Production ML systems: Questions to ask

This lesson focuses on the questions you should ask about your data and model in production systems.

Is each feature helpful?

You should continuously monitor your model to remove features that contribute little or nothing to the model's predictive ability. If the input data for that feature abruptly changes, your model's behavior might also abruptly change in undesirable ways.

Also consider the following related question:

• Does the usefulness of the feature justify the cost of including it?

It is always tempting to add more features to the model. For example, suppose you find a new feature whose addition makes your model's predictions slightly better. Slightly better predictions certainly seem better than slightly worse predictions; however, the extra feature adds to your maintenance burden.

Is your data source reliable?

Some questions to ask about the reliability of your input data:

• Is the signal always going to be available or is it coming from an unreliable source? For example:
  • Is the signal coming from a server that crashes under heavy load?
  • Is the signal coming from humans that go on vacation every August?
• Does the system that computes your model's input data ever change? If so:
  • How often?
  • How will you know when that system changes?

Consider creating your own copy of the data you receive from the upstream process. Then, only advance to the next version of the upstream data when you are certain that it is safe to do so.

Is your model part of a feedback loop?

Sometimes a model can affect its own training data. For example, the results from some models, in turn, become (directly or indirectly) input features to that same model.

Sometimes a model can affect another model. For example, consider two models for predicting stock prices:

• Model A, which is a bad predictive model.
• Model B.

Since Model A is buggy, it mistakenly decides to buy stock in Stock X. Those purchases drive up the price of Stock X. Model B uses the price of Stock X as an input feature, so Model B can come to some false conclusions about the value of Stock X. Model B could, therefore, buy or sell shares of Stock X based on the buggy behavior of Model A. Model B's behavior, in turn, can affect Model A, possibly triggering a tulip mania or a slide in Company X's stock.

Exercise: Check your understanding

Which three of the following models are susceptible to a feedback loop?

A housing-value model that predicts house prices, using size (area in square meters), number of bedrooms, and geographic location as features.

It is not possible to quickly change a house's location, size, or number of bedrooms in response to price forecasts, making a feedback loop unlikely. However, there is potentially a correlation between size and number of bedrooms (larger homes are likely to have more rooms) that may need to be teased apart.

A university-ranking model that rates schools in part by their selectivity—the percentage of students who applied that were admitted.

The model's rankings may drive additional interest to top-rated schools, increasing the number of applications they receive. If these schools continue to admit the same number of students, selectivity will increase (the percentage of students admitted will go down). This will boost these schools' rankings, which will further increase prospective student interest, and so on…

A face-attributes model that detects whether a person is smiling in a photo, which is regularly trained on a database of stock photography that is automatically updated monthly.

There is no feedback loop here, as model predictions don't have any impact on the photo database. However, versioning of the input data is a concern here, as these monthly updates could potentially have unforeseen effects on the model.

A book-recommendation model that suggests novels its users may like based on their popularity (i.e., the number of times the books have been purchased).

Book recommendations are likely to drive purchases, and these additional sales will be fed back into the model as input, making it more likely to recommend these same books in the future.

An election-results model that forecasts the winner of a mayoral race by surveying 2% of voters after the polls have closed.

If the model does not publish its forecast until after the polls have closed, it is not possible for its predictions to affect voter behavior.

A traffic-forecasting model that predicts congestion at highway exits near the beach, using beach crowd size as one of its features.

Some beachgoers are likely to base their plans on the traffic forecast. If there is a large beach crowd and traffic is forecast to be heavy, many people may make alternative plans. This may depress beach turnout, resulting in a lighter traffic forecast, which then may increase attendance, and the cycle repeats.