Julien Taillard

Conception of a dominance-based multi-objective local search in the context of classification rule mining in large and imbalanced data sets

Graphical abstractDisplay Omitted HighlightsFormulation of the classification rule mining problem as a multi-objective problem.Proposal of MOCA-I that deals both with uncertainty, class imbalance and volumetry.Comparison of different MOCA-I based DMLS versions and DMLS 1?* shows better results.Comparison with 13 state-of-the-art classification algorithms.MOCA-I gives shorter and statistically more effective rules than other algorithms. Classification on medical data raises several problems such as class imbalance, double meaning of missing data, volumetry or need of highly interpretable results. In this paper a new algorithm is proposed: MOCA-I (Multi-Objective Classification Algorithm for Imbalanced data), a multi-objective local search algorithm that is conceived to deal with these issues all together. It is based on a new modelization as a Pittsburgh multi-objective partial classification rule mining problem, which is described in the first part of this paper. An existing dominance-based multi-objective local search (DMLS) is modified to deal with this modelization. After experimentally tuning the parameters of MOCA-I and determining which version of DMLS algorithm is the most effective, the obtained MOCA-I version is compared to several state-of-the-art classification algorithms. This comparison is realized on 10 small and middle-sized data sets of literature and 2 real data sets; MOCA-I obtains the best results on the 10 data sets and is statistically better than other approaches on the real data sets.

The benefits of using multi-objectivization for mining pittsburgh partial classification rules in imbalanced and discrete data

A large number of rule interestingness measures have been used as objectives in multi-objective classification rule mining algorithms. Aggregation or Pareto dominance are commonly used to deal with these multiple objectives. This paper compares these approaches on a partial classification problem over discrete and imbalanced data. After performing a Principal Component Analysis (PCA) to select candidate objectives and find conflictive ones, the two approaches are evaluated. The Pareto dominance-based approach is implemented as a dominance-based local search (DMLS) algorithm using confidence and sensitivity as objectives, while the other is implemented as a single-objective hill climbing using F-Measure as an objective, which combines confidence and sensitivity. Results shows that the dominance-based approach obtains statistically better results than the single-objective approach.

Extraction and optimization of classification rules for temporal sequences

This study focuses on the problem of supervised classification on heterogeneous temporal data featuring a mixture of attribute types (numeric, binary, symbolic, temporal). We present a model for classification rules designed to use both non-temporal attributes and sequences of temporal events as predicates. We also propose an efficient local search-based metaheuristic algorithm to mine such rules in large scale, real-life data sets extracted from a hospitals information system. The proposed algorithm, MOSC (Multi-Objective Sequence Classifier), is compared to standard classifiers and previous works on these real data sets and exhibits noticeably better classification performance. While designed with medical applications in mind, the proposed approach is generic and can be used for problems from other application domains.

A scalable biclustering method for heterogeneous medical data

We define the problem of biclustering on heterogeneous data, that is, data of various types (binary, numeric, etc.). This problem has not yet been investigated in the biclustering literature. We propose a new method, HBC (Heterogeneous BiClustering), designed to extract biclusters from heterogeneous, large-scale, sparse data matrices. The goal of this method is to handle medical data gathered by hospitals (on patients, stays, acts, diagnoses, prescriptions, etc.) and to provide valuable insight on such data. HBC takes advantage of the data sparsity and uses a constructive greedy heuristic to build a large number of possibly overlapping biclusters. The proposed method is successfully compared with a standard biclustering algorithm on small-size numeric data. Experiments on real-life data sets further assert its scalability and efficiency.

MOCA-I: Discovering Rules and Guiding Decision Maker in the Context of Partial Classification in Large and Imbalanced Datasets

This paper focuses on the modeling and the implementation as a multi-objective optimization problem of a Pittsburgh classification rule mining algorithm adapted to large and imbalanced datasets, as encountered in hospital data. We associate to this algorithm an original post-processing method based on ROC curve to help the decision maker to choose the most interesting rules. After an introduction to problems brought by hospital data such as class imbalance, volumetry or inconsistency, we present MOCA-I - a Pittsburgh modelization adapted to this kind of problems. We propose its implementation as a dominance-based local search in opposition to existing multi-objective approaches based on genetic algorithms. Then we introduce the post-processing method to sort and filter the obtained classifiers. Our approach is compared to state-of-the-art classification rule mining algorithms, giving as good or better results, using less parameters. Then it is compared to C4.5 and C4.5-CS on hospital data with a larger set of attributes, giving the best results.