Dionisios N. Sotiropoulos

Individualization of music similarity perception via feature subset selection

We address the problem of modeling the subjective perception of similarity between two music files that have been extracted from a music database with use of objective features. We propose the importation of user models in content-based music retrieval systems, which embody the ability of evolving and using different music 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 music similarity of a specific user. Our implementation of the proposed system verifies our hypothesis and exhibits significant improvement in perceived music similarity

The Class Imbalance Problem

We focus on a special category of pattern recognition problems that arise in cases when the set of training patterns is significantly biased towards a particular class of patterns. This is the so-called Class Imbalance Problem which hinders the performance of many standard classifiers. Specifically, the very essence of the class imbalance problem is unravelled by referring to a relevant literature review. It is shown that there exist a wide range of real-world applications involving extremely skewed (imbalanced) data sets and the class imbalance problem stems from the fact that the class of interest occupies only a negligible volume within the complete pattern space.

Artificial Immune System-Based Music Genre Classification

We present a novel approach for the problem of automated music genre classification, which utilizes an Artificial Immune System (AIS)-based classifier. Our inspiration lies in the observation that the natural immune system has the intrinsic property of self/non-self cell discrimination, especially when the non-self (complementary) space of cells is significantly larger than the class of self cells. The AIS-based classifier that we have built is compared with KNN-, RBF- and SVM-based classifiers in various experiments involving music data. We find that the performance of our classifier is similar to that of the other classifiers when tested in multi-class (eg. four class) problems. On the other hand, it exceeds by a significant margin the performance of the other classifiers when tested in two class problems.

Using Time-Series and Sentiment Analysis to Detect the Determinants of Bitcoin Prices

This paper uses time-series analysis to study the relationship between Bitcoin prices and fundamental economic variables, technological factors and measurements of collective mood derived from Twitter feeds. Sentiment analysis has been performed on a daily basis through the utilization of a state-of-the-art machine learning algorithm, namely Support Vector Machines (SVMs). A series of short-run regressions shows that the Twitter sentiment ratio is positively correlated with Bitcoin prices. The short-run analysis also reveals that the number of Wikipedia search queries (showing the degree of public interest in Bitcoins) and the hash rate (measuring the mining difficulty) have a positive effect on the price of Bitcoins. On the contrary, the value of Bitcoins is negatively affected by the exchange rate between the USD and the euro (which represents the general level of prices). A vector error-correction model is used to investigate the existence of long-term relationships between cointegrated variables. This kind of long-run analysis reveals that the Bitcoin price is positively associated with the number of Bitcoins in circulation (representing the total stock of money supply) and negatively associated with the Standard and Poors 500 stock market index (which indicates the general state of the global economy).

Machine Learning Paradigms

Recent major advances in Information Technologies are leading to an entirely new era in the educational process, which is characterized by the development of more engaging and human-like computer-based learning, personalization and incorporation of artificial intelligence techniques. A new research discipline, termed Learning Analytics, is emerging and examines the collection and intelligent analysis of learner and instructor data with the goal to extract information that can render electronic and/or mobile educational systems more personalized, engaging, dynamically responsive and pedagogically efficient. In this volume, internationally established authors are contributing their research ideas and results towards aspects of Learning Analytics with the purpose to (1) measure Student Engagement, to quantify the Learning Experience and to facilitate Self-Regulation, (2) to predict Student Performance (3) to be incorporated in Tools for Building Learning Materials and Educational Courses, and (4) to be used as Tools to support Learners and Educators in Synchronous and Asynchronous e-Learning.

Addressing the Class Imbalance Problem

In this chapter, we investigate the particular effects of the class imbalance problem on standard classifier methodologies and present the various methodologies that have been proposed as a remedy. The most interesting approach within the context of Artificial Immune Systems is the one related to the machine learning paradigm of one-class classification. One-Class Classification problems may be thought of as degenerated binary classification problems in which the available training instances originate exclusively from the under-represented class of patterns.

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

In this paper, we formulate the recommendation problem as a hybrid combination of one-class classification with collaborative filtering. Specifically, we decompose the recommendation problem into a two-level cascade scheme. In the first level, only desirable items are selected for each user from the large amount of all possible items, taking into account only a small portion of his/her available preferences. This is achieved via a one-class classification scheme trained only with positives examples, i.e. only with desirable items for which users have provided a rating value. In the second level, a collaborative filtering approach is applied to assign a rating degree to the items identified at the first level. The efficiency of our approach is analyzed theoretically in terms of best/worst case scenarios and respective lower/upper mean absolute error (MAE) bounds are computed. Moreover, our approach is experimentally tested against pure collaborative and cascade content-based approaches. The results show that our approach outperforms them in terms of MAE and, moreover, the experimental MAE is close to the theoretical lower bound corresponding to the best case scenario. The superiority of our approach is due to the existence of the one class classifier in the first level of the cascade.

Experimental Evaluation of Artificial Immune System-Based Learning Algorithms

In this chapter, we present experimental results to test and compare the performance of Artificial Immune System-based clustering, classification and one-class classification algorithms. The test data are provided via an open access collection of 1000 pieces from 10 classes of western music. This collection has been extensively used in testing algorithms for music signal processing. Specifically, we perform extensive tests on: Music Piece Clustering and Music Database Organization, Customer Data Clustering in an e-Shopping Application, Music Genre Classification, and A Music Recommender System.

SVM-based sentiment classification: a comparative study against state-of-the-art classifiers

Transforming the unstructured textual information contained in various social media streams into useful business knowledge is an extremely difficult computational task, mainly, due to the underlying hard pattern classification problem of sentiment analysis, especially within the context of the Greek language. In this paper, we address the pattern classification problem of sentiment analysis through the utilisation of support vector machines (SVMs). In particular, we conducted an extensive experimental comparison where we tested the aforementioned classifier against a set of state-of-the-art machine learning classifiers on a benchmark dataset originating from the Greek bank sector by collecting data from the streaming API of Twitter that were explicitly referring to the major banks of Greece. Our results present classification accuracy and execution time metrics for each classifier, revealing the superiority of the SVM learning paradigm in assigning patterns to the correct sentiment class.

A comparative study of one-class classifiers in machine learning problems with extreme class imbalance

Classification problems with class imbalance occur when prior probabilities for the data classes differ significantly. The use of one-class classifiers is one of the main approaches to solving such problems. We conduct a comparative study of one-class classification algorithms in classification problems with extreme class imbalance. Emphasis is placed on evaluation of the classificatory accuracy of a one-class classifier based on the Real Valued Negative Selection Algorithm (RVNSA) from Artificial Immune Systems theory, as there are no previous studies focusing on it. Its performance is compared to the performance of 14 alternative classification algorithms which are considered as state of the art in one-class classification problems.