User-Aggregated Biterm Topic Model And It’s Application On Short Text Recommender System

With the rapid growth of mobile Internet and smart cellphone, social media appli-cations like Twitter and Weibo are becoming more popular recently. Mass information is being brought in everyday from thess applications, which causes information over-load problem and makes it difficult for users to find things they are interested. Text recommender system can make personalized suggestion for different users according to their preferences. However, traditional text analysis usually focus on the long text and performs poor on short text context. We propose a User Aggregated Biterm Topic Model to deal with the short text recommendation problem. Specifically, our major contributions include:1. A short text recommendation framework based on the topic model. The user-aggregated topic model can both withdraw user and new text’s topic distribution. The similarity can be used for text sorting in text recommendation.2. User aggregated biterm topic model. It can analyse the topic of user and short text effectually. We also present a Gibbs sampling based model parameters estimating method.3. A short text recommender system. We take advantage of the topic distribution ob-tained by our topic model to make TopN recommendation for users from new short text. Experiments on Weibo and Twitter dataset prove the high quality of topic model and text recommendation both qualitatively and quantificationally.4. A Weibo prototype recommender system. It indicates that our model and frame-workd can be applied to real-world applications properly.