基于内容和兴趣漂移模型的电影推荐算法研究

摘要

针对基于内容的推荐算法中,内容相似度计算精度低、用户兴趣漂移等问题,提出一种结合影评内容相似度和长短期兴趣模型来计算电影相似度的推荐方法.算法利用TextRank、Word2Vec等技术和模型对影评进行关键词抽取和词向量构建,同时基于Word2Vec训练结果进行电影内容相似度计算,一定程度上解决了近义词、网络词等带来的准确率下降问题;然后基于长短期兴趣漂移模型,统计用户对不同内容属性的偏好权重,并随时间窗口动态计算电影相似度矩阵,缓解了用户兴趣随时间漂移而改的问题;最后根据不同推荐策略获得推荐结果.实验结果证明,该算法比对比方法正确率提高了5％左右,同时兴趣模型提取了用户长短期兴趣标签,在工业界及基于标签的算法等场景中都具有很高的实用价值.%Aiming at the problem of content-based recommendation algorithm,such as low accuracy of content similarity and interest drift of users,this paper proposed a recommendation method combing similarity of movie reviews plus short-term and long-term user interest model to calculate the movie similarity.This method used TextRank,Word2Vec and other models to extract the keywords and built the word vector.At the same time,it used the training results of the Word2Vec to calculate the movie content similarity,in this way,the low accuracy caused by the synonyms and Internet vocabulary could be solved to a certain extent.And then,based on the long-short-term interest drift model,it calculated the weights of user's preference content,and dynamically built movie similarity matrix according to the time window,to alleviate the problem of user interest drifting with time.Finally,it obtained the recommended results according to different recommendation strategies.The experimental results show that the proposed method improvcds the accuracy about 5％.Meantime,the interest model extracting the longshort-term interest tags of users,has high practical value in industry and also in the tag-based algorithms.