NonStop Insider

job types

Site navigation

Recent articles



For monthly updates and news.
Subscribe here
NonStop Insider

Bayesian and Dempster-Shafer Models for Combining Multiple Sources of Evidence in In a Fraud Detection System

By Fabrice Daniel, Artificial Intelligence Department of Lusis, Paris, France



lusis May 21 - 1


Combining evidence from different sources can be achieved with Bayesian or Dempster-Shafer methods. The first requires an estimate of the priors and likelihoods while the second only needs an estimate of the posterior probabilities and enables reasoning with uncertain information due to imprecision of the sources and with the degree of conflict between them. This paper describes the two methods and how they can be applied to the estimation of a global score in the context of fraud detection.


Fraud detection mainly relies on expert driven methods that implement a set of rules and data driven approaches implementing machine learning (ML) models. Both provide an estimate (or a score) for a new transaction to be fraudulent.

While each ML model naturally returns a fraud probability, the experts can also attach a probability to each rule. They can also be automatically calculated from the labelled history. Combining them together produces a global score that can be used in a near real time system to rank a set of transactions having the highest probability to be fraudulent. By obtaining this ranking, investigators can concentrate their efforts on the suspect transactions with the highest probability of being true frauds.

lusis May 21The most common approaches for combining scores are summing individual scores or returning the highest score among the trigged rules. This is not entirely satisfactory given that summing scores is equivalent to averaging the probabilities returned by each predictor (rule or model). It also does not take into account the uncertainty of each predictor and the degree of conflict between them.

For the Lusis fraud system, we work on implementing more appropriate approaches.

This paper proposes two ways for addressing this problem. The first is to use Bayesian methods [5]; the second is to combine the scores by using Dempster-Shafer theory [6].


For more information contact
Lusis Payments
315 Montgomery Street #900  San Francisco CA 94104
(+1) 415 829 4577   |