We are given a data set of items, which have certain features, and values for these features (like a vector). We are tasked with categorizing those items into groups. To achieve this, we will use the kMeans algorithm, an unsupervised learning algorithm. Language of implementation is Python.
(It will help if you think of items as points in an n-dimensional space)
The algorithm will categorize the items into k groups of similarity. To calculate that similarity, we will use the euclidean distance as measurement.
The algorithm works as follows: First we initialize k points, called means, randomly. We categorize each item to its closest mean and we update the mean’s coordinates, which are the averages of the items categorized in that mean so far. We repeat the process for a given number of iterations and at the end, we have our clusters.
The “points” mentioned above are called means, because they hold the mean values of the items categorized in it. To initialize these means, we have a lot of options. An intuitive method is to initialize the means at random items in the data set. Another method is to initialize the means at random values between the boundaries of the data set (if for a feature x the items have values in [0,3], we will initialize the means with values for x at [0,3]).
The above algorithm in pseudocode:
Initialize k means with random values For a given number of iterations: Iterate through items: Find the mean closest to the item Assign item to mean Update mean
We receive input as a text file (‘data.txt’). Each line represents an item, and it contains numerical values (one for each feature) split by commas. You can find a sample data set here.
We will read the data from the file, saving it into a list. Each element of the list is another list containing the item values for the features. We do this with the following function:
We want to initialize each mean’s values in the range of the feature values of the items. For that, we need to find the min and max for each feature. We accomplish that with the following function:
The variables minima, maxima are lists containing the min and max values of the items respectively. We initialize each mean’s feature values randomly between the corresponding minimum and maximum in those above two lists:
We will be using the euclidean distance as a metric of similarity for our data set (note: depending on your items, you can use another similarity metric).
### Update Means
To update a mean, we need to find the average value for its feature, for all the items in the mean/cluster. We can do this by adding all the values and then dividing by the number of items, or we can use a more elegant solution. We will calculate the new average without having to re-add all the values, by doing the following:
m = (m*(n-1) + x) / n
where m is the mean value for a feature, n is the number of items in the cluster and x is the feature value for the added item. We do the above for each feature to get the new mean.
You can read more on this technique on this post.
Now we need to write a function to classify an item to a group/cluster. For the given item, we will find its similarity to each mean, and we will classify the item to the closest one.
To actually find the means, we will loop through all the items, classify them to their nearest cluster and update the cluster’s mean. We will repeat the process for some fixed number of iterations. If between two iterations no item changes classification, we stop the process as the algorithm has found the optimal solution.
The below function takes as input k (the number of desired clusters), the items and the number of maximum iterations, and returns the means and the clusters. The classification of an item is stored in the array belongsTo and the number of items in a cluster is stored in clusterSizes.
Finally we want to find the clusters, given the means. We will iterate through all the items and we will classify each item to its closest cluster.
You can find the entire code on my GitHub, along with a sample data set and a plotting function. Thanks for reading.