Filtering irrelevant sequential data out of game session telemetry though similarity collapses

A multitude of game sessions is started every day, generating a huge amount of data that may be useful in many different situations. For this reason, game telemetry is an important trend and feature in modern games, especially for tuning games, quality assurance, testing, and finding game aspects that need to be calibrated. The wealth of tracked information is fundamental for analysis and understanding of events, mistakes, and fluxes of a concrete game session. However, due to game dynamics, the resulting telemetry data may be overwhelming in size, making it difficult to identify sections of interest in the game session. The existing telemetry approaches normally use density clustering techniques for visual analysis, losing the temporal relationship in the process. In order to solve this problem, in this paper we propose three different similarity collapse algorithms based on the classic DBSCAN algorithm, collapsing sequential information of the graph that has similar values or represents the same states. These algorithms allow the game designer to quickly identify relevant information or state transitions without compromising the temporal sequence of events. We implemented this solution over tracked provenance data, which is graph-based, and provide two experimental studies of these algorithms using automatic experimentation and human judges to evaluate each of the proposed algorithms. These experiments show that one of the proposed algorithms is superior to DBSCAN when applied to graphs by better preserving the data semantics when collapsing provenance data. We believe that this kind of approaches will become a trend in the future process of game development.