Extracting and Harnessing Interpretation in Data Mining
Abstract
Machine learning, especially the recent deep learning technique, has aroused significant development to various data mining applications, including recommender systems, misinformation detection, outlier detection, and health informatics. Unfortunately, while complex models have achieved unprecedented prediction capability, they are often criticized as ``black boxes'' due to multiple layers of non-linear transformation and the hardly understandable working mechanism.
To tackle the opacity issue, interpretable machine learning has attracted increasing attentions. Traditional interpretation methods mainly focus on explaining predictions of classification models with gradient based methods or local approximation methods. However, the natural characteristics of data mining applications are not considered, and the internal mechanisms of models are not fully explored. Meanwhile, it is unknown how to utilize interpretation to improve models. To bridge the gap, I developed a series of interpretation methods that gradually increase the transparency of data mining models. First, a fundamental goal of interpretation is providing the attribution of input features to model outputs. To adapt feature attribution to explaining outlier detection, I propose Contextual Outlier Interpretation (COIN). Second, to overcome the limitation of attribution methods that do not explain internal information inside models, I further propose representation interpretation methods to extract knowledge as a taxonomy. However, these post-hoc methods may suffer from interpretation accuracy and the inability to directly control model training process. Therefore, I propose an interpretable network embedding framework to explicitly control the meaning of latent dimensions. Finally, besides obtaining explanation, I propose to use interpretation to discover the vulnerability of models in adversarial circumstances, and then actively prepare models using adversarial training to improve their robustness against potential threats.
My research of interpretable machine learning enables data scientists to better understand their models and discover defects for further improvement, as well as improves the experiences of customers who benefit from data mining systems. It broadly impacts fields such as Information Retrieval, Information Security, Social Computing, and Health Informatics.
Citation
Liu, Ninghao (2021). Extracting and Harnessing Interpretation in Data Mining. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /193073.