Research On Downloading Volume Prediction Based On Sentiment Curves Of Novels

Psychology studies show that people tend to have a better impression of stories which patterns are familiar with them and hate stories that are contrary to their expe-rience.Kurt Vonnegut argues that the sentiment curve of a story is the core value of a novel.Good novels often tend to have similar patterns of sentiment change.In order to better analyze the sentiment changes of the novel,and to use the characteristics of the novels to improve the accuracy of the relevant prediction,we need to generate the sentiment curves of the novel and carry out relevant comparative analysis.The traditional tasks of sentiment analysis lack the cosideration of the overall sen-timent change of the novel.The related work focused on sentiment curves of novels is still relatively rare.The predecessors’work about finding the "pattern" of the sentiment curves of novels may only extract the noise produced by mixing many novels.Their method of generating the sentiment curves of novels also exists the problem of losing information of the origial text because of uneven sampling or over-smoothing.In this dissertation,we propose a kind of method to generate sentiment curves of novels focusing the above problems,and give a downloading volume prediction of novels based on their sentiment curves.First,we propose a new way to generate the sentiment curve of a novel.The choice of sampling window for sentiment scores in this method depends only on the length of the novel text,so that a variable-length sentiment curve of the novel could be generated.Secondly,we introduce Dynamic Time Wrapping from time series analysis to describe the differences between the sentiment curves of the novel and we solve the positive-definite problem when turning Dynamic Time Wrapping distance matrix into a kernel function.Finally,we introduce Gaussian Process to solve the regression problem to predict the downloading volumn of novels using the sentiment curve of novels.The experimental results show that our method of generating sentiment curves of the novels could show the sentiment change of the novel more accurately.We analyze Dynamic Time Wrapping distance measure from both theoretical and experimental as-pects,so that we can better describe the similarities of sentiment curves of novels.We theoretically prove that our method of transforming Dynamic Time Distance into ker-nel function is reasonable.The experimental results also show that our method of using Gaussian Process to predict the download numbers of novels obtains a higher positive correlation.