Proceedings of The 2010 International Conference on Data Mining

Mining with Rare Cases.

Data Mining in the Real World Experiences%2C Challenges%2C and Recommendations.
Abstract- Data mining is used regularly in a variety of industries and is continuing to gain in both popularity and acceptance. However, applying data mining methods to complex real-world tasks is far from straightforward and many pitfalls face data mining practitioners. However, most research in the field tends to focus on the algorithmic issues that arise in data mining and ignores the human element and process issues that are often the cause of these pitfalls. While there are some papers on data mining experiences and lessons learned, they are quite rare, especially in the research community. The purpose of this paper is to begin to fill in the “gap ” between data mining methods and the practice of data mining. This paper describes many of the experiences of the author as a data mining practitioner, highlights the issues that he encountered while in industry, and provides a number of strategies and recommendations for dealing with these issues. This paper should benefit both practitioners and researchers but hopefully, more than anything, it will foster discussion on the practical data mining issues that are all too often ignored.
Are Decision Trees Always Greener on the Open %28Source%29 Side of the Fence%3F
Abstract- This short paper compares the performance of three popular decision tree algorithms: C4.5, C5.0, and WEKA’s J48. These decision tree algorithms are all related in that C5.0 is an updated commercial version of C4.5 and J48 is an implementation of the C4.5 algorithm under the WEKA data mining platform. The purpose of this paper is to verify the explicit or implied performance claims for these algorithms—namely that C5.0 is superior to C4.5 and that J48 mimics the performance of C4.5. Our results are quite surprising and contradict these claims. This is significant because existing work that is based on these claims (e.g., J48 being equivalent to C4.5) may be misleading.
A Combinatorial Fusion Method for Feature Construction.
Abstract- This paper demonstrates how methods borrowed from information fusion can improve the performance of a classifier by constructing (i.e., fusing) new features that are combinations of existing numeric features. The new features are constructed by mapping the numeric values for each feature to a rank and then averaging these ranks. The quality of the fused features is measured with respect to how well they classify minority-class examples, which makes this method especially effective for dealing with data sets that exhibit class imbalance. This paper evaluates our combinatorial feature fusion method on ten data sets, using three learning methods. The results indicate that our method can be quite effective in improving classifier performance.
A Combinatorial Fusion Method for Feature Construction.
Abstract- This paper demonstrates how methods borrowed from information fusion can improve the performance of a classifier by constructing (i.e., fusing) new features that are combinations of existing numeric features. The new features are constructed by mapping the numeric values for each feature to a rank and then averaging these ranks. The quality of the fused features is measured with respect to how well they classify minority-class examples, which makes this method especially effective for dealing with data sets that exhibit class imbalance. This paper evaluates our combinatorial feature fusion method on ten data sets, using three learning methods. The results indicate that our method can be quite effective in improving classifier performance.
Quantification and semi%2Dsupervised classification methods for handling changes in class distribution.
In realistic settings the prevalence of a class may change after a classifier is induced and this will degrade the performance of the classifier. Further complicating this scenario is the fact that labeled data is often scarce and expensive. In this paper we address the problem where the class distribution changes and only unlabeled examples are available from the new distribution. We design and evaluate a number of methods for coping with this problem and compare the performance of these methods. Our quantification-based methods estimate the class distribution of the unlabeled data from the changed distribution and adjust the original classifier accordingly, while our semi-supervised methods build a new classifier using the examples from the new (unlabeled) distribution which are supplemented with predicted class values. We also introduce a hybrid method that utilizes both quantification and semi-supervised learning. All methods are evaluated using accuracy and F-measure on a set of benchmark data sets. Our results demonstrate that our methods yield substantial improvements in accuracy and F-measure.
Guest editorial special issue on utility%2Dbased data mining.

Maximizing classifier utility when there are data acquisition and modeling costs.

Cost%2DSensitive Learning vs. Sampling Which is Best for Handling Unbalanced Classes with Unequal Error Costs%3F

Maximizing classifier utility when training data is costly.

Maximizing classifier utility when training data is costly.

UBDM 2006 Utility%2DBased Data Mining 2006 workshop report.

Data Mining in Telecommunications.
Telecommunication companies generate a tremendous amount of data. These data include call detail data, which describes the calls that traverse the telecommunication networks, network data, which describes the state of the hardware and software components in the network, and customer data, which describes the telecommunication customers. This chapter describes how data  mining can be used to uncover useful information buried within these data  sets. Several data mining applications are described and together they demonstrate that data mining can be used to identify telecommunication fraud, improve marketing effectiveness, and identify network faults.
Mining with Rare Cases.

Report on UBDM%2D05 Workshop on Utility%2DBased Data Mining.

Mining with rarity a unifying framework.
Rare objects are often of great interest and great value. Until recently, however, rarity has not received much attention in the context of data mining. Now, as increasingly complex real-world problems are addressed, rarity, and the related problem of imbalanced data, are taking center stage. This article discusses the role that rare classes and rare cases play in data mining. The problems that can result from these two forms of rarity are described in detail, as are methods for addressing these problems. These descriptions utilize examples from existing research, so that this article provides a good survey of the literature on rarity in data mining. This article also demonstrates that rare classes and rare cases are very similar phenomena---both forms of rarity are shown to cause similar problems during data mining and benefit from the same remediation methods.
Learning When Training Data are Costly The Effect of Class Distribution on Tree Induction.
For large, real-world inductive learning problems, the number of training examples often  must be limited due to the costs associated with procuring, preparing, and storing the data  and/or the computational costs associated with learning from the data. One question of practical  importance is: if n training examples are going to be selected, in what proportion should  the classes be represented? In this article we analyze the relationship between the marginal  class distribution of training data and the performance of classification trees induced from  these data, when the size of the training set is fixed. We study twenty-six data sets and, for  each, determine the best class distribution for learning. Our results show that, for a fixed  number of training examples, it is often possible to obtain improved classifier performance by  training with a class distribution other than the naturally occurring class distribution. For example,  we show that to build a classifier robust to different misclassification costs, a balanced  class distribution generally performs quite well. We also describe and evaluate a budgetsensitive  progressive-sampling algorithm that selects training examples such that the resulting  training set has a good (near-optimal) class distribution for learning.
A Quantitative Study of Small Disjuncts.
Systems that learn from examples often express the learned  concept in the form of a disjunctive description. Disjuncts  that correctly classify few training examples are known as  small disjuncts and are interesting to machine learning researchers  because they have a much higher error rate than  large disjuncts. Previous research has investigated this phenomenon  by performing ad hoc analyses of a small number  of datasets. In this paper we present a quantitative measure  for evaluating the effect of small disjuncts on learning and  use it to analyze 30 benchmark datasets. We investigate the  relationship between small disjuncts and pruning, training set  size and noise, and come up with several interesting results.   
ANSWER Network Monitoring Using Object%2DOriented Rules.
This paper describes ANSWER, the expert system  responsible for monitoring AT&amp;T&#039;s 4ESS switches. These switches are extremely important, since they handle virtually all of AT&amp;T&#039;s long distance traffic. ANSWER is implemented in R++, a rule-based extension to the C++ object-oriented programming language, and is innovative  because it employs both rule-based and object-oriented programming paradigms. The use of object technology in ANSWER has provided a principled way of modeling the 4ESS and of reasoning about failures within the 4ESS. This has resulted in an expert system that is more clearly organized, easily understood and maintainable than its predecessor, which was implemented using the rule-based paradigm alone. ANSWER has been deployed for more than a year and handles all 140 of AT&amp;T&#039;s 4ESS switches and processes over 100,000 4ESS alarms per week. Introduction  Network reliability is of critical concern to AT&amp;T, since its reputation for network reliability has taken many years to ...
The Problem with Noise and Small Disjuncts.
Many systems that learn from examples express the learned concept as a disjunction. Those disjuncts that cover only a few examples are referred to as small disjuncts. The problem with small disjuncts is that they have a much higher error rate than large disjuncts but are necessary to achieve a high level of predictive accuracy. This paper investigates the effect of noise on small disjuncts. In particular, we show that when noise is added to two real-world domains, a significant, and disproportionate number of the total errors are contributed by the small disjuncts; thus, in the presence of noise, it is the small disjuncts that are primarily responsible for the poor predictive accuracy of the learned concept. 1 INTRODUCTION  Systems that learn from examples often express the learned concept as a disjunction. The coverage, or size, of each disjunct is defined as the number of training examples that it correctly classifies (Holte, Acker &amp; Porter, 1989). Small disjuncts are those disjuncts...
Learning to Predict Rare Events in Event Sequences.
Learning to predict rare events from sequences of events with categorical features is an important, real-world, problem that existing statistical and machine learning methods are not well suited to solve. This paper describes timeweaver, a genetic algorithm based machine learning system that predicts rare events by identifying predictive temporal and sequential patterns. Timeweaver is applied to the task of predicting telecommunication equipment failures from 110,000 alarm messages and is shown to outperform existing learning methods. Introduction An event sequence is a sequence of timestamped observations, each described by a fixed set of features. In this paper we focus on the problem of predicting rare events from sequences of events which contain categorical (non-numerical) features. Predicting telecommunication equipment failures from alarm messages is one important problem which has these characteristics. For AT&amp;T, where most traffic is handled by 4ESS switches, the specific ...
Implementing Design Patterns with Object%2DOriented Rules.
ANSWER-4ESS, an operation support system for maintaining and monitoring the 4ESS switches distributed in the AT&amp;T network, recently became the first product to use the R++ languagean extension to C++ that adds support for objectoriented (OO) rules. This article shows how these rules not only implement the intelligent behavior of an expert system, but also facilitate inter-object communication and enable frequently occurring design patterns to be implemented in a way that leads to more understandable, maintainable, and reusable code.  Most of AT&amp;T&#039;s long-distance traffic is controlled by the 140 4ESS switches distributed throughout the AT&amp;T network. Automatic Network Surveillance with Expert Rules (ANSWER) is the operation support system responsible for monitoring and maintaining these switching systems. ANSWER&#039;s main purpose is to perform alarm filtering: It takes as input alarms from the 4ESS switches and outputs an alert once it has identified a problem that requires human interventi...
Learning with Rare Cases and Small Disjuncts.
Systems that learn from examples often create a  disjunctive concept definition. Small disjuncts  are those disjuncts which cover only a few  training examples. The problem with small  disjuncts is that they are more error prone than  large disjuncts. This paper investigates the  reasons why small disjuncts are more error prone  than large disjuncts. It shows that when there are  rare cases within a domain, then factors such as  attribute noise, missing attributes, class noise and  training set size can result in small disjuncts  being more error prone than large disjuncts and  in rare cases being more error prone than  common cases. This paper also assesses the  impact that these error prone small disjuncts and  rare cases have on inductive learning (i.e., on  error rate). One key conclusion is that when low  levels of attribute noise are applied only to the  training set (the ability to learn the correct  concept is being evaluated), rare cases within a  domain are primarily responsible for making  learning difficult.
