We revisit recently proposed algorithms for probabilistic clustering with pair-wise constraints between data points. We evaluate and compare existing techniques in terms of robustness to misspecified constraints. We show that the technique that strictly enforces the given constraints, namely the chunklet model, produces poor results even under a small number of misspecified constraints. We further show that methods that penalize constraint violation are more robust to misspecified constraints but have undesirable local behaviors. Based on this evaluation, we propose a new learning technique, extending the chunklet model to allow soft constraints represented by an intuitive measure of confidence in the constraint.

Dimensionality reduction of a feature set is a common preprocessing step used for pattern recognition and classification applications. Principal Component Analysis (PCA) is one of the popular methods used, and can be shown to be optimal using different optimality criteria. However, it has the disadvantage that measurements from all the original features are used in the projection to the lower dimensional space. This paper proposes a novel method for dimensionality reduction of a feature set by choosing a subset of the original features that contains most of the essential information, using the same criteria as PCA. We call this method Principal Feature Analysis (PFA). The proposed method is successfully applied for choosing the principal features in face tracking and content-based image retrieval (CBIR) problems. Automated annotation of digital pictures has been a highly challenging problem for computer scientists since the invention of computers. The capability of annotating pictures by computers can lead to breakthroughs in a wide range of applications including Web image search, online picture-sharing communities, and scientific experiments. In our work, by advancing statistical modeling and optimization techniques, we can train computers about hundreds of semantic concepts using example pictures from each concept. The ALIPR (Automatic Linguistic Indexing of Pictures - Real Time) system of fully automatic and high speed annotation for online pictures has been constructed. Thousands of pictures from an Internet photo-sharing site, unrelated to the source of those pictures used in the training process, have been tested. The experimental results show that a single computer processor can suggest annotation terms in real-time and with good accuracy.

No abstract available.

Violations of service level objectives (SLO) in Internet services are urgent conditions requiring immediate attention. Previously we explored (I. Cohen et al., 2004) an approach for identifying which low-level system properties were correlated to high-level SLO violations (the metric attribution problem). The approach is based on automatically inducing models from data using pattern recognition and probability modeling techniques. In this paper we extend our approach to adapt to changing workloads and external disturbances by maintaining an ensemble of probabilistic models, adding new models when existing ones do not accurately capture current system behavior. Using realistic workloads on an implemented prototype system, we show that the ensemble of models captures the performance behavior of the system accurately under changing workloads and conditions. We fuse information from the models in the ensemble to identify likely causes of the performance problem, with results comparable to those produced by an oracle that continuously changes the model based on advance knowledge of the workload. The cost of inducing new models and managing the ensembles is negligible, making our approach both immediately practical and theoretically appealing.

We use data mining and machine learning techniques to predict upcoming periods of high utilization or poor performance in enterprise systems. The abundant data available and complexity of these systems defies human characterization or static models and makes the task suitable for data mining techniques. We formulate the problem as one of classification: given current and past information about the system's behavior, can we forecast whether the system will meet its performance targets over the next hour? Using real data gathered from several enterprise systems in Hewlett-Packard, we compare several approaches ranging from time series to Bayesian networks. Besides establishing the predictive power of these approaches our study analyzes three dimensions that are important for their application as a stand alone tool. First, it quantifies the gain in accuracy of multivariate prediction methods over simple statistical univariate methods. Second, it quantifies the variations in accuracy when using different classes of system and workload features. Third, it establishes that models induced using combined data from various systems generalize well and are applicable to new systems, enabling accurate predictions on systems with insufficient historical data. Together this analysis offers a promising outlook on the development of tools to automate assignment of resources to stabilize performance, (e.g., adding servers to a cluster) and allow opportunistic job scheduling (e.g., backups or virus scans).

We present a method for automatically extracting from a running system an indexable signature that distills the essential characteristic from a system state and that can be subjected to automated clustering and similarity-based retrieval to identify when an observed system state is similar to a previously-observed state. This allows operators to identify and quantify the frequency of recurrent problems, to leverage previous diagnostic efforts, and to establish whether problems seen at different installations of the same site are similar or distinct. We show that the naive approach to constructing these signatures based on simply recording the actual ``raw'' values of collected measurements is ineffective, leading us to a more sophisticated approach based on statistical modeling and inference. Our method requires only that the system's metric of merit (such as average transaction response time) as well as a collection of lower-level operational metrics be collected, as is done by existing commercial monitoring tools. Even if the traces have no annotations of prior diagnoses of observed incidents (as is typical), our technique successfully clusters system states corresponding to similar problems, allowing diagnosticians to identify recurring problems and to characterize the ``syndrome'' of a group of problems. We validate our approach on both synthetic traces and several weeks of production traces from a customer-facing geoplexed 24 x 7 system; in the latter case, our approach identified a recurring problem that had required extensive manual diagnosis, and also aided the operators in correcting a previous misdiagnosis of a different problem.

Automatic classification is one of the basic tasks required in any pattern recognition and human computer interaction application. In this paper, we discuss training probabilistic classifiers with labeled and unlabeled data. We provide a new analysis that shows under what conditions unlabeled data can be used in learning to improve classification performance. We also show that, if the conditions are violated, using unlabeled data can be detrimental to classification performance. We discuss the implications of this analysis to a specific type of probabilistic classifiers, Bayesian networks, and propose a new structure learning algorithm that can utilize unlabeled data to improve classification. Finally, we show how the resulting algorithms are successfully employed in two applications related to human-computer interaction and pattern recognition: facial expression recognition and face detection.

Understanding human emotions is one of the necessary skills for the computer to interact intelligently with human users. The most expressive way humans display emotions is through facial expressions. In this paper, we report on several advances we have made in building a system for classification of facial expressions from continuous video input. We use Bayesian network classifiers for classifying expressions from video. One of the motivating factor in using the Bayesian network classifiers is their ability to handle missing data, both during inference and training. In particular, we are interested in the problem of learning with both labeled and unlabeled data. We show that when using unlabeled data to learn classifiers, using correct modeling assumptions is critical for achieving improved classification performance. Motivated by this, we introduce a classification driven stochastic structure search algorithm for learning the structure of Bayesian network classifiers. We show that with moderate size labeled training sets and large amount of unlabeled data, our method can utilize unlabeled data to improve classification performance. We also provide results using the Naive Bayes (NB) and the Tree-Augmented Naive Bayes (TAN) classifiers, showing that the two can achieve good performance with labeled training sets, but perform poorly when unlabeled data are added to the training set.

Recognizing human facial expression and emotion by computer is an interesting and challenging problem. We propose a method for recognizing emotions through facial expressions displayed in video sequences. We introduce a tree-augmented naive-Bayes (TAN) classifier that learns the dependencies between facial features; we also provide an algorithm for finding the best TAN structure. Our person-dependent and person-independent experiments show that using this TAN structure provides significantly better results than using simpler NB-classifiers.

Accurate calculation of the three dimensional shape of an object is one of the classic research areas of computer vision. Many of the existing methods are based on surface normal estimation, and subsequent integration of surface gradients. In general, these methods do not produce valid surfaces due to violation of surface integrability. We introduce a new method for shape reconstruction by integration of valid surface gradient maps. The essence of the new approach is in the strict enforcement of the surface integrability via belief propagation across graphical models. The graphical model is selected in such a way as to extract information from underlying, possibly noisy, surface gradient estimators, utilize the surface integrability constraint, and produce the maximum a-posteriori estimate of a valid surface. We demonstrate the algorithm for two classic shape reconstruction techniques; shape-from-shading and photometric stereo. On a set of real and synthetic examples, the new approach is shown to be fast and accurate, in the sense that shape can be rendered even in the presence of high levels of noise and sharp occlusion boundaries.

We propose a general architecture and implementation for the autonomous assessment of the health of arbitrary service elements, as a necessary prerequisite to self-control. We describe a health engine, the central component of our proposed 'self-awareness and control' architecture. The health engine combines domain independent statistical analysis and probabilistic reasoning technology (Bayesian networks) with domain dependent measurement collection and evaluation methods. The resultant probabilistic assessment enables open, non-hierarchical communications about service element health. We demonstrate the validity of our approach using HP's corporate email service and detecting email anomalies: mail loops and a virus attack.

Person recognition is a fundamental problem faced in any computer vision system. This problem is relatively easy if the frontal view is available, however, it gets intractable in the absence of the frontal view. We have provided a framework, which tries to solve this problem using the top view of the person. A special scenario of "smart conference room" is considered. Although, not much information is available in the top view, we have shown that by making use of DTC and Bayesian networks the output of the various sensors can be combined to solve this problem. The results presented in the end show that we can do person recognition (pose independent) with 96% accuracy for a group of 12 people. For pose dependent case, we have achieved 100% accuracy. Finally we have provided a framework to achieve this in real time.