What to do with “small” data?

By Ahmed El Deeb


Many technology companies now have teams of smart data-scientists, versed in big-data infrastructure tools and machine learning algorithms, but every now and then, a data set with very few data points turns up and none of these algorithms seem to be working properly anymore. What the hell is happening? What can you do about it?

Where do small data come from?

Most data science, relevance, and machine learning activities in technology companies have been focused around “Big Data” and scenarios with huge data sets. Sets where the rows represent documents, users, files, queries, songs, images, etc. Things that are in the thousands, hundreds of thousands, millions or even billions. The infrastructure, tools, and algorithms to deal with these kinds of data sets have been evolving very quickly and improving continuously during the last decade or so. And most data scientists and machine learning practitioners have gained experience is such situations, have grown accustomed to the appropriate algorithms, and gained good intuitions about the usual trade-offs (bias-variance, flexibility-stability, hand-crafted features vs. feature learning, etc.). But small data sets still arise in the wild every now and then, and often, they are trickier to handle, require a different set of algorithms and a different set of skills. Small data sets arise is several situations:

  • Enterprise Solutions: when you try to make a solution for an enterprise of a relatively limited members instead of a single solution for thousands of users, or if you are making a solution which companies instead of individuals are the focus of the experience
  • Time Series: Time is in short supply! Esp. in comparison with users, queries, sessions, documents, etc. This obviously depends on the time unit or sampling rate, but it’s not always easy to increase the sampling rate effectively, and if your ground truth is a daily number, then you have one data point for each day.
  • Aggregate modeling of states, countries, sports teams, or any situation where the population itself is limited (or sampling is really expensive).
  • Modeling of rare phenomena of any kind: Earthquakes, floods, etc.

Small Data problems

Problems of small-data are numerous, but mainly revolve around high variance:

  • Over-fitting becomes much harder to avoid
  • You don’t only over-fit to your training data, but sometimes you over-fit to your validation set as well.
  • Outliers become much more dangerous.
  • Noise in general becomes a real issue, be it in your target variable or in some of the features.

So what to do in these situation?

1- Hire a statistician

I’m not kidding! Statisticians are the original data scientists. The field of statistics was developed when data was much harder to come by, and as such was very aware of small-sample problems. Statistical tests, parametric models, bootstrapping, and other useful mathematical tools are the domain of classical statistics, not modern machine learning. Lacking a good general-purpose statistician, get a marine-biologist, a zoologist, a psychologist, or anyone who was trained in a domain that deals with small sample experiments. The closer to your domain the better. If you don’t want to hire a statistician full time on your team, make it a temporary consultation. But hiring a classically trained statistician could be a very good investment.

2- Stick to simple models

More precisely: stick to a limited set of hypotheses. One way to look at predictive modeling is as a search problem. From an initial set of possible models, which is the most appropriate model to fit our data? In a way, each data point we use for fitting down-votes all models that make it unlikely, or up-vote models that agree with it. When you have heaps of data, you can afford to explore huge sets of models/hypotheses effectively and end up with one that is suitable. When you don’t have so many data points to begin with, you need to start from a fairly small set of possible hypotheses (e.g. the set of all linear models with 3 non-zero weights, the set of decision trees with depth <= 4, the set of histograms with 10 equally-spaced bins). This means that you rule out complex hypotheses like those that deal with non-linearity or feature interactions. This also means that you can’t afford to fit models with too many degrees of freedom (too many weights or parameters). Whenever appropriate, use strong assumptions (e.g. no negative weights, no interaction between features, specific distributions, etc.) to restrict the space of possible hypotheses.

Any crazy model can fit this single data point (drawn from distribution around yellow curve)
As we have more points, less and less models can reasonably explain them together.
The figures were taken from Chris Bishop’s book Pattern Recognition and Machine Learning

3- Pool data when possible

Are you building a personalized spam filter? Try building it on top of a universal model trained for all users. Are you modeling GDP for a specific country? Try fitting your models on GDP for all countries for which you can get data, maybe using importance sampling to emphasize the country you’re interested in. Are you trying to predict the eruptions of a specific volcano? … you get the idea.

4- Limit Experimentation

Don’t over-use your validation set. If you try too many different techniques, and use a hold-out set to compare between them, be aware of the statistical power of the results you are getting, and be aware that the performance you are getting on this set is not a good estimator for out of sample performance.

5- Do clean up your data

With small data sets, noise and outliers are especially troublesome. Cleaning up your data could be crucial here to get sensible models. Alternatively you can restrict your modeling to techniques especially designed to be robust to outliers. (e.g. Quantile Regression)

6- Do perform feature selection

I am not a big fan of explicit feature selection. I typically go for regularization and model averaging (next two points) to avoid over-fitting. But if the data is truly limiting, sometimes explicit feature selection is essential. Wherever possible, use domain expertise to do feature selection or elimination, as brute force approaches (e.g. all subsets or greedy forward selection) are as likely to cause over-fitting as including all features.

7- Do use Regularization

Regularization is an almost-magical solution that constraints model fitting and reduces the effective degrees of freedom without reducing the actual number of parameters in the model. L1 regularization produces models with fewer non-zero parameters, effectively performing implicit feature selection, which could be desirable for explainability of performance in production, while L2 regularization produces models with more conservative (closer to zero) parameters and is effectively similar to having strong zero-centered priors for the parameters (in the Bayesian world). L2 is usually better for prediction accuracy than L1.

L1 Regularization can push most model parameters to zero

8- Do use Model Averaging

Model averaging has similar effects to regularization is that it reduces variance and enhances generalization, but it is a generic technique that can be used with any type of models or even with heterogeneous sets of models. The downside here is that you end up with huge collections of models, which could be slow to evaluate or awkward to deploy to a production system. Two very reasonable forms of model averaging are Bagging and Bayesian model averaging.

Each of the red curves is a model fitted on a few data points
But averaging all these high variance models gets us a smooth output that is remarkably close to the original distribution Pattern Recognition and Machine Learning

9- Try Bayesian Modeling and Model Averaging

Again, not a favorite technique of mine, but Bayesian inference may be well suited for dealing with smaller data sets, especially if you can use domain expertise to construct sensible priors.

10- Prefer Confidence Intervals to Point Estimates

It is usually a good idea to get an estimate of confidence in your prediction in addition to producing the prediction itself. For regression analysis this usually takes the form of predicting a range of values that is calibrated to cover the true value 95% of the time or in the case of classification it could be just a matter of producing class probabilities. This becomes more crucial with small data sets as it becomes more likely that certain regions in your feature space are less represented than others. Model averaging as referred to in the previous two points allows us to do that pretty easily in a generic way for regression, classification and density estimation. It is also useful to do that when evaluating your models. Producing confidence intervals on the metrics you are using to compare model performance is likely to save you from jumping to many wrong conclusions.

Parts of the feature space are likely to be less covered by your data and prediction confidence within these regions should reflect that
Bootstrapped performance charts from ROCR

Summary

This could be a somewhat long list of things to do or try, but they all revolve around three main themes: constrained modeling, smoothing and quantification of uncertainty.

Most figures used in this post were taken from the book “Pattern Recognition and Machine Learning” by Christopher Bishop.