Category: Data Science

John Mount demonstrates an important concept:

Our business goal is to build a model relating attendance to popcorn sales, which we will apply to future data in order to predict future popcorn sales. This allows us to plan staffing and purchasing, and also to predict snack bar revenue.

In the above example data, all dates in August of 2024 are “in the past” (available as training and test/validation data) and all dates in September of 2024 are “in the future” (dates we want to make predictions for). The movie attendance service we are subscribing to supplies

• past schedules
• past (recorded) attendance
• future schedules, and
• (estimated) future attendance.

John’s example scenario covers the problem of future estimations interfering with model quality. Another important scenario is when the past changes. As one example, digital marketing providers (think Google, Bing, Amazon, etc.) will provide you impression and click data pretty quickly, and each day they close the books on a prior day’s data at some normal time. For some of these providers, that prior day’s data is yesterday’s data—on Tuesday, provider X closes the books on Monday’s data and promises that it won’t change after that. But for other providers, they might change data over the course of the next 10 days. This means that the data you’re using for model training might change from under you, and you might never know if you don’t keep track of the actual data you used for training at the time of training.

Comments closed

Peter Ellis finds interesting results with sampling in R:

A week ago I was surprised to read on Thomas Lumley’s Biased and Inefficient blog that when using R’s sample() function without replacement and with unequal probabilities of individual units being sampled:

“What R currently has is sequential sampling: if you give it a set of priorities w it will sample an element with probability proportional to w from the population, remove it from the population, then sample with probability proportional to w from the remaining elements, and so on. This is useful, but a lot of people don’t realise that the probability of element i being sampled is not proportional to w_i”

Read on for a demonstration. H/T R-Bloggers.

Comments closed

Michael Mayer wants to suss out the effects of inputs into a causal forest model:

We use a causal forest [1] to model the treatment effect in a randomized controlled clinical trial. Then, we explain this black-box model with usual explainability tools. These will reveal segments where the treatment works better or worse, just like a forest plot, but multivariately.

Read on for the example, as well as several mechanisms you can use to gauge feature relevance.

Comments closed

Michael Mayer handles missing data:

{missRanger} is a multivariate imputation algorithm based on random forests, and a fast version of the original missForest algorithm of Stekhoven and Buehlmann (2012). Surprise, surprise: it uses {ranger} to fit random forests. Especially combined with predictive mean matching (PMM), the imputations are often quite realistic.

This looks like an interesting package. At first, I thought it was a way of generating predictions outside the boundaries of training data and had concerns—a classic point (limitation?) of random forest as an algorithm is that it will not even try to predict values outside the range of what it sees in training data, so if the largest label is 10 and the smallest is 0, you won’t see a prediction of 11 or 50, no matter how you scale the inputs.

Instead of doing that, missRanger looks like it’s filling in missing data using a clever approach. That’s quite useful for dealing with incomplete data, a really common problem whose good solutions tend to be complex enough that people typically ignore them in favor of simple but less useful solutions like dropping rows altogether.

Comments closed