We propose an Information bottleneck (IB) for Goal representation learning (InfoGoal), a self-supervised method for generalizable goal-conditioned reinforcement learning (RL). Goal-conditioned RL learns a policy from reward signals to predict actions for reaching desired goals. However, the policy would overfit the task-irrelevant information contained in the goal and may be falsely or ineffectively generalized to reach other goals. A goal representation containing sufficient task-relevant information and minimum task-irrelevant information is guaranteed to reduce generalization errors. However, in goal-conditioned RL, it is difficult to balance the tradeoff between task-relevant information and task-irrelevant information because of the sparse and delayed learning signals, i.e., reward signals, and the inevitable task-relevant information sacrifice caused by information compression. Our InfoGoal learns a minimum and sufficient goal representation with dense and immediate self-supervised learning signals. Meanwhile, InfoGoal adaptively adjusts the weight of information minimization to achieve maximum information compression with a reasonable sacrifice of task-relevant information. Consequently, InfoGoal enables policy to generate a targeted trajectory toward states where the desired goal can be found with high probability and broadly explores those states. We conduct experiments on both simulated and real-world tasks, and our method significantly outperforms baseline methods in terms of policy optimality and the success rate of reaching unseen test goals. Video demos are available at infogoal.github.io.

We propose a word sense aware topic model for document classification based on soft orthogonalized topics. An essential problem for this task is to capture word senses related to classes, i.e., class-specific word senses. Traditional models mainly introduce semantic information of knowledge libraries for word sense discovery. However, this information may not align with the classification targets, because these targets are often subjective and task-related. We aim to model the class-specific word senses in topic space. The challenge is to optimize the class separability of the senses, i.e., obtaining sense vectors with (a) high intra-class and (b) low inter-class similarities. Most existing models predefine specific topics for each class to specify the class-specific sense vectors. We call them hard orthogonalization based methods. These methods can hardly achieve both (a) and (b) since they assume the conditional independence of topics to classes and inevitably lose topic information. To this problem, we propose soft orthogonalization for topics. Specifically, we reserve all the topics and introduce a group of class-specific weights for each word to handle the importance of topic dimensions to class separability. Besides, we detect and use highly class-specific words in each document to guide sense estimation. Our experiments on two standard datasets show that our proposal outperforms other state-of-the-art models in terms of accuracy of sense estimation, document classification, and topic modeling. In addition, our joint learning experiments with the pre-trained language model BERT showcased the best complementarity of our model in most cases compared to other topic models.

We propose Predictive Information bottleneck for Goal representation learning (PI-Goal), a self-supervised method for sample-efficient goal-conditioned reinforcement learning (RL). Goal-conditioned RL learns to reach commanded goals with reward signals. A goal could be given in a noisy or abstract form, and thus jeopardizes sample efficiency. Previous methods usually assume that the agent can map a state to an achievable goal. In this work, we consider a setting in which the goal space is unknown to the agent and the agent cannot recognize a goal in a specific state (referred to as a goal state) until the goal is commanded. Our PI-Goal learns a goal representation which contains only the predictive information of a goal state, i.e., the mutual information between a current state and a future state, and guarantees the optimality of the learned policy. Experimental results show that PI-Goal consistently outperforms the baseline methods in tasks with unknown goal spaces, e.g., object manipulation, object search, and embodied question answering.

This paper proposes 18 types of statistical data explanations and three kinds of procedures to investigate credibility in unethical and biased explanations due to exploitation of the 10 instincts proposed by Rosling et al. The explanation "women have lower math scores than men'' accompanied with the averages and the distributions of their scores is an example of such an explanation, as it exploits the gap instinct, i.e., our tendency to divide all kinds of things into two distinct and often conflicting groups. It becomes much less credible if we replace the word "math'' with "English'', even if we keep the data as they are, as the exploitation seems to fail. Our judging procedures are based on phrase embedding and carefully designed comparisons to judge the credibility. The results of our experiments comparing the 18 types with their variants show promising results and clues for further developments.

We propose a hybrid context based topic model for word sense disambiguation in document representation. Document representation is an essential part of various document based tasks, and word sense disambiguation is to capture the distinctions of word senses in the representation. Traditional methods mainly rely on knowledge libraries for data enrichment; however, semantics division for a word may vary from different domain-specific datasets. We aim to discover more particular word semantic differences for each input dataset and handle the disambiguation problem without data enrichment. The challenge for this disambiguation is to (1) divide various senses for each polysemous word while (2) preserve the differences between synonyms. Most of the existing models are either based on separate context clusters or integrating an auxiliary module to specify word senses. They can hardly achieve both (1) and (2) since different senses of a word are assumed to be independent and their intrinsic relationships are ignored. To solve this problem, we estimate a word sense by both the context in which it occurs and the contexts of its other occurrences. Besides, we introduce the “Bag-of-Senses” (BoS) assumption: a document is a multiset of word senses, and the senses are generated instead of the words. Our experiments on three standard datasets show that our proposal outperforms other state-of-the-art methods in terms of accuracy of word sense estimation, topic modeling, and document classification.

Modern Convolutional Neural Networks (CNNs) have been improving the accuracy of person re-identification (re-id) using a large number of training samples. Such a re-id system suffers from a lack of training samples for deployment to practical security applications. To address this problem, we focus on the approach that transfers features of a CNN pre-trained on a large-scale person re-id dataset to a small-scale dataset. Most of the existing CNN feature transfer methods use the features of fully connected layers that entangle locally pooled features of different spatial locations on an image. Unfortunately, due to the difference of view angles and the bias of walking directions of the persons, each camera view in a dataset has a unique spatial property in the person image, which reduces the generality of the local pooling for different cameras/datasets. To account for the camera- and dataset-specific spatial bias, we propose a method to learn camera and dataset-specific position weight maps for discriminative local pooling of convolutional features. Our experiments on four public datasets confirm the effectiveness of the proposed feature transfer with a small number of training samples in the target datasets.

Affinity measure in object tracking outputs a similarity or distance score for given detections. As an affinity measure is typically imperfect, it generally has an uncertain region in which regarding two groups of detections as the same object or different objects based on the score can be wrong. How to reduce the uncertain region is a major challenge for most similarity-based tracking methods. Early mistakes often result in distribution drifts for tracked objects and this is another major issue for object tracking. In this paper, we propose a new offline tracking method called agglomerative hierarchical clustering with ensemble of tracking experts (AHC_ETE), to tackle the uncertain region and early mistake issues. We conduct tracking from certain to uncertain to reduce early mistakes. Meanwhile, we ensemble multiple tracking experts to reduce the uncertain region as the final uncertain region is the intersection of those of all tracking experts. Experiments on the MOT15 and MOT16 datasets demonstrated the effectiveness of our method. The code is publicly available at https://github.com/cyoukaikai/ahc_ete.

Describing the color and textural information of a person image is one of the most crucial aspects of person re-identification (re-id). Although a covariance descriptor has been successfully applied to person re-id, it loses the local structure of a region and mean information of pixel features, both of which tend to be the major discriminative information for person re-id. In this paper, we present novel meta-descriptors based on a hierarchical Gaussian distribution of pixel features, in which both mean and covariance information are included in patch and region level descriptions. More specifically, the region is modeled as a set of multiple Gaussian distributions, each of which represents the appearance of a local patch. The characteristics of the set of Gaussian distributions are again described by another Gaussian distribution. Because the space of Gaussian distribution is not a linear space, we embed the parameters of the distribution into a point of Symmetric Positive Definite (SPD) matrix manifold in both steps. We show, for the first time, that normalizing the scale of the SPD matrix enhances the hierarchical feature representation on this manifold. Additionally, we develop feature norm normalization methods with the ability to alleviate the biased trends that exist on the SPD matrix descriptors. The experimental results conducted on five public datasets indicate the effectiveness of the proposed descriptors and the two types of normalizations.

We propose a GAN-based one-shot generation method on a fine-grained category, which represents a subclass of a category, typically with diverse examples. One-shot generation refers to a task of taking an image which belongs to a class not used in the training phase and then generating a set of new images belonging to the same class. Generative Adversarial Network (GAN), which represents a type of deep neural networks with competing generator and discriminator, has proven to be useful in generating realistic images. Especially DAGAN, which maps the input image to a low-dimensional space via an encoder and then back to the example space via a decoder, has been quite effective with datasets such as handwritten character datasets. However, when the class corresponds to a fine-grained category, DAGAN occasionally generates images which are regarded as belonging to other classes due to the rich variety of the examples in the class and the low dissimilarities of the examples among the classes. For example, it accidentally generates facial images of different persons when the class corresponds to a specific person. To circumvent this problem, we introduce a metric learning with a triplet loss to the bottleneck layer of DAGAN to penalize such a generation. We also extend the optimization algorithm of DAGAN to an alternating procedure for two types of loss functions. Our proposed method outperforms DAGAN in the GAN-test task for VGG-Face dataset and CompCars dataset by 5.6% and 4.8% in accuracy, respectively. We also conducted experiments for the data augmentation task and observed 4.5% higher accuracy for our proposed method over DAGAN for VGG-Face dataset.

In this paper, we propose a new heuristic training procedure to help a deep neural network (DNN) repeatedly escape from a local minimum and move to a better local minimum. Our method repeats the following processes multiple times: randomly reinitializing the weights of the last layer of a converged DNN while preserving the weights of the remaining layers, and then conducting a new round of training. The motivation is to make the training in the new round learn better parameters based on the “good” initial parameters learned in the previous round. With multiple randomly initialized DNNs trained based on our training procedure, we can obtain an ensemble of DNNs that are more accurate and diverse compared with the normal training procedure. We call this framework “retraining”. Experiments on eight DNN models show that our method generally outperforms the state-of-the-art ensemble learning methods. We also provide two variants of the retraining framework to tackle the tasks of ensemble learning in which 1) DNNs exhibit very high training accuracies (e.g., ) and 2) DNNs are too computationally expensive to train.

In this paper, we propose MDNNiSF (Multimodal Deep Neural Network with image Sequence Features) for generating a sentence description of a given video clip. A recently proposed model, S2VT, uses a stack of two LSTMs to solve the problem and demonstrated high METEOR. However, experiments show that S2VT sometimes produces inaccurate sentences, which is quite natural due to the challenging nature of learning relationships between visual and textual contents. A possible reason is that the video caption data were still small for the purpose. We try to circumvent this flaw by integrating S2VT with NeuralTalk2, which is for image captioning and known to generate an accurate description due to its capability of learning alignments between text fragments to image fragments. Experiments using two video caption data, MSVD and MSRVTT, demonstrate the effectiveness of our MDNNiSF over S2VT. For example, MDNNiSF achieved METEOR 0.344, which is 21.5% higher than S2VT, with MSVD.

Describing the color and textural information of a person image is one of the most crucial aspects of person re-identification. In this paper, we present a novel descriptor based on a hierarchical distribution of pixel features. A hierarchical covariance descriptor has been successfully applied for image classification. However, the mean information of pixel features, which is absent in covariance, tends to be major discriminative information of person images. To solve this problem, we describe a local region in an image via hierarchical Gaussian distribution in which both means and covariances are included in their parameters. More specifically, we model the region as a set of multiple Gaussian distributions in which each Gaussian represents the appearance of a local patch. The characteristics of the set of Gaussians are again described by another Gaussian distribution. In both steps, unlike the hierarchical covariance descriptor, the proposed descriptor can model both the mean and the covariance information of pixel features properly. The results of experiments conducted on five databases indicate that the proposed descriptor exhibits remarkably high performance which outperforms the state-of-the-art descriptors for person re-identification.

This paper presents fine-tuned CNN features for person re-identification. Recently, features extracted from top layers of pre-trained Convolutional Neural Network (CNN) on a large annotated dataset, e.g., ImageNet, have been proven to be strong off-the-shelf descriptors for various recognition tasks. However, large disparity among the pre-trained task, i.e., ImageNet classification, and the target task, i.e., person image matching, limits performances of the CNN features for person re-identification. In this paper, we improve the CNN features by conducting a fine-tuning on a pedestrian attribute dataset. In addition to the classification loss for multiple pedestrian attribute labels, we propose new labels by combining different attribute labels and use them for an additional classification loss function. The combination attribute loss forces CNN to distinguish more person specific information, yielding more discriminative features. After extracting features from the learned CNN, we apply conventional metric learning on a target re-identification dataset for further increasing discriminative power. Experimental results on four challenging person re-identification datasets (VIPeR, CUHK, PRID450S and GRID) demonstrate the effectiveness of the proposed features.

We propose a novel probabilistic semi-supervised anomaly detection framework for multi-dimensional systems with high correlation among variables. Our method is able to identify both abnormal instances and abnormal variables of an instance.

One of the effective methodologies for time series classification is to identify informative subsequence patterns in time series and exploit them as discriminative features. Previous studies on this methodology have achieved promising results using a small number of individually selected patterns. However, there remain difficulties in finding a set of related patterns or patterns of a minor class, which can be critical in real-world applications. In this paper, we exploit the sparse learning technique for the support vector machine (SVM) to identify informative and exemplary patterns. We first present a representation of time series as a vector of distances to exemplary patterns. It allows a structural SVM to handle distance space data and function as the nearest neighbor classifier, the combination of which is known to be highly competitive in time series classification. We then extend the zero-norm approximation method for the structural SVM, which can eliminate non-essential patterns from the classification model. The resulting model makes predictions by a simple modified nearest neighbor rule, yet has a strong mathematical support for empirical risk minimization and feature selection. We conduct an empirical study on real-world behavior and sequential data to evaluate the effectiveness of the proposed method and graphically examine the exemplary patterns.

In this paper, we address a classification task under a time-sensitive setting, in which the amount of observation required to make a prediction is viewed as a practical cost. Such a setting is intrinsic in many systems where the potential reward of the action against the predicted event depends on the response time, e.g., surveillance/warning and diagnostic applications. Meanwhile, predictions are usually less reliable when based on fewer observations, i.e., there exists a trade-off between such temporal cost and the accuracy. We address the task as a classification of subsequences in a time series. The goal is to predict the occurrences of events from subsequent observations and to learn when to commit to the prediction considering the trade-off. We propose an ensemble of classifiers which respectively makes predictions based on subsequences of different lengths. The prediction of the ensemble is given by the earliest confident prediction among the individual classifiers. We propose a cutting-plane algorithm for jointly training an ensemble of linear classifiers considering their temporal dependence. We compare the proposed algorithm against conventional approaches over a collection of behavioral trajectory data.

This special issue focuses on Discovery Science (DS), which is a scientific discipline on any discovery process that is mainly approached by computer science. DS first started as a national project in Japan involving more than 100 researchers in 1998 and the project gave birth to a series of international conferences on DS, which have been held successfully every year since 1998. The objective of this special issue is to provide a leading forum for timely, in-depth presentation of recent advances in algorithms, theories and applications in the field of DS. Seven papers, ranging in a spectrum from basic theoretical research to solid application research, are included in this special issue.

Graph clustering, or community detection, is an important task of discovering the underlying structure in a network by clustering vertices in a graph into communities. In the past decades, non-overlapping methods such as normalized cuts and modularit

Behavior analysis using trajectory data presents a practical and interesting challenge for KDD. Conventional analyses address discriminative tasks of behaviors, e.g., classification and clustering typically using the subsequences extracted from the trajectory of an object as a numerical feature representation. In this paper, we explore further to identify the difference in the high-level semantics of behaviors such as roles and address the task in a cross-domain learning approach. The trajectory, from which the features are sampled, is intuitively viewed as a domain, and we assume that its intrinsic structure is characterized by the underlying role associated with the tracked object. We propose a novel hybrid method of spectral clustering and density approximation for comparing clustering structures of two independently sampled trajectory data and identifying patterns of behaviors unique to a role. We present empirical evaluations of the proposed method in two practical settings using real-world robotic trajectories.

This paper addresses an application of anomaly detection from subsequences of time series (STS) to autonomous robots' behaviors. An important aspect of mining sequential data is selecting the temporal parameters, such as the subsequence length and the degree of smoothing. For example in the task at hand, the patterns of the robot's velocity, which is one of its fundamental features, vary significantly subject to the interval for measuring the displacement. Selecting the time scale and resolution is difficult in unsupervised settings, and is often more critical than the choice of the method. In this paper, we propose an ensemble framework for aggregating anomaly detection from different perspectives, i.e., settings of user-defined, temporal parameters. In the proposed framework, each behavior is labeled whether it is an anomaly in multiple settings. The set of labels are used as meta-features of the respective behaviors. Cluster analysis in a meta-feature space partitions anomalous behaviors pertained to a specific range of parameters. The framework also includes a scalable implementation of the instance-based anomaly detection. We evaluate the proposed framework by ROC analysis, in comparison to conventional ensemble methods for anomaly detection.

This paper proposes an Extended Minimum Description Length Principle (EMDLP) for feature-based inductive transfer learning, in which both the source and the target data sets contain class labels and relevant features are transferred from the source domain to the target one. Despite numerous works on this topic, few of them have a solid theoretical framework and are parameter-free. Our EMDLP overcomes these flaws and allows us to evaluate the inferiority of the results of transfer learning with the add-sum of the code lengths of five components: the corresponding two hypotheses, the two data sets with the help of the hypotheses, and the set of the transferred features. We design a code book to build the connections between the source and the target tasks. Extensive experiments using both real and artificial data sets show that EMDLP is robust against noise and performs better on the classification accuracy than the state-of-the-art methods.

Dimensionality reduction has been considered as one of the most significant tools for data analysis. In general, supervised information is helpful for dimensionality reduction. However, in typical real applications, supervised information in multiple source tasks may be available, while the data of the target task are unlabeled. An interesting problem of how to guide the dimensionality reduction for the unlabeled target data by exploiting useful knowledge, such as label information, from multiple source tasks arises in such a scenario. In this paper, we propose a new method for dimensionality reduction in the transfer learning setting. Unlike traditional paradigms where the useful knowledge from multiple source tasks is transferred through distance metric, our proposal firstly converts the dimensionality reduction problem into integral regression problems in parallel. Gaussian process is then employed to learn the underlying relationship between the original data and the reduced data. Such a relationship can be appropriately transferred to the target task by exploiting the prediction ability of the Gaussian process model and inventing different kinds of regularizers. Extensive experiments on both synthetic and real data sets show the effectiveness of our method.

The central challenge in temporal data analysis is to obtain knowledge about its underlying dynamics. In this paper, we address the observation of noisy, stochastic processes and attempt to detect temporal segments that are related to inconsistencies and irregularities in its dynamics. Many conventional anomaly detection approaches detect anomalies based on the distance between patterns, and often provide only limited intuition about the generative process of the anomalies. Meanwhile, model-based approaches have difficulty in identifying a small, clustered set of anomalies.

In real-world data mining applications, one often has access to multiple datasets that are relevant to the task at hand. However, learning from such datasets can be difficult as they are often drawn from different domains, i.e., not identically distributed or differ in class or feature sets. In this paper, we consider the problem of learning the class structures %, unique and shared, of related domains in an unsupervised manner. Its setting generalizes that of information filtering and novelty detection applications which addresses both known and unknown classes. We propose a co-clustering framework for estimating and adapting the class structures of two related domains, {enabling the analyses of shared and unique classes.} We define an objective function using interaction information to take account of the divergence between the corresponding clusters of respective domains. We present an iterative algorithm which alternates object and feature clustering and converges to a local minimum of the objective function. We present empirical results using text benchmarks, comparing the proposed algorithm and combinations of conventional approaches in problems of partitioning documents and detecting unknown topics.

This paper discusses a structure of multi-objective optimization problems, which cause deception for conventional multi-objective genetic algorithms (MOGAs). Further, we propose a distributed multi-objective genetic algorithm (DMOGA), which employs a multiple subpopulation implementation and a replacement scheme based on the information theoretic entropy, to improve the performance of MOGA in such deceptive problems. Several studies have reported that the conventional MOGAs' have difficulties in generating marginal segments of the Pareto front in a combinatorial optimization problems, though structural causes of their behaviors have not yet been thoroughly studied. Our analysis of the conventional MOGAs' behaviors in two test deceptive problems suggests that the use of the local density in the selection causes an implicit bias which results in a premature convergence. DMOGA is a distributed implementation of MOGA, which emphasizes the diversity of the subpopulations by the entropy of the objective functions. This approach alleviates the premature convergence and enables MOGA to effectively generate Pareto fronts for complex objective functions. In a set of simulated experiments, the proposed method generated more comprehensive Pareto fronts than the conventional MOGAs, i.e., NSGA-II and SPEA2 in the deceptive test functions, and also achieved comparable performance in the standard multi-objective benchmarks.

Detection of rare and exceptional occurrences in large- scale databases have become an important practice in the field of knowledge discovery and information retrieval. Many databases include large amount of noise or irrelevant data, whose distribution often overlaps with the subsets of exceptional data containing useful knowledge. This paper addresses the problem of finding a small subset of "minority" data whose distribution overlaps with, but are exceptional to or inconsistent with that of the majority of the database. In such a case, conventional distance-based or density-based approaches in Outlier Detection are ineffective due to their dependence on the structure of the majority or the prerequisite of critical parameters. We formalize the task as an estimation of a model of the minority subset which provides a simple description of the subset and yet maintains divergence from that of the majority. This estimation is formalized as a minimization problem using an information theoretic framework of Rate Distortion theory. We further introduce conditions of the majority to derive an objective function which factorizes the property of the minority and dependence to the structure of the majority. The proposed method shows improvements from conventional approaches in artificial data and a promising result in document retrieval problem.

We proposed a selection scheme called sample-based crowding, which is aimed to improve the performance of genetic algorithms for multimodal optimization in ill-scaled and locally multimodal domains. These domains can be problematic for conventional approaches, but are commonly found in real-world optimization problems. The principle of crowding is to apply a tournament selection to a parent-child pair with a high similarity. In the sample-based crowding, we determine such pairs based on a statistical comparison of the fitness values, which are sampled from the region between the pairs. Further, we take into account the ranks of the parents among the sampled values in the selection process, to determine their indispensability. These measurements are scale-invariant, which enables the proposed method to search a domain without presuming the distance between the optima or the scaling and the correlation of the variables. The proposed approach is evaluated in two benchmark problems with an ill-scaled and a locally multimodal landscape. The proposed method has a substantial advantage in terms of comprehensiveness compared to the conventional approaches, despite the additional cost of evaluations.

This article presents an algorithm that seeks every possible exception rule that violates a commonsense rule and satisfies several assumptions of simplicity. Exception rules, which represent systematic deviation from commonsense rules, are often found interesting. Discovery of pairs that consist of a commonsense rule and an exception rule, resulting from undirected search for unexpected exception rules, was successful in various domains. In the past, however, an exception rule represented a change of conclusion caused by adding an extra condition to the premise of a commonsense rule. That approach formalized only one type of exception and failed to represent other types. To provide a systematic treatment of exceptions, we categorize exception rules into 11 categories, and we propose a unified algorithm for discovering all of them. Preliminary results on 15 real-world datasets provide an empirical proof of effectiveness of our algorithm in discovering interesting knowledge. The empirical results also match our theoretical analysis of exceptions, showing that the 11 types can be partitioned in three classes according to the frequency with which they occur in data. © 2005 Wiley Periodicals, Inc. Int J Int Syst 20: 673–691, 2005.

We propose a method which visualizes irregular multidimensional time-series data as a sequence of probabilistic prototypes for detecting exceptions from medical test data. Conventional visualization methods often require iterative analysis and considerable skill thus are not totally supported by a wide range of medical experts. Our PrototypeLines displays summarized information based on a probabilistic mixture model by using hue only thus is considered to exhibit novelty. The effectiveness of the summarization is pursued mainly through use of a novel information criterion. We report our endeavor with chronic hepatitis data, especially discoveries of interesting exceptions by a nonexpert and an untrained expert.

We propose an incremental subspace clustering method for flexibly detecting hostile accesses to a Web site. Typical log data for Web accesses are huge, contain irrelevant information, and exhibit dynamic characteristics. We overcome these difficulties through data squashing, subspace clustering, and an incremental algorithm. We have improved, by modifying its data squashing functionality, our subspace clustering method SUBCCOM so that it can exploit previous results. Experimental evaluation confirms superiority of our I-SUBCCOM in terms of precision, recall, and computation time.

We report our preliminary endeavour for spiral discovery of exception rules based on discovered pieces of knowledge. An exception rule, which represents a deviational pattern to a general rule, exhibits unexpectedness and is sometimes extremely useful. We have proposed a domain-independent approach for simultaneous discovery of exception rules and their general rules. Exceptions are always interesting to discoverers, as they challenge the existing knowledge and often lead to the growth of knowledge in new directions. We propose a discovery method which exploits pre-discovered pairs of exception rules and their general rules, and apply it to a benchmark data set in knowledge discovery.

This paper presents an efficient algorithm for discovering exception rules from a data set without domain-specific information. An exception rule, which is defined as a deviational pattern to a strong rule, exhibits unexpectedness and is sometimes extremely useful. Previous discovery approaches for this type of knowledge can be classified into a directed approach, which obtains exception rules each of which deviates from a set of user-prespecified strong rules, and an undirected approach, which typically discovers a set of rule pairs each of which represents a pair of an exception rule and its corresponding strong rule. It has been pointed out that unexpectedness is often related to interestingness. In this sense, an undirected approach is promising since its discovery outcome is free from human prejudice and thus tends to be highly unexpected. However, this approach is prohibitive due to extra search for strong rules as well as unreliable patterns in the output. In order to circumvent these difficulties we propose a method based on sound pruning and probabilistic estimation. The sound pruning reduces search time to a reasonable amount, and enables exhaustive search for rule pairs. The normal approximations of the multinomial distributions are employed as the method for evaluating reliability of a rule pair. Our method has been validated using two medical data sets under supervision of a physician and two benchmark data sets in the machine learning community.

Research issues in the domain of AI for design can be organized in three categories: decision making, representation and knowledge handling. In the area of knowledge handling, this paper addresses issues concerning the management of design experience to guide a priori the generation of candidate solutions. The approach is based on keeping the trace of a previous design experience as a hierarchical knowledge base. A level in the hierarchy can be viewed as a level of granularity of the description of the design process. A general framework for defining a partial order function between the granularity levels in the knowledge bases of design expertise is proposed. It is then possible to compute the sets of the elements belonging to smaller granularity levels, which are linked to any component of the hierarchy. Thus, it makes it possible to compute the level in the hierarchy that can be reused without modification for the design of a new product. Computation of the appropriate level is mainly based on matching the data corresponding to the new requirements with these sets. The approach has been tested by using a multiple expert systems structure based on using interactively two systems, an expert system development tool for design, KAUS, and an expert system development tool for diagnosing engineering processes, SUPER. The intrinsic properties of SUPER have also been used for improving the design procedure when qualitative and quantitative knowledge is involved.>