一种结合矩阵分解和深度学习技术的POI推荐模型

A POI Recommendation Model Combining Matrix Factorization and Deep Learning Technology

摘要:

基于位置社交网络(Location-based social network, LBSN)的兴趣点(Point-of-Interest,POI)推荐算法是近年来的研究热点,有效的POI推荐具有极大的经济和社会效益。针对LBSN中的数据稀疏问题、用户细粒度兴趣(即用户的长期和短期兴趣)序列建模问题和联合多种影响因素的POI推荐问题,研究一种结合矩阵分解和带有注意力机制深度学习技术的POI推荐模型(CF-ADNN)。一方面,构造特征矩阵缓解签到数据稀疏问题,通过矩阵分解得到隐藏因子,计算POI的特征向量；另一方面,构建一种带注意力机制的用户细粒度兴趣的序列建模方式,有效学习用户长期和短期的兴趣特征,提高POI推荐精确度；结合上述两种方法,最终得到可以融合多种影响因素的POI推荐模型。并通过对比试验,进一步验证模型的推荐效果。

Abstract:

Point-of-Interest (POI)recommendation algorithm based on the Location-based social network (LBSN) is a hot research topic in recent years. Effective POI recommendation has great economic and social benefits. The POI recommendation model (MF-ADNN) which combines matrix factorization and deep learning with attention mechanism is studied, aiming at the data sparsity problem in LBSN, user’s fine-grained interest (i.e. long-term and short-term interest) sequence modeling problem and the POI recommendation problem of combining various influencing factors. On the one hand, feature matrix is constructed to alleviate the problem of sparse check-in data, and hidden factors are obtained by matrix factorization to calculate POI feature vectors; On the other hand, construct a sequence modeling method of user's fine-grained interest with attention mechanism, effective Learning long-term and short-term interest characteristics of users and improving the accuracy of POI recommendation. Combining the above two methods, the POI recommendation model that can integrate a variety of influencing factors is finally obtained. Through the comparison test, the recommended effect of the model is further verified.