BIGDATA: Small: DA: Collaborative Research: From Data to Users: Providing Interpretable and Verifiable Explanations in Data Mining

点击翻译按钮获取中文摘要

The fruits of data mining pervade every aspect of our lives. We have books and movies recommended; we are given differential pricing for insurance; screened for potential terror threats; diagnosed with various diseases; and targeted for political advertising. The ability to sift through massive data sets with sophisticated algorithms has resulted in applications with impressive predictive power. And yet there is still a gap between what such tools can deliver, and what the users of data mining really need. It is often hard to interpret the answers produced by a learning algorithm, due to its sophistication and the use of large data sets to build models. The results of mining are often "one-size-fits-all", and convincing a user that results are actually relevant to them is difficult. Finally, there is the important problem of validation. As the results of data mining affect more and more of our lives, the more crucial it is that the user be able to validate decisions made on their behalf and that affect them. The common theme tying these issues together is a user-centric perspective on the problems of data mining. Rather than asking "What patterns can be found in this mountain of data?" this work instead asks "What structures in this data affect me?" These issues arise precisely because of the vast amounts of data we now have the ability to mine, and the sophisticated methods at our disposal to analyze this data. In this research, the PIs develop a computational framework and key tools for user-centric data mining. A central theme in this research is the idea of interaction. In both machine learning and in the foundations of complexity theory, interaction has been used to allow a (weaker) entity to probe a much more powerful system and determine answers that it lacks the resources to compute directly itself. The PIs use formal interaction mechanisms both from the perspective of a user interacting with a powerful algorithm, as well as a client interacting with a computing source with access to large data, in order to enable the user to interpret and validate the results of data mining. The goal of this project is to develop a computational framework for user-centric data mining that enables existing users to tailor data analysis to their needs and facilitates the use of data mining in new areas where existing The team proposes interactive mechanisms that start with the results of a learning process and, via interaction with the user, produce an explanation expressed in terms of meaningful features, drawing on ideas from active learning, feature selection, and domain adaptation. 2. Locality: Answers that are relevant. Here, the focus is on providing information that depends more on a user?s local neighborhood, achieved via a new local notion of stability. 3. Verifiability: Answers you can check. The team proposes a framework for the validation of computationally-intensive data mining by the computationally-weak user, with ideas from interactive proof theory and stream algorithms. Tools for analyzing patient medical data have become more sophisticated and individual medical profiles play a far more significant role in diagnosis and treatment.The research examines user-centric data mining via three core primitives (classification, regression and clustering), and studies the three problems of interpreting results, providing local explanations, and validating the results of data mining. Firstly, the research draws on ideas from active learning, feature selection and domain adaptation to build interpretable results via interaction with users. Secondly, it introduces local notions of stability as a way of validating predictions for a specific user. Finally, it develops a general framework for validation of an analysis by a computationally-weak user, by drawing on ideas from the theory of interactive proofs and streaming algorithms.