Using Kmeans clustering to identify 5 cancer types

Performed K-Means clustering on a real world Genome Atlas. This dataset is available from the UC Irvine Machine Learning Repository. Data contains gene expressions from 5 distinct cancer subtypes. There are 881 samples (rows) and 20,531 gene expressions (columns).

Preprocessing

The labels for cancer types are in string format. Hence, I used label encoder to transform them to a numerical format. Since the data has large number of dimensions, I reduced 20,531 columns to 2 columns using PCA. For this, I built pipelines for feauture scaling and performing Principal Component analysis. I have use MinMax scaler so that the data doesnt have to follow a normal distribution.

Modeling using K_means clustering

Next, I built another pipeline for performing Kmeans clustering. Since there are 5 cancer types, I initialized k to be 5 and fit the tranformed data from PCA.

Model Evaluation

In order to evaluate the model, I used Silhouette coefficient and attained a score of 0.51. As ground truth labels are provided in the dataset, I performed ARI to calculate adjusted Rand Score to compare ground truth labels with predicted labels and rand score came to be 0.72.

Visualization

Visualized the clusters in 2D space by using principal components. The model performed good but there was a slight overlap cuased from squishing 20,531 dimensions to 2 dimensions.

Tuning the model

In order to improve the model, I calculated silhouette coefficient and Rand scores for all the principal components ranging from 2 to 11 and plotted the scores using a line chart. Rand score improved to 0.95 when number of principal components is equal to 7.