Improving One-class Recommendation with Multi-tasking on Various Preference Intensities

TL;DR

The paper tackles implicit-feedback one-class recommendation by addressing the unrealistic assumption of fixed user preference intensity. It introduces a multi-tasking framework that learns multiple entity representations corresponding to different preference intensities, enforced through per-task losses $L_l$ and aggregated as $L_{total} = \frac{1}{K} \sum_{l=0}^{K-1} L_l$, while an attentive graph convolutional layer captures high-order user-item relations. Representations across tasks are concatenated to form robust joint embeddings used for ranking, with the model evaluated on Gowalla, Amazon-book, and Yelp2018 showing significant gains in Recall@20 and NDCG@20 over strong baselines. The approach improves robustness and generalization in implicit-feedback recommendations by modeling intensity variability and dynamic neighbor influence, providing a practical uplift for real-world recommender systems.

Abstract

In the one-class recommendation problem, it's required to make recommendations basing on users' implicit feedback, which is inferred from their action and inaction. Existing works obtain representations of users and items by encoding positive and negative interactions observed from training data. However, these efforts assume that all positive signals from implicit feedback reflect a fixed preference intensity, which is not realistic. Consequently, representations learned with these methods usually fail to capture informative entity features that reflect various preference intensities. In this paper, we propose a multi-tasking framework taking various preference intensities of each signal from implicit feedback into consideration. Representations of entities are required to satisfy the objective of each subtask simultaneously, making them more robust and generalizable. Furthermore, we incorporate attentive graph convolutional layers to explore high-order relationships in the user-item bipartite graph and dynamically capture the latent tendencies of users toward the items they interact with. Experimental results show that our method performs better than state-of-the-art methods by a large margin on three large-scale real-world benchmark datasets.

Figures (2)

• Figure 1: Multi-tasking on various preference intensities in learned representation space. We learn multiple vector representation for $u_j$. They lie in the bar-shaped area, indicating different preference intensities.
• Figure 2: An overview of our model. We establish a lookup table to learn user and item vectors in $R_0$. Each attentive graph convolutional layer following it obtains a new set of representation $R_i$ that assumes stronger preference intensities. We calculate the BPR loss of each $R_i$ and average them to form the total loss. This requires each $R_i$ to satisfy the BPR loss constraints.