An important sub-field of machine learning is the inductive formation of a pertinent class description. Given a collection of positive and negative examples of the concept, the aim is to create a description not only capable of correctly classifying the training examples, but one able to be used predictively on unseen examples. This thesisi nvestigatesth e problemo f inductivec onceptf ormation in poorly understood domains.M any well-understoodp roblemse xist wheret he individual attribute-valueuss ed to describee xamplesv ary systematicallyw ith categorym embership.O ften this meanst hat such descriptions are sufficient to identify significant regularities in the concept. In contrastm, anyr eal-worldp roblemsa rep oorly understoodi,. e . examplesa re describedb y a relatively large number of seemingly irrelevant attributes (because expertise is often unavailablet o specify a suitable level of abstractionw hen measurementas re initially recorded). The fundamentaal ssumptionb eing that when combinedi n somew ay, these attributes are complete enough to identify the target concept. This initial language insufficiency,o ften causedb y concealeda ttributei nteractionp resentsp roblemsf or many current induction algorithms which typically reply on uncovering simpler correlations. For all but the simplestp roblems,t he combinatoriale xplosiona ssociatedw ith unconstrained hypothesis generation means that the inductive process must employ more intelligent mechanisms. A two-stage solution is proposed based on first identifying whether the initial problem formulation has the potential to cause difficulties for typical inductive learners. A qualitative measure based on a novel information theoretic function is used to gauge the absence of conditional dependencies between attributes. This approach is compared to other current identification measures, in particular a bias towards misleading estimates of concept difficulty due to irrelevant attributes is addressed. Once the level of attribute interaction has been estimated one of two learning components is selected for acquiring the relevant concept. For low to moderate degrees of attribute interaction, a general-to-specific beam search is utilised. However this mechanism focuses the induction process on the most promising hypotheses by utilising relative assessment measures i. e. the degree with which a specialised hypothesis improves with respect to its constituent parts. This relative improvement becomes increasingly important as conditional dependencies increase. In addition, a pair of relative bounds are calculated for each hypothesis based on the assessmenht euristic used for validation whilst learning. These bounds place limits on the number of negative examples a hypothesis can cover and still outperform its best constituent part. These bounds result in a substantial reduction in the number of poor hypotheses generated during concept formation. For extremel evels of featurei nteraction,a specific-to-generagl reedy searcht echniquei s employed. This approach is more likely to uncover hidden interactions than approaches that begin hypothesisf ormationb asedo n one-dimensionapl rojections. This combination of search direction and a heuristic based on Minimum Description Length, ensures that highly conditional dependenciecsa n be pinpointed. In addition a number of speedup operatorsa re developedw hich curtail the numbero f tentativeh ypothesesg enerateda nd alsor esulti n fewerp roblemsa ssociatedw ith local searchs pacem inima.

published_or_final_version / Industrial Engineering / Master / Master of Philosophy

published_or_final_version / Computer Science / Master / Master of Philosophy

Activities of DNA are to a great extent controlled epigenetically through the internal struc- ture of chromatin. This structure is dynamic and is influenced by different modifications of histone proteins. Various combinations of epigenetic modification of histones pinpoint to different functional regions of the DNA determining the so-called chromatin states. How- ever, the characterization of chromatin states by the DNA sequence properties remains largely unknown. In this study we aim to explore whether DNA sequence patterns in the human genome can characterize different chromatin states. Using DNA sequence motifs we built binary classifiers for each chromatic state to eval- uate whether a given genomic sequence is a good candidate for belonging to a particular chromatin state. Of four classification algorithms (C4.5, Naive Bayes, Random Forest, and SVM) used for this purpose, the decision tree based classifiers (C4.5 and Random Forest) yielded best results among those we evaluated. Our results suggest that in general these models lack sufficient predictive power, although for four chromatin states (insulators, het- erochromatin, and two types of copy number variation) we found that presence of certain motifs in DNA sequences does imply an increased probability that such a sequence is one of these chromatin states.

Twitter with its rising popularity as a micro-blogging website has inevitably attracted attention of spammers. Spammers use myriad of techniques to lure victims into clicking malicious URLs. In this thesis, we present several novel features capable of distinguishing spam accounts from legitimate accounts in real-time. The features exploit the behavioral and content entropy, bait-techniques, community-orientation, and profile characteristics of spammers. We then use supervised learning algorithms to generate models using the proposed features and show that our tool, spAmbush, can detect spammers in real-time. Our analysis reveals detection of more than 90% of spammers with less than five tweets and more than half with only a single tweet. Our feature computation has low latency and resource requirement. Our results show a 96% detection rate with only 0.01% false positive rate. We further cluster the unknown spammers to identify and understand the prevalent spam campaigns on Twitter.

No description available.

No description available.

This thesis provides a comparative study of machine vision (MV) classification techniques for the detection of missing clips on an automotive part known as a cross car beam. This is a difficult application for an automated MV system because the inspection is conducted in an open manufacturing environment with variable lighting conditions. A laboratory test cell was first used to investigate the effect of lighting. QVision, a software program originally developed at Queen’s University, was used to perform a representative inspection task. Solutions with different light sources and camera settings were investigated in order to determine the best possible set up to acquire an image of the part. Feature selection was applied to improve the results of this classification. The MV system was then installed on an industrial assembly line. QVision was modified to detect the presence or absence of four clips and communicate this information to the computer controlling the manufacturing cell. Features were extracted from the image and then a neuro fuzzy (ANFIS) system was trained to perform the inspection. A performance goal of 0% False Positives and less than 2% False Negatives was achieved with the feature based ANFIS classifier. In addition, the problem of a rusty clip was examined and a radial hole algorithm was used to improve performance in this case. In this case, the system required hours to train. Five new classifiers were then compared to the original feature based ANFIS classifier: 1) feature based with a Neural Network, 2) feature based with principle component analysis (PCA) applied and ANFIS, 3) feature based with PCA applied and a Neural Network, 4) Eigenimage based with ANFIS and 5) Eigenimage based with a Neural Network. The effect of adding a Hough rectangle feature and a principle component colour feature was also studied. It was found that the Neural Network classifier performed better than the ANFIS classifier. When PCA was applied the results improved still further. Overall, feature based classifiers had better performance than Eigenimage based classifiers. Finally, it should be noted that these six classifiers required only minutes to train. / Thesis (Master, Mechanical and Materials Engineering) -- Queen's University, 2009-08-07 17:03:10.422

No description available.

No description available.