Jose J. Valero-Mas

An experimental study on rank methods for prototype selection

Prototype selection is one of the most popular approaches for addressing the low efficiency issue typically found in the well-known k-Nearest Neighbour classification rule. These techniques select a representative subset from an original collection of prototypes with the premise of maintaining the same classification accuracy. Most recently, rank methods have been proposed as an alternative to develop new selection strategies. Following a certain heuristic, these methods sort the elements of the initial collection according to their relevance and then select the best possible subset by means of a parameter representing the amount of data to maintain. Due to the relative novelty of these methods, their performance and competitiveness against other strategies is still unclear. This work performs an exhaustive experimental study of such methods for prototype selection. A representative collection of both classic and sophisticated algorithms are compared to the aforementioned techniques in a number of datasets, including different levels of induced noise. Results report the remarkable competitiveness of these rank methods as well as their excellent trade-off between prototype reduction and achieved accuracy.

Prototype generation on structural data using dissimilarity space representation

Data reduction techniques play a key role in instance-based classification to lower the amount of data to be processed. Among the different existing approaches, prototype selection (PS) and prototype generation (PG) are the most representative ones. These two families differ in the way the reduced set is obtained from the initial one: While the former aims at selecting the most representative elements from the set, the latter creates new data out of it. Although PG is considered to delimit more efficiently decision boundaries, the operations required are not so well defined in scenarios involving structural data such as strings, trees, or graphs. This work studies the possibility of using dissimilarity space (DS) methods as an intermediate process for mapping the initial structural representation to a statistical one, thereby allowing the use of PG methods. A comparative experiment over string data is carried out in which our proposal is faced to PS methods on the original space. Results show that the proposed strategy is able to achieve significantly similar results to PS in the initial space, thus standing as a clear alternative to the classic approach, with some additional advantages derived from the DS representation.

Oversampling imbalanced data in the string space

Abstract Imbalanced data is a typical problem in the supervised classification field, which occurs when the different classes are not equally represented. This fact typically results in the classifier biasing its performance towards the class representing the majority of the elements. Many methods have been proposed to alleviate this scenario, yet all of them assume that data is represented as feature vectors. In this paper we propose a strategy to balance a dataset whose samples are encoded as strings. Our approach is based on adapting the well-known Synthetic Minority Over-sampling Technique (SMOTE) algorithm to the string space. More precisely, data generation is achieved with an iterative approach to create artificial strings within the segment between two given samples of the training set. Results with several datasets and imbalance ratios show that the proposed strategy properly deals with the problem in all cases considered.

Selecting promising classes from generated data for an efficient multi-class nearest neighbor classification

The nearest neighbor rule is one of the most considered algorithms for supervised learning because of its simplicity and fair performance in most cases. However, this technique has a number of disadvantages, being the low computational efficiency the most prominent one. This paper presents a strategy to overcome this obstacle in multi-class classification tasks. This strategy proposes the use of Prototype Reduction algorithms that are capable of generating a new training set from the original one to try to gather the same information with fewer samples. Over this reduced set, it is estimated which classes are the closest ones to the input sample. These classes are referred to as promising classes. Eventually, classification is performed using the original training set using the nearest neighbor rule but restricted to the promising classes. Our experiments with several datasets and significance tests show that a similar classification accuracy can be obtained compared to using the original training set, with a significantly higher efficiency.

A Study of Prototype Selection Algorithms for Nearest Neighbour in Class-Imbalanced Problems

Prototype Selection methods aim at improving the efficiency of the Nearest Neighbour classifier by selecting a set of representative examples of the training set. These techniques have been studied in situations in which the classes at issue are balanced, which is not representative of real-world data. Since class imbalance affects the classification performance, data-level balancing approaches that artificially create or remove data from the set have been proposed. In this work, we study the performance of a set of prototype selection algorithms in imbalanced and algorithmically-balanced contexts using data-driven approaches. Results show that the initial class balance remarkably influences the overall performance of prototype selection, being generally the best performances found when data is algorithmically balanced before the selection stage.

A Supervised Classification Approach for Note Tracking in Polyphonic Piano Transcription

Abstract In the field of Automatic Music Transcription, note tracking systems constitute a key process in the overall success of the task as they compute the expected note-level abstraction out of a frame-based pitch activation representation. Despite its relevance, note tracking is most commonly performed using a set of hand-crafted rules adjusted in a manual fashion for the data at issue. In this regard, the present work introduces an approach based on machine learning, and more precisely supervised classification, that aims at automatically inferring such policies for the case of piano music. The idea is to segment each pitch band of a frame-based pitch activation into single instances which are subsequently classified as active or non-active note events. Results using a comprehensive set of supervised classification strategies on the MAPS piano data-set report its competitiveness against other commonly considered strategies for note tracking as well as an improvement of more than in terms of F-measure when compared to the baseline considered for both frame-level and note-level evaluations.

Recognition of Handwritten Music Symbols using Meta-features Obtained from Weak Classifiers based on Nearest Neighbor.

The classification of musical symbols is an important step for Optical Music Recognition systems. However, little progress has been made so far in the recognition of handwritten notation. This paper considers a scheme that combines ideas from ensemble classifiers and dissimilarity space to improve the classification of handwritten musical symbols. Several sets of features are extracted from the input. Instead of combining them, each set of features is used to train a weak classifier that gives a confidence for each possible category of the task based on distance-based probability estimation. These confidences are not combined directly but used to build a new set of features called Confidence Matrix, which eventually feeds a final classifier. Our work demonstrates that using this set of features as input to the classifiers significantly improves the classification results of handwritten music symbols with respect to other features directly retrieved from the image.

Prototype Generation on Structural Data Using Dissimilarity Space Representation: A Case of Study

Data Reduction techniques are commonly applied in instance-based classification tasks to lower the amount of data to be processed. Prototype Selection (PS) and Prototype Generation (PG) constitute the most representative approaches. These two families differ in the way of obtaining the reduced set out of the initial one: while the former aims at selecting the most representative elements from the set, the latter creates new data out of it. Although PG is considered to better delimit decision boundaries, operations required are not so well defined in scenarios involving structural data such as strings, trees or graphs. This work proposes a case of study with the use of the common RandomC algorithm for mapping the initial structural data to a Dissimilarity Space (DS) representation, thereby allowing the use of PG methods. A comparative experiment over string data is carried out in which our proposal is faced to PS methods on the original space. Results show that PG combined with RandomC mapping achieves a very competitive performance, although the obtained accuracy seems to be bounded by the representativity of the DS method.