Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Transactions Representations for Information Management in Banks: Mastering Local, Global, and External Knowledge

Alexandra Bazarova, Maria Kovaleva, Ilya Kuleshov, Evgenia Romanenkova, Alexander Stepikin, Alexandr Yugay, Dzhambulat Mollaev, Ivan Kireev, Andrey Savchenko, Alexey Zaytsev

TL;DR

This study tackles the problem of learning transaction representations for banks that balance global (history-wide) and local (momentary) properties. It benchmarks eight unsupervised approaches spanning generative and contrastive SSL, temporal point processes, and an external-context enrichment technique that aggregates information from other clients. The findings show generative approaches excel at local tasks while contrastive methods perform well on global classification, with external-context aggregation providing substantial local gains (up to $20\%$). A comprehensive evaluation across open datasets (Churn, Default, HSBC, Age) and a real private dataset demonstrates the approach's practical potential for scalable, versatile information management in banking.

Abstract

In today's world, banks use artificial intelligence to optimize diverse business processes, aiming to improve customer experience. Most of the customer-related tasks can be categorized into two groups: 1) local ones, which focus on a client's current state, such as transaction forecasting, and 2) global ones, which consider the general customer behaviour, e.g., predicting successful loan repayment. Unfortunately, maintaining separate models for each task is costly. Therefore, to better facilitate information management, we compared eight state-of-the-art unsupervised methods on 11 tasks in search for a one-size-fits-all solution. Contrastive self-supervised learning methods were demonstrated to excel at global problems, while generative techniques were superior at local tasks. We also introduced a novel approach, which enriches the client's representation by incorporating external information gathered from other clients. Our method outperforms classical models, boosting accuracy by up to 20\%.

Learning Transactions Representations for Information Management in Banks: Mastering Local, Global, and External Knowledge

TL;DR

This study tackles the problem of learning transaction representations for banks that balance global (history-wide) and local (momentary) properties. It benchmarks eight unsupervised approaches spanning generative and contrastive SSL, temporal point processes, and an external-context enrichment technique that aggregates information from other clients. The findings show generative approaches excel at local tasks while contrastive methods perform well on global classification, with external-context aggregation providing substantial local gains (up to ). A comprehensive evaluation across open datasets (Churn, Default, HSBC, Age) and a real private dataset demonstrates the approach's practical potential for scalable, versatile information management in banking.

Abstract

In today's world, banks use artificial intelligence to optimize diverse business processes, aiming to improve customer experience. Most of the customer-related tasks can be categorized into two groups: 1) local ones, which focus on a client's current state, such as transaction forecasting, and 2) global ones, which consider the general customer behaviour, e.g., predicting successful loan repayment. Unfortunately, maintaining separate models for each task is costly. Therefore, to better facilitate information management, we compared eight state-of-the-art unsupervised methods on 11 tasks in search for a one-size-fits-all solution. Contrastive self-supervised learning methods were demonstrated to excel at global problems, while generative techniques were superior at local tasks. We also introduced a novel approach, which enriches the client's representation by incorporating external information gathered from other clients. Our method outperforms classical models, boosting accuracy by up to 20\%.
Paper Structure (58 sections, 7 equations, 9 figures, 7 tables)

This paper contains 58 sections, 7 equations, 9 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Supervised methods for representation learning
  4. Self-Supervised Learning Paradigm
  5. Temporal Point Process Models
  6. Methodology
  7. Global properties' validation methods
  8. Local properties validation methods
  9. Local downstream targets
  10. Next transaction MCC prediction
  11. Challenge in inferring local representations
  12. Strategy for validation of embeddings dynamism
  13. Validation metrics
  14. Classification metrics
  15. Change point detection quality
  16. ...and 43 more sections

Figures (9)

  • Figure 1: The general flow of the application of self-supervised learning to industrial tasks in banking.
  • Figure 2: Scheme of the sliding window approach for local validation.
  • Figure 3: Scemes of the Autoencoder (AE) and Autoregressive (AR) methods.
  • Figure 4: Scheme of the Masked Language Model (MLM) method.
  • Figure 5: General pipeline for obtaining external context vectors.
  • ...and 4 more figures