Precision marketing for financial industry using a PU-learning recommendation method

Abstract

Due to the growing competition among peers, financial companies are in urgent need of simultaneously recommending multiple types of products with data that only contain positive and unlabeled (PU) samples, which poses a crucial challenge for traditional recommendation methods. To fill this gap, we propose a recommendation model based on PU-learning and matrix completion, in which we employ the PU-Logistic model to reduce the prediction bias, and use a “preference matrix” to identify customers’ interest in different products. With this model, we can better understand the homogeneity and heterogeneity of customer preferences and effectively estimate the customer preferences of missing data. Our simulation results indicate that our proposed method has competitive performance in coefficient estimation and prediction and is robust to potential model misspecification. By applying our method to a consumer financial dataset from China between 2019 and 2020, we obtain interpretable findings with competitive prediction performance.