Behaviour Trace-Based Unsupervised Deep Embedding for Online Multiple Product Recommendation in Commercial Banks

Recommender systems (RSs) for products and services stay central to the operations of commercial banks in China. Efficiency, effectiveness and resilence of relevant methods are therefore the main focus. In this paper, a novel Behaviour Trace-based Unsupervised Deep Embedding (BTUDE) architecture has been developed for online recommendation of multiple products in the mobile bank application of a large commercial bank in China (LCB). BTUDE is implemented with a Variational AutoEncoder (VAE) model in which a ResNet1D network and its reverse implementation, modified for 1D time series processing, play the role of encoder and decoder, respectively. In this architecture, Customer behaviour data grouped in many time series are first processed with a novel normalization method, and then fed into the VAE model. As a quantitative form of the customer profile system, BTUDE represented as the latent space embedding generated by VAE is built up independent of scenarios for any specific product recommendation. In recommendation events for each specific product, BTUDE similarity between potential customers and those already possessed the product can be used as the feature to further detect target customers via a two-component Gaussian Mixture Model (GMM). Different types of tests have been carried out. Online A/B tests have been conducted for a money market fund product. Performance for wealth management and life insurance products are also evaluated. The promising results show that BTUDE is effectively adapted to online environment for recommendation shifting among multiple products in a fast pace.