Search

Electroencephalogram (EEG) signal has been widely applied in detecting human emotion. Individual differences limit the generalization in cross-subject classification since the release of emotion will be definitely different across persons even with the same emotion stimuli. Previous research utilizes domain adaptation to solve this problem in a leave-one-subject-out training that is learning common emotion-related features from many subjects and testing on a new subject, which requires abundant labeled data. This paper proposed a novel one-to-one domain adaptation method, the Task-specific Domain Adversarial Neural Network (T-DANN) which transfers knowledge from either one subject to predict on another subject or knowledge from one phase to predict on another phase within the same subject. Therefore, T-DANN is more flexible and requires much less data during training. T-DANN is an adversarial training method which adapts the conditional distribution between domains and adapts classification boundaries between classes simultaneously. Compared with data from different subjects, phase data from the same subject has much deeper correlation thus enhances the prediction of emotion in new phase. We evaluated our method on EEG emotion benchmark dataset SEED. The experiments showed that our proposed method outperformed other baseline methods in cross-subject adaptation. By cross-phase adaptation, our method achieved accuracy that approximately 3% lower than the state-of-the-art method but only used 1/14 labeled data, indicating the priority of our proposed method in real-time application.

We propose new methods for support vector machines using a tree architecture for multi-class classification. In each node of the tree, we select an appropriate binary classifier, using entropy and generalization error estimation, then group the examples into positive and negative classes based on the selected classifier, and train a new classifier for use in the classification phase. The proposed methods can work in time complexity between O(log2 N) and O(N), where N is the number of classes. We compare the performance of our methods with traditional techniques on the UCI machine learning repository using 10-fold cross-validation. The experimental results show that the methods are very useful for problems that need fast classification time or those with a large number of classes, since the proposed methods run much faster than the traditional techniques but still provide comparable accuracy.

Two of the biggest challenges in building models for detecting emotions from electroencephalography (EEG) devices are the relatively small amount of labeled samples and the strong variability of signal feature distributions between different subjects. In this study, we propose a context-generalized model that tackles the data constraints and subject variability simultaneously using a deep neural network architecture optimized for normally distributed subject-independent feature embeddings. Variational autoencoders (VAEs) at the input level allow the lower feature layers of the model to be trained on both labeled and unlabeled samples, maximizing the use of the limited data resources. Meanwhile, variational regularization encourages the model to learn Gaussian-distributed feature embeddings, resulting in robustness to small dataset imbalances. Subject-adversarial regularization applied to the bi-lateral features further enforces subject-independence on the final feature embedding used for emotion classification. The results from subject-independent performance experiments on the SEED and DEAP EEG-emotion datasets show that our model generalizes better across subjects than other state-of-the-art feature embeddings when paired with deep learning classifiers. Furthermore, qualitative analysis of the embedding space reveals that our proposed subject-invariant bi-lateral variational domain adversarial neural network (BiVDANN) architecture may improve the subject-independent performance by discovering normally distributed features.

Emotion recognition based on electroencephalograms has become an active research area. Yet, identifying emotions using only brainwaves is still very challenging, especially the subject-independent task. Numerous studies have tried to propose methods to recognize emotions, including machine learning techniques like convolutional neural network (CNN). Since CNN has shown its potential in generalization to unseen subjects, manipulating CNN hyperparameters like the window size and electrode order might be beneficial. To our knowledge, this is the first work that extensively observed the parameter selection effect on the CNN. The temporal information in distinct window sizes was found to significantly affect the recognition performance, and CNN was found to be more responsive to changing window sizes than the support vector machine. Classifying the arousal achieved the best performance with a window size of ten seconds, obtaining 56.85% accuracy and a Matthews correlation coefficient (MCC) of 0.1369. Valence recognition had the best performance with a window length of eight seconds at 73.34% accuracy and an MCC value of 0.4669. Spatial information from varying the electrode orders had a small effect on the classification. Overall, valence results had a much more superior performance than arousal results, which were, perhaps, influenced by features related to brain activity asymmetry between the left and right hemispheres.

Metric learning algorithms create distance metrics to capture the important relationships among features. These algorithms have been successful in low dimensional data. However, it is a challenge to handle high dimensional problems in real-world applications. This paper applies in high dimensional clustering problems a Reinforcement Learning (RL) based metric filtering approach in Evolutionary Distance Metric Learning (EDML), which relies on an evolutionary approach to optimize its distance metric. The proposed framework called High Dimensional Reinforced (HDR)-EDML has two novelties. The first is using a function approximation RL method called Least-Squares Policy Iteration (LSPI) to create a feature selection control strategy in high-dimensional input space to filter the metric. The second is adopting a two-way information exchange approach between LSPI and EDML Evolutionary Algorithm (EA) in metric learning. In the first way, LSPI will learn and send feedback that will affect the EDML metric creation. In the second way, the evolutionary feature prioritizing of EDML is utilized by LSPI in its learning process. These two novelties aim to reduce the number of selected features while maintaining the clustering performance. HDR-EDML is compared to conventional K-means, Information-Theoretic Metric Learning (ITML), normal EDML, and EDML with l 1 norm regularization for feature selection as baselines. Moreover, 3 different HDR-EDML approaches are examined to explore different ways of information exchange. Results show a significant decrease in the number of features while maintaining accuracy as well as reduced computational time and memory compared to another RL-based filtering method.

We proposed a method to extract causal relations of clusters from multi-dimensional event sequence data, along with another method to detect the changes of the extracted relations over time. The proposed Granger Cluster Sequence Mining (GCSM) algorithm identifies the pairs of spatial data clusters that have causality over time with each other. It extends the Cluster Sequence Mining algorithm, which utilized a statistical inference technique to identify occurrence relation, with a causality based on Granger causality. We also proposed a statistical model to infer the changes over time of each extracted causal relation. With experiments using synthetic data and semi-real data, we confirmed that the algorithm works correctly, and able to extract the embedded causal relations with high F-score and high accuracy of change points.

Brain-Computer Interface (BCI) is a communication tool between humans and systems using electroencephalography (EEG) to predict certain cognitive state aspects, such as attention or emotion. For brainwave recording, there are many types of acquisition devices created for different purposes. The wet system conducts the recording with electrode gel and can obtain high-quality brainwave signals, while the dry system expressly proposes the practical and ease of use. In this paper, we study a comparative study of wet and dry systems using two cognitive tasks: attention and music-emotion. The 3-back task is used as an assessment to measure attention and working memory in attention studies. Comparatively, the music-emotion experiments are used to predict the emotion according to the subject's questionnaires. Our analysis shows the similarities and differences between dry and wet electrodes by calculating the statistical values and frequency bands. Besides, we further study the relative characteristics by conducting the classification experiments. We proposed the end-to-end models of EEG classification, which are constructed by combining EEG-based feature extractors and classification networks. A deep convolution neural network (Deep ConvNet) and a shallow convolution neural network (Shallow ConvNet) were applied as the feature extractor of temporal and spatial filtering from raw EEG signals. The extracted feature is then forwardly conveyed to a long short-term memory ( LSTM ) to learn the dependencies of convolved features and classify attention states or emotional states. Additionally, transfer learning was utilized to improve the performance of the dry system by using transferred knowledge from the wet system. We applied the model not only on our dataset but also on the existing dataset to verify the model performance compared with the baseline techniques and the-state-of-the-art models. Using our proposed model, the result shows the significant differences between accuracy and chance level in attention classification (92.0%, S.D. 6.8%) and SEED dataset's emotion classification (75.3%, S.D. 9.3%).

The performance of human workers can be fluctuated due to changes in the cognitive state during sustained work. Though past researches have made human performance monitoring possible by utilizing physiological signals, little attention has been paid to the context of office works. This research proposes a multimodal approach to estimate office task performance. A transcription typing experiment was conducted to simulate the real working environment while typing speed and error rate represented as performance metrics. Physiological data collected during the experiment, together with conventional machine learning algorithms showed feasibility to accurately predict two levels (good/bad) of task performance. More importantly, a comprehensive comparison between choices of modality suggests that using data from particular sources could gain predictive performance comparable to the complete set of modalities.

This paper proposed a method to detect changes in causal relations over a multi-dimensional sequence of events. Cluster Sequence Mining algorithm was modified to extract causal relations in the form of g-patterns: a pair of clusters of events that have their occurrence time determined by Granger causality. This paper also proposed the pattern time signature, a probabilistic density function of the cluster sequence occurring at any given time. Synthetic data were used for validation. The result shows that the proposed algorithm can correctly identify the changes in causal relations even under noisy data.

In this paper, we propose two modified neural networks based on dual path multi-scale fusion networks (SFANet) and SegNet for accurate and efficient crowd counting. Inspired by SFANet, the first model, which is named M-SFANet, is attached with atrous spatial pyramid pooling (ASPP) and context-aware module (CAN). The encoder of M-SFANet is enhanced with ASPP containing parallel atrous convolutional layers with different sampling rates and hence able to extract multi-scale features of the target object and incorporate larger context. To further deal with scale variation throughout an input image, we leverage the CAN module which adaptively encodes the scales of the contextual information. The combination yields an effective model for counting in both dense and sparse crowd scenes. Based on the SFANet decoder structure, M-SFANet's decoder has dual paths, for density map and attention map generation. The second model is called M-SegNet, which is produced by replacing the bilinear upsampling in SFANet with max unpooling that is used in SegNet. This change provides a faster model while providing competitive counting performance. Designed for high-speed surveillance applications, M-SegNet has no additional multi-scale-aware module in order to not increase the complexity. Both models are encoder-decoder based architectures and are end-to-end trainable. We conduct extensive experiments on five crowd counting datasets and one vehicle counting dataset to show that these modifications yield algorithms that could improve state-of-the-art crowd counting methods. Codes are available at https://github.com/Pongpisit-Thanasutives/Variations-of-SFANet-for-Crowd-Counting., Accepted at ICPR 2020

This is an extension from a selected paper from JSAI2019. With the proliferation of wearable devices and the inflow of new health data, artificial intelligence is expected to revolutionize the field of wellness and health management by providing potential tools for analyzing harmful conditions like prolonged stress. Currently, one of the standard measurements used by medical practitioners to measure stress is heart rate variability (HRV), a set of numerical indices that reflect autonomic balance. However, recent advances in machine learning have shown that learned features tend to outperform hand-crafted features. In this work we propose a more expressive intermediate data representation based on Lomb-Scargle periodograms combined with the feature learning capabilities of deep learning. Using stress data from naturalistic work activities, we tested different shallow and deep learning architectures and show that significant improvements can be achieved compared to traditional HRV indices. Results show that models trained on our spectral-temporal representation significantly outperform models trained on traditional HRV indices for predicting naturalistic work stress.

Brain-Computer Interface (BCI) has been a hot topic and an emerging technology in this decade. It is a communication tool between humans and systems using electroencephalography (EEG) to predicts certain aspects of cognitive state, such as attention or emotion. There are many types of sensors created to acquire the brain signal for different purposes. For example, the wet electrode is to obtain good quality, and the dry electrode is to achieve a wearable purpose. Hence, this paper investigates a comparative study of wet and dry systems using two cognitive tasks: attention experiment and music-emotion experiment. In attention experiments, a 3-back task is used as an assessment to measure attention and working memory. Comparatively, the music-emotion experiments are conducted to predict the emotion according to the user’s questionnaires. The proposed model is constructed by combining a shallow convolutional neural network (Shallow ConvNet) and a long short-term memory (LSTM) network to perform the feature extraction and classification tasks, respectively. This study further proposes transfer learning that focuses on utilizing knowledge acquired for the wet system and applying it to the dry system.

With their rising availability and reliability, wearable devices such as electroencephalograms (EEG) could bring about advancements in personalized mental health monitoring. However, a major roadblock to the adoption of EEG for monitoring of mental health are concerns surrounding accuracy and the many sources of noise inherent to these types of sensitive devices. Combining noise-robust representations and flexible machine learning models could be the key to addressing these major issues. In this work, we use visual EEG representations to take advantage of the adaptive properties of deep learning models in order to model EEG signals during mindfulness meditation. Using a naturalistic dataset gathered from employees of a Japanese company, we attempt to identify and address some of the major issues inherent to acquisition and processing. Specifically, we use a topographic representation of EEG to enable efficient data utilization despite the presence of noisy and missing data. We also use deep model activations to guide the construction of a more practical architecture for this type of input data. Results indicate that shallow but wide architectures with more filters lead to better test performance than deeper models. Specifically, the shallower model realized significant performance gains of >5% compared to ResNet50 while also requiring fewer samples before reaching convergence. Finally, all models using the topographic representation showed good performance despite the inclusion of of samples with noisy and missing data channels.

This paper introduces a hybrid system called R-EDML, combining the sequential decision making of Reinforcement Learning (RL) with the evolutionary feature prioritizing process of Evolutionary Distance Metric Learning (EDML) in clustering aiming to optimize the input space by reducing the number of selected features while maintaining the clustering performance. In the proposed method, features represented by the elements of EDML distance transformation matrices are prioritized. Then a selection control strategy using Reinforcement Learning is learned. R-EDML was compared to normal EDML and conventional feature selection. Results show a decrease in the number of features, while maintaining a similar accuracy level.

People often learn their behavior from its outcome, e.g., success and failure. Also, in the real world, people are not alone and many interactions occur among people every day. To model such learning and interactions, let us consider reinforcement learning agents playing games. Many researchers have studied a lot of reinforcement learning algorithms to obtain good strategies in games. However, most of the algorithms are “suspicious”, i.e., focusing on how to escape from being exploited by greedy opponents. Therefore, it takes long time to establish cooperation among such agents. On the other hand, if the agents are “innocent”, i.e., prone to trust others, they establish cooperation easily but are exploited by acquisitive opponents. In this work, we propose an algorithm that uses two complementary, “innocent” and “suspicious” algorithms in the early and the late stage, respectively. The algorithm allows the agent to cooperate with good associates quickly as well as treat greedy opponents well. The experiments in ten games showed us that the proposed algorithm successfully learned good strategies quickly in nine games.

A good sleep is important for a healthy life. Recently, several consumer sleep devices have emerged on the market claiming that they can provide personal sleep monitoring; however, many of them require additional hardware or there is a lack of scientific evidence regarding their reliability. In this paper we proposed a novel method to assess the sleep quality through sound events recorded in the bedroom. We used subjective sleep quality as training label, combined several machine learning approaches including kernelized self organizing map, hierarchical clustering and hidden Markov model, obtained the models to indicate the sleep pattern of specific quality level. The proposed method is different from traditional sleep stage based method, provides a new aspect of sleep monitoring that sound events are directly correlated with the sleep of a person.

Many research studies on distance metric learning (DML) reiterate that the definition of distance between two data points substantially affects clustering tasks. Recently, variety of DML methods have been proposed to improve the accuracy of clustering by learning a distance metric; however, most of them only perform a linear transformation, which yields insignificant to non-linear separable data. This study proposes a DML method which provides an integration of kernelization technique with Mahalanobis-based DML. Thus, non-linear transformation of the distance metric can be performed. Moreover, a cluster validity index is optimized by an evolutionary algorithm. The empirical results on semi-supervised clustering suggest the promising result on both synthetic and real-world data set.

Students engage in many learning activities outside of class but, it is not easy for them to learn on their own because they also need to identify what activities to perform, decide how long to engage in them, evaluate their progress, shift to other activities if needed and avoid distractions aside from others. This research designed and implemented a learning support tool called Sidekick, which used a retrospective approach to help students analyze and evaluate their own behavior so they can adjust it accordingly. The results showed that students benefitted from understanding their behavior more. It also showed how students' learning behavior changed over time and the differences in the type and amount of change between learning sessions according to students' level of autonomy. Less autonomous students seemed to improve less compared to highly autonomous students however, the system was able to encourage them to recall and self-evaluate which they might not have done without the system.

We propose a method to discover sleep patterns via clustering of sound events recorded during sleep. The proposed method extends the conventional self-organizing map algorithm by kernelization and sequence-based technologies to obtain a fine-grained map that visualizes the distribution and changes of sleep-related events. We introduced features widely applied in sound processing and popular kernel functions to the proposed method to evaluate and compare performance. The proposed method provides a new aspect of sleep monitoring because the results demonstrate that sound events can be directly correlated to an individual's sleep patterns. In addition, by visualizing the transition of cluster dynamics, sleep-related sound events were found to relate to the various stages of sleep. Therefore, these results empirically warrant future study into the assessment of personal sleep quality using sound data.

Adopting a motif, that is the most basic component of a music, has been proposed for generating a chord progression adapted to personal sensibility in consideration of the entire music structure. Personal sensibility models for the motif and chord progression are induced, and chord progression is generated based on the models using an evolutionary computation. Symbiotic evolution is an evolutionary computation algorithm that is characterized by maintaining a partial solution population and a whole solution population. A whole solution is a combination of some partial solutions. As a musical piece can be considered as a combination of motifs, symbiotic evolution is appropriate for the tasks of generating a chord progression. In this paper, we propose how to design populations in symbiotic evolution to generate chord progression in consideration of the entire music structure. Our experimental results show that musical pieces having the target structure are composed.

Automated mental health analysis of stress could lead towards diagnosis tools that can be used in environments such as clinics, schools and corporations. However, attempts at building general models are often limited by the subjectivity of physiological stress responses. This work aims to discover the effects of combining data from physiological signals and psychological context from work activities when building a machine-learned model of mental stress. A software application was built to guide subjects through a monitoring process which allowed pre and post-assessment of psychological context through various stress-related annotation modules including the Cohen Stress Scale and the COPE inventory. Meanwhile, wearable sensors tracked physiological data in the form of heart beats. Tests were performed on this data by building supervised and unsupervised machine-learned models. Results show a general increase in classification performance when psychological context data is integrated into the models. Furthermore, models present similar performance using either questionnaire answers or coping profile scores.

In our society, people engage in a variety of interactions. To analyze such interactions, we consider these interactions as a game and people as agents equipped with reinforcement learning algorithms. Reinforcement learning algorithms are widely studied with a goal of identifying strategies of gaining large payoffs in games; however, existing algorithms learn slowly because they require a large number of interactions. In this work, we constructed an algorithm that both learns quickly and maximizes payoffs in various repeated games. Our proposed algorithm combines two different algorithms that are used in the early and later stages of our algorithm. We conducted experiments in which our proposed agents played ten kinds of games in self-play and with other agents. Results showed that our proposed algorithm learned more quickly than existing algorithms and gained sufficiently large payoffs in nine games.

This paper proposed an efficient generalized clustering method which derives from the k-means algorithm for achieving k-anonymization with good data quality and minimum information loss. We defined the distance function for the three major attribute types: numerical type, categorical type, and structural type. Then we proceeded the method in two stages: preprocessing stage and postprocessing stage. The preprocessing stage is to partitions all records into[n/k] groups, and then add the records that are naturally similar to each other into every group. The postprocessing stage is to add each remaining record into a cluster with respect to which the increment of the information loss is minimal. We experimentally compared our method with other two clustering-based k-anonymization methods. The experiment showed that our method outperforms their method and also ensures the anonymization of data.

The fuel cell is regarded as a highly efficient, low-pollution power generation system. In particular, Solid Oxide Fuel Cell (SOFC) has a high generation efficiency. However, a crucial issue in putting SOFC to practical use is the establishment of a technique for evaluating the deterioration. We previously developed a technique by which to measure the mechanical damage of SOFC using the Acoustic Emission (AE) method. In the present paper, we applied the kernel Self-Organizing Map (SOM), which is an extended neural network model, to produce a cluster map reflecting the similarity of AE events. The obtained map visualized the change in occurrence patterns of similar AE events, revealing four phases of damage progress. The methodology of the present study provides a common foundation for a comprehensive damage evaluation system and a damage monitoring system.

This paper presents a technology for over the air (OTA) updating that allows software updates in consumer devices such as mobile phones and car-navigation systems that are connected to networks, and it adopts NAND flash memories and demand paging technologies. Software updates for these kinds of consumer devices are carried out using the binary difference that minimizes the amount of update data, software construction technologies that minimize the rewriting of flash memories, and fast compression technologies that enables rapid rewriting. This paper focuses on byte pair encoding (BPE), which is one of the most efficient methods of compressing program code in consumer devices that adopts NAND flash memories and demand paging technologies. We describe a fast method of compressing BPE and evaluating it.

This research investigates the use of emotion data derived from analyzing change in activity in the autonomic nervous system (ANS) as revealed by brainwave production to support the creative music compositional intelligence of an adaptive interface. A relational model of the influence of musical events on the listener's affect is first induced using inductive logic programming paradigms with the emotion data and musical score features as inputs of the induction task. The components of composition such as interval and scale, instrumentation, chord progression and melody are automatically combined using genetic algorithm and melodic transformation heuristics that depend on the predictive knowledge and character of the induced model. Out of the four targeted basic emotional states, namely, stress, joy, sadness, and relaxation, the empirical results reported here show that the system is able to successfully compose tunes that convey one of these affective states.

This paper discusses a predictive modeling framework actualized in a learning agent that uses logged tutorial interactions to discover predictive characteristics of students. The agent automatically forms cluster models that are described in terms of student-system interaction attributes, i.e., in terms of the student's knowledge state and behaviour and system's tutoring actions. The agent utilizes the knowledge of its various clusters together with a weighting scheme to learn predictive models of high-level student information, specifically, the time it will take the student to respond to a problem and whether the response is correct, that can be utilized to support individualized adaptation. We investigated utilizing the Self-Organizing Map and AutoClass as clustering algorithms and the naive Bayesian classifier and single layer neural network as weighting algorithms. Empirical results show that by utilizing cluster knowledge the agent's predictions are acceptably strong for response time and accurate at the average for response correctness. Further investigation is needed to validate the scalability of the framework given other datasets and possibly migrate to other approaches that can obtain more meaningful cluster models, detect richer attribute relations, and provide better approximations to further improve prediction of response behaviour for a more informed pedagogical decision-making by the system.

We propose a concept drift detection method utilizing statistical change detection in which a drift detection method and the Page-Hinkley test are employed. Our method enables users to annotate clustering results without constructing a model of drift detection for every input. In our experiments using synthetic data, we evaluated our proposed method on the basis of detection delay and false detection, also revealed relations between the degree of drift and parameters of the method.

To gain a better understanding of how neural ensembles communicate and process information, neural decoding algorithms are used to extract information encoded in their spiking activity. Bayesian decoding is one of the most used neural population decoding approaches to extract information from the ensemble spiking activity of rat hippocampal neurons. Recently it has been shown how Bayesian decoding can be implemented without the intermediate step of sorting spike waveforms into groups of single units. Here we extend the approach in order to make it suitable for online encoding/decoding scenarios that require real-time decoding such as brain-machine interfaces. We propose an online algorithm for the Bayesian decoding that reduces the time required for decoding neural populations, resulting in a real-time capable decoding framework. More specifically, we improve the speed of the probability density estimation step, which is the most essential and the most expensive computation of the spike-sorting-less decoding process, by developing a kernel density compression algorithm. In contrary to existing online kernel compression techniques, rather than optimizing for the minimum estimation error caused by kernels compression, the proposed method compresses kernels on the basis of the distance between the merging component and its most similar neighbor. Thus, without costly optimization, the proposed method has very low compression latency with a small and manageable estimation error. In addition, the proposed bandwidth matching method for Gaussian kernels merging has an interesting mathematical property whereby optimization in the estimation of the probability density function can be performed efficiently, resulting in a faster decoding speed. We successfully applied the proposed kernel compression algorithm to the Bayesian decoding framework to reconstruct positions of a freely moving rat from hippocampal unsorted spikes, with significant improvements in the decoding speed and acceptable decoding error., National Institute of Mental Health (U.S.) (Grant MH-061976), United States. Office of Naval Research (MURI N00014-10-1-0936 Grant), Japan Society for the Promotion of Science (Strategic Young Researcher Overseas Visits Program for Accelerating Brain Circulation)

In a real-time video game, AI-controlled players, called agents, are still inferior to skilled human players on equal footing. In this work, we aim to construct a strong agent enough to fight with skilled human players in a real-time fighting video game. First we investigate the relation between perception speed and performance. From a simulation using two agents one of which has delayed perception, we know that perception speed is a critical factor in performance. Moreover, it means that it is effective to predict the opponent's behavior to enhance the agent. Therefore, we construct an agent that predicts its opponent's position and action in a fighting video game. The agent uses linear extrapolation to predict the position and the k-nearest neighbor method to predict the action. Comparing agents with and without the prediction ability, we see that the predicting agent mostly obtained higher scores than the non-predicting one in fighting with six contestants of a previous competition of the game.

Evolutionary distance metric learning (EDML) is an efficient technique for solving clustering problems with some background knowledge. However, EDML has never been applied to real world applications. Thus, we demonstrate EDML for cluster analysis and visualization of two applications, i.e., a face recognition image dataset and a literature dataset. In the facial image clustering, we demonstrate improvement of the cluster validity index and also analyze the distributions of classes (ages) visualized by a self-organizing map and a K-means clustering with K-nearest neighbor centroids graph. For the literature dataset, we have analyzed the topics (i.e., a cluster of articles) that are the most likely to win the best paper award. Application of EDML to these datasets yielded qualitatively promising visualization results that demonstrate the practicability and effectiveness of EDML.

To have an instructional plan guide the learning process is significant to various teaching styles and an important task in an ITS. Though various approaches have been used to tackle this task, the compelling need is for an ITS to improve on its own the plans established in a dynamic way. We hypothesize that the use of knowledge derived from student categories can significantly support the improvement of plans on the part of the ITS. This means that category knowledge can become effectors of effective plans. We have conceived a Category-based Self-improving Planning Module (CSPM) for an ITS tutor agent that utilizes the knowledge learned from learner categories to support self-improvement. The learning framework of CSPM employs unsupervised machine learning and knowledge acquisition heuristics for learning from experience. We have experimented on the feasibility of CSPM using recorded teaching scenarios.

The predictive technique proposed in this project was initially designed for an indoor smart environment wherein intrusive tracking techniques, such as cameras, mobile phones, and GPS tracking systems, could not be appropriately utilized. Instead, we installed simple motion detection sensors in various areas of the experimental space and observed movements. However, the data collected cannot provide as much information about human mobility as data from a GPS or mobile phone. In this paper, we conducted an exhaustive analysis to determine the predictability of future mobility of people using only this limited dataset. Furthermore, we proposed an aperiodic and periodic predictive technique for long-term human mobility prediction that works well with our limited dataset. The evaluation of the dataset collected of the movement and daily activity in the smart space for three months shows that our model is able to predict future mobility and activities of participants in the smart environment setting with high accuracy – even for a month in advance.

Software product line engineering (SPLE) is a methodology for developing a diversity of software products and software intensive systems at lower costs, in shorter time, and with higher quality. SPLE is widely known to develop industrial products such as embedded system. However, for applying SPLE to develop Web applications which have to provide transactional logics for products such as financial products, the biggest issue is the lack of design method and application architecture to execute concurrently for a whole product line on the same runtime. Therefore development cost still tends to increase with frequent changes of the business requirements. We proposed two noble approaches to apply SPLE each for design method and application architecture. Firstly, for design method, we redefined the design scope of product as Product Domain Model, and designed a UML based meta-model which adds the notations of variability. Secondly, for application architecture, we adopted dependency injection technology to execute transaction logics for product line. We also defined a unit of logics for product line as Instance Product. To generate a plenty of resources regarding Instance Product, we created a generator which inputs Product Domain Model. In this paper, we introduce our approaches and evaluation by the pilot development of banking products. The results show that with our approaches, the issue of Web application development can be solved effectively, as well as the additional issue of maintainability.

Knowledge acquisition with machine learning techniques is a fundamental requirement for knowledge discovery from databases and data mining systems. Two techniques in particular -- inductive learning and theory revision -- have been used toward this end. A method that combines both approaches to effectively acquire theories (regularity) from a set of training examples is presented. Inductive learning is used to acquire new regularity from the training examples; and theory revision is used to improve an initial theory. In addition, a theory preference criterion that is a combination of the MDL-based heuristic and the Laplace estimate has been successfully employed in the selection of the promising theory. The resulting algorithm developed by integrating inductive learning and theory revision and using the criterion has the ability to deal with complex problems, obtaining useful theories in terms of its predictive accuracy.

Students often face difficulty in self-directed learning scenarios (e.g., studying, research) because they need to control many aspects of the learning session. They need to decide what to learn, how long to perform a learning task, when to shift to a different learning task and manage distractions apart from others. We observed from our previous research that self-reflection and self-evaluation helped students manage their own learning. However, majority of the students only evaluated one or two major aspects of the learning session that they think needed to be changed or improved (e.g., need to spend less time in nonlearning related activities, need to focus on only one learning task at a time). If students would look further into their learning session, they would discover more behaviors that also need to be re-evaluated. In this paper we discussed reinforcement learning-based methods for discovering good learning behavior which can be used by future systems to suggest to students possible ways to improve their behavior.

Searching for understandable health information on the Internet remains difficult for most consumers. Every consumer has different health topic familiarity. This diversity may cause misunderstanding because the information presented during health information searches may not fit the consumer’s understanding. This study aimed to develop health topic familiarity prediction models based on the consumer’s searching behavior, how the consumers formulate the query and how they interact with the search results. The experimental results show that Naive Bayes and Sequential Minimal Optimization classifiers achieved high accuracy on the combination of query formulation and search result interaction feature sets in predicting consumer’s health topic familiarity. This finding suggests that health topic familiarity identification based on the query formulation and the search result interaction is feasible and effective.

Automatic systems that compose music adapted to personal sensibility have been proposed. These systems induce a personal sensibility model by using a listener’s emotional impressions of music and compose music based on the model using evolutionary computation algorithms. The objective of this study was to compose music consisting of four parts: introduction, development, turn, and conclusion. We propose a method for generating chord progressions based on a symbiotic evolution algorithm that is characterized by parallel evolutions of both partial and whole solutions. Our experimental results show that the proposed method effectively composes musical pieces having the target structure.

Over the past decade, several inductive logic programming (ILP) systems have been developed. However, the normal ILP system does not have enough power to induce logic programs in some domains. Therefore, various new ILP systems, such as nonmonotonic ILP, that apply new learning tasks, have been proposed. In the present paper, a new learning task, called relational learning from nondichotomizable examples, is proposed for relational domains. The scheme of this task is similar to that of normal ILP, with the exception of the training examples. Normally, conventional ILP uses positive and negative examples in the training process. However, the examples of the proposed learning task are not strictly positive or negative: the training examples have a continuous scale. This new learning task is defined and a new method for learning this task is proposed. A new ILP system, SYNGIP (for SYNthesized system using Genetic programming and Inductive logic Programming), is developed based on this method and the performance of SYNGIP is compared experimentally to that of the conventional ILP method for the task of human feeling acquisition. © 2001 Scripta Technica, Syst Comp Jpn, 32(14): 34–40, 2001

Detecting and diagnosing errors in novice behavior is an important student modeling task. In this paper, we describe MEDD, an unsupervised incremental multistrategy system for the discovery of classes of errors from, and their detection in, novice programs. Experimental results show that MEDD can effectively detect and discover misconceptions and other knowledge-level errors that underlie novice Prolog programs, even when multiple errors are enmeshed together in a single program, and when the programs are presented to MEDD in a different order.