Djamel Abdelkader Zighed

Identifying and Handling Mislabelled Instances

Data mining and knowledge discovery aim at producing useful and reliable models from the data. Unfortunately some databases contain noisy data which perturb the generalization of the models. An important source of noise consists of mislabelled training instances. We offer a new approach which deals with improving classification accuracies by using a preliminary filtering procedure. An example is suspect when in its neighbourhood defined by a geometrical graph the proportion of examples of the same class is not significantly greater than in the database itself. Such suspect examples in the training data can be removed or relabelled. The filtered training set is then provided as input to learning algorithms. Our experiments on ten benchmarks of UCI Machine Learning Repository using 1-NN as the final algorithm show that removal gives better results than relabelling. Removing allows maintaining the generalization error rate when we introduce from 0 to 20% of noise on the class, especially when classes are well separable. The filtering method proposed is finally compared to the relaxation relabelling schema.

Mining Complex Data

Session A1.- Using Text Mining and Link Analysis for Software Mining.- Generalization-Based Similarity for Conceptual Clustering.- Trajectory Analysis of Laboratory Tests as Medical Complex Data Mining.- Session A2.- Conceptual Clustering Applied to Ontologies.- Feature Selection: Near Set Approach.- Evaluating Accuracies of a Trading Rule Mining Method Based on Temporal Pattern Extraction.- Session A3.- Discovering Word Meanings Based on Frequent Termsets.- Quality of Musical Instrument Sound Identification for Various Levels of Accompanying Sounds.- Discriminant Feature Analysis for Music Timbre Recognition and Automatic Indexing.- Session A4.- Contextual Adaptive Clustering of Web and Text Documents with Personalization.- Improving Boosting by Exploiting Former Assumptions.- Discovery of Frequent Graph Patterns that Consist of the Vertices with the Complex Structures.- Session B1.- Finding Composite Episodes.- Ordinal Classification with Decision Rules.- Data Mining of Multi-categorized Data.- ARAS: Action Rules Discovery Based on Agglomerative Strategy.- Session B2.- Learning to Order: A Relational Approach.- Using Semantic Distance in a Content-Based Heterogeneous Information Retrieval System.- Using Secondary Knowledge to Support Decision Tree Classification of Retrospective Clinical Data.- POM Centric Multi-aspect Data Analysis for Investigating Human Problem Solving Function.

Topological comparisons of proximity measures

In many fields of application, the choice of proximity measure directly affects the results of data mining methods, whatever the task might be: clustering, comparing or structuring of a set of objects. Generally, in such fields of application, the user is obliged to choose one proximity measure from many possible alternatives. According to the notion of equivalence, such as the one based on pre-ordering, certain proximity measures are more or less equivalent, which means that they should produce almost the same results. This information on equivalence might be helpful for choosing one such measure. However, the complexity O (n 4 ) of this approach makes it intractable when the size n of the sample exceeds a few hundred. To cope with this limitation, we propose a new approach with less complexity O (n 2 ). This is based on topological equivalence and it exploits the concept of local neighbors. It defines equivalence between two proximity measures as having the same neighborhood structure on the objects. We illustrate our approach by considering 13 proximity measures used on datasets with continuous attributes.

Neighborhood Random Classification

Neighborhood Random Classification Diala Ezzeddine Universit´e de Lyon (Lumi`ere Lyon 2) – Laboratoire ERIC – diala.ezzeddine@univ-lyon2.fr 29 August 2013 D. Ezzeddine (ERIC) Neighborhood Random Classification 29 August 2013 1 / 18 Introduction We propose to use neighborhood graphs in Ensemble Method EM Main purpose : Using neighborhood graphs is a strong alternative In this work, we : Used an EM classifier based on neighborhood, Random Neighborhood Classifier (RNC) Compared RNC to kNN, then to other methods D. Ezzeddine (ERIC) Neighborhood Random Classification 29 August 2013 2 / 18 Introduction Outline 1 Introduction 2 Basic Concepts Classifier Neighborhood Classifiers Partition by Neighborhood Graphs 3 Ensemble Methods 4 Neighborhood Random Classifier 5 Result 6 Conclusion D. Ezzeddine (ERIC) Neighborhood Random Classification 29 August 2013 3 / 18 Introduction Outline 1 Introduction 2 Basic Concepts Classifier Neighborhood Classifiers Partition by Neighborhood Graphs 3 Ensemble Methods 4 Neighborhood Random Classifier 5 Result 6 Conclusion D. Ezzeddine (ERIC) Neighborhood Random Classification 29 August 2013 3 / 18 Introduction Outline 1 Introduction 2 Basic C

Neighborhood random classification

Ensemble methods (EMs) have become increasingly popular in data mining because of their efficiency. These methods(EMs) generate a set of classifiers using one or several machine learning algorithms (MLAs) and aggregate them into a single classifier (Meta-Classifier, MC). Of the MLAs, k-Nearest Neighbors (kNN) is one of the most well-known used in the context of EMs. However, handling the parameter k can be difficult. This drawback is the same for all MLA that are instance based. Here, we propose an approach based on neighborhood graphs as an alternative. Thanks to these related graphs, like relative neighborhood graphs (RNGs) or Gabriel graphs (GGs), we provide a generalized approach with less arbitrary parameters. Neighborhood graphs have never been introduced into EM approaches before. The results of our algorithm : Neighborhood Random Classification are very promising as they are equal to the best EM approaches such as Random Forest or those based on SVMs. In this exploratory and experimental work, we provide the methodological approach and many comparative results.

Comparison of Proximity Measures: A Topological Approach

In many application domains, the choice of a proximity measure affect directly the result of classification, comparison or the structuring of a set of objects. For any given problem, the user is obliged to choose one proximity measure between many existing ones. However, this choice depend on many characteristics. Indeed, according to the notion of equivalence, like the one based on pre-ordering, some of the proximity measures are more or less equivalent. In this paper, we propose a new approach to compare the proximity measures. This approach is based on the topological equivalence which exploits the concept of local neighbors and defines an equivalence between two proximity measures by having the same neighborhood structure on the objects.We compare the two approaches, the pre-ordering and our approach, to thirty five proximity measures using the continuous and binary attributes of empirical data sets.