Mining Correlation Features of User Financial Behavior Based on Attention Mechanism and Dual Channel

Abstract

For the financial field, the amount of data is large and the data forms are complicated. Traditional data processing methods are unsuitable for current big data. Therefore, we propose a financial big data feature mining method based on attention mechanism and dual channel (NDUA) to understand the deep features of financial big data and mine the association features of customers’ financial behavior. First, the pre-trained model Bert is used to vectorize textual data in financial data. Next, a dual-tower model is proposed to predict user behavior, and multiple user behaviors are used as labels. Then, we perform model training and use the trained three sets of models to extract the last fully connected layer as the user behavior feature of the item, and for each user to cluster the user behavior feature representations of all the operated items to further obtain the association of user behavior. Experiments demonstrate that the proposed method can effectively understand the deep features of financial big data and mine the association features of customers’ financial behavior.