Masakatsu Ohta

Detecting Anomalous Events in Ubiquitous Sensor Environments using Bayesian Networks and Nonparametric Regression

We propose a novel anomaly detection method for a heterogeneous sensor environment in a living space where it is hard to analyze the entire mechanism of the environment and we are unlikely to predict irregular events. By using Bayesian networks and nonparametric regression, our method learns the ordinary behaviors of sensor values and examines the degree of anomaly for each observation according to the estimated variance from the results of learning. We applied our method to data collected in an office room equipped with brightness and motion sensors and obtained the plausible sensor relation networks with no or little prior knowledge. We detected some symptoms of anomalous events and determined the causal sensors by using the network structure.

Topological Feature Mining for Rambling Activities

A method for investigating rambling activities of moving objects is proposed. The goal is to construct common metrics used in various environments for characterizing the trajectory followed by rambling objects. Rambling activities are multi-stop,multi-purpose trips with trajectories with many intersections.Mathematical knot theory is introduced to examine the topological relation between intersections. The trajectories in an environment are represented in a vector space consisting of prime knots. Like a prime number, a prime knot is universal; thus, it is possible to compare the features of rambling activities across environments. An experiment using real-world taxi trajectories demonstrated that our method effectively classifies rambling activities according to daytime, nighttime, and a special event.

Extended another memory: understanding everyday lives in ubiquitous sensor environments

A lifetime recording agent that suggests unusual events to a user is proposed. The goal is to create a memory device that supports human memory by filtering, categorizing, and remembering everyday events. In ubiquitous sensor environments, the agent classifies users’ experiences represented by surrounding objects and predicts typical events that a user will experience next. Unusual events are detected by the awareness of different characteristics as the human brain does. If the prediction is incorrect, the actual event is considered to be unusual. A recurrent neural network that autonomously alters its architecture is introduced to perform event prediction. Experiments confirm: (1) a suitable hierarchical level of event categories for a current situation can be obtained by estimating the event prediction performance, that is, the recall rate and (2) rehearsal sequences dynamically generated by the network can substitute for a sequence of actual events. Thus, the agent easily responds to new environments without forgetting previous memories.

Non-negative Matrix Factorization for Inferring Implicit Preference of Potential Customer

An agent model that infers a person’s implicit preference is proposed. The goal is to identify potential customers interested in various products or services of companies. As suggested by the mere exposure effect, the repeated exposure to stimuli increased the subjects’ positive attitude to repeated stimuli. To express a person’s preference, therefore, the agent extracts objects that appear frequently in the images of scenes that the person likely sees. If some extracted objects are related to a company’s business, the person is assumed to be its potential customer. Sparse Non-negative Matrix Factorization (SNMF) is introduced to extract unknown objects appearing in many images. The sparseness imposed on the coefficient matrix is related to the ability of a person to recognize objects, and it is controlled by one parameter. Experiments confirmed: (1) as the number of objects recognized at one time increased, the number of extracted objects increased. On the other hand, as the number decreased, similar objects were assumed to be the same. Thus, it is possible to infer preferences of persons with different levels of abilities for object recognition; (2) the sparseness condition is suitable for detecting multiple objects in one image; and (3) according to the level of the sparseness, the optimal number of objects that should be extracted is efficiently obtained by an adaptive gain control.

A Flexible Personal Data Disclosure Method Based on Anonymity Quantification

In this paper, we propose a method of controlling personal data disclosure based on LooM (Loosely Managed Privacy Protection Method) that prevents a malicious third party from identifying a person when he/she gets context-aware services using personal data. The basic function of LooM quantitatively evaluates the anonymity level of a person who discloses his/her data, and controls the personal-data disclosure according to the level. LooM uses a normalized entropy value for quantifying the anonymity. In this version of the LooM, the disclosure control is accomplished by adding two new functions. One is an abstracting-function that generates abstractions (or summaries) from the raw personal data to reduce the danger that the malicious third party might identify the person who discloses his/her personal data to the party. The other function is a unique-value-masking function that hides the unique personal data in the database. These functions enhance the disclosure control mechanism of LooM. We evaluate the functions using simulation data and questionnaire data. Then, we confirm the effectiveness of the functions. Finally, we show a prototype of a crime-information-sharing service to confirm the feasibility of these functions.