A time series dissimilarity measure based on minimum jump costs
Publication Type:
Journal ArticleSource:
Journal article (Submitted)Keywords:
Time series; Similarity measures; Experimental comparison; ClassificationAbstract:
Time series are ubiquitous, and a measure to assess their similarity is a core part of many computational systems. Classical time series similarity measures based on dynamic time warping or the edit distance remain as the best general-purpose options, but their accuracies still leave much room for improvement. Here we propose a new approach to time series similarity based on the costs of iteratively jumping (or moving) between the sample values of two time series. We show that this approach can be very competitive when compared against the aforementioned classical measures. In fact, extensive experiments show that it can be statistically significantly superior for a number of data sources. Since the approach is also intuitive and computationally simple, we foresee its application as an alternative off-the-shelf tool to be used in time series retrieval, clustering, or classification systems.
Power-law distribution in encoded MFCC frames of speech, music, and environmental sound signals
Publication Type:
Conference PaperSource:
Int. World Wide Web Conf., Workshop on Advances on Music Information Retrieval (AdMIRe), WWW, Lyon, France, p.895-902 (2012)URL:
http://www2012.wwwconference.org/proceedings/forms/companion.htm#8Abstract:
Many sound-related applications use Mel-Frequency Cepstral Coefficients (MFCC) to describe audio timbral content. Most of the research efforts dealing with MFCCs have been focused on the study of different classification and clustering algorithms, the use of complementary audio descriptors, or the effect of different distance measures. The goal of this paper is to focus on the statistical properties of the MFCC descriptor itself. For that purpose, we use a simple encoding process that maps a short-time MFCC vector to a dictionary of binary code-words. We study and characterize the rank-frequency distribution of such MFCC code-words, considering speech, music, and environmental sound sources. We show that, regardless of the sound source, MFCC code-words follow a shifted power-law distribution. This implies that there are a few code-words that occur very frequently and many that happen rarely. We also observe that the inner structure of the most frequent code-words has characteristic patterns. For instance, close MFCC coefficients tend to have similar quantization values in the case of music signals. Finally, we study the rank-frequency distributions of individual music recordings and show that they present the same type of heavy-tailed distribution as found in the large-scale databases. This fact is exploited in two supervised semantic inference tasks: genre and instrument classification. In particular, we obtain similar classification results as the ones obtained by considering all frames in the recordings by just using 50 (properly selected) frames. Beyond this particular example, we believe that the fact that MFCC frames follow a power-law distribution could potentially have important implications for future audio-based applications.
PDFFile:
Melody, bassline and harmony representations for music version identification
Publication Type:
Conference PaperSource:
Int. World Wide Web Conf., Workshop on Advances on Music Information Retrieval (AdMIRe), WWW, Lyon, France, p.887-894 (2012)URL:
http://www2012.wwwconference.org/proceedings/forms/companion.htm#8Abstract:
In this paper we compare the use of different musical representations for the task of version identification (i.e. retrieving alternative performances of the same musical piece). We automatically compute descriptors representing the melody and bass line using a state-of-the-art melody extraction algorithm, and compare them to a harmony-based descriptor. The similarity of descriptor sequences is computed using a dynamic programming algorithm based on nonlinear time series analysis which has been successfully used for version identification with harmony descriptors. After evaluating the accuracy of individual descriptors, we assess whether performance can be improved by descriptor fusion, for which we apply a classification approach, comparing different classification algorithms. We show that both melody and bass line descriptors carry useful information for version identification, and that combining them increases version detection accuracy. Whilst harmony remains the most reliable musical representation for version identification, we demonstrate how in some cases performance can be improved by combining it with melody and bass line descriptions. Finally, we identify some of the limitations of the proposed descriptor fusion approach, and discuss directions for future research.
PDFFile:
Combining two lazy learning methods for classification and knowledge discovery.
Publication Type:
Conference PaperSource:
International Conference on Knowledge Discovery and Information Retrieval, INSTICC, Senart, Paris (2011)Keywords:
Machine Learning; Lazy learning methods; knowledge discovery; classification; medical diagnosisAbstract:
The goal of this paper is to construct a classifier for diagnosing malignant melanoma. We experimented with two lazy learning methods, $k$-NN and \textsf{LID}, and compared their results with the ones produced by decision trees. We performed this comparison because we are also interested on building a domain model that can serve as basis to dermatologists to propose a good characterization of early melanomas. We shown that lazy learning methods have a better performance than decision trees in terms of sensitivity and specificity. We have seen that both lazy learning methods produce complementary results ($k$-NN has high specificity and LID has high sensitivity) suggesting that a combination of both could be a good classifier. We report experiments confirming this point. Concerning the construction of a domain model, we propose to use the explanations provided by the lazy learning methods, and we see that the resulting theory is as predictive and useful as the one obtained from decision trees.
Classification of Melanomas in situ using Knowledge Discovery with Explained CBR
Publication Type:
Journal ArticleSource:
Artificial Intelligence in Medicine, Volume 51, p.12 (2011)Keywords:
knowledge discovery; classification; clustering; CBR; applicationAbstract:
The goal of Knowledge Discovery is to extract knowledge from a set of data. Most common techniques used in knowledge discovery are based on clustering methods whose goal is to analyze a set of objects and to obtain clusters based on the similarity between these objects. A desirable characteristic of clustering results is that they should be easily understandable by domain experts. In this paper we introduce LazyCL, a procedure using a lazy learning method to produce explanations on clusters of unlabeled cases. These explanations are the basis on which experts can perform knowledge discovery. Here we use LazyCL to generate a domain theory for classification of melanomas in situ.
Evaluating reliability and relevance for WOWA aggregation of Sleep Apnea case data
Publication Type:
Conference ProceedingsSource:
Congress of the European Society of Fuzzy Logic and Technology - EUSFLAT '99, Palma de Mallorca, p.283-286 (1999)Keywords:
sleep apnea diagnosis; questionnaire responses; WOWA aggregation; clustering; classification; reliability; relevance.Abstract:
In this article, joint medical and data
analysis expertise is brought to bear using
contrasting data analysis methods and the
WOWA aggregation operator to solve a
difficult medical diagnosis problem, that of
sleep apnea syndrome screening. We
describe a method of calculating the
relevance and reliability weights used by
the WOWA operator.
