Abstract. This paper describes a method for fusing a collection of classifiers where the fusion can compensate for some positive correlation among the classifiers. Specifically, it does not require the assumption of evidential independence of the classifiers to be fused (such as Dempster Shafer’s fusion rule). The proposed method is associative, which allows fusing three or more classifiers irrespective of the order. The fusion is accomplished using a generalized intersection operator (T-norm) that better represents the possible correlation between the classifiers. In addition, a confidence measure is produced that takes advantage of the consensus and conflict between classifiers. 1

In this paper, we describe the experiments that we have carried out during the European Research Project NetProtect II that aims at filtering harmful Web pages in order to protect children. These experiments focus on the combination of classifiers (relying on texts, images and addresses), dealing with heterogeneous classes (bomb-making, drug, pornography, violence) for multimedia documents (composed of both semi-structured text and images). We test and compare different combination formulas (Voting methods, logical methods, k Nearest Neighbors, evidence-based k Nearest Neighbors, Naive Bayes, Artificial Neural Network and Support Vector Machine) on a five thousand webpages database. We present how learning based methods combined to introduction of a priori knowledge on classifiers enable us to get better filtering performances than classical approaches (such as static black/white lists and single classifier).

This article presents a new system for automatically extracting high-level video concepts. The novelty of the approach lies in the feature fusion method. The system architecture is divided into three steps. The first step consists in creating sensors from a low-level (color or texture) descriptor, and a Support Vector Machine (SVM) learning to recognize a given concept (for example, “beach ” or “road”). The sensor fusion step is the combination of several sensors for each concept. Finally, as the concepts depend on context, the concept fusion step models interaction between concepts in order to modify their prediction. The fusion method is based on the Transferable Belief Model (TBM). It offers an appropriate framework for modeling source uncertainty and interaction between concepts. Results obtained on TREC video protocol demonstrate the improvement provided by such a combination, compared to mono-source information. 1.

Motivated by the need to support language-based communication between robots and their human users, as well as grounded symbolic reasoning, this paper presents a learning architecture that can be used by robotic agents for long-term and open-ended category acquisition. In this learning architecture, multiple object representations and multiple classifiers and classifier combinations are used. All learning computations are carried out during the normal execution of the agent, which allows continuous monitoring of the performance of the different classifiers. The measured classification successes of the individual classifiers support an attentional selection mechanism, through which classifier combinations are dynamically reconfigured and a specific classifier is chosen to predict the category of a new unseen object. In the current implementation of this learning architecture, base classifiers follow a memory-based approach, in which misclassified instances are simply added to the instance database. The main similarity measures used in the implementation are based on Euclidean distance and on a multi-resolution matching algorithm. Classifier combinations are based on majority voting and the Dempster-Shafer evidence theory. A simple agent, incorporating these learning capabilities, is used to test the approach. A long-term experiment was carried out having in mind the open-ended nature of category learning. With the help of a human mediator, the agent incrementally learned 68 categories of real world objects visually perceivable through an inexpensive camera.

In this paper we propose a new way to combine classifiers trained and diverged by NCL that leads to superior results in compare with averaging and DTs methods alone. In the proposed method after training classifiers by NCL, DTs and averaging are employed independently to combine them and then the outputs are combined again by averaging. For the second level combination, support vectors are scaled to the [0,1] interval. We show that each method used in the first level of combination yields a better performance in a different part of dataset so that they can complement each other. We did two sets of experiments, one on Satimage dataset and the other one on ORL dataset which showed a better performance for our method.

Abstract—A large experiment on Persian handwritten digits are reported and discussed. In this paper the techniques to combine multiple classifiers based on static structures is investigated. A static structure includes two main strategies to combine result of base classifiers: a) class indifferent methods b) class conscious methods. We establish our model on Decision Template and Dempster Shafer, which are under category of class indifferent method, and compare theirs recognition rate with five of the most famous combining methods of class conscious category. To evaluate our proposed model a real-world database of Persian handwritten digits containing 8600 handwritten digit images is used. Experiments using our database demonstrate that combining result of base classifiers with class indifferent methods indeed are far more effective than combining the result with class conscious methods in Persian handwritten digit recognition. Evaluating the proposed system with 2150 test samples the recognition rate of 91.98 % is achieved. Keywords—Class conscious, Class indifferent, Classifier fusion,

Abstract. Classifier outputs in the form of continuous values have often been combined using linear sum or stacking, but little is generally known about evidential reasoning methods for combining truncated lists of ordered decisions. In this paper we introduce a novel class-indifferent method for combining such a kind of classifier decisions. Specifically we model each output given by classifiers on new instances as a list of ranked decisions that is divided into 2 subsets of decisions, which are represented by triplet-based belief functions and then are combined using Dempster’s rule of combination. We present a formalism for triplet-based belief functions and establish a range of general formulae for combining these beliefs in order to arrive at a consensus decision. In addition we carry out a comparative analysis with an alternative representation dichotomous belief functions on the UCI benchmark data. We also compare our combination method with the popular methods of stacking, boosting, linear sum and majority voting over the same benchmark data to demonstrate the advantage of our approach. 1