Search

A common approach to aggregate classification estimates in an ensemble of decision trees is to either use voting or to average the probabilities for each class. The latter takes uncertainty into account, but not the reliability of the uncertainty estimates (so to say, the "uncertainty about the uncertainty"). More generally, much remains unknown about how to best combine probabilistic estimates from multiple sources. In this paper, we investigate a number of alternative prediction methods. Our methods are inspired by the theories of probability, belief functions and reliable classification, as well as a principle that we call evidence accumulation. Our experiments on a variety of data sets are based on random decision trees which guarantees a high diversity in the predictions to be combined. Somewhat unexpectedly, we found that taking the average over the probabilities is actually hard to beat. However, evidence accumulation showed consistently better results on all but very small leafs., Comment: Preprint version. Appeared in: Discovery Science. 24th International Conference, DS 2021, Halifax, NS, Canada, October 11 to 13, 2021, Proceedings. See https://ds2021.cs.dal.ca/ for further information

A rule set is a type of classifier that, given attributes X, predicts a target Y. Its main advantage over other types of classifiers is its simplicity and interpretability. A practical challenge is that the end user of a rule set does not always know in advance which target will need to be predicted. One way to deal with this is to learn a multi-directional rule set, which can predict any attribute from all others. An individual rule in such a multi-directional rule set can have multiple targets in its head, and thus be used to predict any one of these. Compared to the naive approach of learning one rule set for each possible target and merging them, a multi-directional rule set containing multi-target rules is potentially smaller and more interpretable. Training a multi-directional rule set involves two key steps: generating candidate rules and selecting rules. However, the best way to tackle these steps remains an open question. In this paper, we investigate the effect of using Random Forests as candidate rule generators and propose two new approaches for selecting rules with multi-target heads: MIDS, a generalization of the recent single-target IDS approach, and RR, a new simple algorithm focusing only on predictive performance. Our experiments indicate that (1) using multi-target rules leads to smaller rule sets with a similar predictive performance, (2) using Forest-derived rules instead of association rules leads to rule sets of similar quality, and (3) RR outperforms MIDS, underlining the usefulness of simple selection objectives. ispartof: pages:517-532 ispartof: Lecture Notes in Computer Science vol:12323 pages:517-532 ispartof: 23rd International Conference on Discovery Science location:Online date:19 Oct - 21 Oct 2020 status: Published online

© 2019, Springer Nature Switzerland AG. Time series are ubiquitous, resulting in substantial interest in time series data mining. Clustering is one of the most widely used techniques in this setting. Recent work has shown that time series clustering can benefit greatly from small amounts of supervision in the form of pairwise constraints. Such constraints can be obtained by asking the user to answer queries of the following type: should these two instances be in the same cluster? Answering “yes” results in a must-link constraint, “no” results in a cannot-link. In this paper we present an interactive clustering system that exploits such constraints. It is implemented on top of the recently introduced COBRAS TS method. The system repeats the following steps until a satisfactory clustering is obtained: it presents several pairwise queries to the user through a visual interface, uses the resulting pairwise constraints to improve the clustering, and shows this new clustering to the user. Our system is readily available and comes with an easy-to-use interface, making it an effective tool for anyone interested in analyzing time series data. Code related to this paper is available at: https://bitbucket.org/toon_vc/cobras_ts/src. ispartof: pages:654-657 ispartof: Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases vol:11053 pages:654-657 ispartof: European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases location:Dublin date:10 Sep - 14 Sep 2018 status: published

Multi-directional ensembles of Classification and Regression treeS (MERCS) extend random forests towards multi-directional prediction. The current work discusses different strategies of induction of such a model, which comes down to selecting sets of input and output attributes for each tree in the ensemble. It has been previously shown that employing multi-targets trees as MERCS component models helps reduce both model induction and inference time. In the current work, we present a novel output selection strategy for MERCS component model that takes relatedness between the attributes into account and compare it to the random output selection. We observe that accounting for relatedness between targets has a limited effect on performance and discuss the reasons why it is inherently difficult to improve the overall performance of a multi-directional model by altering target selection strategy for its component models.

© 2018, Springer Nature Switzerland AG. Clustering is ubiquitous in data analysis, including analysis of time series. It is inherently subjective: different users may prefer different clusterings for a particular dataset. Semi-supervised clustering addresses this by allowing the user to provide examples of instances that should (not) be in the same cluster. This paper studies semi-supervised clustering in the context of time series. We show that COBRAS, a state-of-the-art active semi-supervised clustering method, can be adapted to this setting. We refer to this approach as COBRASTS. An extensive experimental evaluation supports the following claims: (1) COBRASTS far outperforms the current state of the art in semi-supervised clustering for time series, and thus presents a new baseline for the field; (2) COBRASTS can identify clusters with separated components; (3) COBRASTS can identify clusters that are characterized by small local patterns; (4) actively querying a small amount of semi-supervision can greatly improve clustering quality for time series; (5) the choice of the clustering algorithm matters (contrary to earlier claims in the literature). ispartof: pages:179-193 ispartof: Proceedings of the 21st International Conferences on Discovery Science vol:11198 LNAI pages:179-193 ispartof: International Conference on Discovery Science location:Cyprus status: published

© Springer Nature Switzerland AG 2018. A simple but non-trivial setting for automating data science is introduced. Given are a set of worksheets in a spreadsheet and the goal is to automatically complete some values. We also outline elements of the Synth framework that tackles this task: Synth-a-Sizer, an automated data wrangling system for automatically transforming the problem into attribute-value format; TacLe, an inductive constraint learning system for inducing formulas in spreadsheets; Mercs, a versatile predictive learning system; as well as the autocompletion component that integrates these systems. ispartof: pages:3-14 ispartof: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) vol:11191 LNCS pages:3-14 ispartof: 17th International Symposium on Intelligent Data Analysis (IDA) location:NETHERLANDS, Hertogenbosch date:24 Oct - 26 Oct 2018 status: published