Abstract: With the wide application of the recommender system in various domains, the cold start problem has gained increasing attention in recent years. For the situation where only text comments of items are available, this paper proposes a set of solution and algorithm that can solve the cold start problem of new users. Firstly, the similarity between the article topic extraction and the feature extraction based on commodity evaluation is found and the latent Dirichlet allocation model is used to extract the features of items innovatively. Then the COLINBA algorithm is proposed which introduces the synergies of neighboring users based on the traditional Bandits strategy. This algorithm adds the neighbors’ similarity weight coefficient which determines the influence degree of the neighbors on the target user and makes the synergy more accurate and effective. The user features are updated according to the feedback of the user and the recommended item features. Finally, based on the two real world datasets, Delicious and Last.fm, this paper experimentally compares COLINBA to the state-of-the-art methods, and the results show that COLINBA offers a significant increase in recommendation performance.

Key words: recommender system, cold start, Bandits, collaborative filtering

摘要： 随着推荐算法在众多领域的广泛应用，冷启动问题得到了越来越多的关注。针对仅可获得老用户对商品文字评价的场景，提出了一套解决用户冷启动问题的方案与算法。首先通过分析发现了文章主题提取与基于商品评价提取特征的相似性，因此引入自然语言处理领域的LDA（latent Dirichlet allocation）生成模型提取商品潜在特征；然后在传统Bandits算法的基础上融入邻居用户的协同作用提出了COLINBA（collaborative filtering context linear Bandits）算法，该算法通过相似度权重因子控制邻居用户对推荐结果的贡献，使得协同作用更加精确有效，推荐完成后根据用户真实反馈以及所推荐商品的特征更新用户特征。最后采用真实数据集Delicious和Last.fm将该算法与该领域的最新方法进行比较，实验结果表明该算法对推荐效果有提升作用。

关键词: 推荐系统, 冷启动, 多臂赌博机, 协同过滤