Abstract:For modern recommender systems, the use of low-dimensional latent representations to embed users and items based on their observed interactions has become commonplace. However, many existing recommendation models are primarily designed for coarse-grained and homogeneous interactions, which limits their effectiveness in two critical dimensions. Firstly, these models fail to leverage the relational dependencies that exist across different types of user behaviors, such as page views, collects, comments, and purchases. Secondly, they struggle to capture the fine-grained latent factors that drive user interaction patterns. To address these limitations, we present a heterogeneous graph collaborative filtering model MixRec that excels at disentangling users' multi-behavior interaction patterns and uncovering the latent intent factors behind each behavior. Our model achieves this by incorporating intent disentanglement and multi-behavior modeling, facilitated by a parameterized heterogeneous hypergraph architecture. Furthermore, we introduce a novel contrastive learning paradigm that adaptively explores the advantages of self-supervised data augmentation, thereby enhancing the model's resilience against data sparsity and expressiveness with relation heterogeneity. To validate the efficacy of MixRec, we conducted extensive experiments on three public datasets. The results clearly demonstrate its superior performance, significantly outperforming various state-of-the-art baselines. Our model is open-sourced and available at: <a class="link-external link-https" href="https://github.com/HKUDS/MixRec" rel="external noopener nofollow">this https URL</a>.
What problem does this paper attempt to address?
This paper attempts to solve two key problems in modern recommendation systems:
1. **Modeling of multi - behavior interaction patterns**: Existing recommendation models are mainly designed to handle coarse - grained and homogeneous user behaviors (such as clicks, purchases, etc.), while ignoring the relationship dependencies between different types of user behaviors (such as page views, favorites, comments, purchases). This limits their ability to capture users' fine - grained preferences.
2. **Capturing of implicit intention factors**: Existing models have difficulty in capturing the fine - grained latent factors that drive user interaction patterns, that is, users' implicit intentions. These intentions may be the specific purposes of users in specific situations, such as shopping for a family gathering or buying when products are on sale.
To solve these problems, the author proposes a new heterogeneous graph collaborative filtering model - MixRec. This model better understands and utilizes users' different behavior patterns by introducing intention decoupling and multi - behavior modeling, and reveals the implicit intentions behind each behavior. Specifically, MixRec achieves this goal in the following ways:
- **Parameterized heterogeneous hypergraph architecture**: MixRec uses a parameterized heterogeneous hypergraph to encode users' behavior - aware latent intentions and generate factorization representations related to specific types of user - item interactions.
- **Contrastive learning paradigm**: A new contrastive learning paradigm is introduced to improve the model's expressiveness on sparse data and relational heterogeneity through self - supervised data augmentation.
To verify the effectiveness of MixRec, the author conducted extensive experiments on three public datasets, and the results show that MixRec significantly outperforms a variety of state - of - the - art baseline models.
### Formula summary
- **Node embedding update formula**:
\[
z^{(u/v)}_{i,k}=\sum_{j\in N_{i,k}}m_{i,j}\odot e_j,\quad\bar{z}^{(u/v)}_i = \sum_{k = 1}^K z^{(u/v)}_{i,k}
\]
- **Hypergraph adjacency matrix generation**:
\[
H^{(u)}_k=Z^{(u)}_k\cdot W^{(u)T}_k,\quad H^{(v)}_k = Z^{(v)}_k\cdot W^{(v)T}_k
\]
- **Hypergraph embedding propagation**:
\[
H^{(u/v)}_k=\delta(\tilde{H}^{(u/v)}\Gamma^{(u/v)}_k)=\delta(\tilde{H}^{(u/v)}\delta(\tilde{H}^{(u/v)T}Z^{(u/v)}_k))
\]
- **Node - level contrastive loss**:
\[
L_n=-\sum_{i = 1}^I\sum_{k = 1}^K\log\frac{\exp(s(eH_{i,k'},eH_{i,k})/\tau)}{\sum_{i'= 1}^I\exp(s(eH_{i',k'},eH_{i,k})/\tau)}
\]
- **Graph - level contrastive loss**:
\[
L_g=-\sum_k\log\frac{\exp(s(\bar{\Gamma}_{k'},\bar{\Gamma}_k))}{\exp(s(\bar{\Gamma}_{k'},\bar{\Gamma}_k))+\exp(s(\bar{\Gamma}'_{k'},\bar{\Gamma}_k))}
\]
Through these methods, MixRec can more accurately capture user intentions and provide better recommendation effects under complex behavior heterogeneity and sparse data conditions.