Abstract: Since the traditional item-based collaborative filtering algorithms only consider the score of historical items when calculating the similarity between items, but they ignore the impact of historical item preferences, therefore, the recommended accuracy is not ideal. To solve this issue, this paper proposes a movie recommendation algorithm combining deep learning and attention mechanism. Firstly based on the implicit feedback obtained, on the feature-level attention frame, starting from the item content feature extraction network, the preference of item features is learned. Then the item feature preferences and project features are weighted to obtain the project content feature vector. Finally, in the item-level feature attention frame, it obtains the final recommendation results through the scores of the item preferences learned by the item content feature vector. The experimental results on MovieLens 100K and MovieLens 1M datasets demonstrate that the proposed algorithm has higher accuracy and recommendation personalization than the traditional algorithms and outperforms the state-of-the-art methods.

Key words: deep learning, neural networks, implicit feedback, attention mechanism, collaborative filtering

摘要： 传统基于项目的协同过滤算法在计算项目之间相似度时只考虑历史项目的评分，而忽略了历史项目偏好对其的影响，以至于推荐精度不够理想。针对此问题，提出了一种融合注意力机制的深度电影推荐算法。根据得到的隐性反馈，在特征级注意力框架上，从项目内容特征提取网络开始，学习项目特征的偏好；将项目特征偏好与项目特征加权得到项目内容特征向量；在项目级特征注意力框架中，通过项目内容特征向量学习对项目偏好的评分，从而产生最终的推荐结果。实验结果表明，提出的推荐算法在MovieLens 100K和MovieLens 1M两个公开数据集上的准确率和推荐个性化较传统算法均有不同程度的提高，表现出较为优越的推荐性能。

关键词: 深度学习, 神经网络, 隐性反馈, 注意力机制, 协同过滤