Fusion - Overview

In our 1st blog, we introduced you to our fully adaptive and data-driven customer behaviour detection system, FUSION, that by construction is 100% Covid-proof, and drastically improves the AML detection process.

The main purpose of FUSION is to provide the user with a fully data-driven methodology to construct segments of customers who exhibit similar (financial) behaviour.

The clusters are derived recursively thanks to a powerful Unsupervised Learning algorithm that scores each customer according to how “normal” their behaviour is within their segment.

Given its intuitive explanatory power, versatility and computational capacity, FUSION is a SaaS that can be used across different industries and for different purposes.

Guiding Principles

FUSION’s algorithm was designed following the principles below:

• The clustering process is fully automated and data-driven to the extent as possible
• By design, the clusters are not mutually exclusive
• The objective is not to spot suspicious behaviour, but to define what behavioural patterns are deemed “normal” for each cluster
• Each customer will be “scored” both at a local level (within each cluster they belong to) and at a global level (aggregating the scores of each cluster they belong to).
• The outputs are easy to interpret and “ready-to-use” for any business need

Data Description

The users of FUSION will be able to upload their dataset onto the platform. Based on our experience, a typical dataset is made up of:

• Demographic and semi-static features: these reflect characteristics such as age, residency, profession…
• Transactional features: these are aggregate transactional features (usually pre-engineered) such as average outflows per product (savings, credit cards…) or spending behaviour (expenses per category such as dining, shopping, travel…)

In case of raw data, we are able to assist you and engineer the features for you.

High-level Clustering Process

• Clusters selection: users are able to define the first dimension of the segmentation process; however the algorithm can also spot the most important features itself.
• Cluster model calibration: customers are clustered via an Isolation Forest unsupervised learning routine. The output contains a normality score and SHAP (SHapley Additive exPlanations) values for each customer within a segment.
• Iteration: using the SHAP values, the algorithm decides which dimension to use to carry on the sub-clustering process. The process will stop once the final clusters has reached a homogenous state.

Main Intuition

The key step is to fit a cluster to an Isolation Forest model, which essentially provides each customer with a similarity score that is a representation of the average number (across the trees) of iterations necessary to isolate a customer from others.

Intuitively, as you can see from the charts above, the higher the score, the more iterations required and, hence, the most similar to other customers in the cluster a customer is.

Once the normality score is assigned, the algorithm computes the SHAP values for each customer: these values determine the marginal impact of each feature on every customer’s similarity score, helping to understand the structure of the cluster. Using these marginal values for ranking the feature importance within each cluster, the algorithm continues to sub-cluster each segment until it reaches a homogenous status.

We will get back to the algorithm functioning in a more detailed blog post in the coming future.

In the next blog, we will first deep dive into FUSION’s applications, in particular, the Financial Services and Retail sectors.

Get in touch ASAP if you want to book your demo of FUSION.