Abstract: Using user behavior data, adopting effective recommendation methods, and offering individualized recommendation methods are the strategy adopted generally by social network platforms, while the effectiveness of recommendation methods is the key that decides the quality of recommendation services. Methods based on matrix decomposition and methods based on collaborating filter are difficult to be promoted and applied on a large scale due to such bottlenecks as difficulty in sparsity and over-fitting. Based on the research on similarity and association between neighboring behaviors in the user behavior sequence, this paper digs the TextRank of internal structural relationship among words, and puts forward a new user behavior recommendation method by incorporating word2vec. Analysis and experiment results show that the new recommendation method is better than the traditional recommendation methods and is improved in each index, which verifies the validity and accuracy of the new method.

Key words: word2vec, recommendation system, non-textual sequences, user behavior, TextRank

摘要： 利用用户行为数据，采用有效推荐方法，提供个性化推荐服务是社交网络平台普遍采用的策略，其中推荐方法的有效性是决定推荐服务质量的关键。基于矩阵分解和基于协同过滤的推荐方法由于存在稀疏性和过拟合问题等瓶颈难以大规模推广应用。在研究用户行为序列中相邻行为之间相似性和关联性的基础上，挖掘词语之间内部结构关系的TextRank，融合word2vec提出新的用户行为推荐方法。分析与实验结果表明：该推荐方法较传统推荐方法，在各项指标上都得到了一定的提升，验证了该方法的有效性和准确性。

关键词: word2vec, 推荐系统, 非文本化序列, 用户行为, TextRank