Synthetic Data for Clustering Experiments

This article begins with a complete, short program rather than with definitions. The program builds a small synthetic data set, creates a k-means clustering model, and trains that model on the data. Read it once before you go on, and notice which line you expect to be the most important one.

The script has three jobs. It obtains data, it builds a model object, and it trains that model. Each job maps to one or two lines. Once you can name the job a line performs, you know which part of the program to inspect when results look wrong.

Reading the script this way means you can answer three questions for any line. Does it create data, create a model, or train the model? Mapping each line to one job turns a short program into three checkpoints. The rest of this article takes each job in turn.

In this script the data is produced by make_blobs, which returns two things, X and y. The variable X holds the two-dimensional points that the clustering step will work on. The variable y holds a second array. The script stores it, but it never passes y to the clustering step. The clustering model in this script sees only X.

The argument random_state=1 controls the random generation so that the same synthetic points are produced on each run. This is important for debugging. If two runs of the same script disagree, the data was not the cause, because the data is fixed by the seed. You can therefore set the data aside as a suspect when you compare runs.

The source describes this step as instantiating the KMeans class and setting the number of clusters you are looking for. In code, that is the single line kmeans = KMeans(n_clusters=3). The class name is KMeans. The keyword argument is n_clusters. The value 3 is the number of clusters the model is asked to find.

Notice what this constructor line does and does not establish. It records the setting that the model should use. It does not yet find any groups. The learning from the data happens in the next step, when fit is called. A learner who expects the object creation line to find the clusters will be surprised, and that surprise is a useful signal that the training step is the place to look.

The source states that the fit method is called on the KMeans object, with the data as an argument. The line kmeans.fit(X) therefore has two parts. The first part is the object whose method runs, which is kmeans. The second part is the data that method receives, which is X, the array produced by make_blobs.

When you trace the script, watch the moment fit runs. Before that line, the model holds only its setting. After it, the model has been trained on the points in X. Any unexpected grouping you see must therefore come from one of three places: the data passed in, the setting given to the constructor, or the training call itself.

When a clustering result does not match what you expected, avoid re-reading the whole script at once. Work through the three jobs in order. The first job whose state differs from your expectation is where the divergence begins.