Discovering metric temporal constraint networks on temporal databases

• Authors: Miguel R. ÁLvarez;Paulo FéLix;PurificacióN CariñEna
• Venue: Artificial Intelligence in Medicine
• Year: 2013

Objective: In this paper, we propose the ASTPminer algorithm for mining collections of time-stamped sequences to discover frequent temporal patterns, as represented in the simple temporal problem (STP) formalism: a representation of temporal knowledge as a set of event types and a set of metric temporal constraints among them. To focus the mining process, some initial knowledge can be provided by the user, also expressed as an STP, that acts as a seed pattern for the searching procedure. In this manner, the mining algorithm will search for those frequent temporal patterns consistent with the initial knowledge. Background: Health organisations demand, for multiple areas of activity, new computational tools that will obtain new knowledge from huge collections of data. Temporal data mining has arisen as an active research field that provides new algorithms for discovering new temporal knowledge. An important point in defining different proposals is the expressiveness of the resulting temporal knowledge, which is commonly found in the bibliography in a qualitative form. Methodology: ASTPminer develops an Apriori-like strategy in an iterative algorithm where, as a result of each iteration i, a set of frequent temporal patterns of size i is found that incorporates three distinctive mechanisms: (1) use of a clustering procedure over distributions of temporal distances between events to recognise similar occurrences as temporal patterns; (2) consistency checking of every combination of temporal patterns, which ensures the soundness of the resultant patterns; and (3) use of seed patterns to allow the user to drive the mining process. Results: To validate our proposal, several experiments were conducted over a database of time-stamped sequences obtained from polysomnography tests in patients with sleep apnea-hypopnea syndrome. ASTPminer was able to extract well-known temporal patterns corresponding to different manifestations of the syndrome. Furthermore, the use of seed patterns resulted in a reduction in the size of the search space, which reduced the number of possible patterns from 2.1x10^7 to 1219 and reduced the number of frequent patterns found from 1167 to 340, thereby increasing the efficiency of the mining algorithm. Conclusions: A temporal data mining technique for discovering frequent temporal patterns in collections of time-stamped event sequences is presented. The resulting patterns describe different and distinguishable temporal arrangements among sets of event types in terms of repetitive appearance and similarity of the dispositions between the same events. ASTPminer allows users to participate in the mining process by introducing domain knowledge in the form of a temporal pattern using the STP formalism. This knowledge constrains the search to patterns consistent with the provided pattern and improves the performance of the procedure.