Mining Financial Transactions for Customers Behavior Prediction

The paper considers the algorithm for solving the business process of an international service that provides opportunities for the loyalty system using a bank card. The main idea of the study is to find a way to determine the missing information of customer transactions. On the basis of textual information about the performed payment (SMS from the bank) contained in the transaction, the presence of a relationship with other features of the data set is investigated. Text data was presented in the form of vectors, using the approaches “bag of words”, TF-IDF, Word2Vec, FastText, on the basis of which the models of logistic regression, Random Forest, FastText were trained. The analysis of the effect of clearing the text (lower case, removing punctuation, numbers) on the quality of the classification was carried out. The increase in quality was shown by the cleaning of punctuation marks, other methods worsened the result. It was possible to achieve acceptable quality models - the best result was shown by the Word2Vec, FastText ~ 95% accuracy models on the test sample, which is 25% of the entire set.