International Journal on Artificial Intelligence Tools

Multi-class classification is one of the most important tasks in machine learning. In this paper we consider two online multi-class classification problems: classification by a linear model and by a kernelized model. The quality of predictions is measured by the Brier loss function. We obtain two computationally efficient algorithms for these problems by applying the Aggregating Algorithms to certain pools of experts and prove theoretical guarantees on the losses of these algorithms. We kernelize one of the algorithms and prove theoretical guarantees on its loss. We perform experiments and compare our algorithms with logistic regression.

Conformal Predictors (CPs) are Machine Learning algorithms that can provide reliable confidence measures to their predictions. In this work, we make use of the Conformal Prediction framework for the assessment of stroke risk based on ultrasound images of atherosclerotic carotid plaques. For this application, images were recorded from 137 asymptomatic and 137 symptomatic plaques (symptoms are Stroke, Transient Ischaemic Attack (TIA), and Amaurosis Fugax (AF)). Two feature sets were extracted from the plaques; the first based on morphological image analysis and the second based on image texture analysis. Both sets were used in order to evaluate the performance of CPs on this problem. Four CPs were constructed using four popular classification methods, namely Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), Naive Bayes Classification (NBC), and k-Nearest Neighbours. The results given by all CPs demonstrate the reliability and importance of the obtained confidence measures on the problem of stroke risk assessment.

This paper presents a decision support system for radiotherapy treatment planning for head, neck and brain cancer. The aim of a treatment plan is to apply radiation to kill tumor cells, while minimizing the damage to healthy tissue and critical organs. Since treatment planning is a complex decision making process that relies heavily on the subjective experience of clinicians, we propose the use of case-based reasoning (CBR), in which problems are solved based on the solutions of similar past problems. This paper focuses on the case retrieval process of a CBR system. The attributes, which describe the cases, are selected by assessing their effect on the performance of the CBR system. We have developed a context sensitive local weighting scheme that assigns weights to attributes based on their value and the values of other attributes in the target case. A novel two phase retrieval mechanism is developed, in which each phase is optimized to retrieve a particular part of the solution. We also present an original use of fuzzy logic in order to represent nonlinearity in the similarity measure. Experiments, which evaluate the similarity measure using real brain cancer patient cases, show promising results.

The continuous increase in demand for new products and services on the market brought the need for systematic improvement of recommendation technologies. Recommender systems proved to be the answer to the data overload problem and an advantage for e-business. Nevertheless, challenges that recommender systems face, like sparsity and scalability, affect their performance in real-world situations where both the number of users and items are high and item rating is infrequent. In this article we propose a cluster based recommendation approach using genetic algorithms. Users are grouped into clusters based on their past choices and preferences and receive recommendations from the other cluster members with the aid of an innovative recommendation scheme called Top-Nvoted items. Similarity between users is computed using the max_norm Pearson coefficient. This is a modified form of the widely used Pearson coefficient and it is used to prevent very active users dominating recommendations. We compare our approach with five well established recommendation methods with the aid of three different datasets. These datasets vary in terms of the number of users, the number of items, and the sparsity of ratings. As a result important conclusions are drawn about the efficiency of each method with respect to scalability and dataset's sparsity.

The NP-hard problem of scheduling jobs on unrelated parallel machines with the objective of minimizing the makespan is addressed. A variable neighborhood descent heuristic hybridized with mathematical programming is proposed. A new constructive heuristic is also introduced to seed the search with a solution residing in a promising region. The strength of the approach lies in the adoption of appropriately chosen optimization criteria that avoid early local optimum entrapment and in the hybridization of the heuristic with mathematical programming for the exploration of the large neighborhood structures. Experimental results on a large set of benchmark problem instances attest to the efficacy of the proposed approach.

Rule engines, business rule management systems and other rule-based systems used today widely utilize methods, techniques and technologies from the era of expert systems. Unfortunately, this doesn't seem to be the case when it comes to explanation facilities. Nowadays, the use of explanation facilities seems more important than ever. Business rule management systems control or constrain the behavior of business processes through business rules and an explanation of the inference process intended for the end user would be more than welcome. An explanation facility framework which was created in order to remedy this situation is presented in this paper. It is written in Java and is supposed to be a generic solution for modern rule-based systems. Besides being free and open-source, it is simple to use and can generate explanations in the form of natural language like sentences. Internationalization is also supported and explanations can be saved as textual, XML or PDF reports.

An area of focus in music improvization is interactive improvization between a human and a computer system in real time. In this paper, we present a musical interactive system acting as a melody continuator. For each musical pattern given by the user, a new one is returned by the system which is built by using general patterns for both pitch and duration stored in its knowledge base. The latter consists of data mining rules extracted from different sets of melodies for different musical styles. The proposed system uses a new music representation scheme which treats separately pitch and duration. Also, it adopts a similarity measure initially developed for clustering categorical data. Moreover, we present experimental results, using Bach's Chorales and Jazz as test inputs, for both assessing the aesthetic quality of the proposed system and comparing it to human results.

This paper provides a comparative survey of natural language understanding (NLU) methodologies used for processing multiple documents. The scope of these methodologies is to generate a text output with reduced information redundancy and increased information coverage. The purpose of this paper is to inform the reader what methodologies exist and their features, capabilities and maturities based on evaluation criteria selected by users and developers. Tables of comparison, at the end of this survey, provide a quick glance of these technical attributes and maturity indicators abstracted from available information in the publications.

In this paper, we focus on multi-criteria decision-making problems. We propose a model based on influence diagrams; this model is able to handle uncertainty, represent interdependencies among the different decision variables and facilitate communication between the decision-maker and the analyst. The particular structure of the proposed model makes it possible to take into account the alternatives described by an attribute set, the decision-maker's characteristics and preferences, and other information (e.g., internal or external factors) that influence the decision. Modeling the decision problem in terms of influence diagrams requires a lot of work to gather expert knowledge. However, once the model is built, it can be easily and efficiently used for different instances of the decision problem. In fact, using our model simply requires entering some basic information, such as the values of internal or external factors and the decision-maker's characteristics. Our model also defines the importance of each criterion in terms of what is known about the decision maker, the quality index and the utility of each alternative.