Dionysios N. Sotiropoulos

MUSIPER: a system for modeling music similarity perception based on objective feature subset selection

We explore the use of objective audio signal features to model the individualized (subjective) perception of similarity between music files. We present MUSIPER, a content-based music retrieval system which constructs music similarity perception models of its users by associating different music similarity measures to different users. Specifically, a user-supplied relevance feedback procedure and related neural network-based incremental learning allows the system to determine which subset of a set of objective features approximates more accurately the subjective music similarity perception of a specific user. Our implementation and evaluation of MUSIPER verifies the relation between subsets of objective features and individualized music similarity perception and exhibits significant improvement in individualized perceived similarity in subsequent music retrievals.

Evaluation of a Cascade Hybrid Recommendation as a Combination of One-Class Classification and Collaborative Filtering

This paper decomposes the problem of recommendation into a two level cascade recommendation scheme which benefits from both content-based and collaborative filtering methodologies. The first level utilizes the content-based features of items in order to incorporate the individualized (subjective) user preferences within the recommendation process. This is achieved through the exploitation of the one-class classification paradigm which provides the means in order to filter out user specific undesirable items. The second level, on the other hand, serves the purpose of assigning particular rating degrees to the user-specific desirable items identified by the first level. The combination of two approaches in a cascade form, mimics the social process when someone has selected some items according to his preferences and asks for opinions about these by others, in order to achieve the best selection. Our experimentation provides significant evidence on the recommendation efficiency of the adapted hybrid approach which outperforms pure content-based and pure collaborative techniques.

Artificial immune system-based classification in class-imbalanced problems

We investigate the effect of the Class Imbalance Problem on the performance of an Artificial Immune System(AIS)-based classification algorithm. Our motivation stems from the fact that the Adaptive Immune System constitutes one of the most sophisticated biological systems which is particularly evolved in order to continuously address an extremely unbalanced pattern classification problem. That is the “self”/“non-self” discrimination process, consisting in classifying any cell as “self” or “non-self”. Our experimentation indicates that the AIS-based classification paradigm has the intrinsic properly in dealing more efficiently with highly skewed datasets than standard pattern classification algorithms such as the Support Vector Machines (SVMs). Specifically, the experimental results presented in this paper provide justifications concerning the superiority of AISbased classification in identifying instances from the minority class.

Artificial Immune System-based Classification in Class-Imbalanced Image Classification Problems

In this paper, we compare the performance of Artificial Immune System (AIS)-based classification algorithms to the performance of Gaussian kernel-based Support Vector Machines (SVM) in problems with a high degree of class imbalance. Our experimentation indicates that the AIS-based classification paradigm has the intrinsic properly of dealing more efficiently with highly skewed datasets. Specifically, our experimental results indicate that AIS-based classifiers identify instances from the minority class quite efficiently.

A Music Recommender Based on Artificial Immune Systems

In this paper, we address the recommendation process as a one-class classification problem based on content features and a Negative Selection (NS) algorithm that captures user preferences. Specifically, we develop an Artificial Immune System (AIS) based on a Negative Selection Algorithm that forms the core of a music recommendation system. The NS-based learning algorithm allows our system to build a classifier of all music pieces in a database and make personalized recommendations to users. This is achieved quite efficiently through the intrinsic property of NS algorithms to discriminate “self-objects” (i.e. music pieces of user’s like) from “non self-objects”, especially when the class of non self-objects is vast when compared to the class of self-objects and the examples (samples) of music pieces come only from the class of self-objects (music pieces of user’s like). Our recommender has been fully implemented and evaluated and found to outperform state of the art recommender systems based on support vector machines methodologies.

A mathematical analysis of the Genetic-AIRS classification algorithm

This paper presents the inception and the basic concepts of a hybrid classification algorithm called Genetic-AIRS [1]. Genetic-AIRS, is a combination of the Artificial Immune Resource System (AIRS) algorithm witch uses evolutionary computation techniques. An analysis is presented to determine the final algorithm architecture and parameters. The paper also includes an experimental evaluation on various publicly available datasets of Genetic-AIRS vs AIRS.

A Back End System for Digital Image Library Organization Based on Concept Learning

We present a back-end system which organizes digital image libraries according to a user-defined concept. The concept is extracted from a set of images that the user submits to the system as its representative instances. A relevance feedback procedure, implemented with SVM-based incremental learning algorithms, tunes a classifier that discriminates between concept-relevant and concept-irrelevant images. The system is fully developed and experimentally tested with success.

Individualization of Content-Based Image Retrieval Systems via Objective Feature Subset Selection

We explore the use of objective features to model the subjective perception of similarity between two images that have been extracted from an image database. We present a Content-based Image Retrieval system which evolves and uses different image similarity measures for different users. Specifically, a user-supplied relevance feedback procedure allows the system to determine which subset of a set of objective features approximates more efficiently the subjective image similarity of a specific user. Our implementation and evaluation of the system verifies our hypothesis and exhibits significant improvement in perceived image similarity.