JIDECA: Jointly Improved Deep Embedded Clustering for Android activity
Choi Sung-Min, Hyeon-Tae Seo, Yo-Sub Han · 2023
Recently, clustering research using deep neural networks to learn latent vectors is growing rapidly. In particular, people are interested in developing clustering methods based on the result of deep embedded clustering (DEC). Among several clustering applications, we examine Android activity clustering using the Rico dataset. We propose a Jointly Improved Deep Embedded Clustering for Android activity (JIDECA) method for better activity clustering. We generate various activity latent vectors using a DNN autoencoder and a CNN autoencoder with the Rico dataset. Simultaneously, JIDECA learns latent vectors and performs clustering with local structure preservation for each modality. In addition, JIDECA uses cross-modality alignment loss to make each single modality similar. Our experimental results show that JIDECA outperforms both single modal methods and multimodal methods for real activity images.