Multiview graph dual-attention deep learning and contrastive learning for multi-criteria recommender systems

TL;DR

The paper tackles the inadequacy of single-criterion recommender systems for items with multi-dimensional attributes by introducing D-MGAC, a framework that models multi-criteria ratings on a multi-edge bipartite graph using Dual Multiview Graph Attention. It combines local criterion-specific attention with global cross-view pooling and incorporates anchor-based local and global contrastive learning to produce robust embeddings for recommendations. Empirical results on Yahoo!Movies and BeerAdvocate demonstrate improved rating prediction accuracy (lower MAE/RMSE) and stability, with ablations confirming the importance of global attention and cross-criterion learning. While TripAdvisor data highlight challenges with noise and heterogeneity, the approach shows strong promise for scalable, nuanced multi-criteria recommendations and is made reproducible via public code and data resources.

Abstract

Recommender systems leveraging deep learning models have been crucial for assisting users in selecting items aligned with their preferences and interests. However, a significant challenge persists in single-criteria recommender systems, which often overlook the diverse attributes of items that have been addressed by Multi-Criteria Recommender Systems (MCRS). Shared embedding vector for multi-criteria item ratings but have struggled to capture the nuanced relationships between users and items based on specific criteria. In this study, we present a novel representation for Multi-Criteria Recommender Systems (MCRS) based on a multi-edge bipartite graph, where each edge represents one criterion rating of items by users, and Multiview Dual Graph Attention Networks (MDGAT). Employing MDGAT is beneficial and important for adequately considering all relations between users and items, given the presence of both local (criterion-based) and global (multi-criteria) relations. Additionally, we define anchor points in each view based on similarity and employ local and global contrastive learning to distinguish between positive and negative samples across each view and the entire graph. We evaluate our method on two real-world datasets and assess its performance based on item rating predictions. The results demonstrate that our method achieves higher accuracy compared to the baseline method for predicting item ratings on the same datasets. MDGAT effectively capture the local and global impact of neighbours and the similarity between nodes.

Figures (6)

• Figure 1: Single-Criterion recommender system
• Figure 2: Multi-Criteria recommender system
• Figure 4: In Stage 1, we construct a bipartite multi-edge graph based on users $\{ u_1, u_2, u_3 \}$, items $\{ i_1, i_2, i_3, i_4 \}$, and criteria $\{ c_1, c_2, c_3, c_4 \}$. Each edge in the graph represents the rating of an item based on a specific criterion. In Stage 2, we define a view for each criterion by constructing L-BGNN matrices, where each matrix captures the relationship between users and items based on the respective criterion. The matrix $B$ represents the user-item bipartite adjacency, while $B^T$ denotes the transposed relationship, connecting items back to users. Each $c_i$ corresponds to a specific criterion, highlighting how user ratings are influenced by various factors. In Stage 3, we generate embeddings using dual Multi-Graph Attention Networks (MGAT), which incorporate both local and global attention mechanisms. Stage 4 focuses on defining anchor points and applying Local and Global Contrastive Learning (CL). Stage 5 emphasises fusing the embeddings to generate recommendations. Finally, in Stage 6, we evaluate the proposed model by calculating the mean of error metrics, including the mean Absolute Error (MAE) and Root Mean Square Error (RMSE), alongside the standard deviation (Std) over 30 runs. Additionally, we predict item ratings using a Support Vector Machine (SVM) model based on the training and test sets.
• Figure 5: Evaluating the proposed model: first, the dataset is prepared and the embedding (for the whole dataset) is computed. Then, split the data into training and testing sets. Next, embeddings are extracted, and the SVR model is trained on the training set, followed by making predictions for both the training and testing sets. The model's performance is then evaluated using metrics such as RMSE, MAE, and standard deviation, calculated over 30 runs.
• Figure 6: The effect of varying $\lambda$ values on MAE for the Yahoo!Movies and BeerAdvocate datasets, under different parameter settings for $\alpha$ and $\beta$. The parameters $\alpha$ and $\beta$ represent weights for balancing different components of the model, with four configurations explored: $\alpha = 0.1, \beta = 0.1$, $\alpha = 0.5, \beta = 0.5$, $\alpha = 0.5, \beta = 0.1$, and $\alpha = 0.1, \beta = 0.5$. Each curve shows how the MAE varies as $\lambda$, the regularization weight, changes across multiple values from 0.2 to 0.9.

Theorems & Definitions (4)

• Definition 1: Multi-Criteria Recommender System
• Definition 2: Bipartite graph
• Definition 3: L-BGNN Adjacency Matrix
• Definition 4: Multiview Network Embedding