结合强化学习和用户短期行为的新闻推荐算法

NEWS RECOMMENDATION ALGORITHM COMBINING REINFORCEMENT LEARNING AND USER SHORT-TERM BEHAVIOR

针对传统的协同过滤推荐算法仅根据用户历史评分矩阵进行推荐,存在矩阵稀疏和无法动态观察用户兴趣变化的问题,提出一种将用户短期行为和强化学习相结合的新闻推荐方法。将新闻文本向量化后,通过聚类提取类别特征,再根据强化学习中的状态、动作和奖励等概念,以Double DQN算法为框架来建立推荐模型,利用循环神经网络近似动作值函数来进行计算。最后在财新网的真实新闻浏览数据集上对提出的算法进行验证,对比传统算法,实验结果表明,提出的算法在推荐准确率、召回率等指标上都有明显提高,能够更加有效地进行推荐。

The traditional collaborative filtering recommendation algorithm only makes recommendations based on the user history score matrix,which has the problems of sparse matrix and inability to dynamically observe user interest changes.A news recommendation method that combines user short-term behavior and reinforcement learning is proposed.After the news text was vectorized,the category features were extracted through clustering.Based on the concepts of state,action and reward in reinforcement learning,the Double DQN algorithm was used as the framework to establish a recommendation model,and the recurrent neural network was used to approximate the action value function for calculation.The proposed algorithm was verified on the real news browsing data set of Caixin.Compared with the traditional algorithm,the experimental results show that the proposed algorithm has significantly improved the recommendation precision rate,recall rate and other indicators,and can perform more effectively.