In my previous article, I spoke about how with the mainstreaming of Machine Learning and Deep Learning, it is becoming increasingly difficult to decipher how machines make decisions. Progressive improvements in algorithmic capability are accompanied by greater abstraction of intelligence; raising questions on whether we can open this black box to have a better sense of how decisions are made. The need for identifying the critical parameters for a decision goes beyond curiosity. As covered in the previous article, this helps build comfort that the predictive process seems meaningful and does not use inappropriate factors for decision-making. (E.g. factors that might be considered discriminatory in nature)
The objective of building a causal factor prediction model is to provide guidance around key parameters that influenced the decision for a specific observation – thus enabling targeted investigation. This is different from identifying parameters influencing the decision on an aggregate basis, and not targeted to individual data points. Essentially, the intent is not to come up with a clear rule book (the whole idea of creating Machine Learning models and ensembles was to handle real world scenarios that don’t lend themselves to straight forward rules).
The broad approach to building causal factor prediction model is driven by four key steps:
- Extract the decision criteria from the anomaly detection model. Our governing principle while developing the Causal Factor Prediction model is that we do not allow the factor prediction to influence anomaly detection, as we do not want to impact the accuracy of the prediction
- Gather data on reasons and attributes used for prediction (for sample cases) and develop the causal factor space. This would essentially represent subjective judgement from domain experts, around why certain observations were marked as anomalous
- Gather data from the predictive model around the decision flow taken up for each observation and build an initial view of causal prediction for the data point – and compare with subjective judgement from experts
- Iterate the causal factor prediction model parameters such that it best aligns with predictions through subjective judgement
As an example, in the case of media spend anomaly detection, anomalies are determined by analyzing approximately 30 attributes ranging from advertisement duration, day of week, nature of channel, slot duration, time slot etc. Based on discussions with subject matter experts, it could be gleaned that certain observations were deemed suspicious based primarily on 2 factors: day of the week and product category. Likewise, another set of observations were deemed suspicious based on unusual patterns around the price charged by a channel for the time slot during which the commercial was aired.
This approach has enabled us to identify anomalies and their possible causes with a fair degree of accuracy. Also, in alignment with investigator preference, our models are conservative and tend to err towards including more factors into causal factor prediction than missing relevant factors. As mentioned earlier, our intent with this is not to achieve the deterministic accuracy of rule regime, but to provide a broad direction of what attributes shaped the decision and hence should be focused on.
This approach has enabled business teams and investigators to respond to Machine Learning models with greater confidence – as there is more visibility compared to the erstwhile black box. In addition to guiding expert investigation, this helps in providing an explanation of the decision made providing a greater sense of control.