Faced with the explosive growth of mobile applications, how to recommend mobile applications accurately and efficiently for users to choose their desirable and interesting mobile applications, which has become a challenging issue nowadays. To solve this problem, we propose a topic attention mechanism and FMs based mobile application recommendation method. Firstly, it uses LSA to obtain the global topic of mobile application description text. Then, the local semantic representations of mobile application are trained by BiLSTM model. Secondly, as for the global topic information and local semantic information in the content representation of mobile application description text, attention mechanism is performed to distinguish the contribution degree of different words and gain their weight values. Thirdly, the classification and prediction of mobile application are completed by using the softmax activation function through a full connection layer. Finally, based on user’s searching requirement, it exploits factorization machines to combine the various features of the classified mobile applications to rank and recommend the user’s expected mobile application with higher predicted score. The evaluation is conducted on a real and open dataset Mobile App Store, and the experimental results indicate that the performance of the proposed approach is better than other baseline methods in terms of precision, recall, F1-score, MAE, RMSE, and AUC.

中文翻译：

---

基于主题关注机制和分解机的移动应用推荐方法

面对移动应用的爆炸性增长，如何准确，有效地推荐移动应用，以使用户选择自己想要的和有趣的移动应用，这已成为当今的一个难题。为了解决这个问题，我们提出了一种主题关注机制和基于FM的移动应用推荐方法。首先，它使用LSA获取移动应用程序描述文本的全局主题。然后，通过BiLSTM模型训练移动应用程序的本地语义表示。其次，针对移动应用程序描述文本内容表示中的全局主题信息和局部语义信息，采用注意力机制来区分不同单词的贡献度并获得其权重值。第三，移动应用的分类和预测是通过整个连接层使用softmax激活功能完成的。最后，根据用户的搜索需求，它利用分解机将分类的移动应用程序的各种功能组合在一起，从而对具有较高预测分数的用户预期的移动应用程序进行排名和推荐。评估是在真实且开放的数据集Mobile App Store上进行的，实验结果表明，在精度，召回率，F1得分，MAE，RMSE和AUC方面，所提出方法的性能优于其他基准方法。它利用分解机将分类的移动应用程序的各种功能组合在一起，从而以较高的预测分数对用户的预期移动应用程序进行排名和推荐。评估是在真实且开放的数据集Mobile App Store上进行的，实验结果表明，在精度，召回率，F1得分，MAE，RMSE和AUC方面，所提出方法的性能优于其他基准方法。它利用分解机将分类的移动应用程序的各种功能组合在一起，从而以较高的预测分数对用户的预期移动应用程序进行排名和推荐。评估是在真实且开放的数据集Mobile App Store上进行的，实验结果表明，在精度，召回率，F1得分，MAE，RMSE和AUC方面，所提出方法的性能优于其他基准方法。