Mining Frequent Data Streams of High Dimensionality with a Case Study in Digital Games

Abstract

In a data stream with many attributes and a high frequency of events, how to spot clusters of similar events? How to identify outlying events, and sort them according to how much they differ from the main data patterns? Can it be performed in real time? For example, based on the actions of users over time in an online digital game management system, how to identify clusters of users/players that share similar preferences, so to assist targeted marketing and the development of new games? How to automatically identify spammers, bots, hackers and famous players, aimed at banning users in the first three categories as well as understanding why those in the last category are famous? Today, there exists a need for accurate, fast and scalable algorithms able to extract patterns in real time from streams with many attributes and a high frequency of events, such as the many streams that are constantly collected by web systems and automated sensors from a variety of modern applications. This project aims to reduce the aforementioned issue focused on clustering and outlier detection in streams with thousands of new events received per second, each one described by tens/hundreds of attributes. The new methodologies and algorithms to be developed will be validated in a case study with the analysis of streams from a digital game management system that controls online matches worldwide, as well as the creation/publication of new content for the game Super Mario Maker, from the developer Nintendo Company Ltd.. (AU)