Research On Hot News Click Rate Prediction And Topic Evolution

Under the background of the freedom of the public opinion,hot news is more likely to become a focus of public debate and contradiction.Hot news click rate prediction and mining the evolution of news are significant in news media.In government affairs,the objective news click rate prediction have easy access to the development trend of events in time.In news filtering application,users are supposed to be presented with hot news only in the top of the list,but it is also essential to offer the evolution of the hot topic.In the thesis,hot news click rate prediction and mining the evolution of the hot topic act as two problems to be solved.The main contributions are as follows:(1)A hot news click prediction algorithm has been proposed based on Grey Verhulst model and Extreme learning machine.Historical click rate data yields to random and fluctuant noise which are inevitable.However,the proposed model can filter the noise with effect and maintain the distribution of click rate data at the same time.Experiments demonstrate the average accuracy has increased by 7%compared with the state-of-the-art methods.(2)A feature representation method is proposed inspired by the multi-semantic fusion strategy between the news pictures and texts.Feature representation is the first step in the process of mining the evolution of the hot topic.Although the news pictures implicit vast information related to the news events,they are easier to be ignored compared with the news texts.The classification probability values achieved by the deep convolution neural networks are taken as the semantic feature of the news picture and the TD-IDF vector is considered to be the feature of the news text.Thus,the syncretic feature containing the pictures and texts of news.Experiments indicate classification accuracy has increased by 2.4%compared with the single texts feature.(3)On the base of(2),a multi-label classification algorithm is put forward for mining evolution of news.Establishing a model to measure the similarity of news between time t and t +1 is the core of tracking evolutionary trajectory.Therefore,the thesis introduces the label co-occurrence as the basis of a similarity measure.Experiments results show that the news evolution trace is easy to understand.