Burstiness has been one of the most important criteria for extracting topics and events from documents posted on social media. Recently, researchers are focusing on extracting geolocal topics and events from such social documents because of the increasing number of geo-annotated documents (e.g., Geo-tagged tweets on Twitter). In our previous work, we developed a method for identifying local temporal burstiness to detect local hot keywords considering the users' location. The previous method is based on Kleinberg's temporal burst detection algorithm, which presupposes that the rate of posting remains constant. However, this leads to a difference in bursty periods depending on public awareness. To address this issue, in this paper, we propose a novel method for identifying local temporal burstiness by using the MACD-histogram-based temporal burst detection algorithm. The MACD-histogram-based temporal burst detection algorithm is based on the trend analysis of stock prices. To compare the proposed method with the previous method, we conducted experiments using actual burst detection in geo-tagged documents. The experiments revealed that the proposed method can identify local temporal burstiness on the basis of public awareness.

An emerging topic in social media is the increase in the number of geo-annotated documents, which include not only posted time but also posted location. Social media users have been transmitting information about things they witnessed themselves in their daily life through such geo-annotated (georeferenced) documents. Georeferenced documents are usually related to not only personal topics but also local topics and events. Therefore, identifying high-density areas associated with local “attractive” topics in georeferenced documents is one of the most important challenges in many application domains. In this study, we propose a novel density-based spatial clustering algorithm called the (ε,σ)- density-based adaptive spatial clustering algorithm for identifying high-density areas in which geo-related local topics in georeferenced documents are located. The (ε,σ)-density-based adaptive spatial clustering algorithm can identify local high-density areas by using adaptive spatial clustering criteria.

Burstiness is the simplest but the most robust criterion for detecting topics and events in online documents. Online documents are referred to as document streams because they have a temporal order. Kleinberg's temporal burst detection algorithm is the most successful algorithm for detecting bursty periods related to a topic- or event-related keyword. Kleinberg's temporal burst detection algorithm aims to find certain time periods in which a keyword occurs at a high frequency. In recent times, large-scale online documents are increasingly common on social media. Therefore, speed-up of burst-detection processing is one of the most important issues in this era of big data. In this paper, we propose a novel parallelization model, called the hybrid parallelization model with a hidden I/O thread, to enable the parallel processing of Kleinberg's temporal burst detection algorithm on a multi-core CPU. In a multi-core CPU environment, I/O latency is a critical issue for improving the performance of a parallelization model. To automatically hide the I/O latency, the proposed parallelization model utilizes speculative I/Os. The results of experiments using actual large-scale document streams show that the proposed parallelization model performs well compared with a conventional parallelization model.

Proteins are important biochemical compounds that have biogenic functions for biological activities. The three-dimensional structures of proteins are closely related to its biological functions, and therefore, techniques for comparing them have been studied. Many of these techniques for comparing protein structures are based on protein structure alignment, which is one of the most effective methods. CMO (Contact Map Overlap) is formulated as combinatorial optimization to find the optimal structure alignments. In this paper, we propose a novel heuristic using Modified Extremal Optimization (MEO) for CMO. Our MEO-based heuristic is characterized by three features. First, the proposed heuristic uses MEO for alternation generations. Second, an initial solution is created by dynamic programming (DP). Third, state transition is executed using the best admissible move strategy.

With an increasing interest in social media, a large number of Web images have been created on the Internet. Therefore, extracting useful knowledge from a large-scale set of Web images has become a new type of challenge. In this paper, we focus on Web images that have been posted onto the Internet through social media sites. The main objective of this study is to extract the events and track the topics of a document stream that includes Web images. To address this challenge, this paper proposes a novel method for burst detection in a Web-image document stream. The proposed method integrates a clustering technique with Kleinberg's burst detection. To evaluate the proposed method, we used actual tweets from Twitter users. The experimental results show that the proposed method can extract the events and track the topics related to Web images posted on social media sites.

Recently, we have been able to create multimedia contents easily without individual recourse to each medium. Often one needs to select data to match one's impressions from a medium with which one is unfamiliar. We propose a retrieval method for music using "impression" words. The characteristic of the proposed method is the use of the "affective value" of the musical piece based on results of psychological research on music. In order to develop the affective value, we examine the relationship between the affective value and musical structure by applying machine learning. We present the results of this examination and discuss a method for producing affective values.

A method is presented for indexing terms in a knowledge-base retrieval-by-unification (RBU) system. The term is a well-defined structure which represents knowledge using variables. RBU operations are an extension of relational database operations using unification and backtracking to retrieve terms from term relations. The term indexing proposed uses hashing and trie structures to reduce the number of comparisons between elements of a search condition and of an object term relation. Unification on a trie structure is suited to backtracking bindings of variables. The search and updating speed of an RBU prototype is measured to evaluate the indexing method. This method is effective for fast term retrieval for a large number of similar and varied form terms. The overhead for maintaining indexes in updating is low.>