Article Preview
Top1. Introduction
The application Scenario of this paper is an APP called “Wei-Mi”. It focuses on recommending we-media contents of WeChat, WeiBo, ZhiHu to users. The mainly topic of this paper is we-media article recommendation.
Along with the development of Internet, we-medias such as WeChat official accounts have been developed very well. We-media is becoming one of the main information sources of the public. According to the interim results report of Tencent released in 2016, WeChat's monthly active users (MAU) has reached 806 million. Number of WeChat official accounts has reached more than 12 million. There are tens of millions of WeChat official accounts (Penguin intelligence, 2015), and they produce a lot of articles every day. It’s hard for users to find their target articles in vast amounts of consultations just rely on themselves. A recommendation system is needed to solve this problem.
Traditional article recommendation algorithms like Collaborative filtering recommendation (Rush A M, 2015), Content based recommendation (Werbos P J., 1990) and Most popular recommendation. The Collaborative filtering recommendation is combined with user-based collaborative filtering and item-based collaborative filtering. Generally, article recommendation is based on user-based collaborative filtering, which is to get a user set S that is similar to the target user u and recommend what do users in the set S like but user u don’t know yet. The Content-based recommendation is to recommends new articles that are similar to articles the user liked before.
The Most-popular recommendations are typically based on the popularity of articles (read number, like number). This algorithm recommends popular articles to users.
There are many kinds of article recommendations system, such as a news recommendation system, scientific literature recommendation system and E-mail recommendation system (Rush A M, 2015). But We-media users show some new characteristics when they read. Firstly, users tend to read more personalized articles compared with the news recommendation which pay more attention to holding popularity and timeliness of news. Secondly, we-media reading shows new characteristics of fragmentation reading and speed reading. Thirdly, according to statistics, most of the We-media users tends to read the latest articles. But there are not enough historical data on new articles.
In this paper, to cover these new characteristics, we put forward a recommendation algorithm of we-media articles based on topic model, Latent Dirichlet Allocation (LDA), and deep learning algorithm, Recurrent Neural Networks (RNNs).
This paper starting from the simple intuitive notion of preserving information. Section 3 introduce the basics knowledge of this paper. To analyze the reading characteristics of We-media users, we also used actual data to measure the user's reading behavior in section 2. Section 3 shows the LDA model (Blei D M, 2003) and the traditional LDA article recommendation algorithm (XiangLiang., 2016), the RNNs algorith (Medsker L R, 2001) and its improved version LSTM algorithm (Sundermeyer M, 2012). All of those lead to Section 5 to motivate the LDA-LSTM recommendation algorithm. Section 4 describes how the LDA–LSTM recommendation algorithm is implemented in the We-media article recommendation system. This section expounds the structure of the LDA-LSTM algorithm and describes the training method. It presents experiments that show considerably outperforms of LDA-LSTM algorithm. We also show experimental results of Random recommendation, Most-popular recommendation and Collaborative-filtering recommendation (Linden G, 2003). The result shows that the LDA-LSTM algorithm has obvious improvement on the precision and recall rate compare with 3 other algorithms. And it performs better than the other three algorithms on new articles. Section 6 summarizes our findings and describe the outlook of our work.