John Z. Zhang

Enhancing multi-label music genre classification through ensemble techniques

In the field of Music Information Retrieval (MIR), multi-label genre classification is the problem of assigning one or more genre labels to a music piece. In this work, we propose a set of ensemble techniques, which are specific to the task of multi-label genre classification. Our goal is to enhance classification performance by combining multiple classifiers. In addition, we also investigate some existing ensemble techniques from machine learning. The effectiveness of these techniques is demonstrated through a set of empirical experiments and various related issues are discussed. To the best of our knowledge, there has been limited work on applying ensemble techniques to multi-label genre classification in the literature and we consider the results in this work as our initial efforts toward this end. The significance of our work has two folds: (1) proposing a set of ensemble techniques specific to music genre classification and (2) shedding light on further research along this direction.

A Perceptual Study on Music Segmentation and Genre Classification

Abstract Music Information Retrieval (MIR) is an interdisciplinary area that is engaged in the retrieval of information from music. It includes various tasks, such as music classification, clustering, perception and cognition, etc. In this article, we report our recent perceptual studies on segmentation and genre classification, two indispensable steps in the MIR process. Segmentation attempts to capture ‘drastic’ changes in music and provides a basis for further perceptual and computational analysis while genre classification amounts to separating music into different groups such that each group uniformly represents a music genre. Our perceptual study considers various related issues. The goal of this work is to (1) explore and deepen our understanding of the relationship between perceptual surface and perceptual structure of music through segmentation by human subjects and (2)reveal and demonstrate the multi-label nature of genre classification.

Where to Build a Door

A room is a simple polygon with a prespecified point, called the door, on its boundary. Search starts at the door, and must detect all intruders that may be in the room, while making sure that no intruder escapes through the door during the search. Depending on where the door is placed, the intruders may be able to avoid detection. We present an efficient algorithm that can determine all the intervals on the boundary where the door should be placed in order for the polygon to be searchable by two guards on the boundary who keep mutual visibility, or a single searcher with a flashlight. Our algorithm works in O(n log n) time, where n is the number of vertices of the given polygon

A linear-time algorithm for finding all door locations that make a room searchable

A room is a simple polygon with a prespecified point, called the door, on its boundary. Search may be conducted by two guards on the boundary who keep mutual visibility at all times, or by a single boundary searcher with a flashlight. Search starts at the door, and must detect any intruder that was in the room at the time the search started, preventing the intruder from escaping through the door. A room may or may not be searchable, depending on where the door is placed or no matter where the door is placed. We want to find all intervals on the boundary where the door can be placed for the resultant room to be searchable. It is known that this problem can be solved in O(n log n) time, if the given polygon has n sides. We improve this complexity to O(n).

A dynamic composite approach for evaluating association rules

Association mining is one the many tasks in data mining. In this paper, we consider the problem of evaluating association rules, an integral post process in association mining. In the literature, different interestingness measures have been proposed to evaluate association rules. Given an association mining task, measures are selected according to a set of user-specified properties. However, in practice, due to the subjectivity and imperfection in property specifications, it is a non-trivial task to make appropriate measure selections. In our work, we propose a novel approach that dynamically evaluates association rules according to a composite and collective effect of multiple measures. In essence, our approach makes use of neural networks along with back-propagation learning capability to determine the relative importance of measures in evaluating association rules. The effectiveness of our approach is shown through a set of empirical simulations. To the best of our knowledge, this is the first time that neural networks are applied to evaluating association rules.

Clustering-based genre prediction on music data

In this work, we study the problem of genre prediction on music data. The prediction is based on a genre map, which is constructed from clustering training music data. We make use of a novel algorithm which captures the structural distances from music data and achieves a high clustering accuracy. Preliminary experiments are conducted and discussed.

Finding the minimum-distance schedule for a boundary searcher with a flashlight

Consider a dark polygonal region in which intruders move freely, trying to avoid detection. A robot, which is equipped with a flashlight, moves along the polygon boundary to illuminate all intruders. We want to minimize the total distance traveled by the robot until all intruders are detected in the worst case. We present an O(nlogn) time and O(n) space algorithm for optimizing this metric, where n is the number of vertices of the given polygon. This improves upon the best known time and space complexities of O(n2) and O(n2), respectively. The distance graph plays a critical role in our analysis and algorithm design.

Evaluating Association Rules by Quantitative Pairwise Property Comparisons

Evaluating association rules is an integral post process in association rule mining. Association rules are examined by measures for their interestingness. Different interestingness measures have been proposed. Given an association rule mining task, measures are assessed and selected against a set of user-specified properties. However, in practice, due to the subjectivity and imperfection in property specifications, it is a non-trivial task to make appropriate measure selection. In this work, we propose a novel measure selection approach that makes use of the Analytic Hierarchy Process (AHP), a scheme for making complex decisions. Our approach captures a user¡¯s desired requirements quantitatively in an application domain to assess interestingness measures. It detects inconsistencies in property specifications, and is invariant to the number of association rules to be evaluated. The effectiveness of our approach is shown through case studies.

Surveillance of a polygonal area by a mobile searcher from the boundary: Searchability testing

We study the surveillance of a polygonal area by a robot, which is equipped with a flashlight and moves along the polygon boundary. Its aim is to illuminate any intruder who can move faster than the moving flashlight beam, trying to avoid detection. We propose an O(n)-time algorithm for testing if it is possible for such a robot to always detect any intruder in a given polygon, where n is the number of vertices of the given polygon. This improves upon the best previous time complexity of O(n log n).

On Searchability and LR-Visibility of Polygons

Imagine that intruders are in a dark polygonal room and move at a finite but unbounded speed, trying to avoid detection. Polygon search problem asks whether a polygon is searchable, i.e., no matter how intruders move, searcher(s) can always detect them. A polygon is LR-visible if there exist two boundary points such that the two polygonal chains divided by them are mutually weakly visible. We explore the relationship between the searchability and LR-visibility of a polygon. Our result can be used as a preprocessing step in designing algorithms related to polygon search.