Temporal Pattern Mining in Dynamic Environments Andreas D. Lattner Dynamic scenes with many different objects and interrelations changing over time demand complex representations. The identification of frequent patterns and prediction rules in such scenes would be very valuable as associations in the data could be discovered or a system’s performance could even be improved by utilizing the new information in the behavior decision process. In this work, a novel approach to temporal pattern mining in dynamic environments has been proposed. 1 Introduction Many domains feature a dynamic characteristic and it would be useful to learn temporal patterns, e.g., in logistics, sports, and medicine. In the logistics domain, for instance, there might be many different kinds of objects like different transport vehicles (e.g., trucks, ships, or planes), different actors or organizations (e.g., storages, transport companies, manufacturers), highways or tracks, reloading points, etc. Different events can occur like traffic jams, weather events, break down of a transport vehicle, or delay of some goods. It would be valuable to identify repeating patterns that lead to certain situations in order to predict a traffic jam or a delay and initiate some counter-actions in order to avoid financial loss or penalty payments. Such a pattern could be, for instance: If the traffic density is medium and increasing on highway X on a Friday afternoon and the weather is rainy, it is likely that there will follow a traffic jam on highway Y. In the requirements analysis, different demands on representation formalisms for dynamic scenes as well as for patterns to be mined from dynamic scenes have been identified from a soccer scenario of the RoboCup simulation league. Among other demands the representation of objects and relations and the temporal validity of these relations have been detected to be crucial. As various actions and events in dynamic scenes can occur concurrently, it is important that the representation also supports this concurrency. A comprehensive investigation of the state of the art has led to a number of relevant approaches covering parts of the requirements. Particularly of interest were the approaches dealing with association rule mining, especially extensions for sequential, temporal, or relational data. WARMR [3] can be used to mine relational association rules but has no direct means to represent the temporal dimension. The work of Höppner [5] which addresses rule learning from interval-based temporal data uses interval relations in order to represent temporal interrelations among states in the input data. A third relevant approach by Lee [6] allows for mining first-order sequential rules but does not support concurrently occurring intervals of relations. 2 Mining Temporal Patterns Based on the defined requirements and on the analyzed approaches, the temporal pattern mining approach MiTemP has been developed. It can mine temporal patterns from time intervalbased relational data and additional conceptual information about objects and their interrelations in dynamic scenes. The relevant concepts of the approach have been defined formally. Besides formal definitions of dynamic scenes and their schemata as well as definitions for patterns and prediction rules, a partial order has been defined for the generalization relation between patterns. Following ideas of Lee [6], an optimal refinement operator has been defined that guarantees a complete and non-redundant generation of frequent patterns from the given representation. In the conceptual chapter of this thesis, it has also been shown how WARMR can be used in order to mine temporal patterns albeit it generates many redundant ones. The pattern mining approach is based on the association rule mining algorithm Apriori [1]. It starts with the most general (empty) pattern and successively performs refinement operations to those patterns that still exceed the minimal frequency threshold. Five refinement operations have been set up: lengthening, temporal refinement, unification, concept refinement, and instantiation. Concept restrictions can represent the information that specific variables can only be bound to instances of certain concepts in the concept hierarchy. In order to represent temporal relations in patterns, a new concise set of mutual exclusive and jointly exhaustive interval relations has been set up by combining ideas from Allen’s and Freksa’s interval relations [2, 4]. For the interval relations (before/after, older/younger & contemporary, head-to-head), a composition table and an temporal reasoning algorithm has been set up. The reason for the new set of interval relations was to reduce complexity and to focus on important relations for prediction rules. However, the set of interval relations can be easily replaced without changing the mining algorithm, e.g., by using Allen’s interval relations [2]. It can easily happen that a huge number of patterns is generated. Therefore, different means to restrict the relevant pattern space have been introduced. It is possible to disable single refinement types (e.g., no instantiation) or to restrict the maximal refinement level. If it is known before mining that only certain patterns with some predicates are of interest, a bias can be set up consisting of partial conjunctive patterns. If such a bias is defined, the patterns that are inconsistent with this bias are filtered out during pattern mining. Another way to reduce the number of patterns is a selection of patterns to be refined at each refinement level. In the MiTemP implementation, a random selection of n patterns can be done before the refinement is performed. Fig. 1 illustrates the time interval-based dynamic scene rep- developed temporal pattern mining approach as well as the prediction rule generation. Using soccer matches from the 2D and 3D simulation league, prediction rules with average accuracies of 70.18% and 63.46% on unseen data could be generated. Of course, it is possible to filter out prediction rules with higher accuracies (by increasing the minimal confidence). The advantage of MiTemP in comparison to WARMR has been shown; bringing in the knowledge about temporal relations and concepts can reduce the number of patterns to be generated (and checked during pattern matching) significantly. In an example, WARMR has generated more than 42000 patterns while MiTemP has only created 7280 at refinement level seven. Acknowledgment Figure 1: Pattern and prediction rule generation resentation and the pattern generation and matching. The area marked by the dashed rectangle represents a sliding time window; matches of temporal patterns are restricted to fit into such time window with a given extent. 3 Prediction Rule Generation In order to utilize the identified frequent patterns for prediction, an algorithm for prediction rule generation has been introduced. Prediction rules are like other association rules with the restriction that the consequence of the rule must be past the precondition. As it has been shown in this thesis, this has the advantage of only linear effort w.r.t. the conjunctive pattern size. Due to the anti-monotonicity property of the frequency, the prediction rule generation procedure can even be stopped if the minimal confidence value cannot be reached any more. Thus, an efficient method for creating prediction rules from temporal patterns has been introduced. An example for a prediction rule is: ((pass(X, Y ), pass(Y, X) score(X)), {(1, 2, bef ore), (1, 3, bef ore), (2, 3, bef ore)}, {(X, object), (Y, object)}). This prediction rule says that after X passed to Y and Y passed to X, X will score (temporal restrictions are represented by the indices of predicate pairs and interval relations). The conceptual restriction {(X, object), (Y, object)} says that X and Y must be instances of the concept object. Furthermore, different criteria for the evaluation of prediction rules have been set up: frequency, confidence, information value (J-measure), predicate preference, size, and specifity. The different criteria are put together by a weighted sum to an overall evaluation value. If certain aspects should not be considered in the measure, the corresponding weights can be set to zero. I would like to thank my doctoral advisor Prof. Dr. Otthein Herzog for his continuous support and his motivating as well as inspiring comments while I have been writing this dissertation. I would also like to express my gratitude to Prof. Dr. Stefan Wrobel for many helpful comments and the evaluation of the thesis as the second referee. Furthermore, I want to thank my (partly former) colleagues from the Artificial Intelligence Research Group (AG-KI) at the Universität Bremen for many exciting discussions about their as well as my research. References [1] [2] [3] [4] [5] [6] Contact Dr. Andreas D. Lattner Johann Wolfgang Goethe-Universität (FB 12) Information Systems and Simulation P.O. Box 11 19 32, D-60054 Frankfurt am Main E-mail: lattner@informatik.uni-frankfurt.de Bild 4 Results The learning approach has been implemented in XSB Prolog. The evaluation consists of four parts: a simple example for illustration purposes, a number of experiments with synthetical data, a comparison of WARMR and MiTemP, and the generation of prediction rules from RoboCup soccer matches. The experiments have shown the successful implementation of the R. Agrawal and R. Srikant. Fast algorithms for mining association rules. In Proceedings of the 20th International Conference on Very Large Data Bases, p. 487-499, September 1994. J. F. Allen. Maintaining knowledge about temporal intervals. Communications of the ACM, 26(11):832-843, November 1983. L. Dehaspe and H. Toivonen. Discovery of frequent DATALOG patterns. Data Mining and Knowledge Discovery, 3(1):7-36, March 1999. C. Freksa. Temporal reasoning based on semi-intervals. Artificial Intelligence, 54(1-2):199-227, 1992. F. Höppner. Knowledge Discovery from Sequential Data. PhD thesis, Technische Universität Braunschweig, 2003. S. D. Lee. Constrained Mining of Patterns in Large Databases. PhD thesis, Albert-Ludwigs-Universität Freiburg, 2006. Andreas D. Lattner studied computer science at the Universität Bremen. Afterwards he joined the Center for Computing Technologies (TZI, Universität Bremen) where he received a doctor’s degree in 2007. He is now working at the chair for Information Systems and Simulation, Universität Frankfurt. His research interests include machine learning, temporal pattern mining, and multi-agent systems.