Abstract To develop smart home technology designed to analyze the activity of residents based on the logs of installed sensors, an activity model tailored to individuals must be constructed from less privacy-invasive sensors to avoid interference in daily life. Unsupervised machine learning techniques are desirable to automatically construct such models without costly data annotation, but their application has not yet been sufficiently successful. In this study, we show that an activity model can be effectively estimated without activity labels via the Dirichlet multinomial mixture (DMM) model. The DMM model assumes that sensor signals are generated according to a Dirichlet multinomial distribution conditioned on a single unobservable activity and can capture the burstiness of sensors, in which even sensors that rarely fire may fire repeatedly after being triggered. We demonstrate the burstiness phenomenon in real data using passive infrared ray motion sensors. For such data, the assumptions of the DMM model are more suitable than the assumptions employed in models used in previous studies. Moreover, we extend the DMM model so that each activity depends on the preceding activity to capture the Markov dependency of activities, and a Gibbs sampler used in the model estimation algorithm is also presented. An empirical study using publicly available data collected in real-life settings shows that the DMM models can discover activities more correctly than the other models and expected to be used as a primitive activity extraction tool in activity analysis.

We herein propose a domain-independent topic segmentation algorithm for free-form multi-party meeting recordings. The advantage of the proposed algorithm is that topical and lexical knowledge, which are difficult to adapt to the target meeting before speech recognition and topic segmentation, are not required. For an errorful sequence of phonemes obtained using a continuous phoneme recognizer, the proposed algorithm exhaustively analyzes the occurrence pattern of subsequences of phonemes and partitions the sequence into segments with coherent patterns. An empirical study on the ICSI Meeting Corpus has indicated that it performs comparably to lexical-cohesion-based text segmenters applied to human transcripts. Furthermore, the performance of the text segmenters applied to LVCSR output decreases significantly when keywords are not included in the lexicon. This suggests that, for the purpose of obtaining topical structure, the phoneme sequence segmenter could be more robust than text segmenters with LVCSR.

The goal of the present paper is to explore the feasibility of a topic segmentation method without using large vocabulary continuous speech recognition (LVCSR). The proposed method is domain-independent in the sense that it is not constrained by vocabulary and does not require training data. For a sequence of sub-word units obtained using a continuous sub-word recognizer, the proposed method merges similar adjacent parts of the sequence in an agglomerative manner to produce a hierarchical cluster tree. The proposed method uses a string kernel to efficiently compute the similarity between two strings of sub-word units based on the frequencies of any sub-strings appearing in the strings. By carefully excluding the influence of the sub-strings that are irrelevant to the topic of interest, topically coherent clusters are formed without linguistic knowledge. An empirical study on a Japanese news speech corpus shows that the method performs better than a topic segmenter using LVCSR.

This paper concerns the classification task in discrete attribute spaces, but considers the task in a more fundamental framework: the learning of Boolean functions. The purpose of this paper is to present a new learning algorithm for Boolean functions called Boolean kernel classifier (BKC) employing capacity control using Boolean kernels. BKC uses support vector machines (SVMs) as learning engines and Boolean kernels are primarily used for running SVMs in feature spaces spanned by conjunctions of Boolean literals. However, another important role of Boolean kernels is to appropriately control the size of its hypothesis space, to avoid overfitting. After applying a SVM to learn a classifier f in a feature space H induced by a Boolean kernel, BKC uses another Boolean kernel to compute the projections f/sup k/ of f onto a subspace H/sub k/ of H spanned by conjunctions with length at most k. By evaluating the accuracy of f/sup k/ on training data for any k, BKC can determine the smallest k such that f/sup k/ is as accurate as f and learn another f' in H/sub k/ expected to have lower error for unseen data. By an empirical study on learning of randomly generated Boolean functions, it is shown that the capacity control is effective, and BKC outperforms C4.5 and naive Bayes classifiers.