Agnieszka Dardzińska

Action Rules Discovery Based on Tree Classifiers and Meta-actions

Action rules describe possible transitions of objects from one state to another with respect to a distinguished attribute. Early research on action rule discovery usually required the extraction of classification rules before constructing any action rule. Newest algorithms discover action rules directly from a decision system. To our knowledge, all these algorithms assume that all attributes are symbolic or require prior discretization of all numerical attributes. This paper presents a new approach for generating action rules from datasets with numerical attributes by incorporating a tree classifier and a pruning step based on meta-actions. Meta-actions are seen as a higher-level knowledge (provided by experts) about correlations between different attributes.

Solving Failing Queries through Cooperation and Collaboration

Sometime Query Answering Systems (QAS) for a Distributed Autonomous Information System (DAIS) may fail by returning the empty set of objects as an answer for a query q. Systems in DAIS can be incomplete, have hierarchical attributes, and the semantics of attributes and their values may differ between sites. Also, if there are no objects in S matching q, the query may fail when submitted to S. Alternatively, QAS for S may try to relax the query q as it was proposed in T. Gaasterland (IEEE Expert, 12(5), 1997, 48–59), P. Godfrey (International Journal of Cooperative Information Systems, 6(2), 1997, 95–149) and W. Chu et al. (Journal of Intelligent Information Systems, 6(2/3), 1996, 223–259). It means that q can be replaced by a new more general query. Clearly, the goal is to find possibly the smallest generalization of q which will not fail in S. Smaller generalizations guarantee higher confidence in objects returned by QAS. Such QAS is called cooperative (only one site is involved). Queries may also fail in S when some of the attributes listed in q are outside the domain of S. To resolve this type of queries, assuming that S is a part of DAIS, we may extract definitions of such attributes from information systems residing at some of the remote sites for S and next use them to approximate q in S. In order to do that successfully, we assume that all involved systems have to agree on the ontology of some of their common attributes Z.W. Raś and A. Dardzińska (Information Systems International Journal, 29(1), 2004, 47–58; Proceedings of FQAS 2004 Conference, LNCS/LNAI No. 3055, 2004, pp. 125–136); Z.W. Raś and S. Joshi, Fundamenta Informaticae Journal, 30(3/4), 1997, 313–324. QAS based on the above strategy is called collaborative (minimum two sites are involved). Similarly, a query may fail in S when the granularity of an attribute used in q is finer than the granularity of the same attribute in S. This paper shows how to use collaboration and cooperation approach to solve failing queries in DAIS assuming that attributes are hierarchical. Some aspects of a collaboration strategy dealing with failing query problem for non-hierarchical attributes have been presented in Z.W. Raś and A. Dardzińska (Information Systems International Journal, 29(1), 2004, 47–58; Proceedings of FQAS 2004 Conference, LNCS/LNAI No. 3055, 2004, pp. 125–136).

Action rules discovery, a new simplified strategy

A new strategy for discovering action rules (or interventions) is presented in this paper. The current methods [14], [12], [8] require to discover classification rules before any action rule can be constructed from them. Several definitions of action rules [8], [13], [9], [3] have been proposed. They differ in the generality of their classification parts but they are always constructed from certain pairs of classification rules. Our new strategy defines the classification part of an action rule in a unique way. Also, action rules are constructed from single classification rules. We show how to compute their confidence and support. Action rules are used to reclassify objects. In this paper, we propose a method for measuring the level of reclassification freedom for objects in a decision system.

Data Security and Null Value Imputation in Distributed Information Systems

Distributed Information System (DIS) is seen as a collection of autonomous in-formation systems which can collaborate with each other. This collaboration can be driven by requests for knowledge needed to predict what values should replace null values in missing or incomplete attributes. Any incompleteness in data can be seen either as the result of a partial knowledge about properties of objects stored in DIS or some attributes might be just hidden from users because of the security reason. Clearly, in the second case, we have to be certain that the missing values can not be predicted from the available data by chase, distributed chase or any other null value imputation method. Let us assume that an attributes d is hidden at one of the sites of DIS, denoted by S and called a client. With a goal to reconstruct this hidden attribute, a request for a definition of this attribute can be sent by S to some of its remote sites (see [15]). These definitions stored in a knowledge-base KB can be used by Chase algorithm (see [4, 6]) to impute missing attribute values describing objects in S. In this paper we show how to identify these objects and what additional values in S have to be hidden from users to guarantee that initially hidden attribute values in S can not be properly predicted by Distributed Chase.

System ADReD for discovering rules based on hyperplanes

Abstract The decision table describing n objects in terms of k classification attributes and one decision attribute can be seen as a collection of n points in k -dimensional space. Each point is classified either as positive or negative. The goal of this paper is to present an efficient strategy for constructing possibly the smallest number of hyperplanes so each area surrounded by them contains a group of points, mostly of the same type (either positive or negative). A threshold value given by user, uniquely defines what we mean by mostly . The strategy presented in (Proceedings of the Symposium on Methods of AI (AI-METH2003), Silesian University of Technology, Gliwice, Poland, 123) shows how to construct a possibly smallest number of pairs of hyperplanes, surrounding any dense cluster of objects, whose intersection is a line orthogonal to and intersecting with one of the axes. In this paper that constraint is softened and these hyperplanes are built more independently. The main procedure starts with partitioning all negative objects into dense clusters. The same step is repeated for all positive objects also dividing them into dense clusters. To learn a negative rule, we take all objects in one of this negative clusters jointly with all positive objects. The algorithm, presented in this paper, constructs a minimal number of hyperplanes needed to build classification part of a rule describing this negative cluster. The same procedure is repeated for all the remaining negative clusters. Rules describing positive clusters are constructed the same way.

Discriminant feature analysis for music timbre recognition and automatic indexing

The high volume of digital music recordings in the internet repositories has brought a tremendous need for a cooperative recommendation system to help users to find their favorite music pieces. Music instrument identification is one of the important subtasks of a content-based automatic indexing, for which authors developed novel new temporal features and built a multi-hierarchical decision system S with all the low-level MPEG7 descriptors as well as other popular descriptors for describing music sound objects. The decision attributes in S are hierarchical and they include Hornbostel-Sachs classification and generalization by articulation. The information richness hidden in these descriptors has strong implication on the confidence of classifiers built from S. Rule-based classifiers give us approximate definitions of values of decision attributes and they are used as a tool by content-based Automatic Indexing Systems (AIS). Hierarchical decision attributes allow us to have the indexing done on different granularity levels of classes of music instruments. We can identify not only the instruments playing in a given music piece but also classes of instruments if the instrument level identification fails. The quality of AIS can be verified using precision and recall based on two interpretations: user and system-based [16]. AIS engine follows system-based interpretation.

Cascade Classifiers for Hierarchical Decision Systems

Hierarchical classifiers are usually defined as methods of classifying inputs into defined output categories. The classification occurs first on a low-level with highly specific pieces of input data. The classifications of the individual pieces of data are then combined systematically and classified on a higher level iteratively until one output is produced. This final output is the overall classification of the data. In this paper we follow a controlled devise type of approach. The initial group of classifiers is trained using all objects in an information system S partitioned by values of the decision attribute d at its all granularity levels (one classifier per level). Only values of the highest granularity level (corresponding granules are the largest) are used to split S into information sub-systems where each one is built by selecting objects in S of the same decision value. These sub-systems are used for training new classifiers at all granularity levels of its decision attribute. Next, we split each sub-system further by sub-values of its decision value. The obtained tree-structure with groups of classifiers assigned to each of its nodes is called a cascade classifier. Given an incomplete information system with a hierarchical decision attribute d, we consider the problem of training classifiers describing values of d at its lowest granularity level. Taking MIRAI database of music instrument sounds [16], as an example, we show that the confidence of such classifiers can be lower than the confidence of cascade classifiers.

Tree-Based Algorithms for Action Rules Discovery

One of the main goals in Knowledge Discovery is to find interesting associations between values of attributes, those that are meaningful in a domain of interest. The most effective way to reduce the amount of discovered patterns is to apply two interestingness measures, subjective and objective. Subjective measures are based on the subjectivity and understandability of users examining the patterns. They are divided into actionable, unexpected, and novel. Because classical knowledge discovery algorithms are unable to determine if a rule is truly actionable for a given user [1], we focus on a new class of rules [15], called E-action rules, that can be used not only for automatic analysis of discovered classification rules but also for hints of how to reclassify some objects in a data set from one state into another more desired one. Actionability is closely linked with the availability of flexible attributes [18] used to describe data and with the feasibility and cost [23] of desired re-classifications. Some of them are easy to achieve. Some, initially seen as impossible within constraints set up by a user, still can be successfully achieved if additional attributes are available. For instance, if a system is distributed and collaborating sites agree on the ontology [5], [6] of their common attributes, the availability of additional data from remote sites can help to achieve certain re-classifications of objects at a server site [23]. Action tree algorithm, presented in this paper, requires prior extraction of classification rules similarly as the algorithms proposed in [15] and [17] but it guarantees a faster and more effective process of E-action rules discovery. It was implemented as system DEAR 2.2 and tested on several public domain databases. Support and confidence of E-action rules is introduced and used to prune a large number of generated candidates which are irrelevant, spurious, and insignificant.

Knowledge discovery based query answering in hierarchical information systems

The paper concerns failing queries in incomplete Distributed Autonomous Information Systems (DAIS) based on attributes which are hierarchical and which semantics at different sites of DAIS may differ. Query q fails in an information system S, if the empty set of objects is returned as an answer. Alternatively, query q can be converted to a new query which is solvable in S. By a refinement of q, we mean a process of replacing q by a new relaxed query, as it was proposed in [2], [7], and [8], which is similar to q and which does not fail in S. If some attributes listed in q have values finer than the values used in S, then rules discovered either locally at S or at other sites of DAIS are used to assign new finer values of these attributes to objects in S. Queries may also fail in S when some of the attributes listed in q are outside the domain of S. To resolve this type of a problem, we extract definitions of such attributes at some of the remote sites for S in DAIS and next use them to approximate q in S. In order to do that successfully, we assume that all involved information systems have to agree on the ontology of some of their common attributes [14], [15], [16]. This paper shows that failing queries can be often handled successfully if knowledge discovery methods are used either to convert them to new queries or to find finer descriptions of objects in S.

Rule Discovery by Axes-Driven Hyperplanes Construction

Decision table with k classification attributes and one decision attribute, describing n objects, is seen in this paper as a collection of n points in k-dimensional space where each point is assigned either to a positive or negative class. The problem posed in this paper is to find an efficient strategy for constructing possibly the smallest number of hyperplanes so each area surrounded by them contains only one class of points (either positive or negative). The strategy, presented in [3], shows how to surround a cluster of objects by possibly smallest number of pairs of hyperplanes both originating from the same line which is either orthogonal or parallel to one of the axes. In this paper that constraint is softened and such hyperplanes do not have to originate from the same line. Initially, we consider all negative objects and partition them into clusters which are dense. We repeat the same step for all positive objects also dividing them into dense clusters. Next we take intersection of these two partitions seen as partitions in the initial space containing all objects. This intersection is subtracted from both the positive and negative partition. The remaining sub-clusters in the positive and negative partition form the final set of positive and negative clusters. In the last step of our algorithm we surround these clusters by minimal number of hyperplanes and use these hyper-planes to construct rules describing these clusters. The overall support and confidence of rules created that way is usually much higher than the confidence and support of rules created using hyperplanes parallel to axes (C4.5, Rosetta).