Interactions are patterns between several attributes in data that cannot be inferred from any subset of these attributes. While mutual information is a well-established approach to evaluating the interactions between two attributes, we surveyed its generalizations as to quantify interactions between several attributes. We have chosen McGill’s interaction information, which has been independently rediscovered a number of times under various names in various disciplines, because of its many intuitively appealing properties. We apply interaction information to visually present the most important interactions of the data. Visualization of interactions has provided insight into the structure of data on a number of domains, identifying redundant attributes and opportunities for constructing new features, discovering unexpected regularities in data, and have helped during construction of predictive models; we illustrate the methods on numerous examples. A machine learning method that disregards interactions may get caught in two traps: myopia is caused by learning algorithms assuming independence in spite of interactions, whereas fragmentation arises from assuming an interaction in spite of independence.

This paper is about constructing confidence bands around ROC curves. We first introduce to the machine learning community three band-generating methods from the medical field, and evaluate how well they perform. Such confidence bands represent the region where the "true" ROC curve is expected to reside, with the designated confidence level. To assess the containment of the bands we begin with a synthetic world where we know the true ROC curve---specifically, where the class-conditional model scores are normally distributed. The only method that attains reasonable containment out-of-the-box produces non-parametric, "fixed-width" bands (FWBs). Next we move to a context more appropriate for machine learning evaluations: bands that with a certain confidence level will bound the performance of the model on future data. We introduce a correction to account for the larger uncertainty, and the widened FWBs continue to have reasonable containment. Finally, we assess the bands on 10 relatively large benchmark data sets. We conclude by recommending these FWBs, noting that being non-parametric they are especially attractive for machine learning studies, where the score distributions (1) clearly are not normal, and (2) even for the same data set vary substantially from learning method to learning method.

Much work has studied the effect of different treatments of missing values on model induction, but little work has analyzed treatments for the common case of missing values at prediction time. This paper first compares several different methods—predictive value imputation, the distributionbased imputation used by C4.5, and using reduced models—for applying classification trees to instances with missing values (and also shows evidence that the results generalize to bagged trees and to logistic regression). The results show that for the two most popular treatments, each is preferable under different conditions. Strikingly the reduced-models approach, seldom mentioned or used, consistently outperforms the other two methods, sometimes by a large margin. The lack of attention to reduced modeling may be due in part to its (perceived) expense in terms of computation or storage. Therefore, we then introduce and evaluate alternative, hybrid approaches that allow users to balance between more accurate but computationally expensive reduced modeling and the other, less accurate but less computationally expensive treatments. The results show that the hybrid methods can scale gracefully to the amount of investment in computation/storage, and that they outperform imputation even for small investments.

This dissertation is an investigation into how ambiguity should be classified for authors and readers of text, and how this process can be automated. Usually, authors and readers disambiguate ambiguity, either consciously or unconsciously. However, disambiguation is not always appropriate. For instance, a linguistic construction may be read differently by different people, with no consensus about which reading is the intended one. This is particularly dangerous if they do not realise that other readings are possible. Misunderstandings may then occur. This is particularly serious in the field of requirements engineering. If requirements are misunderstood, systems may be built incorrectly, and this can prove very costly. Our research uses natural language processing techniques to address ambiguity in requirements. We develop a model of ambiguity, and a method of applying it, which represent a novel approach to the problem described here. Our model is based on the notion that human perception is the only valid criterion for judging ambiguity. If people perceive very differently how an ambiguity should be read, it will cause misunderstandings. Assigning a preferred reading to it is therefore unwise. In

Software defect prediction strives to improve software quality and testing efficiency by constructing predictive classification models from code attributes to enable a timely identification of fault-prone modules. Several classification models have been evaluated for this task. However, due to inconsistent findings regarding the superiority of one classifier over another and the usefulness of metric-based classification in general, more research is needed to improve convergence across studies and further advance confidence in experimental results. We consider three potential sources for bias: comparing classifiers over one or a small number of proprietary data sets, relying on accuracy indicators that are conceptually inappropriate for software defect prediction and cross-study comparisons, and, finally, limited use of statistical testing procedures to secure empirical findings. To remedy these problems, a framework for comparative software defect prediction experiments is proposed and applied in a large-scale empirical comparison of 22 classifiers over 10 public domain data sets from the NASA Metrics Data repository. Overall, an appealing degree of predictive accuracy is observed, which supports the view that metric-based classification is useful. However, our results indicate that the importance of the particular classification algorithm may be less than previously assumed since no significant performance differences could be detected among the top 17 classifiers.

Abstract not found

This paper presents a new structure-based interest region detector called Principal Curvature-Based Regions (PCBR) which we use for object class recognition. The PCBR interest operator detects stable watershed regions within the multi-scale principal curvature image. To detect robust watershed regions, we “clean ” a principal curvature image by combining a grayscale morphological close with our new “eigenvector flow ” hysteresis threshold. Robustness across scales is achieved by selecting the maximally stable regions across consecutive scales. PCBR typically detects distinctive patterns distributed evenly on the objects and it shows significant robustness to local intensity perturbations and intra-class variations. We evaluate PCBR both qualitatively (through visual inspection) and quantitatively (by measuring repeatability and classification accuracy in real-world object-class recognition problems). Experiments on different benchmark datasets show that PCBR is comparable or superior to state-of-art detectors for both feature matching and object recognition. Moreover, we demonstrate the application of PCBR to symmetry detection.

In this paper, we present an algorithm for multidimensional vector regression on data that are highly uncertain and nonlinear, and then apply it to the problem of indoor location estimation in a wireless local area network (WLAN). Our aim is to obtain an accurate mapping between the signal space and the physical space without requiring too much human calibration effort. This location estimation problem has traditionally been tackled through probabilistic models trained on manually labeled data, which are expensive to obtain. In contrast, our algorithm adopts Kernel Canonical Correlation Analysis (KCCA) to build a nonlinear mapping between the signal-vector space and the physical location space by transforming data in both spaces into their canonical features. This allows the pairwise similarity of samples in both spaces to be maximally correlated using kernels. We use a Gaussian kernel to adapt to the noisy characteristics of signal strengths and a Matérn kernel to sense the changes in physical locations. By using real data collected in an 802.11 wireless LAN environment, we achieve accurate location estimation for pervasive computing while requiring a much smaller set of labeled training data than previous methods.

Abstract. Mining of data streams must balance three evaluation dimensions: accuracy, time and memory. Excellent accuracy on data streams has been obtained with Naive Bayes Hoeffding Trees—Hoeffding Trees with naive Bayes models at the leaf nodes—albeit with increased runtime compared to standard Hoeffding Trees. In this paper, we show that runtime can be reduced by replacing naive Bayes with perceptron classifiers, while maintaining highly competitive accuracy. We also show that accuracy can be increased even further by combining majority vote, naive Bayes, and perceptrons. We evaluate four perceptron-based learning strategies and compare them against appropriate baselines: simple perceptrons, Perceptron Hoeffding Trees, hybrid Naive Bayes Perceptron Trees, and bagged versions thereof. We implement a perceptron that uses the sigmoid activation function instead of the threshold activation function and optimizes the squared error, with one perceptron per class value. We test our methods by performing an evaluation study on synthetic and real-world datasets comprising up to ten million examples. 1

Abstract. Logistic Model Trees have been shown to be very accurate and compact classifiers [8]. Their greatest disadvantage is the computational complexity of inducing the logistic regression models in the tree. We address this issue by using the AIC criterion [1] instead of crossvalidation to prevent overfitting these models. In addition, a weight trimming heuristic is used which produces a significant speedup. We compare the training time and accuracy of the new induction process with the original one on various datasets and show that the training time often decreases while the classification accuracy diminishes only slightly. 1