基于变分循环自动编码器的协同推荐方法

COLLABORATIVE RECOMMENDATION APPROACH BASED ON VARIATIONAL RECURRENT AUTO-ENCODER

基于概率矩阵分解的协同过滤是推荐系统中应用最广泛的方法。它通过学习用户-商品评分矩阵的两个低维近似矩阵来做推荐。但是在评分矩阵极其稀疏的情况下,概率矩阵分解的推荐准确性就会下降。为了缓解这个问题,提出一种基于变分循环自动编码器的概率矩阵分解方法,该方法综合考虑商品描述文本和评分矩阵,先将商品的描述文本编码成一个特征向量,然后将该特征向量融合到概率矩阵分解模型中来缓解稀疏问题。该方法在编码商品特征向量时,考虑了商品内容的上下文信息和语义信息,并且该特征向量服从高斯分布。在两个真实数据集上的验证结果表明:我们的模型与其他模型相比较,在评分矩阵极其稀疏的情况下,能更有效地预测用户感兴趣的商品列表,提高推荐准确性。

Collaborative filtering based on the probabilistic matrix factorization is widely used in the recommender system, which makes recommendation by learning two appropriate matrix based on the user-item rating matrix. However, when the rating matrix is sparse, the accuracy of probabilistic matrix factorization will decline. To address this problem, we propose a probabilistic matrix factorization based on the variational recurrent auto-encoder that takes both the rating matrix and content of item into account, it encodes the content information of item into a representation vector, and then combine vector with probabilistic matrix factorization method to ease the data sparseness. Consequently, our approach can capture then contextual and semantic information. Besides, the vector obeys the Gaussian distribution. Our evaluation on the two real dataset show that our approach outperforms the other models when the rating matrix is extremely sparse, our approach can improve the rating prediction accuracy.