Abstract:Abstract In recent years, graph neural networks (GNNs) have been widely applied in recommender systems. However, existing recommendation algorithms based on GNNs still face challenges in node aggregation and feature extraction processes because they often lack the ability to capture the interactions between users and items, as well as users’ multiple intentions. This hinders accurate understanding of users’ needs. To address the aforementioned issues, we propose a recommendation model called multi-intent-aware interactive graph convolutional network (Multi-IAIGCN). This model is capable of integrating multiple user intents and adopts an interactive convolution approach to better capture the information on the interaction between users and items. First, before the interaction between users and items begins, user intents are divided and mapped into a graph. Next, interactive convolutions are applied to the user and item trees. Finally, by aggregating different features of user intents, predictions of user preferences are made. Extensive experiments on three publicly available datasets demonstrate that Multi-IAIGCN outperforms existing state-of-the-art methods or can achieve results comparable to those of existing state-of-the-art methods in terms of recall and NDCG, thus verifying the effectiveness of Multi-IAIGCN.
What problem does this paper attempt to address?
### Problems the Paper Attempts to Solve
This paper aims to address the challenges present in existing Graph Neural Network (GNN)-based recommendation algorithms during node aggregation and feature extraction processes. Specifically, current recommendation algorithms often lack the ability to capture interactions between users and items as well as the multiple intentions of users, leading to an inaccurate understanding of user needs. To overcome these issues, the authors propose a new recommendation model—Multi-Intention Aware Interaction Graph Convolutional Network (Multi-IAIGCN). This model integrates multiple user intentions and employs an interaction convolution method to better capture the interaction information between users and items.
### Main Contributions
1. **Multi-Intention Capture Module**: This module integrates and partitions multiple user intentions, mapping them to different intention trees. It uses an attention mechanism to accurately capture user intention needs, thereby eliminating the ambiguity of user intentions.
2. **Interaction Intention Aware Convolution Method**: This method enhances the node feature aggregation process by adding interaction signals between multiple user intention trees and item trees. It uses an interaction convolution method to locally aggregate features within a single user intention tree and the corresponding item tree.
3. **Scalable Experiments**: Experimental results show that Multi-IAIGCN outperforms other benchmark models in terms of recall rate and NDCG metrics, significantly improving recommendation performance.
### Related Work
1. **Collaborative Filtering**: Collaborative filtering is one of the most successful techniques in recommendation systems, predicting future interests based on users' interests in certain items. Existing methods include matrix factorization, deep neural networks, etc.
2. **Intention Recommendation**: A key issue in recommendation systems is effectively modeling user behavior and extracting sufficient embedding signals to recommend items that match user intentions. Existing methods extract user intention features through behavioral data and improve the interpretability of recommendation systems.
3. **Graph Convolution-Based Recommendation**: Although deep learning performs well on Euclidean data, it is limited when dealing with non-Euclidean data. Graph Convolutional Networks (GCN) excel in handling graph-structured data, but existing methods have limitations in aggregating high-order neighbor features.
### Model Architecture
1. **Embedding Module**: This module encodes historical user-item interaction data, mapping user interaction sequences to multiple graph structures and item interaction sequences to a single graph structure.
2. **Graph Convolution Module**: This module uses IA-GCN to perform convolution operations on user intentions and items, aggregating user intention features and item features using user-item interaction data.
3. **Intention Attention Module**: This module fuses the feature representations of multiple user intentions through an attention mechanism, reducing the ambiguity of user intentions.
4. **Prediction Module**: This module outputs the final prediction results, performing fully connected operations only on the user tree to avoid precision loss.
### Experimental Results and Analysis
1. **Datasets**: Experiments were conducted on three public datasets (Yelp2018, Gowalla, Amazon-Book), with 80% of each dataset used for training and the remaining part for testing.
2. **Evaluation Metrics**: Average Recall@20 and NDCG@20 were used as evaluation metrics.
3. **Hyperparameter Settings**: The model was implemented under the PyTorch framework, with a fixed embedding size of 64, using the Adam optimizer, a learning rate of 1e-4, and an L2 regularization coefficient of 1e-4.
4. **Baseline Models**: Comparisons were made with baseline models such as BPRMF, GC-MC, PinSage, GRU4REC, NGCF, and LightGCN.
### Conclusion
By introducing multi-intention awareness and interaction convolution methods, Multi-IAIGCN significantly improves the performance of recommendation systems, particularly in capturing multiple user intentions and enhancing recommendation accuracy.