belonging to interval $[0, 1]$ and $E + F + G = 1$ . $Inher (child_url)$ is the relevance score inherited by the current node from its parent node, $Nei (child_url)$ is the anchor text of the current page sublink and the relevance between the anchor text context and the topic, and $P R (child_url)$ is the PR value of the current page sublink. Finally, it is determined whether the result of $Poten_Sco (child_url)$ is greater than the set threshold $χ$ . The determination of $χ$ can be expressed by equation (4).

χ = \frac{\sum_{i = 1}^{m} u r l_{i}}{m}

(4)

In equation (4), $u r l_{i}$ is the theme relevance score of the $i$ th link. If $Poten_Sco (child_url)$ is greater than $χ$ , the algorithm puts the webpage link into the theme webpage queue. On the contrary, the algorithm iterates again and continues selecting web links that are greater than $χ$ until the end.

Construction of information extraction method based on DOM and target webpage localization

To achieve accurate analysis of users’ emotions, it not only needs the target page positioning method to locate the webpage users want to visit but also needs to extract the information inside the located webpage. Therefore, the study designs an information extraction method based on the document object model (DOM). The extraction process of this method is shown in Figure 2.

FIG. 2.

Comprehensive process of information extraction based on webpage positioning (the icons sourced from https://iconpark.oceanengine.com/official).

Figure 2 illustrates the extraction process comprising webpage preprocessing, topic information template generation, and template information extraction. Preprocessing involves modifying special symbols and removing unrelated tags and code from target webpages located via an improved PR algorithm. For topic information template generation, DOM is employed to cluster information from target products such as computers.²⁰ DOM provides a tree-structured representation of HTML/XML documents with dynamic operation capabilities.²¹ The process utilizes a multidimensional semantic dictionary to annotate DOM semantics, detailed in Figure 3.

FIG. 3.

Multidimensional semantic DOM semantic annotation process. DOM, document object model.

As shown in Figure 4, the annotation process traverses each text node in the DOM tree to match attributes (e.g., “product name,” “delivery method,” “price”) against predefined sets. Successfully matched nodes are added to the information table. The effectiveness of information extraction is evaluated by the structural semantic entropy degree, as shown in equation (5).

B (Z) = - \sum_{i = 1}^{q} u_{i} l o g (u_{i})

(5)

FIG. 4.

Basic structure of BiLSTM and structure of the hybrid attention mechanism. BiLSTM, bidirectional long short-term memory.

In equation (5), $B (Z)$ is the semantic structure entropy, $u_{i}$ is the probability that a node belongs to a specific category, and $q$ is the number of annotated categories. If the calculated $B (Z)$ of a node is greater than the set threshold, the semantic topic information of that node will be saved to the database, and its path will be recorded using extensible markup language to construct a template. For information extraction, the system first checks the template library for existing matches; if absent, a new template is created using DOM-based webpage matching. Unmatched webpages are added to the sample set to continuously improve the template library.

Design of e-commerce network search system based on target webpage localization and sentiment analysis recommendations

After locating the target webpage and using the DOM algorithm to extract the targeted webpage information, it is necessary to accurately analyze the user comments and emotions contained in the extracted information to improve the accuracy of e-commerce network search. Therefore, the study first uses latent Dirichlet allocation (LDA) to integrate the topic words in consumer reviews with semantically similar topic words into a topic interface.²² LDA is an unsupervised learning probabilistic topic model that has advantages such as unsupervised learning ability and strong interpretability of probabilistic models. It is widely used in fields such as text classification and topic extraction.²³ LDA can automatically identify themes such as “price,” “quality,” and “logistics” from a large number of product reviews.²⁴ The number of topics is determined by perplexity $K H (O_{test})$ . As described later in equation (6), perplexity is used to determine the number of topics.²⁵

K H (O_{test}) = \exp (- \frac{\sum \log ρ (v)}{\sum_{l = 1}^{U} N_{l}})

(6)

In equation (6), $O_{test}$ is the test set, $ρ (v)$ is the marginal probability of word $v$ appearing in the document, $U$ is the total number of comments in the document, $N_{l}$ is the total number of vocabulary in the document, and $l$ is the document index. After integrating themes through the LAD model, the study uses BiLSTM and a hybrid attention mechanism for sentiment classification of themes. The basic structure and hybrid attention mechanism structure of BiLSTM are shown in Figure 4.²⁶

In Figure 4a, $S_{x - 1}$ , $S_{x}$ , and $S_{x + 1}$ of the input layer are the input data of time steps $x - 1$ , $x$ , and $x + 1$ . ${\overset{⃖}{h}}_{x - 1}$ , ${\overset{⃖}{h}}_{x}$ , and ${\overset{⃖}{h}}_{x + 1}$ of the backward LSTM layer are the backward propagating LSTM hidden states for capturing backward information in the sequence. ${\overset{⃖}{h}}_{x - 1}$ , ${\overset{⃖}{h}}_{x}$ , and ${\overset{⃖}{h}}_{x + 1}$ of the forward LSTM layer are the forward propagating LSTM hidden states for capturing reverse information in the sequence. $S_{x - 1}$ , $S_{x}$ , and $S_{x + 1}$ of the output layer are the output data of time steps $x - 1$ , $x$ , and $x + 1$ , respectively. From Figure 4a, BiLSTM is capable of simultaneously accessing information from both preceding and subsequent text when processing sequence data through two layers of LSTM, one processing time series from the past to the future, and the other processing time series that never came to the past. In Figure 4b, ${S r_{1}, S r_{2}, \dots, S r_{c}}$ is the set of eigenvectors, where $c$ is the number and $w^{x}$ is the weight. From Figure 4b, the mechanism process first calculates the contribution $QGGX$ of the emotional tendency of the subject words in the feature vector, and the calculation expression is shown in equation (7).

QGGX = attention ({S r_{1}, S r_{2}, \dots S r_{c}}, w^{x}) = \sum_{i = 1}^{c} b_{i} S r_{i}

(7)

In equation (7), $b_{i}$ is the probability distribution of word sharing degree in the feature vector sequence. $sim$ is the similarity function used to calculate the similarity between the topic word vector and the BiLSTM output sequence vector ${S r_{1}, S r_{2}, \dots, S r_{c}}$ . $Y$ is the parameter matrix, and $ϕ$ is the bias. Extracting topic words from the text and segmenting the text into sequences centered around these topic words, followed by applying the second and third attention calculations on both sides of the topic words, enables the capture of the contextual relevance of the topic words. The contextual representations of the topic words on both sides can be expressed using equation (8).

{\begin{matrix} L b = attention (({S r_{1}, S r_{2}, \dots, S r_{i - 1}}, w^{x})) \\ R b = attention (({S r_{i + 1}, S r_{i + 2}, \dots, S r_{c}}, w^{x})) \end{matrix}

(8)

In equation (8), $L b$ and $R b$ are the contextual representations to the left and right of the topic word. ${S r_{1}, S r_{2}, \dots, S r_{i - 1}}$ is the left sequence vector, and ${S r_{i + 1}, S r_{i + 2}, \dots, S r_{c}}$ is the right sequence vector. Second, the gate control mechanism is used to link the results of the attention mechanism, which can be represented by equation (9).

L X = QGGX ⊙ M K + M K_{L b} ⊙ L b + R b ⊙ M K_{R b}

(9)

In equation (9), $L X$ represents the connection result; $M K$ , $M K_{L b}$ , and $M K_{R b}$ represent the gating mechanism; and their calculation expressions are shown in equation (10).

{\begin{matrix} M K \propto \exp (J Z_{1} Q GGX + Y_{1} h_{x} + b_{1}) \\ M K_{L b} \propto \exp (J Z_{2} L b + Y_{2} h_{x} + b_{2}) \\ M K_{R b} \propto \exp (J Z_{3} R b + Y_{3} h_{x} + b_{3}) \end{matrix}

(10)

In equation (10), $Y_{1}$ , $Y_{2}$ , and $Y_{3}$ are parameter matrices, $J Z_{1}$ , $J Z_{2}$ , and $J Z_{3}$ are parameter matrices, $b_{1}$ , $b_{2}$ , and $b_{3}$ are bias vectors, and $M K + M K_{L b} + M K_{R b} = 1$ . Finally, the position attention mechanism is utilized to obtain the relative position information of each element in the sequence. The integration process of position attention and gating is shown in Table 1.

Table 1.

Pseudo-code of the gate unit integration process for position attention mechanism and BiLSTM

The integration process of location attention mechanism and BiLSTM gate control units
# Initialize BiLSTM parameters (forward & backward)for direction in [‘forward’, ‘backward’]: for gate in [‘f’, ‘i’, ‘c’, ‘o’]: # forget, input, cell, output gates W_{gate}_{direction} = initialize_weights() U_{gate}_{direction} = initialize_weights() b_{gate}_{direction} = initialize_bias() # Initialize hidden states and memory units h_{direction} = [zeros()] * T c_{direction} = [zeros()] * T for direction in [‘forward’, ‘backward’] # BiLSTM computationfor direction, step in [(‘forward’, range(T)), (‘backward’, range(T-1, −1, −1))]: for t in step: h_prev = h_{direction}[t-1] if direction == ‘forward’ else h_{direction}[t + 1] c_prev = c_{direction}[t-1] if direction == ‘forward’ else c_{direction}[t + 1] f = sigmoid(W_f_{direction} @ h_prev + U_f_{direction} @ input_sequence[t] + b_f_{direction}) i = sigmoid(W_i_{direction} @ h_prev + U_i_{direction} @ input_sequence[t] + b_i_{direction}) c_tilde = tanh(W_c_{direction} @ h_prev + U_c_{direction} @ input_sequence[t] + b_c_{direction}) c_{direction}[t] = f * c_prev + i * c_tilde o = sigmoid(W_o_{direction} @ h_prev + U_o_{direction} @ input_sequence[t] + b_o_{direction}) h_{direction}[t] = o * tanh(c_{direction}[t]) # Merge forward and backward hidden states bi_hidden_states = [concatenate(h_forward[t], h_backward[t]) for t in range(T)] # Position attention mechanism position_encodings = [position_encoding(t) for t in range(T)] attention_scores = [dot_product(bi_hidden_states[t], position_encodings[t]) for t in range(T)] attention_weights = softmax(attention_scores) context_vector = sum([attention_weights[t] * bi_hidden_states[t] for t in range(T)]) # Final output final_output = some_activation_function(context_vector)

The integration process of location attention mechanism and BiLSTM gate control units

# Initialize BiLSTM parameters (forward & backward)for direction in [‘forward’, ‘backward’]:
for gate in [‘f’, ‘i’, ‘c’, ‘o’]: # forget, input, cell, output gates
W_{gate}_{direction} = initialize_weights()
U_{gate}_{direction} = initialize_weights()
b_{gate}_{direction} = initialize_bias()
# Initialize hidden states and memory units
h_{direction} = [zeros()] * T
c_{direction} = [zeros()] * T for direction in [‘forward’, ‘backward’]
# BiLSTM computationfor direction, step in [(‘forward’, range(T)), (‘backward’, range(T-1, −1, −1))]:
for t in step:
h_prev = h_{direction}[t-1] if direction == ‘forward’ else h_{direction}[t + 1]
c_prev = c_{direction}[t-1] if direction == ‘forward’ else c_{direction}[t + 1]
f = sigmoid(W_f_{direction} @ h_prev + U_f_{direction} @ input_sequence[t] + b_f_{direction})
i = sigmoid(W_i_{direction} @ h_prev + U_i_{direction} @ input_sequence[t] + b_i_{direction})
c_tilde = tanh(W_c_{direction} @ h_prev + U_c_{direction} @ input_sequence[t] + b_c_{direction})
c_{direction}[t] = f * c_prev + i * c_tilde
o = sigmoid(W_o_{direction} @ h_prev + U_o_{direction} @ input_sequence[t] + b_o_{direction})
h_{direction}[t] = o * tanh(c_{direction}[t])
# Merge forward and backward hidden states
bi_hidden_states = [concatenate(h_forward[t], h_backward[t]) for t in range(T)]
# Position attention mechanism
position_encodings = [position_encoding(t) for t in range(T)]
attention_scores = [dot_product(bi_hidden_states[t], position_encodings[t]) for t in range(T)]
attention_weights = softmax(attention_scores)
context_vector = sum([attention_weights[t] * bi_hidden_states[t] for t in range(T)])
# Final output
final_output = some_activation_function(context_vector)

Thus, the construction of the hybrid attention mechanism is completed. By combining the results obtained from BiLSTM and LDA with the hybrid attention mechanism, a sentiment analysis recommendation model is constructed, as shown in Figure 5.

FIG. 5.

Sentiment analysis recommendation model integrating BiLSTM and LDA with hybrid attention mechanism (the icons sourced from https://iconpark.oceanengine.com/official). LDA, latent Dirichlet allocation.

The model first preprocesses review texts and LDA-extracted topic words, including cleaning, noise reduction, and sentiment annotation. These are transformed into topic word vectors and text vectors via word vector matrix. BiLSTM then extracts sentiment features from these vectors, which are fed into a hybrid attention mechanism to capture relationships between topic words and contextual information, yielding refined sentiment representations. Finally, a fully connected layer with Softmax determines sentiment orientation. Based on these processes, an e-commerce search system integrating target webpage positioning and sentiment analysis recommendation is constructed, with functional architecture shown in Figure 6.

FIG. 6.

Functional architecture of e-commerce search system (the icons sourced from https://iconpark.oceanengine.com/official).

Figure 6 illustrates the system architecture comprising three modules: query processing, sentiment analysis recommendation, and webpage localization with information extraction. The query module receives user requests and processes them via Solr full-text search. The sentiment analysis module employs BiLSTM-LDA-attention integration to analyze user sentiment and recommend relevant products. The webpage positioning module uses an improved PR algorithm for target page location, followed by DOM-based extraction of topic-relevant information. These modules collaborate to identify highly relevant pages, extract information, and generate sentiment-based recommendations.

Results

Performance analysis of sentiment analysis model

To evaluate the proposed BiLSTM-LDA-attention model, experiments were conducted using bidirectional encoder representations combined with convolutional neural network (BERT-CNN), ERNIE and recurrent convolutional neural network (ERNIE-RCNN), and convolutional neural network-gated recurrent unit (CNN-GRU). These models were selected for their proven effectiveness in NLP tasks and architectural diversity—BERT-CNN for deep bidirectional encoding, ERNIE-RCNN for entity-enhanced representation, and CNN-GRU for sequential processing. All models were trained under identical conditions on a combined dataset from Amazon (publicly available at https://www.heywhale.com/mw/dataset/5a698cdeafceb51770d610b4) and scraped JD.com/Taobao computer sales pages (2000 pages, 68,791 entries), spanning January 2023 to December 2024 with balanced category distribution (40% electronics, 30% clothing, 20% household, 10% others).

Key experimental settings: LSTM hidden/state dimensions 128, learning rate 0.001, 300 iterations, 300-dim vectors, Adam optimizer, batch size 25, max sentence length 100, attention weight dimension 128, L2 regularization 0.001, dropout 0.5, LDA perplexity 7. Fivefold cross-validation was employed for parameter tuning.^27–28 Data preprocessing included deduplication, missing value removal, tag cleaning, and min–max normalization. Evaluation metrics comprised accuracy, precision, recall, F1 score, mean squared error (MSE), and root mean square error (RMSE). The experimental environment is detailed in Table 2.

Table 2.

Experimental environment

Parameter names	Parameter
Processor	Intel Core i9-10900K
Main frequency	5.3 GHz
Internal memory	32 GB
Hard disk capacity	500 GB
GPU	NVIDIA GeForce RTX 3080
Matlab version	Matlab 2021a
Deep-learning framework	PyTorch 2.2.0 + CUDA 12.1 + cuDNN 8.9
Python version	Python 3.13.1
Data analysis software	Spss26.0

In the above environment, the study first conducted comparative experiments on the average accuracy and average precision of each model, and the experimental results are shown in Figure 7.

FIG. 7.

Comparison results of precision rate and accuracy rate. (a) Average precision comparison results; (b) Average accuracy comparison results.

As shown in Figure 7, the proposed model achieved average precision of 98.88% and accuracy of 97.69%, outperforming BERT-CNN (92.14%, 94.21%), ERNIE-RCNN (89.59%, 90.02%), and CNN-GRU (87.09%, 86.84%). This superior performance stems from the BiLSTM-attention integration, which captures key emotional cues more precisely to provide high-quality inputs for recommendations. The MSE and RMSE comparisons are shown in Figure 9.

As shown in Figure 8, the proposed model achieved the lowest MSE (1.027) and RMSE (0.476), compared to BERT-CNN (5.821, 2.746), ERNIE-RCNN (4.473, 2.489), and CNN-GRU (2.874, 1.423). These lower error rates reflect the model’s superior ability to precisely analyze sentiment, supporting reliable recommendations. The F1 scores and recall rates are compared in Figure 9.

FIG. 8.

Comparison results of average MSE and average RMSE. MSE, mean squared error; RMSE, root mean square error. (a) MSE comparison results; (b) RMSE comparison results.

Figure 9 shows the proposed model achieved 98.92% recall and 97.78% F1 score, surpassing BERT-CNN (93.75%, 94.26%), ERNIE-RCNN (88.94%, 84.86%), and CNN-GRU (91.24%, 86.58%). The high recall reflects the model’s strong ability to identify relevant documents, while the F1 score indicates balanced precision and recall—both critical for reliable sentiment analysis in recommendation systems. Table 3 presents the statistical analysis results verifying experimental reliability.

FIG. 9.

Comparison results of F1 scores and recall rates. (a) Recall comparison results; (b) F1 score comparison results.

Table 3.

Statistical analysis results

Indicator	Research (%)	BERT-CNN (%)	ERNIE-RCNN (%)	CNN-GRU (%)	p Value
Average precision rate	98.88 ± 0.021	92.14 ± 0.012	89.59 ± 0.021	87.09 ± 0.031	<0.001
Average accuracy rate	97.69 ± 0.017	94.21 ± 0.021	90.02 ± 0.011	86.84 ± 0.014	<0.001
Average MSE	1.027 ± 0.026	5.821 ± 0.015	4.473 ± 0.017	2.874 ± 0.019	<0.001
Average RMSE	0.476 ± 0.014	2.746 ± 0.022	2.489 ± 0.016	1.423 ± 0.017	<0.001
Recall rate	98.92 ± 0.011	93.75 ± 0.027	88.94 ± 0.014	91.24 ± 0.022	<0.001
F1 points	97.78 ± 0.019	94.26 ± 0.018	84.863 ± 0.013	86.58 ± 0.012	<0.001

BERT-CNN, bidirectional encoder representations combined with convolutional neural network; CNN-GRU, convolutional neural network-gated recurrent unit; ERNIE-RCNN, ERNIE and recurrent convolutional neural network.

From Table 3, the p value of each experiment was less than 0.001, indicating that the performance difference between the proposed model and other models was statistically significant. The error range was extremely small. The experimental results were reliable and trustworthy. A total of 1247 reviews were selected for Dell XPS 14 computer, and positive, negative, neutral, and other emotions were analyzed and recommended for testing. The accuracy of sentiment analysis recommendations is shown in Figure 10.

FIG. 10.

Results of recommendation accuracy in sentiment analysis. (a) Research; (b) BERT-CNN; (c) ERNIE-RCNN; (d) CNN-GRU.

The proposed sentiment analysis recommendation system achieved accuracies of 97.9%, 98.4%, 98.3%, and 98.6% across the four sentiment categories (Fig. 10a), while the best comparison models BERT-CNN, ERNIE-RCNN, and CNN-GRU reached only 91.8%, 90.8%, and 89.1%, respectively (Fig. 10b–d). Notably, the lowest accuracy of the proposed model exceeds the highest accuracy of any baseline, demonstrating superior and more consistent performance across all emotional categories.

Performance analysis of e-commerce network search system

After validating the performance of the sentiment analysis recommendation model proposed in the study, a performance comparison analysis was conducted on the proposed e-commerce network search system. The comparative system includes an e-commerce search system based on BERT-CNN, an e-commerce search system based on ERNIE-RCNN, and an e-commerce search system based on the CNN-GRU model. The comparison results of search speed and CPU occupancy among various systems are shown in Figure 11.

FIG. 11.

Comparison results of search speed and CPU usage. (a) Search speed comparison results; (b) Comparison results of CPU usage rates.

Figure 11a shows the proposed system achieved 87.6 ms average search time, lower than BERT-CNN (127.7 ms), ERNIE-RCNN (118.3 ms), and CNN-GRU (138.2 ms). Figure 11b indicates its CPU occupancy was 44.68%, surpassing BERT-CNN (54.76%), ERNIE-RCNN (49.88%), and CNN-GRU (58.97%). These results demonstrate the proposed system’s superior performance in both search speed and CPU occupancy compared to all baseline systems. The comparison of the missed search rate and user satisfaction of each system is shown in Figure 12.

FIG. 12.

Comparison results of missed search rate and satisfaction rate. (a) Research; (b) BERT-CNN; (c) ERNIE-RCNN; (d) CNN-GRU.

Figure 12a shows the proposed system achieved 1.42% missed search rate, lower than BERT-CNN (5.34%), ERNIE-RCNN (7.47%), and CNN-GRU (9.88%). Figure 12b indicates its satisfaction rate reached 99.34%, surpassing BERT-CNN (93.27%), ERNIE-RCNN (87.22%), and CNN-GRU (81.38%). These results demonstrate the proposed system’s superior performance in both missed search rate and user satisfaction.

Discussion

This study conducted a performance comparison analysis between the proposed sentiment analysis recommendation model and the e-commerce network search system. The proposed sentiment analysis model achieved 98.88% precision, 97.69% accuracy, 98.92% recall, and 97.78% F1 score, with an MSE of 1.027 and an RMSE of 0.476. These metrics significantly outperform BERT-CNN, ERNIE-RCNN, and CNN-GRU, demonstrating the effectiveness of the BiLSTM architecture in capturing contextual dependencies for emotion recognition. This finding aligns with Agbesi et al.,²⁹ who similarly reported improved accuracy through BiLSTM enhancement. Furthermore, the notably lower error rates suggest that the hybrid attention mechanism successfully directs focus toward sentiment-bearing phrases, refining recommendation precision. This observation is consistent with Chen et al.,³⁰ confirming attention mechanisms’ utility in fine-grained sentiment analysis.

Regarding system performance, the proposed approach achieved 87.6 ms average search time alongside 44.68% CPU occupancy, 1.42% missed search rate, and 99.34% user satisfaction. The reduced search time indicates that integrating TF-IDF with Shark-Search and PR algorithms optimizes webpage location efficiency without sacrificing result quality, corroborating Liu et al.³¹ Meanwhile, the high fitting coefficient and low missed search rate reflect the DOM algorithm’s accuracy in topic extraction. The comprehensive performance gains—spanning retrieval speed, resource utilization, and recommendation accuracy—validate the synergistic design of combining improved PR-based positioning with LDA-BiLSTM-attention sentiment analysis, a conclusion supported by Shang et al.³² The main limitation of this study is the four-category emotion classification, which may not fully capture the complexity of real-world emotional expression, such as subtle distinctions between surprise and disappointment. Future work can explore multimodal sentiment analysis by integrating image or audio data or incorporate pretrained Transformer models like BERT for more fine-grained classification. In addition, the computational complexity of the proposed model requires optimization for resource-constrained environments through techniques such as pruning, distributed computing, or knowledge distillation to ensure scalability and real-time performance.

Conclusion

To address low accuracy in sentiment analysis for e-commerce search, this study proposes an improved PR algorithm integrating TF-IDF and Shark-Search for webpage positioning, combined with DOM-based topic extraction. A sentiment analysis recommendation model employing LDA, BiLSTM, and hybrid attention mechanism is also constructed. Experimental results demonstrate the proposed model achieves 98.88% precision, 1.027 MSE, 0.476 RMSE, and 97.78% F1 score, significantly outperforming BERT-CNN, ERNIE-RCNN, and CNN-GRU. The integrated search system achieves 87.6 ms average search time, 44.68% CPU occupancy, 1.42% missed detection rate, and 99.34% satisfaction rate, surpassing all comparison systems. These results confirm the proposed approach effectively improves e-commerce search efficiency and user satisfaction.

The innovation of this research lies in three aspects. First, an improved PR algorithm integrating Shark-Search and TF-IDF was proposed for webpage positioning, considering both link relationships and content relevance to improve accuracy and efficiency. Second, a sentiment analysis recommendation model combining BiLSTM with attention mechanism was established to better capture emotional information in text. Finally, these components were unified into an e-commerce search system that simultaneously ensures high-precision webpage positioning and fine-grained sentiment recommendation, unlike existing methods that treat them as independent tasks.

The Innovative Technical Advantage Mechanism of TF-IDF + Shark Search + PR Combination and the Synergy Effect Among Each Module

The integration of TF-IDF, Shark Search, and PR combines text content understanding, webpage positioning optimization, and search quality assurance into a unified solution for e-commerce recommendation systems. The system comprises three complementary modules: query processing, sentiment analysis recommendation using BiLSTM-LDA-attention, and webpage location with improved PR for topic-relevant information extraction. This collaboration enables simultaneous consideration of link structure and user emotions, outperforming traditional methods that neglect either structural information or emotional tendencies.

Footnotes

Author’s Contributions

Y.-C.H.: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing—original draft, Writing—review & editing.

Author Disclosure Statement

The author declares that there is no conflict of interest.

Funding Information

The research is supported by the 2024 Research Project on Higher Education in China, AI-empowered Curriculum Reform in Universities: Exploring a New Task-based Foreign Language Teaching Model, (No. 24WY0302).

Abbreviations Used

References

1. Sanbella

, Van Versie

, Audiah

. Online marketing strategy optimization to increase sales and e-commerce development: An integrated approach in the digital age. SABDA 2024;3(1):54–66; doi: 10.33050/sabda.V3i1.492

2. Salem

, Aimeur

, Hage

. “A multi-party agent for privacy preference elicitation. Proc. AIA 2023;1(2):98–105; doi: 10.47852/bonviewAIA2202514

3. Modi

, Singh

. Understanding online consumer behavior at e-commerce portals using eye-gaze tracking. International Journal of Human–Computer Interaction 2023;39(4):721–742; doi: 10.1080/10447318.2022.2047318

4. Yu

, Liu

, Wang

, et al. Modeling and analyzing logic vulnerabilities of e-commerce systems at the design phase. IEEE Trans Syst Man Cybern, Syst 2023;53(12):7719–7731; doi: 10.1109/TSMC.2023.3299605

5. Subramaniam

. AI-based e-commerce search optimization. JEC 2025;4(6):307–316; doi: 10.5281/zenodo.15767084

6. Wei

, Zhang

, Yang

, et al. A click conversion rate model of e-commerce platforms aiming at effective data sparse. Ieee TETC 2024;8(2):1744–1755; doi: 10.1109/TETCI.2024.3356909

7. An

, Kim

, Lee

, et al. Topic Modeling-Based Framework for Extracting Marketing Information from E-Commerce Reviews. IEEE Access 2023;11(1):135049–135060; doi: 10.1109/ACCESS.2023.3337808

8. Yang

, Wang

, Wei

, et al. Efficient algorithms for personalized pagerank computation: A survey. IEEE Trans Knowl Data Eng 2024;36(9):4582–4602; doi: 10.48550/arXiv.2403.05198

9. Dhaka

, Nagpal

. WoM-based deep BiLSTM: Smart disease prediction model using WoM-based deep BiLSTM classifier. Multimed Tools Appl 2023;82(16):25061–25082; doi: 10.1007/s11042-023-14336-x

10.

10. Srinivas

, Gowda

. A page rank-based analytical design of effective search engine optimization. IJ-AI 2025;14(1):73–82; doi: 10.11591/ijai.V14.I1.Pp73-82

11.

11. Lubis

, Prasiweningrum

. Implementation of PageRank algorithm for visualization and weighting of keyword networks in scientific papers. J Appl Data Sci 2023;4(4):382–391; doi: 10.47738/jads.V4i4.138

12.

12. Murugaiyan

, Uyyala

. Aspect-based sentiment analysis of customer speech data using deep convolutional neural network and BiLSTM. Cogn Comput 2023;15(3):914–931; doi: 10.1007/s12559-023-10127-6

13.

13. Wankhade

, Annavarapu

CSR

, Abraham

. “MAPA BiLSTM-BERT: Multi-aspects position aware attention for aspect-level sentiment analysis. TJSC 2023;79(10):11452–11477; doi: 10.1007/s11227-023-05112-7

14.

14. Zhao

, Mo

, Fan

. POS-ATAEPE-BiLSTM: An aspect-based sentiment analysis algorithm considering part-of-speech embedding. Appl Intell 2023;53(22):27440–27458; doi: 10.1007/s10489-023-04952-3

15.

15. Dongwei

, Jin

, Zhen

. Reliability modeling for failure correlations to a CNC lathe subsystem based on Pagerank algorithm. Quality Eng 2024;36(2):336–349; doi: 10.1080/08982112.2023.2208664

16.

16. Shi

, Chen

, Cheng

, et al. PB-Raft: A byzantine fault tolerance consensus algorithm based on weighted PageRank and BLS threshold signature. Peer-to-Peer Netw Appl 2025;18(1):1–16; doi: 10.1007/s12083-024-01876-8

17.

17. Peng

, Ouyang

, Gui

, et al. Equipment fault propagation path identification based on unstable points detection. IEEE Trans Ind Inf 2024;20(2):1596–1606; doi: 10.1109/TII.2023.3280309

18.

18. Yilahun

, Hamdulla

. “Entity extraction based on the combination of information entropy and TF-IDF,”. IJRIS 2023;15(1):71–78; doi: 10.1504/ijris.2023.128371

19.

19. Silalahi

, Ginting

. Rekomendasi Berita Berkaitan dengan Menerapkan Algoritma text mining dan TF-IDF. bulletincsr 2023;3(4):276–282; doi: 10.47065/bulletincsr.V3i4.266

20.

20. Zuo

, Samangouei

, Zhou

, et al. FMGS: Foundation model embedded 3D gaussian splatting for holistic 3D scene understanding. Int J Comput Vis 2025;133(2):611–627; doi: 10.1007/s11263-024-02183-8

21.

21. Li

, Lu

, Du

. Deciphering DOM-metal binding using EEM-PARAFAC: Mechanisms, challenges, and perspectives. Environ Sci Pollut Res Int 2024;31(10):14388–14405; doi: 10.1007/s11356-024-32072-z

22.

22. Wheeler

, Cohen

, Wang

. A comparison of latent semantic analysis and latent Dirichlet allocation in educational measurement. J Educ Behav STA 2024;49(5):848–874; doi: 10.3102/10769986231209446

23.

23. Zrelli

, Rejeb

, Abusulaiman

, et al. Drone applications in logistics and supply chain management: A systematic review using latent Dirichlet allocation. Arab J Sci Eng 2024;49(9):12411–12430; doi: 10.1007/s13369-023-08681-0

24.

24. Batool

, Byun

. Enhanced sentiment analysis and topic modeling during the pandemic using automated latent Dirichlet allocation. IEEE Access 2024;12:81206–81220; doi: 10.1109/ACCESS.2024.3411717

25.

25. Rejeb

, Rejeb

, Appolloni

, et al. Navigating the landscape of public–private partnership research: A novel review using latent Dirichlet allocation. IJPSM 2025;38(2):213–237; doi: 10.1108/IJPSM-09-2023-0263

26.

26. Ma

, Huang

, Wang

, et al. Prediction of remaining parking spaces based on EMD-LSTM-BiLSTM neural network. J Traffic Transp ENG 2025;12(1):201–213; doi: 10.1016/j.Jtte.2023.01.004

27.

27. Woźniak

, Wieczorek

, Siłka

. “BiLSTM deep neural network for imbalanced medical data of IoT systems. FGCS 2023;141(1):489–499; doi: 10.1016/j.future.2022.12.004

28.

28. Naheliya

, Redhu

, Kumar

. Bi-directional long short term memory neural network for short-term traffic speed prediction using gravitational search algorithm. Int J ITS Res 2024;22(2):316–327; doi: 10.1007/s13177-024-00398-0

29.

29. Agbesi

, Chen

, Ukwuoma

, et al. Multichannel 2D-CNN attention-based BiLSTM method for low-resource ewe sentiment analysis. JDSIS 2023;3(1):67–77; doi: 10.47852/bonviewJDSIS32021512

30.

30. Chen

, Lu

, Wang

. Self-attention mechanism prior to modality fusion for multimodal sentiment analysis. Multimedia SYST 2025;31(4):1–14; doi: 10.1007/s00530-025-01868-5

31.

31. Liu

, Wang

, Zhong

, et al. A new focused crawler using an improved Tabu search algorithm incorporating ontology and host information. Front Inform Technol Electron Eng 2023;24(6):859–875; doi: 10.1631/FITEE.2200315

32.

32. Shang

, Shi

, et al. Enhancing e-commerce recommendation systems with deep learning-based sentiment analysis of user reviews. IInt. J. Eng. Manag. Res 2024;14(4):19–34; doi: 10.5281/zenodo.13221409