Shareh: explainable knowledge graph-based Arabic recommender system

KG-based XRSs

To represent user–item interactions, KG-based XRSs utilize semantic and relational data from KGs. These systems identify higher-order dependencies and provide explanations for recommendations. RippleNet (Wang et al., 2018), propagates the user’s preference from historical interests along the path in the KG; as a result, the need for manually constructed paths is reduced. It enhances explainability by monitoring user–item interactions, leading to notable improvements in recommendation accuracy across domains, including books, films, and news articles. The knowledge graph attention network (KGAT) (Wang et al., 2019a) integrates graph neural networks (GNNs) with an attention-driven embedding propagation mechanism to provide enhanced higher-order connectivity and nuanced user–item representations. KGAT improves accuracy and explainability by dynamically selecting useful relationships. Meanwhile, the knowledge-aware path recurrent network (KPRN) (Wang et al., 2019b) utilizes recurrent neural networks (RNNs) to represent sequential dependencies across entities and relations. It employs path-based reasoning and weighted pooling to choose explanatory paths based on relevance.

Rule-based Recommender System (RuleRec) (Lyu et al., 2020) integrates rule induction with neural recommendations to enhance path-based reasoning. It derives multi-hop relational patterns from KGs to provide visible and generalizable recommendations. The rule learning module first links items with associated entities in an external KG. Next, it summarizes explainable rules, which is in the form of meta-paths in the KG. Hierarchical Attention Graph Convolutional Network for Explainable Recommendation (HAGERec) (Yang & Dong, 2020) improves this concept by placing a hierarchical attention mechanism inside a heterogeneous KG; it facilitates bidirectional information flow and selective entity sampling to augment click-through rate prediction and top-K recommendations. The route-enhanced recurrent network (Huang et al., 2021) extends this enhancement by integrating an RNN encoder for sequential route semantics with an entropy-based encoder to effectively differentiate path significance. It outperforms other models, such as RippleNet and KPRN, across various ranking metrics.

Similarly, Ekar (Song et al., 2022) conceptualizes recommendation as a sequential decision-making process. It adopts reward optimization and long short-term memory (LSTM) networks to enhance user–item path encoding, resulting in considerable improvements in hit ratio relative to previous models. In the medical field, the multi-head graph attention network (Jin et al., 2023) integrates traditional Chinese medicine principles with contemporary pharmacological insights and employs meta-paths and GATs to enhance information propagation. Although these models enhance explainability in RSs, they depend heavily on KG embeddings. This dependence limits their capacity to contextualize user preferences through textual content.

Review-based XRSs

Review-based XRSs increase explainability and accuracy by using semantic insights obtained from user reviews. A notable method is Reliable Recommendation with Review-level Explanations (RRRE) (Lyu et al., 2021), which concurrently forecasts user ratings and review trust scores and eliminates faulty reviews to enhance recommendation robustness. The model employs an attention mechanism to emphasize high-quality evaluations and uses deep learning methods for concurrent user–item representation learning. Neural Explainable Recommender based on Attributes and Reviews (NERAR) (Liu et al., 2020) increases recommendation accuracy and explainability by integrating review- and attribute-based reasoning. It utilizes XGBoost for attribute extraction, Time-aware Gated Recurrent Unit (T-GRU) for modeling dynamic user preferences, and convolutional neural network (CNN) for item review analysis and incorporates an attention mechanism to emphasize contextually pertinent features.

Context-Aware Explanation based on Supervised Attention for Recommendation (CAESAR) (Li, Chen & Dong, 2021) improves context-aware RSs by offering dynamic explanations that respond to changing user settings, such as location and seasonality. It employs a multilevel attention mechanism and multitask learning to extract explicit contextual characteristics from user reviews and produce personalized explanations.

Despite these advancements, review-based XRSs still encounter challenges, such as textual inconsistencies, subjective biases in user reviews, and noise in review embeddings. These may affect recommendation quality and system stability.

Hybrid approaches

Hybrid methodologies in XRSs integrate the structured reasoning capabilities of KGs with the semantic depth of user-generated reviews for alleviating the limitations of each distinct technique. An example is the sentiment-aware knowledge graph (SAKG) (Park et al., 2022), which incorporates sentiment-aware reasoning into KG-based RS. SAKG is constructed by collecting sentiment indicators from user reviews and ratings and incorporating them into the KG framework as interaction (like or dislike). The model combines reinforcement learning and sentiment-aware policy learning to traverse sentiment-enhanced paths, thereby generating interpretable and user-centric recommendations.

Similarly, KANN (Liu & Miyazaki, 2023) enhances hybrid explainability by aggregating knowledge entities from user reviews and modeling semantic relationships with TransE-based KG embeddings. The model uses a dual-level attention mechanism. Specifically, outer attention focuses on explicit user–item relationships, and inner attention captures contextual interactions inside reviews. Although models such as SAKG and KANN attempt to integrate graphs with reviews, they still do not comprehensively capture user preferences.

By contrast, ERUR (Liu et al., 2023), Drs-KGRT (Liu & Zhao, 2023), and Graph Semantic Enhanced Knowledge-based Recommender System (GSE-Krs) (Spillo et al., 2023) increase recommendation accuracy by integrating KG data with user reviews. However, they fail to provide interpretable rationales, resulting in user ambiguity about the reasons for certain item recommendations. Drs-KGRT merges KG-based entity representations with Bidirectional Encoder Representations from Transformers (BERT) for semantic analysis, ERUR develops learned social graphs to improve user embeddings, and GSE-Krs integrates GNN-based embeddings with Transformer-based sentence encodings. They enhance robust recommendations but lack a definitive emphasis on explanation transparency.

The Shareh model overcomes these limitations by offering a robust hybrid methodology that merges KG-based reasoning with semantic insights obtained from reviews to produce entirely interpretable recommendations. Table 1 shows a comparative analysis of current XRSs and their detailed data sources and interpretability features.

User similarity relations

In Shareh, user similarity was computed based on the semantic similarity of their shared book reviews. Cosine similarity was applied to sentence embeddings generated by AraBERT, which encodes reviews into dense vector representations. This method captures nuanced semantic relationships and topical alignment, connecting users who discuss similar themes or contexts, irrespective of lexical overlap. Before the similarity scores were computed, the review dataset underwent preprocessing operations to reduce noise as follow:

• Removal of non-Arabic words—eliminating foreign terms and symbols.

• Filtering hashtags and hyperlinks—removing social media artifacts.

• Normalization—standardizing Arabic orthography (e.g., converting “أ“, “إ“, ”آ” to “ا“).

• Diacritics and elongation removal—stripping unnecessary vocalization.

• Punctuation and special character removal—excluding symbols, such as ?, (), !, @, #, and $.

• Summarization of long reviews—for reviews that exceed 250 tokens, the MT5 model was applied to generate concise summaries while preserving key semantic information.

We employed two sentence similarity approaches to compute user similarity:

• Fine-tuned AraBERT: In this approach, AraBERT was fine-tuned on the Mawdoo3 dataset (11,997 samples) to learn contextual similarity patterns and then evaluated on the MSRvid dataset (1,010 samples) to measure sentence similarity performance. This enabled the model to capture deep semantic relationships between reviews.

• AraBERT with Cosine Similarity: This approach uses AraBERT to build sentence embeddings, which are then compared using cosine similarity to measure similarity between sentences.

This dual evaluation facilitates a comprehensive assessment of sentence similarity and supports the selection of the most suitable model for constructing user–user relations within the knowledge graph. The comparison focused on F1-score, scalability, and computational efficiency to determine which approach provides the most accurate and reliable estimation of semantic similarity between users based on their review content. Table 5 summarizes the performance of both models by comparing their F1-scores and computation times. Although fine-tuned AraBERT exhibited enhanced accuracy, it incurred high computing expenses, making it impractical for extensive applications. Given the extensive dataset (156,500 evaluations from 4,957 books), AraBERT and cosine similarity were selected for Shareh because they considerably reduced the processing time while maintaining a commendable F1-score (0.92).

User similarity scores above 0.8 were retained to ensure robust and reliable connections. This threshold was selected based on preliminary experiments, which demonstrated that a value of 0.8 effectively balanced coverage and noise reduction, producing connections that were both meaningful and stable. The confidence score further refines user–user relationships by differentiating between levels of interaction activity. Users with high interaction activity (many shared reviews) tend to establish strong similarity links with other highly active users, as the abundance of evidence supports a high level of confidence in their shared preferences. Similarly, users with limited interactions can still form similarity relations among themselves. By contrast, links between highly active and sparsely active users naturally yield the weakest confidence scores. This stratification ensures that similarity connections in the graph are not only semantically meaningful but also aligned with the degree of user engagement, thereby improving the reliability of the constructed knowledge graph. Formally, the confidence score between two users’ u₁ and u₂ is defined as shown in Eq. (6):

(6) $$\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{f}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}(u_1,u_2)=\frac{{\displaystyle Numberofsharedreviews}}{max(numberofreviewswrittenby{u}_1,numberofreviewswrittenby{u}_2)}.$$

Based on preliminary experiments and multiple validation runs, a confidence threshold of ≥ 0.4 was adopted, as it provided the best balance between retaining sufficient coverage of user–user relations and minimizing noisy or unreliable connections. Lower thresholds introduced excessive noise by connecting weakly related users, whereas higher thresholds significantly reduced graph connectivity and coverage.

Explainable Arabic knowledge graph for shareh

ArBKG is a heterogeneous KG designed to capture relationships among books, authors, genres, publishers, and other book data derived from platforms, such as Arabic Goodreads (https://www.kaggle.com/datasets/z3rocool/arabic-goodreads-reviews) and Abjjad (https://www.abjjad.com/).

A sample of the constructed ArBKG is provided in Appendix A, The construction of a comprehensive KG for the Shareh model requires the integration of ArBKG with user relationship results obtained from the similarity model. Table 6 shows a comprehensive overview of the resultant explainable Arabic knowledge graph (XArKG), including its architecture and the incorporation of several relationship elements.

This framework guarantees multifaceted reasoning by integrating content-driven and collaborative filtering indicators. To capture higher-order relationships, XArKG uses meta-paths. These paths guide information propagation and reasoning, allowing the system to explain recommendations.

Figure 2 illustrates the structure of the resulting XArKG, explicitly depicting the relationships between users, items, authors, and genres. The graph integrates both content-based (genre, author) and collaborative (similarity, rating) relations, thereby supporting multifaceted reasoning and enhancing the explainability.

Performance metrics

We assessed performance using the metrics defined in section ‘Evaluation Metrics’ (MSE, RMSE, and MAE). Three experiments are being conducted to evaluate the influence of different information propagation strategies.

• Content-based graph modeling (using only item attributes).

• Collaborative-based modeling (leveraging user–user similarities).

• Shareh (combining content-based and collaborative signals).

Such comparison allowed us to assess the individual and combined effects of content and collaborative filtering and obtain insights into the optimal recommendation strategy. The results are summarized in Table 8. Among the approaches, content-based filtering alone produced the highest RMSE (0.7047), indicating that relying solely on the item attributes limited the predictive power of the model. Collaborative filtering remarkably enhanced performance and reduced MSE from 0.4966 to 0.2501, revealing the importance of user–user relationships in improving accuracy. Shareh ensures balance between content-based and collaborative filtering techniques. These results highlight the need for integrating content and user interaction data to develop robust, comprehensible RSs.

Comparison with related work

Table 9 presents a comparative evaluation of multiple Arabic recommender systems, all tested on the BRAD dataset, which is focused on book recommendations. Baseline values were taken directly from the original studies all of which used the same dataset. The results show that the Shareh model achieved the lowest RMSE (0.5001) and MAE (0.4998) among the compared methods. Previous models, including those developed by Meddeb, Maraoui & Zrigui (2021a) and Khalifeh & Al-Mousa (2021), reported RMSE values between 0.800 and 0.9928, with only one of them providing MAE results.

The improved performance of Shareh can be attributed to its unified integration of knowledge graph reasoning with semantic information extracted from user reviews. Unlike earlier approaches that treated reviews and structured data separately, Shareh constructs a combined graph that jointly represents books, user interactions, and user similarity relations. This design reduces sparsity and bias by leveraging both collaborative and content-based signals, while the meta-path-guided GAT enables the model to propagate information selectively, filtering noise and emphasizing semantically meaningful relations. These mechanisms collectively explain why Shareh achieves superior accuracy and interpretability.

Statistical analysis of pooling methods

We evaluated three pooling techniques, namely, mean pooling, sum pooling, and weighted sum pooling, to further refine the model’s aggregation strategy. One-way analysis of variance (ANOVA) was implemented to determine if the differences between the pooling methods were statistically significant (Park et al., 2023). The results, which are given in Table 10, reveal a statistically significant difference (p < 0.05) for MSE and RMSE, but not for MAE.

A post-hoc Tukey’s honestly significant difference (HSD) test was conducted to identify significant differences among the pooling methods. The results are shown in Table 11 and Fig. 3. The results showed that weighted sum pooling performed worse than mean pooling and sum pooling, with significantly higher MSE and RMSE (p-values < 0.05). The fact that the difference between mean pooling and sum pooling was not statistically significant (p-value = 0.0743) suggests that both methods performed similarly. Figure 3’s boxplots visually confirm that mean pooling consistently produces the lowest error values, proving its effectiveness in reducing prediction discrepancies.

Explainability recommendation cases

We assessed two recommendation instances from the system’s outcomes, as summarized in Table 12, to present the explainability features of the Shareh model. The examples demonstrated the system’s capability to offer varied, multifaceted justifications for recommendations by utilizing recognized meta-paths inside the explainable Arabic knowledge graph.

In Case 1 (Fig. 4), the system proposed a book (ID: 16031620) to a user (ID: 17198221) and recognized many meta-paths to substantiate the recommendation. The most significant contribution was from the M1 meta-path (importance: 0.5874), which reflected the user’s preference for particular authors. This importance score was dynamically computed through the GAT attention mechanism, where the cumulative product of attention weights along the path quantifies its overall contribution. The supplementary contributions comprised M2 (content-based, significance: 0.1856), which indicates genre inclinations, and M3 (collaborative, significance: 0.2270), which employs user–user relationships. The system prioritized the most significant meta-path when elucidating its advice to improve the clarity of its recommendations, thereby highlighting the content-based aspect of the filtering in this case.

Case 2 involved an alternative explanatory framework, where the system recommended a book (ID: 29381852) to a user (ID: 5843452) based exclusively on collaborative filtering indicators. A solitary meta-path (M3, significance: 1.0000) linked the user to the recommended book on the basis of only user–user similarities.

This differentiation between scenarios highlights the flexibility of the system in elucidating suggestions obtained from accessible user data. The system ensures that recommendations are accurate and comprehensible through content-based and collaborative filtering, and it customizes explanations for individual users in accordance with their interaction history and relevant contextual data.

Model fidelity

Model fidelity is important for assessing the explainability of recommendation systems because it quantifies the ratio of recommendations that are supported by explanations (Peake & Wang, 2018; Zhang & Chen, 2020). This study evaluated the reliability of the Shareh system to establish an interpretable rationale for its recommendations. Fidelity was defined as the proportion of recommendation cases that yielded a valid explanation and computed as described in Eq. (7):

(7) $$\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{l}\mathrm{F}\mathrm{i}\mathrm{d}\mathrm{e}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{Numberofsuccessfullyexplainedrecommendationpairs}{Totalnumberofrecommendationpairs}\times 100\mathrm{\%}.}$$

Our assessment, which employed empirical data, showed that explanations were effectively produced for 15,014 out of 15,050 user–item recommendation combinations, resulting in a faithfulness score of 99.76%. The system effectively provided explanations for nearly all recommendations, confirming its reliability and interpretability.