Deep Clustering for Mixed-type Data with Frequency Encoding and Doubly Weighted Cross Entropy Loss
Deogho Choi, Daniel H. Chae, Woo-Yeon Kim, Jihong Kim, Janghoon Yang, Jitae Shin · 2022
Clustering algorithm is unsupervised learning that groups a set of data into distinctive classes according to the similarity between each data sample. Most of previous researches have focused on improving K-prototypes or training proper numerical representations of categorical features using autoencoder. But in this research, we investigate that applying frequency encoding to categorical features can be sufficiently effective. Furthermore, we propose doubly weighted cross entropy loss, DW-CE loss, to find optimal cluster centroid by training fully connected layer. The experiment with two mixed-type datasets, credit approval and heart disease, from UCI repository shows that the proposed clustering with frequency encoding and DW-CE loss provides better performance than existing state of the arts methods in most of cases.