Beyond Understanding and Prediction: Data Mining for Action
Lists
Only you see this
Abstract
Association analysis and prediction are two major tasks in data mining, and they represent two foremost objectives: data exploration for understanding and model construction for prediction. Data mining is known as a process to convert raw data to useful information --- knowledge. However, what do we do with the knowledge discovered from data? We will need knowledge to enable actions, such as preventing diseases in health care, taking actions to retain customers, and etc. One important attribute of knowledge is actionability. In order to be acted on, the knowledge must encode causal relationships to imply the mechanisms of the systems under consideration. Causal inference is a sophisticated topic that spans multiple disciplines, computer science, statistics, medicine, economics, and social science, to name a few. There are a number of well-established frameworks for causal inference in data with assumptions. Unfortunately, the assumptions are not directly testable in data, and hence there are limited off-the-shelf data mining methods for causal discovery. Many practitioners still use conventional machine learning methods for the tasks which actually requires causal inference [1]. It is desirable to have some simple data mining tools to explore causal relationships without or with few assumptions. The discoveries may not be proved causal, but are high quality candidates excluding many non-causal or spurious relationships. This will be a significant step forward in actionable data mining since it is well known that association rule mining generates too many spurious relationships and a classification model often provides a correct prediction based on non-interpretable (or even wrong) evidence. The spurious relationships and non-interpretable or wrong evidence hinder many data mining applications, especially in medical and social science areas. We have been using some well-known causal inference principles to discover causal relationships and building causally interpretable data mining models, such as causal rules [2] and causal decision trees [3], and applying causal discovery methods to real world biological problems [4, 5]. In this talk, I will discuss some of our exploratory work in this promising direction.
BibTeX
@inproceedings{li2017beyond,
author = {Jiuyong Li},
title = {{Beyond Understanding and Prediction: Data Mining for Action}},
booktitle = {The Web Conference},
year = {2017},
doi = {10.1145/3041021.3053407},
}Loading related content...