Abstract:

Due to the high sparsity of the user-POI（Point of Interest） check-in data, the traditional recommendation algorithm does not perform well in the recommendation of the point of interest. Therefore, a point of interest recommendation algorithm that combines text and image semantic information is proposed. In the recommendation process, the interpretability of user comments on ratings and the descriptiveness of image information on the appearance of points of interest are considered, and text and image assisted interest point recommendation is fully utilized. Firstly, the convolutional neural network is used to mine the comment text and image semantic information related to users and points of interest, and then construct the user-text semantic feature matrix, the interest point-image semantic feature matrix, and finally merge the user-interest rating matrix, based on the probability. Matrix decomposition constructs a unified recommendation model. Experiments show that the algorithm effectively alleviates the recommendation performance problem caused by the sparseness of the check-in data, and is superior to the mainstream method in terms of MAE（Mean Absolute Error） and RMSE（Root Mean Square Error）.

Key words: user reviews, image information, point of interest, matrix factorization, neural network

摘要：

用户-兴趣点签到数据的高度稀疏性让传统的推荐算法的推荐效果大打折扣。基于此，提出评论文本和图像语义信息融合的兴趣点推荐新算法。该算法同时考虑用户评论对评分数据的可解释性和图像语义信息对兴趣点外观的描述性，充分利用评论文本和图像数据辅助用户偏好特征和兴趣点属性特征的学习。使用神经网络抽取与用户和兴趣点相关的评论文本和图像语义特征，分别建模用户-文本语义特征关系、兴趣点-图像语义特征关系，将两种关系与用户-兴趣点评分矩阵进行融合，基于概率矩阵分解构建统一的推荐模型。在Yelp数据集上实验表明，该算法有效地缓解了签到数据稀疏性带来的推荐准确性问题，在MAE和RMSE两项指标上均优于主流方法。

关键词: 评论文本, 图像信息, 兴趣点推荐, 矩阵分解, 神经网络