Strategies for Reducing Risks of Inconsistencies in Access Control Policies.
Abstract—Managing access control policies is a complex task. We argue that much of the complexity is unnecessary and mostly due to historical reasons. There are number of legacy policy specification languages that all have limitations of some kind. These limitations have forced policy implementers to use certain styles of writing policies, often resulting in inconsistencies. The detection and resolution of these inconsistencies has been widely researched and many solutions have been found. This paper highlights new possibilities for avoiding inconsistencies, drawing on the expressive power allowed in the condition field of rules in modern languages such as XACML. In particular, we show that making use of this expressive power has many advantages—it allows organizations to considerably reduce the number of policies and rules required to protect company assets; it provides improved views and summaries of related policies; and it allows increased scalability of analysis tools, such as tools that detect inconsistencies and tools that perform audits to verify compliance to regulations. Such tools are increasingly important in the current environment where the number of regulations governing company security continues to grow. In addition, we show how our user-friendly representation for the XACML language facilitates the use of complex conditions by increasing their readability. This increased readability has the additional benefit of allowing non-technical users to better understand the implementation of their policies. These factors all contribute to a lower risk of inconsistencies in policies. Keywords-access control management; policy conflict detection; policy management workbench; XACML I.
Offensive Language Detection Using Multi%2Dlevel Classification.

Robustness of Classifiers to Changing Environments.

Annotation Concept Synthesis and Enrichment Analysis.

Annotation Concept Synthesis and Enrichment Analysis.



Classification of Dreams Using Machine Learning.

Improving Co%2Dtraining with Agreement%2DBased Sampling.

Data Privacy From Technology to Economics.

A Tree%2DBased Decision Model to Support Prediction of the Severity of Asthma Exacerbations in Children.

A Cryptographic Solution for Private Distributed Simple Meeting Scheduling.

Classifying Biomedical Abstracts Using Committees of Classifiers and Collective Ranking Techniques.

Active Learning with Automatic Soft Labeling for Induction of Decision Trees.



Privacy and Data Mining New Developments and Challenges.

Parameterized Contrast in Second Order Soft Co%2Doccurrences A Novel Text Representation Technique in Text Mining and Knowledge Extraction.

Image Analysis and Machine Learning How to Foster a Stronger Connection%3F

A Non%2Dtechnical User%2DOriented Display Notation for XACML Conditions.
Abstract. Ideally, access control to resources in complex IT systems ought to be handled by business decision makers who own a given resource (e.g., the pay and benefits section of an organization should decide and manage the access rules to the payroll system). To make this happen, the security and database communities need to develop vendor-independent access management tools, useable by decision makers, rather than technical personnel detached from a given business function. We have developed and implemented such tool, based on XACML. The XACML is an important emerging tool for managing complex access control applications. As a formal notation, based on an XML schema representing the grammar of a given application, XACML is precise and non-ambiguous. But this very property puts it out of reach of non-technical users. We propose a new notation for displaying and editing XACML rules that is independent of XML, and we develop an editor for it. Our notation combines a tree representation of logical expressions with an accessible natural language layer. Our early experience indicates that such rules can be grasped by non-technical users wishing to develop and control rules for accessing their own resources. Keywords: Access control, notation, rule editor, XACML.
Foundations of Intelligent Systems
The aim of this paper is to introduce the theoretical foundations of an approach for intelligent systems development. Derived from semiotics, a classic discipline in human sciences, the theory developed provides a mathematical framework for the concept of knowledge and for knowledge processing. As a result, a new perspective to study and to develop intelligent systems emerges. A taxonomy of elementary types of knowledge is proposed based on the classification of types of signs in semiotics, followed by a another classification of knowledge from the point of view of application in cognitive systems. In addition, we propose the mathematical definition of objects, objects systems and objects networks, to model mathematically the different types of knowledge described. The symbiosis of such key concepts introduces a computational paradigm to develop and implement intelligent systems, called here computational semiotics. 
Proper Model Selection with Significance Test.
Abstract. Model selection is an important and ubiquitous task in machine learning. To select models with the best future classification performance measured by a goal metric, anevaluation metric is often used to select the best classification model among the competing ones. A common practice is to use the same goal and evaluation metric. However, in several recent studies, it is claimed that using an evaluation metric (such as AUC) other than the goal metric (such as accuracy) results in better selection of the correct models. In this paper, we point out a flaw in the experimental design of those studies, and propose an improved method to test the claim. Our extensive experiments show convincingly that only the goal metric itself can most reliably select the correct classification models. 1
Discriminative parameter learning for Bayesian networks.
Bayesian network classifiers have been widely used for classification problems. Given a fixed Bayesian network structure, parameters learning can take two different approaches: generative and discriminative learning. While generative parameter learning is more efficient, discriminative parameter learning is more effective. In this paper, we propose a simple, efficient, and effective discriminative parameter learning method, called Discriminative Frequency Estimate (DFE), which learns parameters by discriminatively computing frequencies from data. Empirical studies show that the DFE algorithm integrates the advantages of both generative and discriminative learning: it performs as well as the state-of-the-art discriminative parameter learning method ELR in accuracy, but is significantly more efficient. 1.
Generation of Globally Relevant Continuous Features for Classification.

How to Prevent Private Data from being Disclosed to a Malicious Attacker.

Privacy%2DPreserving Naive Bayesian Classification over Horizontally Partitioned Data.


Machine Learning ECML 2007


Formal correctness of conflict detection for firewalls.

A Concept%2DBased Framework for Retrieving Evidence to Support Emergency Physician Decision Making at the Point of Care.

Using Secondary Knowledge to Support Decision Tree Classification of Retrospective Clinical Data.

Privacy%2DPreserving Data Mining in Electronic Surveys.

Privacy%2Dpreserving multi%2Dparty decision tree induction.

Privacy%2Dpreserving support vector machine classification.

Privacy%2Dpreserving collaborative association rule mining.
Abstract. This paper introduces a new approach to a problem of data sharing among multiple parties, without disclosing the data between the parties. Our focus is data sharing among parties involved in a data mining task. We study how to share private or confidential data in the following scenario: multiple parties, each having a private data set, want to collaboratively conduct association rule mining without disclosing their private data to each other or any other parties. To tackle this demanding problem, we develop a secure protocol for multiple parties to conduct the desired computation. The solution is distributed, i.e., there is no central, trusted party having access to all the data. Instead, we define a protocol using homomorphic encryption techniques to exchange the data while keeping it private. Key Words: Privacy, security, association rule mining. 1
Unsupervised Named%2DEntity Recognition Generating Gazetteers and Resolving Ambiguity.
Abstract. In this paper, we propose a named-entity recognition (NER) system that addresses two major limitations frequently discussed in the field. First, the system requires no human intervention such as manually labeling training data or creating gazetteers. Second, the system can handle more than the three classical named-entity types (person, location, and organization). We describe the system’s architecture and compare its performance with a supervised system. We experimentally evaluate the system on a standard corpus, with the three classical named-entity types, and also on a new corpus, with a new named-entity type (car brands). 1
Beyond the Bag of Words A Text Representation for Sentence Selection.

Learning and Evaluation in the Presence of Class Hierarchies Application to Text Categorization.

Evaluating Misclassifications in Imbalanced Data.
Abstract. Evaluating classifier performance with ROC curves is popular in the machine learning community. To date, the only method to assess confidence of ROC curves is to construct ROC bands. In the case of severe class imbalance with few instances of the minority class, ROC bands become unreliable. We propose a generic framework for classifier evaluation to identify a segment of an ROC curve in which misclassifications are balanced. Confidence is measured by Tango’s 95%-confidence interval for the difference in misclassification in both classes. We test our method with severe class imbalance in a two-class problem. Our evaluation favors classifiers with low numbers of misclassifications in both classes. Our results show that the proposed evaluation method is more confident than ROC bands. 1
A Crypto%2DBased Approach to Privacy%2DPreserving Collaborative Data Mining.

Privacy%2DPreserving Collaborative Data Mining.
2006) aims to bring together researchers from both industry and academia with diverse backgrounds: data mining, machine learning, database, statistical analysis, and application knowledge to foster interactions, to propose new ideas, to identify promising technologies, to create a forum for discussing recent advances, to better understand the practical challenges in applications, and to inspire new research directions. Many real-world applications deal with huge amounts of temporal data. Examples include alarms/events and performance measurements generated by distributed computer systems and by telecommunication networks, the web server logs, online transaction logs, financial data, workflow process logs, and sensor data collected from sensor networks. Conventionally, temporal data is classified to either categorical event streams or numerical time series and both types have been intensively studied in data mining and statistics. However, several previously less emphasized aspects of temporal data have proven their importance in emerging applications and posed several challenges calling for more research. The major topics of the workshop include but are not limited to: Temporal data benchmarking Temporal pattern discovery Temporal data clustering Anomaly and change detection of streaming data Prediction for temporal data Temporal data characterization and analysis Statistical analysis of temporal data Accommodating domain knowledge in the temporal mining process, Complexity, efficiency and scalability of temporal data mining algorithms, Content-based search, retrieval for temporal data, Process Mining,
Parallelizing Feature Selection.
Feature selection is an important, but computationally expensive component in many computer vision algorithms. In this report we examine a simple yet effective method of parallelizing a feature selection algorithm based on the Sum of Squared Differences technique. Our analysis and experimental results show that this approach to feature selection lends itself well to parallelization. Moreover, we show that using the Parallel Virtual Machine function library we get better performance on a network of workstations than on the CM5 supercomputer.  1 INTRODUCTION  1. Introduction.  The automatic selection of features in an image is an important component of many computer vision applications, such as tracking, stereo, and depth systems. A feature is defined as a unique grouping of pixels in a local region of an image. While features are important, finding them is computationally intensive. The Sum of Squared Differences (SSD) technique has an asymptotic running time of O(n  4  m  2  ) where t...
Inferring and revising theories with confidence analyzing bilingualism in the 1901 canadian census.

Privacy%2DSensitive Information Flow with JML.
Abstract. In today’s society, people have very little control over what kinds of personal data are collected and stored by various agencies in both the private and public sectors. We describe an approach to addressing this problem that allows individuals to specify constraints on the way their own data is used. Our solution uses formal methods to allow developers of software that processes personal data to provide assurances that the software meets the specified privacy constraints. In the domain of privacy, it is often not sufficient to express properties of interest as a relation between the input and output of a program as is done for general program correctness. Here we consider a stronger class of properties that allows us to express constraints on information flow. In particular, we can express that an algorithm does not leak any information from particular “sensitive ” values. We describe a general methodology for expressing this kind of information flow property as Hoare-style program verification judgments. We begin with the Java Modelling Language (JML), which is a behavioral interface specification language designed for Java, and we extend the language to include new concepts and keywords for expressing such properties. We use the Krakatoa tool which starts from JMLannotated Java programs, generates proof obligations in the Coq Proof Assistant, and helps to automate their proofs. We extend the Krakatoa tool to understand our extensions to JML and to generate the new form of required proof obligations. We illustrate our method on several data mining algorithms implemented in Java. 1
PEEP%2D An Information Extraction base approach for Privacy Protection in Email.

Privacy Compliance Enforcement in Email.

Privacy%2DPreserving Collaborative Association Rule Mining.
Abstract. This paper introduces a new approach to a problem of data sharing among multiple parties, without disclosing the data between the parties. Our focus is data sharing among parties involved in a data mining task. We study how to share private or confidential data in the following scenario: multiple parties, each having a private data set, want to collaboratively conduct association rule mining without disclosing their private data to each other or any other parties. To tackle this demanding problem, we develop a secure protocol for multiple parties to conduct the desired computation. The solution is distributed, i.e., there is no central, trusted party having access to all the data. Instead, we define a protocol using homomorphic encryption techniques to exchange the data while keeping it private. Key Words: Privacy, security, association rule mining. 1
Building k%2Dnearest neighbor classifiers on vertically partitioned private data.

Private Mining of Association Rules.
Abstract. This paper introduces a new approach to a problem of data sharing among multiple parties, without disclosing the data between the parties. Our focus is data sharing among two parties involved in a data mining task. We study how to share private or confidential data in the following scenario: two parties, each having a private data set, want to collaboratively conduct association rule mining without disclosing their private data to each other or any other parties. To tackle this demanding problem, we develop a secure protocol for two parties to conduct the desired computation. The solution is distributed, i.e., there is no central, trusted party having access to all the data. Instead, we define a protocol using homomorphic encryption techniques to exchange the data while keeping it private. All the parties are treated symmetrically: they all participate in the encryption and in the computation involved in learning the association rules. Key Words: Privacy, security, association rule mining. 1
STochFS A Framework for Combining Feature Selection Outcomes Through a Stochastic Process.

PEEP%2D Privacy Enforcement in Email Project.
Breaching information privacy is a critical problem where legal remedies intervene only after the fact rather than prevent it. This paper presents an organizational privacy compliance engine that monitors outgoing emails to detect breaches of a privacy policy in an organization. The PEEP system employs email content analysis techniques to extract information and their ownership. Access to the extracted information is verified by privacy rules assisted by an ontology-based model to represent information disclosure privileges. This paper addresses the issues of, first, the information extraction techniques from email, and second, the implementation of an ontology-based model of multilevel disclosure privileges to represent privacy rules. We experiment with the PEEP system on real life emails in an academic environment to detect breaches of privacy in emails. Our results report an F-score of 71.7% of privacy violations detection.
Privacy in Data Mining Using Formal Methods.

Privacy Preserving K%2Dnearest Neighbor Classification.
This paper considers how to conduct k-nearest neighbor classification in the following scenario: multiple parties, each having a private data set, want to collaboratively build a k-nearest neighbor classifier without disclosing their private data to each other or any other parties. Specifically, the data are vertically partitioned in that all parties have data about all the instances involved, but each party has its own view of the instances- each party works with its own attribute set. Because of privacy constraints, developing a secure framework to achieve such a computation is both challenging and desirable. In this paper, we develop a secure protocol for multiple parties to conduct the desired computation. All the parties participate in the encryption and in the computation involved in learning the k-nearest neighbor classifiers 1.
Privacy%2DPreserving Multi%2DParty Decision Tree Induction.

Bayesian Network Induction With Incomplete Private Data.


Privacy%2DPreserving Collaborative Association Rule Mining.
Abstract. This paper introduces a new approach to a problem of data sharing among multiple parties, without disclosing the data between the parties. Our focus is data sharing among parties involved in a data mining task. We study how to share private or confidential data in the following scenario: multiple parties, each having a private data set, want to collaboratively conduct association rule mining without disclosing their private data to each other or any other parties. To tackle this demanding problem, we develop a secure protocol for multiple parties to conduct the desired computation. The solution is distributed, i.e., there is no central, trusted party having access to all the data. Instead, we define a protocol using homomorphic encryption techniques to exchange the data while keeping it private. Key Words: Privacy, security, association rule mining. 1

Email Classification with Temporal Features.
Abstract. We propose a novel solution to the email classification problem: the integration of temporal information with the traditional content-based classification approaches. We discover temporal relations in an email sequence in the form of temporal sequential patterns and embed the discovered information into contentbased learning methods. The new heterogeneous classification system shows a good performance reducing the classification error by up to 22%. 1
A formal approach to using data distributions for building causal polytree structures.%2
(DoD) under DARPA/NASA subcontract NAG2-593 administered by the NASA
Predictive Software Models.
Keywords: With the increased use of powerful, performance-optimized hardware components in embedded systems, timing prediction is getting more and more complex. Thus while the execution speed of software is generally increasing, it is getting more and more difficult (if not infeasible) to perform an accurate and safe timing analysis of software that runs on those high-end embedded computer systems. This paper presents a very rigid software execution model for building distributed hard real-time subsystems that are time predictable. The software model is based on the time-triggered communication model. It uses a purely timetriggered input-output interface and relies on single-path code (code that is free from input-data dependent control flow) in both the operating system and application software. Tasks are only preempted at pre-planned task preemption points and a simple clock synchronization keeps the operations of the hard realtime subsystem in synchrony with the real-time environment. The proposed execution model yields software that is time-predictable by construction. Verifying temporal correctness and tracing the timing behavior of this software is trivial. Real-time systems, time-triggered architecture, determinism, time predictability. 1.
A formal approach to using data distributions for building causal polytree structures.%2
(DoD) under DARPA/NASA subcontract NAG2-593 administered by the NASA
Filtering Multi%2DInstance Problems to Reduce Dimensionality in Relational Learning.
Abstract. Attribute-value based representations, standard in today’s data mining systems, have a limited expressiveness. Inductive Logic Programming provides an interesting alternative, particularly for learning from structured examples whose parts, each with its own attributes, are related to each other by means of first-order predicates. Several subsets of FOL with different expressive power have been proposed in ILP. The challenge lies in the fact that the more expressive the subset of FOL the learner works with, the more critical the dimensionality of the learning task. The Datalog language is expressive enough to represent realistic learning problems when data is given directly in a relational database, making it a suitable tool for data mining. Consequently, it is important to elaborate techniques that will dynamically decrease the dimensionality of learning tasks expressed in Datalog, just as Feature Subset Selection (FS) techniques do it in attribute-value learning. The idea of re-using these techniques in ILP runs immediately into a problem as ILP examples have variable size and do not share the same set of literals. We propose here the first paradigm that brings Feature Subset Selection to the level of ILP, in languages at least as expressive as Datalog. The main idea is to first perform a change of representation, which approximates the original relational problem by a multi-instance problem. The representation obtained as the result is suitable for FS techniques which we adapted from attribute-value learning by taking into account some of the characteristics of the data due to the change of representation. We present the simple FSS proposed for the task, the requisite
Guest Editorial.
of vital signs, and medical emergency teams
Inductive Logic Programming
 
Applying data mining to software maintenance records.

Case Authoring from Text and Historical Experiences.

Enhancing Caching in Distributed Databases Using Intelligent Polytree Representations.

Mining the Maintenance History of a Legacy Software System.

Generalized Features Their Application to Classification.

Feature Subset Selection and Inductive Logic Programming.
In this paper we argue that chess databases have a significant potential as a test-bed for techniques in the area of Knowledge Discovery in Databases (KDD). Conversely, we think that research in Artificial Intelligence has not yet come up with reasonable solutions for the knowledge representation and reasoning problems that are posed by knowledgebased computer chess programs, and consequently argue that KDD techniques could be useful for the advancement of various types of knowledgebased computer chess systems. Although we cannot present any concrete results, we hope to outline some fruitful directions for further research and exchange of ideas between the KDD and computer chess communities.  1 Introduction  Knowledge Discovery in Databases (KDD) or Data Mining is a rapidly growing research area which focuses on the discovery of useful and understandable pieces of knowledge from databases [Frawley et al., 1992; Fayyad et al., 1996] . On the other hand, the rapid increase in computing p...
A Dynamic Approach to Dimensionality Reduction in Relational Learning.

Privacy%2DOriented Data Mining by Proof Checking.
This paper shows a new method which promotes ownership of data by people about whom the data was collected. The data owner may preclude the data from being used for some purposes, and allow it to be used for other purposes. We show an approach, based on checking the proofs of program properties, which implements this idea and provides a tool for a verifiable implementation of the Use Limitation Principle.
Advances in Artificial Intelligence
A framework that combines neural processing, psychological modelling, and Artificial Intelligent (AI) applications is proposed as a means of making steady progress towards the resolution of the problems of the Turing test, understanding neural behaviour, understanding psychological behaviour, and developing useful AI systems. By a principled process of integrating models from these domains, it is hoped that steady progress will be made on a variety of research fronts culminating with a system that can pass the Turing test.
Email classification with co%2Dtraining.
The main problems in text classification are lack of labeled data, as well as the cost of labeling the unlabeled data. We address these problems by exploring co-training - an algorithm that uses unlabeled data along with a few labeled examples to boost the performance of a classifier. We experiment with co-training on the email domain. Our results show that the performance of co-training depends on the learning algorithm it uses. In particular, Support Vector Machines significantly outperforms Naive Bayes on email classification.
Supporting Software Maintenance by Mining Software Update Records.
This paper describes the application of inductive methods to data extracted from both source code and software maintenance records. We would like to extract relations that indicate which files in, a legacy system, are relevant to each other in the context of program maintenance. We call these relations Maintenance Relevance Relations. Such a relation could reveal existing complex interconnections among files in the system, which may in turn be useful in comprehending them. We discuss the methodology we employed to extract and evaluate the relations. We also point out some of the problems we encountered and our solutions for them. Finally, we present some of the results that we have obtained.
GENEX a tool for testing in ILP.

Supporting maintenance of legacy software with data mining techniques.

Relational Learning with Transfer of Knowledge Between Domains.


The Software for Cultures and the Cultures in Software.
Abstract-Software is viewed as an artifact which interacts with cultures of societies in which it functions. Software manufacturers make efforts to adapt the appearance of their products to aesthetic and historical values of the markets in which they are sold (“software for cultures”). It is well known that software embeds behavioral and organizational principles that are culture-determined (“cultures in software”). Internet and e-commerce bring these phenomena into the fore of the debate on societal implications of Information Technology. The paper argues for a research agenda on the multifaceted interactions between software and culture. I.

A Formalism for Building Causal Polytree Structures Using Data Distributions.

Feature Engineering for Text Classification.
Most research in text classification has used the “bag of words ” representation of text. This paper examines some alternative ways to represent text based on syntactic and semantic relationships between words (phrases, synonyms and hypernyms). We describe the new representations and try to justify our suspicions that they could have improved the performance of a rule-based learner. The representations are evaluated using the RIPPER rule-based learner on the Reuters-21578 and DigiTrad test corpora, but on their own the new representations are not found to produce a significant performance improvement. Finally, we try combining classifiers based on different representations using a majority voting technique. This step does produce some performance improvement on both test collections. In general, our work supports the emerging consensus in the information retrieval community that more sophisticated Natural Language Processing techniques need to be developed before better text representations can be produced. We conclude that for now, research into new learning algorithms and methods for combining existing learners holds the most promise.
Machine Learning Method for Software Quality Model Building.
Software quality prediction can be cast as a concept learning problem. In this paper, we discuss the full cycle of an application of Machine Learning to software quality prediction. As it often happens in real-life applications, significant part of the project was devoted to activities outside the learning process: data acquisition, feature engineering, labeling of the examples, etc. We believe that in projects that reach out to real data (rather than rely on the prepared data sets from the existing repositories), these activities often decide about the success or a failure of the project. The method proposed here is applied to a set of real-life COBOL programs and some discussion on the results is presented.  Keywords: Learning applications, software metrics, software quality, symbolic learning.  1. Introduction  Lack of adequate tools to evaluate and estimate software quality is one of the main challenges in Software Engineering. Historically, there have been numerous attempts to dev...
Experiments with Learning Parsing Heuristics.
Any large language processing software  relies in its operation on heuristic decisions  concerning the strategy of processing.  These decisions are usually &#034;hard-wired&#034;  into the software in the form of handcrafted  heuristic rules, independent of the  nature of the processed texts. We propose  an alternative, adaptive approach in which  machine learning techniques learn the rules  from examples of sentences in each class.  We have experimented with a variety of  learning techniques on a representative instance  of this problem within the realm of  parsing. Our approach lead to the discovery  of new heuristics that perform significantly  better than the current hand-crafted heuristic.  We discuss the entire cycle of application  of machine learning and suggest a  methodology for the use of machine learning  as a technique for the adaptive optimisation  of language-processing software.  1 Introduction  Any language processing program---in our case, a top-down parser which outputs only t...
Predicate Invention from a Few Examples.

A Normalization Method for Contextual Data Experience from a Large%2DScale Application.
This paper describes a pre-processing technique to normalize  contextually-dependent data before applying Machine Learning algorithms. Unlike many previous methods, our approach to normalization does not assume that the learning task is a classification task. We propose  a data pre-processing algorithm which modifies the relevant attributes  so that the effects of the contextual attributes on the relevant attributes  are cancelled. These effects are modeled using a novel approach, based  on the analysis of variance of the contextual attributes. The method is  applied on a massive data repository in the area of aircraft maintenance. 
Machine Learning for the Detection of Oil Spills in Satellite Radar Images.
. During a project examining the use of machine learning techniques for oil spill detection, we encountered several essential questions that we believe deserve the attention of the researchcommunity. We use our particular case study to illustrate such issues as problem formulation, selection of evaluation measures, and data preparation. We relate these issues to properties of the oil spill application, suchasitsimbalanced class distribution, that are shown to be common to many applications. Our solutions to these issues are implemented in the Canadian Environmental Hazards Detection System (CEHDS), which is about to undergo field testing.  Keywords: Inductive learning, classification, radar images, methodology  1. Introduction  In this paper we describe an application of machine learning to an importantenvironmental  problem: detection of oil spills from radar images of the sea surface. Wecover the application cycle from the problem formulation phase to the delivery of a system for fie...
Learning When Negative Examples Abound.
. Existing concept learning systems can fail when the negative  examples heavily outnumber the positive examples. The paper discusses  one essential trouble brought about by imbalanced training sets  and presents a learning algorithm addressing this issue. The experiments  (with synthetic and real-world data) focus on 2-class problems with examples  described with binary and continuous attributes.  1 Introduction  The specific problem addressed here is learning from imbalanced training sets where examples from one class heavily outnumber examples from the other class. Highly imbalanced training sets occur in applications where the classifier is to detect an infrequent, albeit important, event: a fraudulent telephone call (Fawcett and Provost, 1996), an unreliable telecommunications customer (Ezawa, Singh, and Norton, 1996), or a rare diagnosis such as the thyroid diseases in the UCI repository. Extremely imbalanced classes also arise in information retrieval (Lewis and Catlett, 1994). ...
Addressing the Curse of Imbalanced Training Sets One%2DSided Selection.
Adding examples of the majority class to the  training set can have a detrimental effect on  the learner&#039;s behavior: noisy or otherwise unreliable  examples from the majority class can  overwhelm the minority class. The paper discusses  criteria to evaluate the utility of classifiers  induced from such imbalanced training  sets, gives explanation of the poor behavior  of some learners under these circumstances,  and suggests as a solution a simple technique  called one-sided selection of examples.  1 Introduction  The general topic of this paper is learning from examples described by pairs [(x; c(x)], where x is a vector of attribute values and c(x) is the corresponding concept label. For simplicity, we consider only problems where c(x) is either positive or negative, and all attributes are continuous. Since Fisher (1936), this task has received plenty of attention from statisticians as well as from researchers in artificial neural networks, AI, and ML. A typical scenario assumes the e...
Improving Image Classification by Combining Statistical
. Evidence for image classification can be considered to come from two sources: traditional statistical information derived algorithmically from image data, and modelbased evidence arising from previous expertise and experience in a given application domain. This paper presents a study of classification techniques based on both these sources (traditional algorithmic and model-based), and illustrates how they can be combined. A prototype image classification system, called Cabaress, has been constructed which implements these methods. We evaluate Cabaress as applied to the problem of identifying crops in agricultural fields, based on classifying image segments extracted from radar image data. Our results demonstrate this mixed-method approach can achieve improved classificational accuracy.  1 Introduction  Following the successful launch of RADARSAT, the research activities in the application of synthetic aperture radar (SAR) have been increased significantly. Synthetic apreture radar i...
Planning and Learning in a Natural Resource Information System.

Efficient Induction of Recursive Prolog Definitons.

A WordNet%2Dbased Algorithm for Word Sense Disambiguation.
We present an algorithm for automatic word sense disambiguation, based on lexical knowledge contained in WordNet and on the results of surface-syntactic analysis. The algorithm is part of a system that analyzes texts in order to acquire knowledge in the presence of as little pre-coded semantic knowledge as possible. On the other hand, we want to make the best use of public-domain information sources such as WordNet. Rather than depend on large amounts of hand-crafted knowledge or statistical data from large corpora, we use syntactic information and information in WordNet and minimize the need for other knowledge sources in the word sense disambiguation process. We propose to guide disambiguation by semantic similarity between words and heuristic rules based on this similarity. The algorithm has been applied to the Canadian Income Tax Guide. Test results indicate that even on a relatively small text the proposed method produces correct noun meaning more than 72% of the time. 1 Introduct...
Guest Editors%27 Introduction Environmental Applications of AI.

Machine Learning and Planning for Data Management in Forestry.

Reuse of Modular Software with Automated Comment Analysis.

Inverting Implication with Small Training Sets.
. We present an algorithm for inducing recursive clauses using inverse implication (rather than inverse resolution) as the underlying generalization method. Our approach applies to a class of logic programs similar to the class of primitive recursive functions. Induction is performed using a small number of positive examples that need not be along the same resolution path. Our algorithm, implemented in a system named CRUSTACEAN, locates matched lists of generating terms that determine the pattern of decomposition exhibited in the (target) recursive clause. Our theoretical analysis defines the class of logic programs for which our approach is complete, described in terms characteristic of other ILP approaches. Our current implementation is considerably faster than previously reported. We present evidence demonstrating that, given randomly selected inputs, increasing the number of positive examples increases accuracy and reduces the number of outputs. We relate our approach to similar re...
Learning Recursive Relations with Randomly Selected Small Training Sets.
We evaluate CRUSTACEAN, an inductive logic programming algorithm that uses inverse implication to induce recursive clauses from examples. This approach is well suited for learning a class of self-recursive clauses, which commonly appear in logic programs, because it searches for common substructures among the examples. However, little evidence exists that inverse implication approaches perform well when given only randomly selected positive and negative examples. We show that CRUSTACEAN  learns recursive relations with higher accuracies than GOLEM, yet with reasonable efficiency. We also demonstrate that increasing the number of randomly selected positive and negative examples increases its accuracy on randomly selected test examples, increases the frequency in which it outputs the target relation, and reduces its number of outputs. We also prove a theorem that defines the class of logic programs for which our approach is complete. 1 MOTIVATION  This paper extends our previous work (Ah...
Reuse of Modular Software with Automated Comment Analysis.

Learning Domain Theories using Abstract Beckground Knowledge.

Using Qualitative Models to Guide Inductive Learning.
This paper presents a method for using qualitative  modds to guide inductive learning.
Enhancing Reuse of Smalltalk Methods by Conceptual Clustering.

Constructive Inductive Logic Programming.
Automating the construction of semantic grammars is a difficult and interesting problem for machine learning. This paper shows how the semantic-grammar acquisition problem can be viewed as the learning of search-control heuristics in a logic program. Appropriate control rules are learned using a new first-order induction algorithm that automatically invents useful syntactic and semantic categories. Empirical results show that the learned parsers generalize well to novel sentences and out-perform previous approaches based on connectionist techniques.  Introduction  Designing computer systems to &#034;understand&#034; natural language input is a difficult task. The laboriously hand-crafted computational grammars supporting natural language applications are often inefficient, incomplete and ambiguous. The difficulty in constructing adequate grammars is an example of the &#034;knowledge acquisition bottleneck&#034; which has motivated much research in machine learning. While numerous researchers have studied ...
A Case%2DBased Approach to Software Reuse.

Sub%2Dunification A Tool for Efficient Induction of Recursive Programs.

Integrating EBL with automatic Text Analysis.

Explanation%2Dbased Learning Helps Acquire Knowledge from Natural Language Texts.

The Importance of Causal Structure and Facts in Evaluating Explanations.

Explanation%2DBased Learning with Incomplete Theories A Three%2Dstep Approach.

An Application of Explanation%2DBased Learning to Protocol Conformance Testing.

LEW Learning by Watching.

NEGOPLAN An Inference%2DBased Negotiation Support Tool.

Learning Procedural Knowledge in the EBG Context.

Knowledge acquisition by incremental learning from problem%2Dsolution pairs.

Measuring Quality of Concept Descriptions.
An important aspect of any learning method is an evaluation of the learned knowledge, in particular, an evaluation of the plausibility and usefulness of concept descriptions that are being created. This paper presents a new, general method for evaluating concept descriptions.
Representing and Acquiring Imprecise and Context%2Ddependent Concepts in Knowledge%2DBased Systems.

Learning by Watching An Incremental Machine Learning Method that Acquires Rules by Conceptual Clustering.

A Logic%2DBased Tools for Negotiation Support.

Copying of Dynamic Structures in a Pascal Environment.

Prograph A Preliminary Report.

A Logic%2DBased Knowledge Source System for Natural Language Document.

Implementation Strategies for Plan%2DBased Deduction.

Exponential Improvement of Efficient Backtracking A Strategy for Plan%2DBased Deduction.

Exponential Improvement of Efficient Backtracking data Structure and Implementation.

On the Completeness of a Set of Transformations Optimizing Linear Programs.

An Experimental Investigation of Geschke%27s Method of Global Program Optimization.

