Have you ever wanted to make immediate use of unstructured data while bypassing the hassle of tagging it all? This can be done through collaborative behavioral profiling. Collaborative behavioral profiles are created by applying a text analytics approach to real-time, adaptive, unsupervised learning. These collaborative behavioral profiles can be used to identify anomalies in customer behavior without the need to apply tags to the data. For example, each action a customer can take is translated into a unique symbol; this creates a unique string of symbols (defined by a fixed action in a symbol dictionary) which defines each customer’s behavior and tracks changes in real time as new events occur.
Bayesian Learning to Derive Archetypes
We use Bayesian learning to derive archetypes in the latent feature space of symbol loadings. These archetypes act to boil the symbol history into a set of real-time adapting behavioral archetypes. These archetypes help us understand customers based on latent features to recognize which other customers they are most similar too, and to detect if an individual’s actions deviate from what’s expected.
It is important to note here that archetypes, in this example, are not predefined customer segmentations. Instead, each customer is represented as a mixture of archetypes, not fixed into a single rigid classification. The probabilistic interpretation is important because it captures probability densities associated with the different archetype assignments; they are not rigidly assigned. This mixture can be updated in real time based on new transactions and other customer information, so every unique customer’s collaborative profile is changing as their individual data evolves.
Using Archetypes to Flag Behaviors
Archetypes are powerful concepts as they are latent features that have physically interpretable meaning and value. We often use images of archetypes to represent actions and behaviors that are typical of certain people who are strongly aligned with a single archetype. When something occurs that causes an individual to deviate from their assigned archetype distribution, an interesting fraud identification application arises; you can see an example of this in the chart below. This individual’s spending habits changed, raising a red flag in the unsupervised adaptive model. This indicates that their behavior is out of sorts with the historical behavioral archetypes/latent features.
The Bayesian learning algorithm of the “words” to archetypes is unsupervised with no targets (meaning there are no tags that indicate fraud or not fraud). Nonetheless, it can detect anomalies. See how the distribution of percentages changes drastically in September? This shift from the established archetype distribution is detected by the unsupervised analytics to cause an alert of deviant behavior. It doesn’t need to occur monthly, or weekly, or daily, in fact, in FICO fraud applications it occurs in real-time as the event occurs.
Outlier detection adds another dimension of learning in addition to traditional supervised analytic methods based on tags. After clustering customers in the archetype space to identify which are similar, we can see if an individual strays from their peers in this archetype space; this indicates deviance. Given the real-time, adaptive, self-learning nature of the techniques, the model can further adapt to population changes while providing real-time, recursive quantile estimations of features. This can even occur with no offline historical dataset storage.
These unsupervised methods provide tremendous flexibility to solve business problems where lack of historical data or lack of modeling tags bring supervised model development to a stand-still. FICO uses this technology in Cyber Analytics, AML Analytics, Marketing Analytics, and Financial Crime Analytics; this technology is patent pending. Check out the Analytics Community and follow me on Twitter @ScottZoldi for more discussion about the possibilities machine learning brings.