9
$\begingroup$

Machine learning texts describing algorithms such as gradient boosting machines or neural networks often comment that these models are good at prediction, but this comes at the price of a loss of explainability or interpretability. Conversely, single decision trees and classical regression models are labelled as good at explanation, but giving a (relatively) poor prediction accuracy compared to more sophisticated models such as random forests or SVMs. Are there machine learning models commonly accepted as representing a good tradeoff between the two? Is there are any literature enumerating the characteristics of algorithms which allow them to be explainable? (This question was previously asked on cross-validated)

$\endgroup$
3
$\begingroup$

Is there are any literature enumerating the characteristics of algorithms which allow them to be explainable?

The only literature I am aware of is the recent paper by Ribero, Singh and Guestrin. They first define explainability of a single prediction:

By “explaining a prediction”, we mean presenting textual or visual artifacts that provide qualitative understanding of the relationship between the instance’s components (e.g. words in text, patches in an image) and the model’s prediction.

The authors further elaborate on what this means for more concrete examples, and then use this notion to define the explainability of a model. Their objective is to try and so-to-speak add explainability artificially to otherwise intransparent models, rather than comparing the explainability of existing methods. The paper may be helpful anyway, as tries to introduce a more precise terminology around the notion of "explainability".

Are there machine learning models commonly accepted as representing a good tradeoff between the two?

I agree with @Winter that elastic-net for (not only logistic) regression may be seen as an example for a good compromise between prediction accuracy and explainability.

For a different kind of application domain (time series), another class of methods also provides a good compromise: Bayesian Structural Time Series Modelling. It inherits explainability from classical structural time series modelling, and some flexibility from the Bayesian approach. Similar to logistic regression, the explainability is helped by regression equations used for the modelling. See this paper for a nice application in marketing and further references.

Related to the Bayesian context just mentioned, you may also want to look at probabilistic graphical models. Their explainability doesn't rely on regression equations, but on graphical ways of modelling; see "Probabilistic Graphical Models: Principles and Techniques" by Koller and Friedman for a great overview.

I'm not sure whether we can refer to the Bayesian methods above as a "generally accepted good trade-off" though. They may not be sufficiently well-known for that, especially compared to the elastic net example.

$\endgroup$
  • $\begingroup$ Now that I've had more of a chance to consider the linked paper by Ribeiro et al., I'd like to say that Section 2 'The Case for Explanation' contains something of a useful definition of 'explainability', and does a decent job of outlining its importance, and as such, deserves to be widely read within the Data Science community. $\endgroup$ – Robert de Graaf May 25 '16 at 0:25
  • $\begingroup$ Although the premise of my question was not accepted on CV, @SeanEaster helped me with this useful link: jstage.jst.go.jp/article/bhmk1974/26/1/26_1_29/_article $\endgroup$ – Robert de Graaf May 26 '16 at 0:07
3
$\begingroup$

Are there machine learning models commonly accepted as representing a good tradeoff between the two?

I assume that by being good at prediction you mean being able to fit nonlinearities present in the data while being fairly robust to overfitting. The tradeoff between interpretability and being able to predict those nonlinearities depends on the data and question asked. There really is no free lunch in data science and no single algorithm can be considered to be the best for any set of data (and the same applies for interpretability).

The general rule should be that the more algorithms you know the better it is for you as you can adopt to your specific needs more easily.

If I had to pick my favorite for classification task that I often use in business environment I would pick elastic-net for logistic regression. Despite strong assumption about the process that generates the data it can easily adopt to data thanks to the regularization term maintaining its interpretability from basic logistic regression.

Is there are any literature enumerating the characteristics of algorithms which allow them to be explainable?

I would suggest you to pick a well written book that describes the commonly used machine learning algorithms and thier pros and cons in diffrent scenarios. An example of such book can be The Elements of Statistical Learning by T. Hastie, R. Tibshirani and J. Friedman

$\endgroup$
  • 3
    $\begingroup$ TBH it was my frustration at that exact text -which uses the word 'interpretable' many times in relation to different models, and at one stages says '...data mining application require interpretable models. It is not enough to simply produce predictions' (section 10.7), without my being able to find material on how to identify an interpretable model - which prompted the question. Though I was and am loathe to appear critical of such a highly regarded text. Similarly TIbshirani's paper introducing the LASSO lists 'interpretable' as one of its virtues without saying what 'interpretable' is. $\endgroup$ – Robert de Graaf May 23 '16 at 10:40
1
$\begingroup$

Possibly see my answer regarding the unreasonable effectiveness of ensembles, and the tradeoffs on explanation versus prediction. Minimum Message Length (MML, Wallace 2005) gives a formal definition of explanation in terms of data compression, and motivates the expectation that explanations generally fit without overfitting, and good explanations generate good, generalizable predictions. But it also touches on the formal theory why ensembles will predict better -- a result going back to (Solomonoff 1964) on optimal prediction and intrinsic to fully Bayesian approaches: integrate over the posterior distribution, don't just pick the mean, median, or mode.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.