Figure 1.
Block diagram of the proposed method.
Table 1.
The list of Symbols.
| Symbol | Meaning |
|---|---|
| U | Set of users, |
| I | Set of items/products |
| ru,i | Ground-truth rating (1–5) given by user u to item i |
| R | Set of observed (u, i, ru,i) tuples |
| d | Embedding dimension for word vectors and hidden states |
| wt | Word embedding at position t in a review sentence (wt ∈ Rd) |
| T | Number of tokens in a review or aspect span |
| A = {a1, …, am} | Set of predefined or discovered aspects for a product category |
| su,i(a) | Sentiment score for aspect ‘a’ in user u's review of item i |
| Rating predicted by Bi-LSTM + attention sentiment module | |
| h | Hidden state |
| xv | Input feature vector for graph node v (user and item) |
| ξ ⊆ U × I | Set of observed user–item interactions |
| G = (V,ξ) | User-item bipartite graph; V = U ∪ I |
| Node embedding of v at GAT layer l | |
| ⊙ | Element-wise product |
| ∥ | Vector concatenation |
Figure 2.
Schematic diagram for BiLSTM-CRF.
Figure 3.
Bi-LSTM with attention network.
Table 2.
Dataset Structure.
| Field Name | Description |
|---|---|
| review_id | Unique ID for the review |
| product_id | Unique ID of the product (ASIN) |
| product_title | Title of the product |
| product_category | Product category (e.g., Electronics, Books) |
| star_rating | Rating score from 1 to 5 |
| helpful_votes | Number of helpful votes received |
| total_votes | Total number of votes (helpful + not helpful) |
| vine | Whether the review is part of the Vine program |
| verified_purchase | Indicates if the user purchased the product |
| review_headline | Review summary or title |
| review_body | Full review text |
| review_date | Date when the review was posted |
| customer_id | Anonymized unique user ID |
| marketplace | Marketplace where the review was posted (e.g., US) |
Table3.
Dataset Details.
| Category | # Reviews | Considered Reviews |
|---|---|---|
| Electronics | ~ 3,093,869 | 5000 |
| Books | ~ 10,319,090 | 5000 |
| Apparel | ~ 1,238,194 | 5000 |
| Automotive | ~ 1,421,419 | 5000 |
| Home | ~ 3,446,184 | 5000 |
| Beauty | ~ 2,020,564 | 5000 |
| Grocery | ~ 1,805,949 | 5000 |
| Tools | ~ 1,927,040 | 5000 |
| Software | ~ 341,296 | 5000 |
| Music | ~ 1,138,510 | 5000 |
Figure 4.
Rating distribution for 10 product categories.
Figure 5.
Confusion matrix for three classes.
Figure 6.
Loss vs. Epochs.
Figure 7.
Accuracy vs. Epochs.
Figure 8.
Training confusion matrix.
Table 4.
Precision and Recall for all category (Training).
| Category | Precision | Recall |
|---|---|---|
| Electronics | 0.9943 | 0.9930 |
| Books | 0.9917 | 0.9937 |
| Apparel | 0.9933 | 0.9890 |
| Automotive | 0.9930 | 0.9950 |
| Home | 0.9927 | 0.9933 |
| Beauty | 0.9937 | 0.9923 |
| Grocery | 0.9957 | 0.9937 |
| Tools | 0.9943 | 0.9940 |
| Software | 0.9917 | 0.9957 |
| Music | 0.9907 | 0.9913 |
| Average | 0.993 | 0.993 |
Figure 9.
Testing confusion matrix.
Table 5.
Precision and Recall for all category (Testing).
| Category | Precision | Recall |
|---|---|---|
| Electronics | 0.9861 | 0.9900 |
| Books | 0.9900 | 0.9890 |
| Apparel | 0.9930 | 0.9920 |
| Automotive | 0.9900 | 0.9900 |
| Home | 0.9920 | 0.9870 |
| Beauty | 0.9890 | 0.9920 |
| Grocery | 0.9940 | 0.9880 |
| Tools | 0.9871 | 0.9920 |
| Software | 0.9919 | 0.9840 |
| Music | 0.9792 | 0.9880 |
| Average | 0.9890 | 0.9890 |
Figure 10.
True and predicted rating with prediction error.
Figure 11.
MAE for 10 categories with average.
Figure 12.
RMSE for 10 categories with average.
Figure 13.
Hitrate vs. K.
Figure 14.
NDCG vs. K.
Table 6.
Scalability Analysis of the Proposed Framework Using Different Numbers of Reviews per Category (Mean ± SD over Five Independent Runs).
| Reviews per Category | Precision | Recall | RMSE |
|---|---|---|---|
| 5,000 | 0.9892 ± 0.0008 | 0.9892 ± 0.0009 | 0.42 ± 0.02 |
| 10,000 | 0.9893 ± 0.0007 | 0.9893 ± 0.0008 | 0.40 ± 0.02 |
| 20,000 | 0.9894 ± 0.0007 | 0.9894 ± 0.0007 | 0.38 ± 0.02 |
| 50,000 | 0.9895 ± 0.0006 | 0.9895 ± 0.0006 | 0.35 ± 0.01 |
| 100,000 | 0.9896 ± 0.0005 | 0.9896 ± 0.0005 | 0.32 ± 0.01 |
Table 7.
Ablation Study of the Proposed Framework.
| Model Configuration | Precision | Recall | RMSE |
|---|---|---|---|
| BiLSTM | 0.9526 ± 0.0041 | 0.9514 ± 0.0043 | 0.91 ± 0.04 |
| BiLSTM + CRF | 0.9639 ± 0.0033 | 0.9631 ± 0.0035 | 0.76 ± 0.03 |
| GAT | 0.9728 ± 0.0026 | 0.9720 ± 0.0027 | 0.64 ± 0.03 |
| GAT + BiLSTM | 0.9814 ± 0.0018 | 0.9808 ± 0.0019 | 0.53 ± 0.02 |
| GAT + BiLSTM + CRF (Proposed) | 0.9892 ± 0.0008 | 0.9890 ± 0.0009 | 0.42 ± 0.02 |
Table 8.
Comparison study e- commerce recommendation system.
| Author | Method | Dataset | Performances measured |
|---|---|---|---|
| Cai et al. [20] | Deep CGSR | Amazon e-commerce dataset | Accuracy-89% |
| Wang et al. [21] | DRS-TC | Trip Advisor dataset | RMSE-2.67 |
| Amazon review dataset | RMSE-0.49 | ||
| Elahi et al. [22] | Youtube Ranker and DFM | Video games dataset | Hit rate-4.01, 3.75; Precision-91%, 92% |
| Digital music dataset | Hit rate-14.32, 9.68; Precision-98.4%, 98.6% | ||
| Bellar et al. [23] | BERT & neural network models | Woman Clothing Reviews from Kaggle | Accuracy-93% |
| Shang et al. [24] | Sentiment aware neural collaborative filtering model | Amazon e-commerce dataset | MSE-3.79 |
| Karabila et al. [25] | BERT-collaborative filtering | Amazon e-commerce dataset | Accuracy-91% |
| Di et al. [26] | DGFedRS | Amazon e-commerce dataset | Accuracy-94% |
| Proposed method | Bi-LSTM + Attention | Amazon review dataset | RMSE-0.42, MAE-0.32, hit rate @ 10 =0.7, NDCG@10 =0.47 |