融合多种数据信息的餐馆推荐模型

Restaurant Recommendation Model with Multiple Information Fusion

餐馆推荐可以利用用户的签到信息、时间上下文、地理上下文、餐馆属性信息以及用户的人口统计信息等挖掘用户的饮食偏好,为用户生成餐馆推荐列表.为了更加有效地融合这些数据信息,提出一种融合了多种数据信息的餐馆推荐模型,该模型首先利用签到信息和时间上下文构建“用户-餐馆-时间片”的三维张量,同时利用其他数据信息挖掘若干用户相似关系矩阵和餐馆相似关系矩阵;然后,在概率张量分解的基础上同时对这些关系矩阵进行分解,并利用 BPR 优化准则和梯度下降算法进行模型求解;最后得到预测张量,从而为目标用户在不同时间片生成相应的餐馆推荐列表.通过在两个真实数据集上的实验结果表明:相比于目前存在的餐馆推荐模型,所提出的模型有着更好的推荐效果和可接受的运行时间,并且缓解了数据稀疏性对推荐效果的影响.

Restaurant recommendation can leverage check-ins, time, location, restaurant attributes, and user demographics to dig user’s dining preference, and recommend a list of restaurants for each user. In order to fuse these data information more effectively, this study proposes a restaurant recommendation model with multiple information fusion. Firstly, this model constructs a three-dimensional tensor by using check-ins and time context, and digs some users’ similar relation matrices and restaurants’ similar relation matrices from additional data information. Secondly, these relation matrices and tensor are decomposed simultaneously. Then, Bayesian personalized ranking optimization criterion method (BPR Opt) and gradient descent algorithm are adopted to solve the model parameters. Finally, the proposed model generates a corresponding restaurant candidate list for target user at different time by calculating predicted tensor. A comprehensive experimental study is conducted on two real-world datasets. The experimental results not only validate the efficacy of the proposed model, which outperforms the current restaurant recommendation model and effectively alleviates influence of the data sparsity on recommendation performance, but also evaluate the efficiency of the proposed model, which has acceptable running time.