International Journal on Artificial Intelligence Tools

Diffusion geometry offers a fresh perspective on multi-scale information analysis, which is critical to multiagent systems that need to process massive data sets. A recent study has shown that when the "diffusion distance" concept is applied to human decision experiences, its performance on solution synthesis can be significantly better than using Euclidean distance. However, as a data set expands over time, it can quickly exceed the processing capacity of a single agent. In this paper, we proposed a multi-agent diffusion approach where a massive data set is split into several subsets and each diffusion agent only needs to work with one subset in diffusion computation. We conducted experiments with different splitting strategies applied to a set of decision experiences. The result indicates that the multi-agent diffusion approach is beneficial, and it is even possible to benefit from using a larger group of diffusion agents if their subsets have common experiences and pairly-shared experiences. Our study also shows that system performance could be affected significantly by the splitting granularity (size of each splitting unit). This study paves the road for applying the multi-agent diffusion approach to massive data analysis.

Problems of motion coordination of multiple entities are addressed in this paper. These problems are dealt on the abstract level where they can be viewed as tasks of constructing a spatial-temporal plan for a set of identical mobile entities. The entities reside in a certain environment where they can move. Each entity need to reach a given goal position supposed it is starting from some initial position. The most abstract formal representations of coordinated motion problems are known as “pebble motion on a graph” and “multi-robot path planning”. The existent algorithms for pebble motion and multi-robot problems were suspected of generating solutions containing redundancies. This hypothesis eventually confirmed in this work. We present several techniques for identifying and eliminating redundancies from solutions generated by existent algorithms. An extensive experimental evaluation was performed and it showed that the quality of generated solutions can be improved up to the order of magnitude. We also identify parameters characterizing instances of problems where a significant improvement is expectable.

Despite the widespread usage of the evaluation mediums for online services by the clients, there is a requirement for a trust evaluation tool that provides the clients with the degree of trustworthiness of the service providers. Such a tool can provide increased familiarity with unknown third party entities, e.g. service providers, especially when those entities neither project completely trustworthy nor totally untrustworthy behaviour. Indeed, developing some metrics for trust evaluation under uncertainty can come handy, e.g., for customers interested in evaluating the trustworthiness of an unknown service provider throughout queries to other customers of unknown reliability.

In this research, we propose an evaluation metric to estimate the degree of trustworthiness of an unknown agent, say a^D, through the information acquired through a group of agents who have interacted with agent a^D. This group of agents is assumed to have an unknown degree of reliability. In order to tackle the uncertainty associated with the trust of these set of unknown agents, we suggest to use possibility distributions. Later, we introduce a new certainty metric to measure the degree of agreement in the information reported by the group of agents in A on agent a^D. Fusion rules are then used to measure an estimation of the agent a^D’s degree of trustworthiness. To the best of our knowledge, this is the first work that estimates trust, out of empirical data, subject to some uncertainty, in a discrete multi-valued trust domain. Finally, numerical experiments are presented to validate the proposed tools and metrics.

In this paper, we address the problem of automatically segmenting non-rigid pedestrians in still images. Since this task is well known difficult for any type of model or cue alone, a novel approach utilizing shape, puzzle and appearance cues is presented. The major contribution of this approach lies in the combination of multiple cues to refine pedestrian segmentation successively, which has two characterizations: (1) a shape guided puzzle integration scheme, which extracts pedestrians via assembling puzzles with constraint of a shape template; (2) a pedestrian refinement scheme, which is fulfilled by optimizing an automatically generated trimap that encodes both human silhouette and skeleton. Qualitative and quantitative evaluations on several public datasets verify the approach's effectiveness to various articulated bodies, human appearance and partial occlusion, and that this approach is able to segment pedestrians more accurately than methods based only on appearance or shape cue.

A strong assumption to derive generalization guarantees in the standard PAC framework is that training (or source) data and test (or target) data are drawn according to the same distribution. Because of the presence of possibly outdated data in the training set, or the use of biased collections, this assumption is often violated in real-world applications leading to different source and target distributions. To go around this problem, a new research area known as Domain Adaptation (DA) has recently been introduced giving rise to many adaptation algorithms and theoretical results in the form of generalization bounds. This paper deals with self-labeling DA whose goal is to iteratively incorporate semi-labeled target data in the learning set to progressively adapt the classifier from the source to the target domain. The contribution of this work is three-fold: First, we provide the minimum and necessary theoretical conditions for a self-labeling DA algorithm to perform an actual domain adaptation. Second, following these theoretical recommendations, we design a new iterative DA algorithm, called GESIDA, able to deal with structured data. This algorithm makes use of the new theory of learning with (ε,γ,τ)-good similarity functions introduced by Balcan et al., which does not require the use of a valid kernel to learn well and allows us to induce sparse models. Finally, we apply our algorithm on a structured image classification task and show that self-labeling domain adaptation is a new original way to deal with scaling and rotation problems.

In protein tertiary structure prediction, assessing the quality of predicted models is an essential task. Over the past years, many methods have been proposed for the protein model quality assessment (QA) and selection problem. Despite significant advances, the discerning power of current methods is still unsatisfactory. In this paper, we propose two new algorithms, CC-Select and MDS-QA, based on multidimensional scaling and k-means clustering. For the model selection problem, CC-Select combines consensus with clustering techniques to select the best models from a given pool. Given a set of predicted models, CC-Select first calculates a consensus score for each structure based on its average pairwise structural similarity to other models. Then, similar structures are grouped into clusters using multidimensional scaling and clustering algorithms. In each cluster, the one with the highest consensus score is selected as a candidate model. For the QA problem, MDS-QA combines single-model scoring functions with consensus to determine more accurate assessment score for every model in a given pool. Using extensive benchmark sets of a large collection of predicted models, we compare the two algorithms with existing state-of-the-art quality assessment methods and show significant improvement.

Information on the World Wide Web is congested with large amounts of news contents. Recommending, filtering, and summarization of Web news have become hot topics of research in Web intelligence, aiming to find interesting news for users and give concise content for reading. This paper presents our research on developing the Personalized News Filtering and Summarization system (PNFS). An embedded learning component of PNFS induces a user interest model and recommends personalized news. Two Web news recommendation methods are proposed to keep tracking news and find topic interesting news for users. A keyword knowledge base is maintained and provides real-time updates to reflect the news topic information and the user's interest preferences. The non-news content irrelevant to the news Web page is filtered out. A keyword extraction method based on lexical chains is proposed that uses the semantic similarity and the relatedness degree to represent the semantic relations between words. Word sense disambiguation is also performed in the built lexical chains. Experiments on Web news pages and journal articles show that the proposed keyword extraction method is effective. An example run of our PNFS system demonstrates the superiority of this Web intelligence system.

We present a new approach for analyzing topic models using visual analytics. We have developed TopicView, an application for visually comparing and exploring multiple models of text corpora, as a prototype for this type of analysis tool. TopicView uses multiple linked views to visually analyze conceptual and topical content, document relationships identified by models, and the impact of models on the results of document clustering. As case studies, we examine models created using two standard approaches: Latent Semantic Analysis (LSA) and Latent Dirichlet Allocation (LDA). Conceptual content is compared through the combination of (i) a bipartite graph matching LSA concepts with LDA topics based on the cosine similarities of model factors and (ii) a table containing the terms for each LSA concept and LDA topic listed in decreasing order of importance. Document relationships are examined through the combination of (i) side-by-side document similarity graphs, (ii) a table listing the weights for each document's contribution to each concept/topic, and (iii) a full text reader for documents selected in either of the graphs or the table. The impact of LSA and LDA models on document clustering applications is explored through similar means, using proximities between documents and cluster exemplars for graph layout edge weighting and table entries. We demonstrate the utility of TopicView's visual approach to model assessment by comparing LSA and LDA models of several example corpora.

A burst is an unusually large number of events occurring within a certain period of time. Various burst analysis methods over multiple window size have been proposed. However, real-time burst analysis is difficult owing to the huge amount of computation. This article proposes a method to analyze a burst in real-time. The proposed method reduces computation by avoiding redundant data updates of no event occurrence, and by suppressing data within a certain period in the case of emergent increase of events. In addition, a level and a rank of a burst can be analyzed simultaneously. The effectiveness of the proposed method is evaluated by experiments with real data. The result shows that the proposed method is more effective in analyzing real data, and more efficient in handling a large amount of events.

Software metrics (features or attributes) are collected during the software development cycle. Metric selection is one of the most important preprocessing steps in the process of building defect prediction models and may improve the final prediction result. However, the addition or removal of program modules (instances or samples) can alter the subsets chosen by a feature selection technique, rendering the previously-selected feature sets invalid. Very limited research have been done considering both stability (or robustness) and defect prediction model performance together in the software engineering domain, despite the importance of both aspects when choosing a feature selection technique. In this paper, we test the stability and classification model performance of eighteen feature selection techniques as the magnitude of change to the datasets and the size of the selected feature subsets are varied. All experiments were conducted on sixteen datasets from three real-world software projects. The experimental results demonstrate that Gain Ratio shows the least stability while two different versions of ReliefF show the most stability, followed by the PRC- and AUC-based threshold-based feature selection techniques. Results also show that the signal-to-noise ranker performed moderately in terms of robustness and was the best ranker in terms of model performance. Finally, we conclude that while for some rankers, stability and classification performance are correlated, this is not true for other rankers, and therefore performance according to one scheme (stability or model performance) cannot be used to predict performance according to the other.

One of the most important types of signal found in the area of machine condition monitoring/prognostic health monitoring (MCM/PHM) is the vibration signal, a type of waveform. Many time-frequency domain techniques have been proposed to interpret such signals, including wavelet packet decomposition (WPD). Previous work has shown how to extend the WPD algorithm to operate on streaming signals, but the number of output variables becomes exponential in the number of levels of decomposition, hindering data mining in limited-memory environments. Feature selection techniques, well understood in other areas of data mining, can be used to greatly reduce the number of output variables and speed up the machine learning algorithms. This paper presents a case study comparing two versions of WPD both with and without feature selection, demonstrating that removing most of the features produced by the WPD does not impair its performance within the context of MCM/PHM.

Because of the scarcity and diversity of outliers, it is very difficult to design a robust outlier detector. In this paper, we first propose to use the maximum margin criterion to sift unknown outliers, which demonstrates superior performance. However, the resultant learning task is formulated as a Mixed Integer Programming (MIP) problem, which is computationally hard. Therefore, we alter the recently developed label generating technique, which efficiently solves a convex relaxation of the MIP problem of outlier detection. Specifically, we propose an effective procedure to find a largely violated labeling vector for identifying rare outliers from abundant normal patterns, and its convergence is also presented. Then, a set of largely violated labeling vectors are combined by multiple kernel learning methods to robustly detect outliers. Besides these, in order to further enhance the efficacy of our outlier detector, we also explore the use of maximum volume criterion to measure the quality of separation between outliers and normal patterns. This criterion can be easily incorporated into our proposed framework by introducing an additional regularization. Comprehensive experiments on toy and real-world data sets verify that the outlier detectors using the two proposed criteria outperform existing outlier detection methods. Furthermore, our models are employed to detect corporate credit risk and demonstrate excellent performance.

Evidence of malicious insider activity is often buried within large data streams, such as system logs accumulated over months or years. Ensemble-based stream mining leverages multiple classification models to achieve highly accurate anomaly detection in such streams, even when the stream is unbounded, evolving, and unlabeled. This makes the approach effective for identifying insider threats who attempt to conceal their activities by varying their behaviors over time. This paper applies ensemble-based stream mining, supervised and unsupervised learning, and graph-based anomaly detection to the problem of insider threat detection. It demonstrates that the ensemble-based approach is significantly more effective than traditional single-model methods, supervised learning outperforms unsupervised learning, and increasing the cost of false negatives correlates to higher accuracy. Future work will consider a wider range of tunable parameters in an effort to further reduce false positives, include a more sophisticated polling algorithm for weighting better models, and implement parallelization to lower runtimes to more rapidly detect emerging insider threats.