融合主题模型和卷积神经网络的APP推荐研究

APP Recommendation Study Based on Fusion Topic Model and Convolution Neural Network

[目的/意义]将主题模型与卷积神经网络进行结合,以实现APP的个性化推荐,并缓解评分数据稀疏性问题。[方法/过程]提出TMCNN模型,针对文本内容,通过用户和APP双通道的卷积神经网络获取卷积语义特征,同时使用LDA模型获取主题特征,并与用户和APP的数值特征组合,从而预测用户对APP的评分,进而推荐。[结果/结论]通过360手机助手数据集的测试,从RMSE,召回率,NDCG三个指标进行分析,TMCNN模型不仅具有良好的评分预测效果,而且APP的推荐结果也相对较好。同时,TMCNN模型也丰富了APP推荐的研究方法。[局限]没有考虑APP的权限信息,评论信息的有用性,以及TMCNN模型的优化函数有待改进。

[Purpose/significance]This paper combines topic model and convolutional neural network to achieve APP personalized recommendation and mitigate rating data sparsity issues.[Method/process] The paper proposes TMCNN model. For the text content,the paper obtains convolution semantic features through users and APP dual-channel convolution neural network. At the same time, LDA is used to acquire topic features. Then, the paper connects them with numerical features of users and APP. Finally,the paper uses these features to predict the user’s rating on APP,and then recommend.[Result/conclusion]Through the test on 360 mobile assistant dataset,TMCNN model does well in rating prediction and APP recommendation results from the perspective of RMSE,recall and NDCG indicators. Meanwhile,TMCNN enriches the research methods of APP recommendation.[Limitations]The paper does not consider permission information of APP,the usefulness of comment information,and the optimization function of TMCNN model also needs to be improved.