The Prediction Model Based On Sentiment And Autoregression

With the advent of Web2.0, customers have changed their way to express opinions. They begin to post comments of products at merchant websites and share their experience via reviews. Different from the description of products, these reviews are written entirely based on the willingness of consumers, and can influence the dec is ion-making process of others. This effect can be observed by some easy-to-measure economic variables, suchas sales performance or product prices. As online commerce activity continues to grow, review mining has received a great deal of attention.In this paper, taking book reviews as a case study, we focus on the mining of sentiments from reviews and investigating ways to use such information to predict sales. In addition, past sales performance as another indicator of future sales reflects the influence and trends of the market, and plays an important role in prediction. The significance of this work lies in pricing, marketing scheme, and so on. In general, we focus our work on the following three aspects.(1) The study is begin by mining sentiments from reviews. First, we construct the sentiment word dictionary, and then present a method of sentiment analysis based on the dictionary and TF-IDF method, as integrating the sentiment factor into autoregressive model, we last propose the prediction model ARES, an autoregressive emotion-sensitive model. Experiments indicate the predictive power of ARES shows good performance.(2) However, due to the complex presentations of natural languages, opinions usually come in multi-aspects, and TF-IDF cannot capture the full view of the sentiments. Therefore, we propose a Latent Sentiment Language (LSL) Model to address this challenge, in which sentiment-language model and sentiment-LDA are used to capture the explicit and implicit sentiment information respectively. Subsequently, we explore ways to use such information to predict product sales, and to generate an SAR, a sentiment autoregressive model. And extensive experiments prove the superior performance of the SAR model.(3) Besides, for the quality factor, we improve the SAR into SQAR by considering two features, length and qualifiers, of the review. And the experiments show the influence of the review quality in sales prediction.