An early warning system that can provide actionable insights is of great strategic value to any enterprise. Today, the alerting mechanisms are getting increasingly sophisticated - driven by the volume and variety of data and ability of algorithms to make sense of it. This is enabling organizations move to a proactive stance around identifying suspicious events. In this blog, I would like to cover how the increasing sophistication of algorithms is giving rise to a new set of challenges.
Just to set the context, let's take a step back - Traditional outlier detection is based on domain expert's understanding of the area and converting their knowledge and experience into rules, that help making a decision. These rules are extremely easy to understand; one would check the conditions prescribed by the checklist and accordingly proceed with the relevant course of action. The challenges with this approach, as articulated in my earlier blog, are high number of false positives, unwieldly exception handling and lack of self-calibration to discern new patterns. These elements become increasingly tougher with evolving business situations, unique nature of every fraud and multiplying data volumes.
The new Machine Learning models overcome some of these challenges: they can handle a huge number of parameters, weights are assigned by the algorithm, they can learn from feedback and can continuously evolve. Ensemble models - which combine the decisions made by several underlying ML models to produce improved results have a higher order challenge, as the decision is made through aggregation of multiple models (e.g. Random Forests, bagging, boosting etc.). Building these ML models requires domain knowledge for selecting features from the given data during the feature engineering process.
The rapidly developing field of Deep Learning is resolving this challenge by managing the feature engineering as well; leading to the intelligence layer being further abstracted and hence even greater loss of intuition into why a particular decision was made. As a broad philosophy, we are increasingly finding that it is not enough to have clarity of the overall model; it is equally important to know the rationale for the determination for each specific case.
Let us consider an example of why the knowing the rationale for each case is vital:
An insurance firm employs Machine Learning algorithms for identifying fraudulent claims which leads to efficiency gains and catches the frauds that previously went unnoticed. The algorithm is trained with historic data and performs well. While it executes and identifies potential cases for investigation, the investigators demand cues on why particular claims were flagged (which would've been evident in rule-based or scorecard based regimes). The management too, wants a better insight into the prediction model and the factors that contributed to the decision.
There are various reasons why transparency into the black box model is desired:
- Users are uncomfortable with a decision if the criteria used are unclear
- To focus attention in investigation (provide indicators around areas to focus on, especially in situations with a large number of causal factors)
- Concern that inappropriate parameters are used for decision making (e.g. race, gender, age etc.)
This is a consistent theme in our discussions with multiple organizations, be it insurance, retail or manufacturing. We believe development of 'causal factor prediction' model would help provide guidance on why a decision was made. I will talk about the approach taken to build the causal factor prediction model and the results achieved in my subsequent posts.