Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Differentiable Inductive Logic Programming for Fraud Detection

Boris Wolfson, Erman Acar

TL;DR

Although the scalability of DILP is a well-known issue, it is shown that with some data curation such as cleaning and adjusting the tabular and numerical data to the expected format of background facts statements, it becomes much more applicable.

Abstract

Current trends in Machine Learning prefer explainability even when it comes at the cost of performance. Therefore, explainable AI methods are particularly important in the field of Fraud Detection. This work investigates the applicability of Differentiable Inductive Logic Programming (DILP) as an explainable AI approach to Fraud Detection. Although the scalability of DILP is a well-known issue, we show that with some data curation such as cleaning and adjusting the tabular and numerical data to the expected format of background facts statements, it becomes much more applicable. While in processing it does not provide any significant advantage on rather more traditional methods such as Decision Trees, or more recent ones like Deep Symbolic Classification, it still gives comparable results. We showcase its limitations and points to improve, as well as potential use cases where it can be much more useful compared to traditional methods, such as recursive rule learning.

Differentiable Inductive Logic Programming for Fraud Detection

TL;DR

Although the scalability of DILP is a well-known issue, it is shown that with some data curation such as cleaning and adjusting the tabular and numerical data to the expected format of background facts statements, it becomes much more applicable.

Abstract

Current trends in Machine Learning prefer explainability even when it comes at the cost of performance. Therefore, explainable AI methods are particularly important in the field of Fraud Detection. This work investigates the applicability of Differentiable Inductive Logic Programming (DILP) as an explainable AI approach to Fraud Detection. Although the scalability of DILP is a well-known issue, we show that with some data curation such as cleaning and adjusting the tabular and numerical data to the expected format of background facts statements, it becomes much more applicable. While in processing it does not provide any significant advantage on rather more traditional methods such as Decision Trees, or more recent ones like Deep Symbolic Classification, it still gives comparable results. We showcase its limitations and points to improve, as well as potential use cases where it can be much more useful compared to traditional methods, such as recursive rule learning.

Paper Structure

This paper contains 45 sections, 26 equations, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Method: $\partial$ILP
  4. Clause Generations
  5. Pipeline
  6. Adjustment of $\partial$ILP to use tabular data
  7. Rule size
  8. Program Template
  9. SQL query generator
  10. Data and the Experimental Setup
  11. Explorative Data Analysis
  12. Feature engeneering
  13. Missing values treatment
  14. Test, train, validation
  15. Data scaling
  16. ...and 30 more sections

Figures (4)

  • Figure 1: Pipeline $\partial$ILP. The dataset is any tabular dataset. Binarizer converts numerical values to binary values and creates sets of facts $P_e$. Program Template defines a set of clauses to generate. The parameter $rules$ in the program template is a set of rule templates for each intensional predicate including Target. Generated High-level rules can be translated to an SQL query afterward
  • Figure 2: Fraudulent (Orange) and Valid (Blue) transaction density distributions. Dashed red and blue lines represent medians, and solid lines represent averages of fraudulent and valid transactions
  • Figure 3: Fraudulent (Orange) and Valid (Blue) transaction count plot per type of transaction
  • Figure 4: Fraud chain example: The transaction from agent 16051 to 16086, highlighted in yellow, is not a fraud chain, while the one in red, from 16086 to agent 16014, is.