Beyond Similarity: Relation Embedding with Dual Attentions for Item-based Recommendation

TL;DR

This paper introduces REDA, a deep learning model that uses relation embeddings and dual attention mechanisms to improve item-based recommendations, especially in sparse data scenarios.

Contribution

REDA employs a novel relation embedding scheme and dual attentions to better capture user preferences and mitigate overfitting in recommendation systems.

Findings

REDA outperforms state-of-the-art methods on real-world datasets.

The dual attention mechanism effectively captures users' latent preferences.

REDA is especially effective with sparse data.

Abstract

Given the effectiveness and ease of use, Item-based Collaborative Filtering (ICF) methods have been broadly used in industry in recent years. The key of ICF lies in the similarity measurement between items, which however is a coarse-grained numerical value that can hardly capture users' fine-grained preferences toward different latent aspects of items from a representation learning perspective. In this paper, we propose a model called REDA (latent Relation Embedding with Dual Attentions) to address this challenge. REDA is essentially a deep learning based recommendation method that employs an item relation embedding scheme through a neural network structure for inter-item relations representation. A relational user embedding is then proposed by aggregating the relation embeddings between all purchased items of a user, which not only better characterizes user preferences but also alleviates the data sparsity problem. Moreover, to capture valid meta-knowledge that reflects users' desired latent aspects and meanwhile suppress their explosive growth towards overfitting, we further propose a dual attentions mechanism, including a memory attention and a weight attention. A relation-wise optimization method is finally developed for model inference by constructing a personalized ranking loss for item relations. Extensive experiments are implemented on real-world datasets and the proposed model is shown to greatly outperform state-of-the-art methods, especially when the data is sparse.