Zbigniew W. Raś

Multi-Label Classification of Emotions in Music

This paper addresses the problem of multi-label classification of emotions in musical recordings. The testing data set contains 875 samples (30 seconds each). The samples were manually labelled into 13 classes, without limits regarding the number of labels for each sample. The experiments and test results are presented.

Advances in Machine Learning II

This is the second volume of a large two-volume editorial project we wish to dedicate to the memory of the late Professor Ryszard S. Michalski who passed away in 2007. He was one of the fathers of machine learning, an exciting and relevant, both from the practical and theoretical points of view, area in modern computer science and information technology. His research career started in the mid-1960s in Poland, in the Institute of Automation, Polish Academy of Sciences in Warsaw, Poland. He left for the USA in 1970, and since then had worked there at various universities, notably, at the University of Illinois at Urbana Champaign and finally, until his untimely death, at George Mason University. We, the editors, had been lucky to be able to meet and collaborate with Ryszard for years, indeed some of us knew him when he was still in Poland. After he started working in the USA, he was a frequent visitor to Poland, taking part at many conferences until his death. We had also witnessed with a great personal pleasure honors and awards he had received over the years, notably when some years ago he was elected Foreign Member of the Polish Academy of Sciences among some top scientists and scholars from all over the world, including Nobel prize winners. Professor Michalskis research results influenced very strongly the development of machine learning, data mining, and related areas. Also, he inspired many established and younger scholars and scientists all over the world. We feel very happy that so many top scientists from all over the world agreed to pay the last tribute to Professor Michalski by writing papers in their areas of research. These papers will constitute the most appropriate tribute to Professor Michalski, a devoted scholar and researcher. Moreover, we believe that they will inspire many newcomers and younger researchers in the area of broadly perceived machine learning, data analysis and data mining. The papers included in the two volumes, Machine Learning I and Machine Learning II, cover diverse topics, and various aspects of the fields involved. For convenience of the potential readers, we will now briefly summarize the contents of the particular chapters.

Action rules mining

Action rules assume that attributes in a database are divided into two groups: stable and flexible. In general, an action rule can be constructed from two rules extracted earlier from the same database. Furthermore, we assume that these two rules describe two different decision classes and our goal is to reclassify objects from one of these classes into the other one. Flexible attributes are essential in achieving that goal because they provide a tool for making hints to a user about what changes within some values of flexible attributes are needed for a given group of objects to reclassify them into a new decision class. A new subclass of attributes called semi‐stable attributes is introduced. Semi‐stable attributes are typically a function of time and undergo deterministic changes (e.g., attribute age or height). So, the set of conditional attributes is partitioned into stable, semi‐stable, and flexible. Depending on the semantics of attributes, some semi‐stable attributes can be treated as flexible and the same new action rules can be constructed. These new action rules are usually built to replace some existing action rules whose confidence is too low to be of any interest to a user. The confidence of new action rules is always higher than the confidence of rules they replace. Additionally, the notion of the cost and feasibility of an action rule is introduced in this article. A heuristic strategy for constructing feasible action rules that have high confidence and possibly the lowest cost is proposed.

HOW TO SUPPORT CONSENSUS REACHING USING ACTION RULES: A NOVEL APPROACH

We consider a consensus reaching process in a group of individuals meant as an attempt to make preferences of the individuals more and more similar, that is, getting closer and closer to consensus. We assume a general form of intuitionistic fuzzy preferences and a soft definition of consensus that is basically meant as an agreement of a considerable (e.g., most, almost all) majority of individuals in regards to a considerable majority of alternatives. The consensus reaching process is meant to be run by a moderator who tries to get the group of individuals closer and closer to consensus by argumentation, persuasion, etc. The moderator is to be supported by some additional information, exemplified by more detailed information on which individuals are critical as, for instance, they are willing to change their testimonies or are stubborn, which pairs of options make the reaching of consensus difficult, etc. In this paper we extend this paradigm proposed and employed in our former works with the use of a novel data mining tool, so called action rules which make it possible to more clearly indicate and suggest to the moderator with which experts and with respect to which option it may be expedient to deal. We show the usefulness of this new approach.

Extracting emotions from music data

Music is not only a set of sounds, it evokes emotions, subjectively perceived by listeners. The growing amount of audio data available on CDs and in the Internet wakes up a need for content-based searching through these files. The user may be interested in finding pieces in a specific mood. The goal of this paper is to elaborate tools for such a search. A method for the appropriate objective description (parameterization) of audio files is proposed, and experiments on a set of music pieces are described. The results are summarized in concluding chapter.

Action rules discovery: system DEAR2, method and experiments

Subjective measures, used to model interestingness of rules, are user-driven, domain-dependent, and include unexpectedness, novelty and actionability (Adomavicius and Tuzhilin 1997, Liu et al. 1997, Silberschatz and Tuzhilin 1995). Liu et al. (1997) define a rule as actionable, if a user can do an action to his/her advantage based on that rule. Their notion of actionability is too vague and leaves the door open to a number of different interpretations. Raś and Wieczorkowska (2000) assume that actionability has to be expressed in terms of attributes that are present in the database. They have introduced a new class of rules (called action rules) that are constructed from certain pairs of association rules extracted from that database. A conceptually similar definition of an action rule was proposed independently by Geffner and Wainer (1998). Action rules have been investigated further in Raś and Gupta (2002), Raś and Tsay (2003), Raś et al. (2005) and Tzacheva and Raś (2004). In order to construct action rules, it is required that attributes in a database are divided into two groups: stable and flexible. Flexible attributes are used in a decision rule as a tool for making hints to a user what changes within some of their values are needed to reclassify a group of objects from one decision class into another one. Two strategies for generating action rules are presented. The first one, implemented as system DEAR, generates action rules from certain pairs of association rules. The second one, implemented as system DEAR2, is based on a tree structure that partitions the set of rules, having the same decision value, into equivalence classes each labelled by values of stable attributes (two rules belong to the same equivalence class, if values of their stable attributes are not conflicting each other). Now, instead of comparing all pairs of rules, only pairs of rules belonging to some of these equivalence classes are compared to construct action rules. This strategy significantly reduces the number of steps needed to generate action rules in comparison to DEAR system.

ARAS: action rules discovery based on agglomerative strategy

Action rules can be seen as logical terms describing knowledge about possible actions associated with objects which is hidden in a decision system. Classical strategy for discovering them from a database requires prior extraction of classification rules which next are evaluated pair by pair with a goal to build a strategy of action based on condition features in order to get a desired effect on a decision feature. An actionable strategy is represented as a term r=[(ω) Λ (α → β)] ⇒ [ϕ →ψ]r=[(ω) Λ (α → β)] ⇒ [ϕ →ψ], where ω, α, β, φ, and ψ are descriptions of objects or events. The term r states that when the fixed condition ω is satisfied and the changeable behavior (α → β) occurs in objects represented as tuples from a database so does the expectation (φ →ψ). This paper proposes a new strategy, called ARAS, for constructing action rules with the main module resembling LERS [6]. ARAS system is more simple than DEAR and its time complexity is also lower.

Ontology-based distributed autonomous knowledge systems

Traditional query processing usually requires that users fully understand the database structure and content to issue a query. Due to the complexity of the database applications and the variety of user needs, the so-called global queries are introduced which traditional query answering systems cannot handle. Query posed to a database D is global if minimum one of its attributes is missing in D while it occurs in other databases. Definitions of a missing attribute in D can be extracted from other databases and shared with D. To handle semantics inconsistencies between the same attributes used at different sites, task ontologies are used as a communication bridge between them. These inconsistencies can be caused either by different granularity levels or by different interpretations of the same attribute. As the final outcome of this research, a rough query answering system based on distributed data mining is presented.

Action rule extraction from a decision table: ARED

In this paper, we present an algorithm that discovers action rules from a decision table. Action rules describe possible transitions of objects from one state to another with respect to a distinguished attribute. The previous research on action rule discovery required the extraction of classification rules before constructing any action rule. The new proposed algorithm does not require pre-existing classification rules, and it uses a bottom up approach to generate action rules having minimal attribute involvement.

Action rules discovery system DEAR_3

E-action rules, introduced in [8], represent actionability knowledge hidden in a decision system. They enhance action rules [3] and extended action rules [4], [6], [7] by assuming that data can be either symbolic or nominal. Several efficient strategies for mining e-action rules have been developed [6], [7], [5], and [8]. All of them assume that data are complete. Clearly, this constraint has to be relaxed since information about attribute values for some objects can be missing or represented as multi-values. To solve this problem, we present DEAR_3 which is an e-action rule generating algorithm. It has three major improvements in comparison to DEAR_2: handling data with missing attribute values and uncertain attribute values, and pruning outliers at its earlier stage.