Exploratory Setting for Unsupervised Learning

Unsupervised learning is easy to describe in the abstract, but a working data scientist needs a more useful question: what are these methods actually for? In practice, unsupervised algorithms show up in two main roles. The first is as a tool for looking at data itself, helping a person build an understanding of what a dataset contains. The second is as a helper inside a larger pipeline, where the output of an unsupervised step feeds into a supervised algorithm. This article walks through both roles using one hypothetical scenario, and it ends with the pitfalls that catch people out.

The scenario works like this. A data scientist receives a table of unlabeled customer records from a small online shop. Nobody has tagged the rows with categories or outcomes. The data scientist wants to know what the table contains before deciding anything. Later in the same project, a second team wants to predict whether a customer will return, and it plans to use a supervised algorithm for that task. Both situations involve unsupervised methods, but they use them in different ways. Keep this scenario in mind as the article proceeds.

A large share of unsupervised work happens in an exploratory setting. Here the goal is not to hand a decision over to an automated system. Instead, a data scientist wants to understand the data better, and the unsupervised algorithm acts as a lens for that understanding. The results inform the person's judgment about the dataset. This is why these methods are typically used on their own, rather than as one component of a larger automatic system.

The two columns of the diagram use the same kind of algorithm, but the purpose differs. In the exploratory column, the person sits at the center and the algorithm's output is judged by that person. In the automated column, the algorithm is one part of a chain, and the surrounding system consumes what it produces. Returning to the scenario, the data scientist who is examining the shop's customer table is working in the exploratory column. Nothing downstream is waiting on the result, so the value lies in what the data scientist learns from it.

The second common application treats an unsupervised algorithm as a preprocessing step. Rather than feeding the raw data straight into a supervised algorithm, the pipeline first passes the data through an unsupervised step that learns a new representation of it. The supervised algorithm then works with this transformed version of the data. The important thing to track is the direction of information flow. The data enters in its original form, the unsupervised step reshapes it, and the reshaped data is what the supervised algorithm learns from.

Applying this to the second team in the scenario, they would not send the raw customer table directly to the return-prediction model. They would insert an unsupervised step that produces a new representation of the customer records, and the prediction model would learn from that representation instead.

The reason to add an unsupervised step to a supervised pipeline is that a new representation of the data can change how well the supervised algorithm performs. The source identifies two kinds of benefit. One is accuracy: a new representation can sometimes improve the accuracy of supervised algorithms. The other is resource use: a new representation can lead to reduced memory and time consumption. The before-and-after view below shows how these outcomes branch from the same change in representation.

For the customer scenario, this means the second team should not assume that a new representation of the customer records will make the return-prediction model more accurate, faster, and smaller all at once. It should compare the model with and without the unsupervised step, and record accuracy, memory use, and time use as three separate results.

Unsupervised algorithms appear in two main roles. As an exploratory tool, they help a data scientist build an understanding of a dataset, and their results inform that person's judgment. As a preprocessing step, they produce a new representation of the data that a supervised algorithm learns from. The benefits of that new representation, whether in accuracy, memory use, or time use, must be measured separately because none of them is guaranteed.