Skip to main content

Assessment of Maching Learning Classifiers

I heard another interesting episode of the Data Skeptic podcast. They were discussing how a classifier could be assessed (episode 121). Many machine learning models are so complex that a human being can't really interpret the meaning of the model. This can lead to problems.

They gave an example of a problem where they had a bunch of posts from two discussion boards. One was atheist and the other board was composed of Christians. They tried to classify each post as being from one or the other board. There was one poster who posted heavily on the Christian board. His name was Keith. Sadly, the model learned that if the person who was posting was named Keith, then they were Christian. The problem is that this isn't very useful for prediction. It's an artifact of the input data. Even cross-validation would eliminate this problem. A human being can see the issue, but a model can't.

In any event, the proposed solution was to build interpretable models in local areas of the predictions. If a series of these local models make sense to a human being, then it is likely that the overall model is a good classifier for prediction problems.

I thought this was interesting as it really shows the need for theory in all kinds of modeling problems. There ought to be some substantive theory in which at least elements of the model are grounded.

I also had the opportunity to re-read Groves, Cialdini, and Couper and the decision to participate in a survey. There is some theory development on the survey response process, but there is also room to grow in this area. I wouldn't want an emphasis on using machine learning techniques for weight development or other purposes to distract us from the need to keep working on the theory development problem as well.

Comments

  1. This comment has been removed by a blog administrator.

    ReplyDelete
  2. This comment has been removed by a blog administrator.

    ReplyDelete
  3. This comment has been removed by a blog administrator.

    ReplyDelete

Post a Comment

Popular posts from this blog

Tailoring vs. Targeting

One of the chapters in a recent book on surveying hard-to-reach populations looks at "targeting and tailoring" survey designs. The chapter references this paper on the use of the terms among those who design health communication. I thought the article was an interesting one. They start by saying that "one way to classify message strategies like tailoring is by the level of specificity with which characteristics of the target audience are reflected in the the communication." That made sense. There is likely a continuum of specificity ranging from complete non-differentiation across units to nearly individualized. But then the authors break that continuum and try to define a "fundamental" difference between tailoring and targeting. They say targeting is for some subgroup while tailoring is to the characteristics of the individual. That sounds good, but at least for surveys, I'm not sure the distinction holds. In survey design, what would constitute

What is Data Quality, and How to Enhance it in Research

  We often talk about “data quality” or “data integrity” when we are discussing the collection or analysis of one type of data or another. Yet, the definition of these terms might be unclear, or they may vary across different contexts. In any event, the terms are somewhat abstract -- which can make it difficult, in practice, to improve. That is, we need to know what we are describing with those terms, before we can improve them. Over the last two years, we have been developing a course on   Total Data Quality , soon to be available on Coursera. We start from an error classification scheme adopted by survey methodology many years ago. Known as the “Total Survey Error” perspective, it focuses on the classification of errors into measurement and representation dimensions. One goal of our course is to expand this classification scheme from survey data to other types of data. The figure shows the classification scheme as we have modified it to include both survey data and organic forms of d

An Experimental Adaptive Contact Strategy

I'm running an experiment on contact methods in a telephone survey. I'm going to present the results of the experiment at the FCSM conference in November. Here's the basic idea. Multi-level models are fit daily with the household being a grouping factor. The models provide household-specific estimates of the probability of contact for each of four call windows. The predictor variables in this model are the geographic context variables available for an RDD sample. Let $\mathbf{X_{ij}}$ denote a $k_j \times 1$ vector of demographic variables for the $i^{th}$ person and $j^{th}$ call. The data records are calls. There may be zero, one, or multiple calls to household in each window. The outcome variable is an indicator for whether contact was achieved on the call. This contact indicator is denoted $R_{ijl}$ for the $i^{th}$ person on the $j^{th}$ call to the $l^{th}$ window. Then for each of the four call windows denoted $l$, a separate model is fit where each household is assu