Accuracy has long been the central quest of machine learning models and those who train them — textbooks are full of algorithms compared solely in terms of their accuracy on particular training sets, and contain little advice on how to measure or achieve other desirable qualities. However, in recent years there has been a realisation that accuracy alone isn’t enough. Bias can kill a model’s usefulness, and so can opacity and complexity.

The form of bias that gets the most attention is probably where models inadvertently discriminate against people, such as when models to predict crime re-enforce police prejudices. Although it is correct that these sorts of biased models receive attention and they need improvement, there are many other areas where bias is a problem. Another example is where commercial models limit their own usefulness through biases.

A credit card company that is trying to expand its market share within a demographic isn’t well served by a default model that says that demographic is the biggest risk factor. The manager of a store in a particular location trying to increase sales may not be well served by a model where the location is the most important predictor of sales given she can’t easily relocate the store.

Fairness (or lack of bias) and explainability are both receiving more attention as desirable model characteristics, but there are still few standard methods to measure them. Although they are different characteristics, they have a link in the sense that to understand whether a model is fair or unbiased some degree of transparency is required, and greater explainability gives greater understanding on fairness.

Some modelers may feel that by saying that we need transparency to be able to achieve fairness we are making a trade-off with accuracy, as we are making it harder to justify using a black box method such as a GBM or a neural network. This is not the case. As discussed in the paper ‘The Secrets of Machine Learning’ by Carlson and Rudin, it’s very frequently overlooked that models perform at the top level of accuracy while still being transparent and explainable — as long as that’s the intention from the beginning, and sufficient care is taken.

The same idea flows through to fairness as well. If you need a fair and unbiased model, you need to choose that path from the beginning. The same article makes the point that if you need to ensure that your model is unbiased and explainable you need to have this objective from the beginning. A suggested way to ensure fairness is to define what fairness means for your model’s context, and make this a goal of model training from the beginning of the process.

A similar principle applies for interpretability, which is heavily dependent on the choice of algorithm. Certain algorithms essentially mean that the resultant model is almost guaranteed to never be interpretable. Other algorithms provide a massive headstart towards building a highly interpretable model.

#data-science #machine-learning #predictive-modeling #deep learning

Beyond Accuracy, Robert de Graaf’s book
1.10 GEEK