The Fifth Elephant
The Fifth Elephant

The Fifth Elephant Pune edition

Call for Proposals

All submissions

Chetan Paulzagade

Rethinking Segmentation in AML Using Peer-Based Anomaly Detection

Submitted Jan 12, 2026

Indicate the track in which your submission fits: Track 1 AI in Software Development Life Cycle (SDLC) Type of submission: 30 mins talk

Introduction

Traditional anomaly detection systems rely heavily on manually defined segmentation—customer types, business categories, or risk bands—before any modeling begins. This assumes that similarity can be explicitly defined upfront using business rules and observable proxies. In practice, such segmentation is often coarse, static, and biased, leading to missed risks and brittle models. In this session, we argue that segmentation is not a prerequisite for anomaly detection, but rather an emergent property of it. We present a peer-conditioned, multivariate contextual anomaly detection framework where entities are evaluated relative to dynamically discovered behavioral peers instead of predefined segments.

The approach combines large-scale nearest-neighbor geometry, robust statistics, and local density comparison to learn similarity directly from data. For each entity, the model retrieves its closest peers, computes peer-relative deviations and isolation scores, and aggregates risk across behavioral contexts. Segmentation emerges implicitly as peer groups on a continuous behavioral manifold, while anomalies are defined contextually within these groups. The session will walk through the conceptual design, algorithmic building blocks, scalability considerations, and explainability aspects of this approach, illustrating how anomaly detection and segmentation become two projections of the same learned structure.

Takeaways

Segmentation and anomaly detection need not be separate steps; both can emerge from a single peer-based modeling framework that learns similarity directly from data.

Peer-conditioned anomaly detection produces more adaptive risk signals and more intuitive explanations than rule-based segmentation approaches, especially at large scale.

Audience

Data scientists and ML engineers working on anomaly detection, fraud, AML, or risk modeling

Architects and technical leads designing large-scale data science systems

Product managers and domain experts interested in moving beyond rule-based segmentation

Speakers

Chetan Paulzagade, Data Science Manager, Nice Actimize R&D
Sachin Dixit, Head of Architecture, Nice Actimize
R&D

All submissions

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

The Fifth Elephant
The Fifth Elephant
Jumpstart better data engineering and AI futures