关键词: 地图个性化推荐, 专题地图产品检索, 深度学习, 负采样, Item2Vec, CBOW模型, 用户事件行为, 隐性评分

Abstract:

Establishing a user preference recommendation model suitable for thematic map product search is one of the effective ways to improve the quality of the thematic map products. In the thematic map product recommendation scenario, there are serious problems of content cold-start and sparse comment data. The existing recommendation algorithms cannot recommend thematic map products with different features for specific types of users, resulting in users' limited preference for obtaining preference information from the thematic maps. Hence, this paper presents a user preference recommendation method based on the combination of CBOW with Negative Sampling and Iten2Vec based on Word2Vec. Firstly, calculating implicit ratings of the interaction behavior data in the user behavior log, to replace sparse user ratings in thematic disaster scenarios; Secondly, extracting context-aware feature information of central thematic map based on CBOW model with Negative Sampling. By controlling the ratio of positive and negative samples to 1:2, the prediction accuracy of the potential score of the target thematic map is improved; Finally, mapping Thematic CMaps with user behavior characteristics information to vector space via Item2Vec, calculating the user's similarity matrix to the thematic map and completing recommendations based on user preference. Test results on thematic map scoring experiment dataset Thematic CMaps and four validation dataset MovieLens show that, compared with the four traditional recommendation algorithm of LFM, Personal Rank, Content Based, and SVD, this proposed method can effectively improve the precision potential scoring, and the highest recommending performance is 27.85%. Compared with Item2Vec with Huffman sampling method and YouTubeNet two neural network recommendation algorithms, the score prediction accuracy has improved to a certain extent, and the recommendation performance has been continuously improved, reaching the maximums of 2.97% and 5.78%. Taking the singular value decomposition (SVD) of the classic algorithm as an example, in the increasing data subset after the segmentation of MovieLens-20M dataset, the score prediction accuracy and performance of the method used in this paper are better than SVD method.

Key words: map personalized recommendation, thematic map products retrieval, deep learning, negative sampling method, Item2Vec, CBOW model, user event behavior, implicit ratings