We consider semi-supervised learning, learning task from both labeled and unlabeled instances and in particular, self-training with decision tree learners as base learners. We show that standard decision tree learning as the base learner cannot be effective in a self-training algorithm to semi-supervised learning. The main reason is that the basic decision tree learner does not produce reliable probability estimation to its predictions. Therefore, it cannot be a proper selection criterion in self-training. We consider the effect of several modifications to the basic decision tree learner that produce better probability estimation than using the distributions at the leaves of the tree. We show that these modifications do not produce better performance when used on the labeled data only, but they do benefit more from the unlabeled data in self-training. The modifications that we consider are Naive Bayes Tree, a combination of No-pruning and Laplace correction, grafting, and using a distance-based measure. We then extend this improvement to algorithms for ensembles of decision trees and we show that the ensemble learner gives an extra improvement over the adapted decision tree learners.

Crisis response organizations operate in very dynamic environments, in which it is essential for responders to acquire all information critical to their task execution in time. In reality, the responders are often faced with information overload, incomplete information, or a combination of both. This hampers their decision-making process, workflow, situational awareness and, consequently, effective execution of collaborative crisis response. Therefore, getting the right information to the right person at the right time is of crucial importance. The task of processing all data during crisis response situations and determining for whom at a particular moment the information is relevant is not straightforward. When developing an information system to support this task, some important challenges have to be taken into account. These challenges relate to the structure and truthfulness of the used data, the assessment of information relevance, and the dissemination of relevant information in time. While methods and techniques from big data can be used to collect and integrate data, machine learning can be used to build a model for relevance assessments. An example implementation of such a framework of big data is the TAID software system that collects and integrates data communicated between first responders and may send information to crisis responders that were not addressed in the initial communication. As an example of the impact of TAID on crisis response, we show its effect in a simulated crisis response scenario.

This paper presents a method for finding anomalies in gas consumption that can identify causes of wasting energy. Our approach is to use historical data on local weather, building usage and gas consumption, to predict the gas consumption for a particular day and time. The prediction is a combination of auto-regression and artificial neural networks and anomalies, relatively large deviations from the predicted gas consumption values, are detected. These can point to incorrect settings of controls, faults in installations or incorrect use of the building.

TERESA is a socially intelligent semi-autonomous telepresence system that is currently being developed as part of an FP7-STREP project funded by the European Union. The ultimate goal of the project is to deploy this system in an elderly day centre to allow elderly people to participate in social events even when they are unable to travel to the centre. In this paper, we present an overview of our progress on TERESA. We discuss the most significant scientific and technical challenges including: understanding and automatically recognizing social behaviour; defining social norms for the interaction between a telepresence robot and its users; navigating the environment while taking into account social features and constraints; and learning to estimate the social impact of the robot’s actions from multiple sources of feedback. We report on our current progress on each of these challenges, as well as our plans for future work.

A theory of “megacitation” is introduced and used in an experiment to demonstrate how a qualitative scholarly book review can be converted into a weighted bibliometric indicator. We employ a manual human-coding approach to classify book reviews in the field of history based on reviewers' assessments of a book author's scholarly credibility (SC) and writing style (WS). In total, 100 book reviews were selected from the American Historical Review and coded for their positive/negative valence on these two dimensions. Most were coded as positive (68% for SC and 47% for WS), and there was also a small positive correlation between SC and WS ( r  = 0.2). We then constructed a classifier, combining both manual design and machine learning, to categorize sentiment-based sentences in history book reviews. The machine classifier produced a matched accuracy (matched to the human coding) of approximately 75% for SC and 64% for WS. WS was found to be more difficult to classify by machine than SC because of the reviewers' use of more subtle language. With further training data, a machine-learning approach could be useful for automatically classifying a large number of history book reviews at once. Weighted megacitations can be especially valuable if they are used in conjunction with regular book/journal citations, and “libcitations” (i.e., library holding counts) for a comprehensive assessment of a book/monograph's scholarly impact.

In this paper we present an analysis of the application of the two most important types of similarity measures for moving object trajectories in machine learning from vessel movement data. These similarities are applied in the tasks of clustering, classification and outlier detection. The first similarity type are alignment measures, such as dynamic time warping and edit distance. The second type are based on the integral over time between two trajectories. Following earlier work we define these measures in the context of kernel methods, which provide state-of-the-art, robust algorithms for the tasks studied. Furthermore, we include the influence of applying piecewise linear segmentation as pre-processing to the vessel trajectories when computing alignment measures, since this has been shown to give a positive effect in computation time and performance. In our experiments the alignment based measures show the best performance. Regular versions of edit distance give the best performance in clustering and classification, whereas the softmax variant of dynamic time warping works best in outlier detection. Moreover, piecewise linear segmentation has a positive effect on alignments, due to the fact that salient points in a trajectory, especially important in clustering and outlier detection, are highlighted by the segmentation and have a large influence in the alignments. Based on our experiments, integral over time based similarity measures are not well-suited for learning from vessel trajectories.

A theory of megacitation is introduced and used in an experiment to demonstrate how a qualitative scholarly book review can be converted into a weighted bibliometric indicator. We employ a manual human-coding approach to classify book reviews in the field of history based on reviewers' assessments of a book author's scholarly credibility (SC) and writing style (WS). In total, 100 book reviews were selected from the American Historical Review and coded for their positive/negative valence on these two dimensions. Most were coded as positive (68% for SC and 47% for WS), and there was also a small positive correlation between SC and WS (r=0.2). We then constructed a classifier, combining both manual design and machine learning, to categorize sentiment-based sentences in history book reviews. The machine classifier produced a matched accuracy (matched to the human coding) of approximately 75% for SC and 64% for WS. WS was found to be more difficult to classify by machine than SC because of the reviewers' use of more subtle language. With further training data, a machine-learning approach could be useful for automatically classifying a large number of history book reviews at once. Weighted megacitations can be especially valuable if they are used in conjunction with regular book/journal citations, and libcitations (i.e., library holding counts) for a comprehensive assessment of a book/monograph's scholarly impact.

In this paper we present GAMYGDALA, an emotional appraisal engine that enables game developers to easily add emotions to their Non-Player Characters (NPC). Our approach proposes a solution that is positioned between event coding of affect, where individual events have predetermined annotated emotional consequences for NPCs, and a full blown cognitive appraisal model. Instead, for an NPC that needs emotions the game developer defines goals and annotates game events with a relation to these goals. Based on this input, GAMYGDALA produces an emotion for that NPC according to the well-known OCC model. In this paper we provide evidence for the following: GAMYGDALA provides black-box Game-AI independent emotion support, is efficient for large numbers of NPCs, and is psychologically grounded.

We present an algorithm for multiclass semi-supervised learning, which is learning from a limited amount of labeled data and plenty of unlabeled data. Existing semi-supervised learning algorithms use approaches such as one-versus-all to convert the multiclass problem to several binary classification problems, which is not optimal. We propose a multiclass semi-supervised boosting algorithm that solves multiclass classification problems directly. The algorithm is based on a novel multiclass loss function consisting of the margin cost on labeled data and two regularization terms on labeled and unlabeled data. Experimental results on a number of benchmark and real-world datasets show that the proposed algorithm performs better than the state-of-the-art boosting algorithms for multiclass semi-supervised learning, such as SemiBoost (Mallapragada et al., 2009) and RegBoost (Chen and Wang, 2011).

In this paper, we consider the multiclass semi-supervised classification problem. A boosting algorithm is proposed to solve the multiclass problem directly. The proposed multiclass approach uses a new multiclass loss function, which includes two terms. The first term is the cost of the multiclass margin and the second term is a regularization term on unlabeled data. The regularization term is used to minimize the inconsistency between the pair wise similarity and the classifier predictions. It assigns the soft labels weighted with the similarity between unlabeled and labeled examples. We then derive a boosting algorithm, named CD-MSSBoost, from the proposed loss function using coordinate gradient descent. The derived algorithm is further used for learning optimal similarity function for a given data. Our experiments on a number of UCI datasets show that CD-MSSBoost outperforms the state-of-the-art methods to multiclass semi-supervised learning.

In this paper we present an unsupervised learning approach to detect meaningful job traffic patterns in Grid log data. Manual anomaly detection on modern Grid environments is troublesome given their increasing complexity, the distributed, dynamic topology of the network and heterogeneity of the jobs being executed. The ability to automatically detect meaningful events with little or no human intervention is therefore desirable. We evaluate our method on a set of log data collected on the Grid. Since we lack a priori knowledge of patterns that can be detected and no labelled data is available, an unsupervised learning method is followed. We cluster jobs executed on the Grid using Affinity Propagation. We try to explain discovered clusters using representative features and we label them with the help of domain experts. Finally, as a further validation step, we construct a classifier for five of the detected clusters and we use it to predict the termination status of unseen jobs.

In this paper we present a new Multiclass semi-supervised learning algorithm that uses a base classifier in combination with a similarity function applied to all data to find a classifier that maximizes the margin and consistency over all data. A novel multiclass loss function is presented and used to derive the algorithm. We apply the algorithm to animal behavior recognition from accelerometer data. Animal-borne accelerometer data are collected from free-ranging animals and then labeled by a human expert. The resulting data are used to train a classifier. However, labeling is not easy from accelerometer data only and it is often not feasible to observe animals fitted with an accelerometer. All current approaches to this behavior recognition task use supervised or unsupervised learning. Since unlabeled data are easy to acquire and collect, a semi-supervised approach seems appropriate and reduces the human efforts for labeling. Experiments with accelerometer data collected from free-ranging gulls and benchmark UCI datasets show that the algorithm is effective and compares favorably with existing algorithms for multiclass semi-supervised learning.

We present an algorithm for multiclass Semi-Supervised learning which is learning from a limited amount of labeled data and plenty of unlabeled data. Existing semi-supervised algorithms use approaches such as one-versus-all to convert the multiclass problem to several binary classification problems which is not optimal. We propose a multiclass semi-supervised boosting algorithm that solves multiclass classification problems directly. The algorithm is based on a novel multiclass loss function consisting of the margin cost on labeled data and two regularization terms on labeled and unlabeled data. Experimental results on a number of UCI datasets show that the proposed algorithm performs better than the state-of-the-art boosting algorithms for multiclass semi-supervised learning.

Bridging the gap between low-level features and semantics is a problem commonly acknowledged in the Multimedia community. Event modeling can fill this gap by representing knowledge about the data at different level of abstraction. In this paper we present the Simple Event Model (SEM) and its application in a Maritime Safety and Security use case about Situational Awareness, where the data also come as low-level features (of ship trajectories). We show how we abstract over these low-level features, recognize simple behavior events using a Piecewise Linear Segmentation algorithm, and model the resulting events as instances of SEM. We aggregate web data from different sources, apply deduction rules, spatial proximity reasoning, and semantic web reasoning in SWI-Prolog to derive abstract events from the recognized simple events. The use case described in this paper comes from the Dutch Poseidon project.

In this paper we present a machine learning framework to analyze moving object trajectories from maritime vessels. Within this framework we perform the tasks of clustering, classification and outlier detection with vessel trajectory data. First, we apply a piecewise linear segmentation method to the trajectories to compress them. We adapt an existing technique to better retain stop and move information and show the better performance of our method with experimental results. Second, we use a similarity based approach to perform the clustering, classification and outlier detection tasks using kernel methods. We present experiments that investigate different alignment kernels and the effect of piecewise linear segmentation in the three different tasks. The experimental results show that compression does not negatively impact task performance and greatly reduces computation time for the alignment kernels. Finally, the alignment kernels allow for easy integration of geographical domain knowledge. In experiments we show that this added domain knowledge enhances performance in the clustering and classification tasks.

Recently Semi-Supervised learning algorithms such as co-training are used in many application domains. In co-training, two classifiers based on different views of data or on different learning algorithms are trained in parallel and then unlabeled data that are classified differently by the classifiers but for which one classifier has large confidence are labeled and used as training data for the other. In this paper, a new form of co-training, called Ensemble-Co-Training, is proposed that uses an ensemble of different learning algorithms. Based on a theorem by Angluin and Laird that produces an approximately correct identification with high probability for reliable examples, we propose a criterion for finding a subset of high-confidence predictions and error rate for a classifier in each iteration of the training process. Experiments show that the new method in almost all domains gives better results than the other methods.