面向用户冷启动的动态推荐模型研究

The recommendation system can effectively improve user experience based on the user's historical records, thereby solving the long-tail problem and creating great commercial value. However, the system has to deal with a large number of new users without historical behavior and face serious cold-start problems. In the long-term research on cold-start, we found that the existing methods are mainly aimed at recovering the rating matrix, and lack of research on dynamic recommender systems with huge application value. For this reason, this proposal will focus on the user cold-start in the dynamic recommendation system and focus on three core problems. (1) For new users without information, it is difficult to produce effective personalized recommendations. It is proposed to establish a user cold-start dynamic recommendation model framework based on arrival time. (2) To solve the challenge of mining the relations between user attribute and new user interest, it is proposed to establish a user cold-start dynamic recommendation model framework based on user attributes. (3) Aiming at the challenge that the dynamic relationship between social relationships and new user interests is difficult to reason, it is proposed to establish a user cold-start dynamic recommendation model framework based on social relationships. Through the above three researches, an effective user cold-start theory and model framework for dynamic recommendation systems will be produced.

推荐系统能根据用户的历史记录有效提升用户体验，进而解决商品长尾问题，创造巨大商业价值。然而，系统每天要处理大量没有历史行为的新用户，面临严重的用户冷启动问题。课题组在长期的冷启动问题研究工作中发现现有方法都主要针对用户评分矩阵恢复问题，缺乏对具有巨大应用价值的动态推荐系统方面的研究。为此，本项目将针对动态推荐系统中的用户冷启动问题，围绕三个核心问题进行研究。(1)针对无附加信息的新用户难以产生有效的个性化推荐的问题，拟建立一套基于到达时间的用户冷启动动态推荐模型框架。(2)针对用户属性信息与新用户兴趣之间隐含关系挖掘困难的问题，拟建立一套基于属性的用户冷启动动态推荐模型框架。(3)针对社交关系与新用户兴趣之间的动态影响关系难以推理的问题，拟建立一套基于社交关系的用户冷启动动态推荐模型框架。通过以上三方面的研究，将产生一套有效针对用户冷启动的动态推荐理论与模型框架。

随着生产力的发展，用户面对的商品越来越多，导致难以从海量的商品中获得感兴趣的部分。推荐系统能通过分析用户的历史购买、浏览、社交关系、个性化属性等信息分析用户的潜在兴趣，同时分析商品的历史购买记录与描述信息获得商品的潜在特征，通过分析两者的匹配度将商品推送给用户。然而，用户的购买能力有限、浏览目的变化大、数据多源异构等原因导致推荐数据稀疏、噪声大、动态变化、信息融合困难等问题。.本课题主要从动态数据分析、信息融合、数据聚类等方面进行了研究：在数据动态分析方面分别研究了动态融合社交信息的推荐算法、基于路径的链路预测方法和基于膨胀卷积的环境声波分类；在数据融合方面研究了基于社交行为融合的自注意力推荐方法、基于关联的多模态数据分类方法、基于演化计算的模糊融合方法、模糊单调决策树模型；在数据聚类方面研究了基于样本稳定性的聚类集成方法。.项目相关成果形成一定影响力的理论与方法，发表SCI论文6篇，中文核心3篇，申请发明专利2项，获得软件著作权2项，形成软件系统2套，联合培养3名博士生（2名在读），2名硕士生毕业（1名在读）。