Feature Selection Research On Financial Text Understanding

A large number of financial text data on the Internet, i.e., financial news, research report and stock BBS, contain rich information which having close relationship with many financial events or stock price trends. How to make machine automatically understand these massive financial text data and providing people with valuable information has become a very valuable work.Choose the appropriate set of feature words is a fundamental and insurmountable link of financial text data understanding. The appropriate set of feature words should not only have good classification ability, but also has a good stability(i.e., low sensitivity to variational train data.).In this paper, roughing a large set of candidate features is selected roughly on the basis of the specific financial text comprehension tasks firstly. Then, apply SVM-RFE and the feature selection algorithm based on random forest to select features from the candidate set and analyze their shortages in stability; Then, a combined feature selection algorithm based on training data perturbation is proposed. We compares it with feature selection methods uncombined in the stability of feature selection and classification performance; The empirical results show that in the condition of that the combined method gets close classification accuracy with uncombined methods, the combined method improves the stability of feature selection and reduces the standard deviation of classification accuracy, showing the effectiveness and superiority of the combined method. In addition, this paper also analyses the methods determining the size of the optimal feature subset and proposes a feature subset size determination algorithm based on training accuracy.