Seeing… or not seeing

This post originally appeared on Applied Insights’ blog. Foviance acquired Applied Insights in November 2008, with Neil Mason joining us as Director of Analytical Consulting. As part of this acquisition, we’ve incorporated Applied Insights’ blog into our own.

When we think about how to evaluate a predictive model the first thing we typically think of is how accurately does that model predict against the (unseen) test data. More often than not though when we develop models our business/research customers want more than that. They want to know how the algorithm got to the predictions i.e. they want to understand the model.

The more transparent predictive methods don’t just predict they also reveal the patterns that underlie them. The two main benefits of this are that

1. Subject Matter Experts (SMEs) typically on the business/research side – can assess the model’s validity by viewing these patterns, for example as rules or formulae. This way they can see if the inherent relationships make sense. Do they see any potential anomalies in the data that we didn’t pick up when we previously explored it?
2. And of course the patterns themselves may reveal useful insights. We often find specific segments of interest; demographic groups who have a higher propensity to convert through a given channel, or re-purchasers who have short, but potentially interesting and valuable, buying cycles.

The bottom line is that when we can see what a model is doing we can glean much more from it than the likelihood that the outcome of interest (convert, attrite, default, etc.) will happen.

To be frank most of our projects are like this. This is where Decision Tree methods often win out because the output let’s us visually explore the data to both understand the model and to examine other potential patterns of interest. They may not necessarily give us the most accurate predictions but often the SMEs care more about understanding than predicting. This is a classic trade-off in PA.

There are exceptions to this. The alternative view is that accuracy is paramount and it could be that the winning model is opaque. Neural Network models are a case in point. Depending on the software you are using you might see a ranked list of fields which contribute to the prediction along with the prediction itself and perhaps an associated confidence level. Even if the final network is displayed it doesn’t necessarily explain much more.

For the most part these are the two most typical scenarios however we are currently designing a 3^rd type – where opaqueness is the main objective (together with an acceptable level of predictive accuracy of course). We’re talking to a government department who don’t want to have to send sensitive data out and who don’t want our models to reveal any of that information either. So the gist of our approach is that we’ll develop black-box models on our data and let them deploy them on their database. They’ll give us addresses and predictive scores in return but in so doing we won’t know why a particular address was selected.

Anyone living in the UK will understand the political backdrop to this as there have been various high profile cases of data going AWOL (here is the latest one). We are hoping that a somewhat unorthodox application of Predictive Analytics might help the UK government provide a valuable public service without further compromising the confidentiality of its citizens. There’s many a slip twixt the cup and the lip mind you – we’ll keep you posted…