Machine learning models have been widely applied in data mining due to their unprecedented prediction capabilities. However, machine learning is often criticized as a “black box” due to its opacity. To tackle the issue, interpretation techniques are needed to understand the working mechanism of models. Interpreting machine learning models, especially deep models, is a challenging problem in data mining because: (1) the definition is vague for interpretation, (2) the structures and information processing paths are convoluted for complex models. I propose to tackle the problem in three aspects. First, given a black-box model, a fundamental requirement of interpretation is to attribute its prediction to the important features. The obtained interpretation could be utilized to improve model robustness. Second, I propose to understand the global latent representations learned by the model to extract structural knowledge. Third, I develop interpretable network embedding models via disentangling latent representations. 

Speaker: Mr Ninghao LIU 
Date: 20 January 2021 (Wed)
Time: 10:00am – 11:00am
Poster: Click here