基于用户浏览轨迹的商品推荐方法研究

发布时间：2018-04-19 11:08

本文选题：个性化推荐 + 浏览轨迹　；参考：《北京化工大学》2016年硕士论文

【摘要】：随着电子商务网站迅速发展,推荐系统在这些网站中得到了广泛的应用。目前应用最广泛的个性化推荐算法是协同过滤推荐算法,但是此方法存在稀疏矩阵与冷启动问题。基于用户浏览记录来预测用户偏爱是缓解这些问题的一个重要研究方向。根据用户在电子商务网站的访问日志,提取出用户的浏览序列,即用户浏览轨迹,然后以用户浏览轨迹为基础,挖掘用户偏爱,并建立偏爱模型,为用户推荐偏爱商品,解决因为缺少历史购买及评分记录引起的冷启动问题。目前通过分析用户浏览轨迹为用户推荐商品的方法主要从浏览轨迹中商品与下一个商品关系角度出发考虑,而本课题从浏览轨迹中被浏览商品与最终被购买商品关系角度出发,提出两种推荐方法,一种基于购买转移关系,另一种基于商品特征趋势。基于购买转移关系的推荐方法,依据浏览轨迹中商品与最终被购买商品转移关系,同时考虑用户浏览轨迹与购买记录的时效性,对过时数据采用衰减策略,另外考虑浏览轨迹中商品的顺序,对轨迹中浏览距离不同转移关系应用不同的权重,构建购买转移概率模型,向用户推荐商品；基于商品特征趋势的推荐方法根据商品特征属性统计用户浏览轨迹中商品特征趋势,构建Markov特征趋势模型,当新用户在线浏览商品时,根据Markov特征趋势模型和用户当前浏览轨迹中可变商品特征预测用户偏爱商品的特征集,查找最符合这些特征的商品推荐给当前用户,该方法也同时考虑历史数据的时效性和特征集在浏览轨迹中的顺序性。实验证明,这两种方法相对于已有的基于用户浏览路径的方法有较好改进,都取得了比较好的推荐效果,提高了推荐算法的准确度与召回率,在一定程度上解决了新用户冷启动问题和新物品冷启动问题。
[Abstract]:With the rapid development of e-commerce websites, recommendation system has been widely used in these websites.At present, the most widely used personalized recommendation algorithm is collaborative filtering recommendation algorithm, but this method has the problem of sparse matrix and cold start.Predicting user preference based on user browsing records is an important research direction to alleviate these problems.According to the user's visit log in the E-commerce website, the user's browsing sequence is extracted, that is, the user's browsing track, and then based on the user's browsing track, the user's preference is excavated, and the preference model is established to recommend the preferred products for the user.Resolve cold startup problems due to lack of historical purchase and rating records.At present, the method of recommending goods for users by analyzing the user's browsing path is mainly considered from the perspective of the relationship between the goods in the browsing path and the next commodity, while this topic starts from the perspective of the relationship between the goods being browsed and the goods that are finally purchased.Two recommendation methods are proposed, one is based on purchase transfer relationship, the other is based on commodity feature trend.Based on the recommendation method of the purchase transfer relationship, according to the relationship between the goods in the browsing track and the final purchased goods, and considering the timeliness of the user's browsing track and purchase record, the attenuation strategy is adopted for the outdated data.In addition, considering the order of the goods in the browsing path, applying different weights to the different transfer relationships of the browsing distance in the trajectory, the model of purchase transfer probability is constructed, and the goods are recommended to the user.According to the commodity feature attribute, the Markov feature trend model is constructed according to the commodity feature attribute. When the new user browses the product online, it constructs the Markov feature trend model.According to the Markov feature trend model and the variable commodity feature in the current browsing path of the user, the feature set of the user's preference is predicted, and the products most suitable for these features are found to be recommended to the current user.The method also considers the timeliness of historical data and the sequence of feature sets in browsing trajectories.Experimental results show that the two methods are better than the existing methods based on user browsing path, and both have better recommendation effect, and improve the accuracy and recall rate of the recommendation algorithm.To a certain extent, the new user cold start problem and new items cold start problem.
【学位授予单位】：北京化工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】