Improvement Graph Convolution Collaborative Filtering with Weighted Addition Input

Tin T. Tran, Václav SnáselACIIDS 2022

TL;DR

This paper proposes a novel Graph Convolutional Collaborative Filtering (WiGCN) model that enhances recommendation accuracy by incorporating a weighted influence matrix of users. By learning the varying degrees of influence users have on each other's preferences, WiGCN aims to capture richer collaborative signals than traditional methods.

What Problem Does This Paper Address?

Traditional collaborative filtering (CF) methods for recommendation systems often rely on user-item interaction matrices. While effective, these methods can overlook the nuanced social dynamics that influence user preferences. For instance, a user's decision to interact with an item might be influenced by the opinions or behaviors of their friends or other users whose tastes they value more. Existing graph-based methods, like GCN, can capture relationships, but they often treat all connections equally. This paper addresses the gap of not explicitly modeling the varying influence users have on each other, which can lead to a loss of valuable information in the recommendation process.

Methods

The core of this paper is the introduction of the Weighted input GCN (WiGCN) model. WiGCN builds upon existing Graph Neural Network (GNN) approaches for recommendation systems. The key innovation is the incorporation of a weighted influence matrix that captures the degree to which users influence each other. This is in contrast to standard GNNs that might treat all user-user relationships equally.

Here's a breakdown of the methodology:

  • Bi-partite Graph Representation: The system models user-item interactions as a bi-partite graph, where users and items are nodes, and interactions are edges.
  • Weighted User References Matrix ($W_u$): This matrix is derived from the user-item interaction matrix and captures the commonality of items interacted with by pairs of users. This implicitly represents a form of "social influence" or shared taste.
  • Weighted Influence Matrix ($\Delta$): This is the novel component. It's constructed using the weighted user references matrix and aims to explicitly model the varying influence users have on each other. The paper suggests this matrix acts as a "moderator of signals" during the message-passing process in the GNN.
  • Graph Convolutional Network (GCN) Architecture: The model utilizes a GCN framework to propagate information through the graph. This involves extracting signals from user-item interactions and aggregating them to refine user and item embeddings.
  • Propagation Process: The model employs a message-passing mechanism where signals are extracted from user-item pairs and then aggregated to update user representations. The weighted influence matrix is integrated into this signal extraction and aggregation process.
  • Loss Function: The paper uses Bayesian Personalized Ranking (BPR) loss, a common choice for implicit feedback datasets, to optimize the model.

Datasets

The proposed WiGCN model was evaluated on three well-known benchmark datasets:

  • Gowalla: A location-based social networking service dataset.
  • Amazon-book: A dataset containing product reviews and metadata from Amazon books.
  • Yelp2018: A dataset from Yelp, an online directory for local businesses.

The authors used a 10-core setting for all datasets, ensuring that each user had at least ten interactions. 20% of interactions were randomly selected as the test set, with the remaining 80% used for training.

Key Insights and Ideas

  • User Influence Varies: The core insight is that not all user-to-user relationships are equal in their impact on preferences. Explicitly modeling this varying influence can significantly improve recommendation accuracy.
  • Implicit Social Signals: Even without explicit social network data (like friendships), patterns in user-item interactions can reveal implicit social signals or shared tastes that can be leveraged. The weighted user references matrix captures this.
  • WiGCN's Advantage: By incorporating the weighted influence matrix, WiGCN can better capture the nuances of collaborative signals, leading to more accurate predictions compared to models that treat all interactions uniformly.
  • Propagation Effectiveness: The paper suggests that their proposed propagation process, informed by the weighted influence, is effective in reducing the loss faster than some baseline methods.

Key Results

The paper reports that their proposed WiGCN model generally outperforms state-of-the-art models like GCMC, NGCF, and LightGCN across the tested datasets in terms of precision, recall, and NDCG@20.

  • Overall Performance: WiGCN achieved better scores in precision, recall, and NDCG@20 on the Gowalla and Amazon-book datasets.
  • Yelp2018 Anomaly: On the Yelp2018 dataset, WiGCN's performance was comparable to or slightly worse than GCMC in some metrics. The authors suspect this might be due to Yelp2018's dataset structure, possibly involving user categorization that GCMC's decomposition approach handles well.
  • Loss Reduction: Figure 3 demonstrates that WiGCN reduces the loss faster per epoch compared to MF and NGCF, indicating a more efficient learning process.

Limitations and Future Work

  • Dataset Specificity: The authors acknowledge that the performance on Yelp2018 might be influenced by the dataset's specific characteristics, suggesting that further investigation into how different dataset structures affect the model is needed.
  • Tuning Parameters: Determining optimal parameters for the learning model and initializing embedding layers remains a challenge.
  • Future Work - Explicit Social Graphs: The paper suggests exploring the replacement of the implicit "trust weight graph" with explicit social relation graphs to see if it yields even better recommendation results.
  • Future Work - User Clustering: Investigating user clustering based on interests, habits, or behaviors could lead to more effective collaborative filtering.
  • Future Work - Data Preparation: The paper highlights the importance of data preparation and learning from multiple information sources, implying that future work could focus on more sophisticated data integration techniques.

Summary updated

BibTeX
@inproceedings{tran2022improvement,
  author = {Tin T. Tran and Václav Snásel},
  title = {{Improvement Graph Convolution Collaborative Filtering with Weighted Addition Input}},
  booktitle = {ACIIDS},
  year = {2022},
}
Figure