Improving the Applicability of Variational Deep Embedding in Unsupervised Large-Scale Data Clustering
Zhu, Wenfei · Uppsala University Publications (Uppsala University) · 2020
The purpose of the thesis is to apply deep clustering (DC) on King'splayer segmentation. To that end we propose six crucial properties a DCneeds to meet in the context of big data applicability. We implement ourmethod based on VaDE (Variational Deep Embedding) together with fourimprovements to meet the six criteria, the method is called S3VaDE, asimple, stable and scalable VaDE. The experiments investigate theaccuracy, stability and scalability between S3VaDE and VaDE on threebenchmark datasets. The results show that S3VaDE outperformed state-ofthe-art. In the thesis, we also demonstrate how to do model selection byvisualizing latent space. We then apply S3VaDE on King's dataset andinterpret the clusters with three KPIs, player engagement, skill leveland monetization. The analysis shows that the clusters are balanced andinterpretable. The investigation further shows that the model is stableduring fine-tune.