Interpretability Vs Accuracy in Operations Research and Management Science Community

Question

This question might be somewhat general and not completely relevant to this forum but I think here is the most relevant place to ask the question.

Currently, deep learning, RL and generally black-box approaches are gaining much attention and many practitioners and academia are using these problems to solve their optimization problems. For example, in real-world prediction problems, LSTM artificial neural networks (ANNs) are common which are black-box algorithms with good accuracy but in many cases, there is no proof of convergence. However, in the field of OR/MS, the most common approaches are Time Series approaches like ARIMA. etc. Or, for example, decision trees are more popular in OR/MS because they are interpretable but with lower accuracy compared to deep learning. In this situation, AI algorithms enable researchers to use different sources of data like historical data, crawling webpages, reading news, etc. and use all of them to predict while conventional approaches in OR/MS do not let us to this.

To my view, the computer science community is using approaches that are more acceptable in the industry and more applicable while OR/MS is sacrificing applicability in real-world cases to solve problems with convergence proof. For example, many papers use linear regression with unknown coefficients because they can prove that their algorithm can converge to the true values of coefficients (if the true model is also linear).

These are my views based on papers published in Management Science, Operations Research, and M&SOM. Some people may consider this question a subjective one and wants to close this question but it is really confusing question that I cannot find the justification for some times.

Answer 1

[Fair warning: old guy rant follows.]

I'm not sure that the computer science community's reliance on ML models (to the extent that generalization holds) is necessarily a good thing.

1. I've seen references to published research (can't recall any details now) on retrofitting interpretability to ML models, ostensibly because industry people are uncomfortable with the models if they cannot understand at some level the model logic.
2. A growing issue in industry (and government) is model bias. My impression is that model bias is easier to detect and either fix or justify (as not really being bias) with interpretable models.
3. ML models can be prone to overfitting. With an interpretable model, you can get a clue that overfitting occurred when the interpretation defies common sense / logic / the "sniff" test. With an ML model, you can put in penalty terms or something to try to combat overfitting, but it's a guessing game (add a lasso term and hope for the best).
4. Sometimes, in the real world, you don't have a ton of (reliable) data. My trust in a statistical model is based on a combination of the plausibility of the assumptions (normality, IID observations, whatever) and having a "decent" sample size. My trust in an ML model (to the extent I have any) is based on it being trained on a really, really big sample.

Answer 2

Many people seem to think that, but we do nothing of the sort in OR - our process is much simpler than people think.

We build and solve a model, and if the results make sense, we are pretty much done. If the results don't make sense we typically know why. Crucially, we can know whether the issue is in the data or not.

The reason OR has a reputation for being difficult and time consuming is because problems that describe the real world well take skill to create and even more skill to solve if a solver can't handle them out-of-the-box.

ML is optimisation. The difference is in how the models are created, and the skill level required to produce something that is usable.

Keep in mind that "classical" mathematical optimisation is simply engineering, and has been used in the industry with great success for many decades.

It is useful to remember that ML & optimisation are not mutually exclusive. Either can solve problems the other one can't solve at all. The advent of ML has actually triggered massive growth in OR because the two technology stacks are incredibly powerful when used in conjunction.