Kyeong-Jin Oh

Social network analysis in a movie using character-net

There have been various approaches to analyzing movie stories using social networks. Social network analysis is an effective means to extract semantic information from movies. Movie analysis through social relationships among characters can support various types of information retrieval better than audio-visual feature analysis. The relationships among characters form the main structure of the story. Therefore, through social network analysis among characters, movie story information such as the major roles and the corresponding communities can be determined. Progression of most movie stories is done by characters, and the scriptwriter or director narrates the story and relationships among characters using character dialogs. A dialog has a direction and time that supplies information. Therefore, the dialog is better for constructing social networks of characters than the co-appearance. Additionally, through social networks using the dialog, we can extract accurate movie stories such as classification of major, minor or extra roles, community clustering, and sequence detection. To achieve this, we propose a Character-net that can represent the relationships between characters using dialogs, and a method that can extract the sequences via clustering communities composed of characters. Our experiments show that our proposed method can efficiently detect sequences.

Personalized advertisement system using social relationship based user modeling

The influence of social relationships has received considerable attention in recommendation systems. In this paper, we propose a personalized advertisement recommendation system based on user preference and social network information. The proposed system uses collaborative filtering and frequent pattern network techniques using social network information to recommend personalized advertisements. Frequent pattern network is employed to alleviate cold-start and sparsity problems of collaborative filtering. For the social relationship modeling, direct and indirect relations are considered and relation weight between users is calculated by using six degrees of Kevin Bacon. Weight ‘1’ is given to those who have connections directly, and weight ‘0’ is given to those who are over six steps away and hove no relation to each other. According to a research of Kevin Bacon, everybody can know certain people through six depths of people. In order to improve prediction accuracy, we apply social relationship to user modeling. In our experiments, advertisement information is collected and item rating and user information including social relations are extracted from a social network service. The proposed system applies user modeling between collaborative filtering and frequent pattern network model to recommend advertisements according to user condition. User’s types are composed with combinations of both techniques. We compare the performance of the proposed method with that of other methods. From the experimental results, a proposed system applying user modeling using social relationships can achieve better performance and recommendation quality than other recommendation systems.

Social filtering using social relationship for movie recommendation

Traditional recommendation systems provide appropriate information to a target user after analyzing user preferences based on user profiles and rating histories. However, most of people also consider the friends opinions when they purchase some products or watch the movies. As social network services have been recently popularized, many users obtain and exchange their opinions on social networks. This information is reliable because they have close relationships and trust each other. Most of the users are satisfied with the information. In this paper, we propose a recommendation system based on advanced user modeling using social relationship of users. For the user modeling, both direct and indirect relations are considered and the relation weight between users is calculated by using six degrees of Kevin Bacon. From the experimental results, our proposed social filtering method can achieve better performance than a traditional user-based collaborative filtering method.

U2Mind: visual semantic relationships query for retrieving photos in social network

This research is to investigate a method that enables social networks to provide a semi-automatic system. The system will allow users to organize their target photos, using the concept of ownership attributes that describe the relationships between objects in the photos. In this paper, we propose formulating a visual semantic relationships query for photo retrieval. A Visual Semantic Relationship Query interface helps users describe their perspectives about the desired photo in a semantic manner. In the ranking process, by interpreting both concepts and relationships, a users query is transformed into a SPARQL, which is then sent to the JOSEKI server, and the returned photos are evaluated in terms of relevance to each photo. The experimental results demonstrate the effectiveness of the proposed system.

Ontology-driven visualization system for semantic searching

Technical manuals are very diverse, ranging from software to commodities, general instructions and technical manuals that deal with specific domains such as mechanical maintenance. Due to the vast amount of documentation, finding the information is a tedious and time consuming task, especially for the mechanics. It is also difficult to grasp relationships among contents in manuals. Many researchers have adopted ontology to solve these problems and semantically represent contents of manuals. However, if ontology becomes very large and complex, it is not easy to work with ontology. Visualization has been an effective way to grasp and manipulate ontology. In this research, we propose a new ontology model to represent and retrieve contents from the manuals. We have also designed a visualization system based on our proposed ontology. In order to model the ontology, we have analyzed aircraft maintenance process, extracted the concepts and defined relationships between concepts. After modeling ontology schema, all instances of ontology are created by instance creator. From here, raw data of maintenance manuals are preprocessed to well-formed format. Next, we create a set of rule mapping well-formed document and ontology schema. For the Component class, instance creator uses a classifier to separate all parts into Component and Primitive part class. If population task is complete, validity of data for created instances will be checked by JENA engine. The inference process will create inferred triples based on the ontology schema, and then the triples are saved into a triple repository. Our system then will use this triples repository to search necessary information and visualize the search results. We use the Prefuse toolkit to visualize the search results. With this, the mechanics can intuitively grasp the relationship between maintenance manuals using the provided information. This will allow the mechanics to easily obtain information for given tasks, reduce their time to search related information and understand the information through visualization.

Automatic indexing of cooking video by using caption-recipe alignment

In recent years, the internet use has led to rapid increases in multimedia resources. In this regard, handling multimedia data has become an important issue in information retrieval fields, and multimedia indexing plays an essential role in handling multimedia data. Textual metadata plays an important role in describing and analyzing multimedia resources. However, because only few multimedia resources provide textual metadata, there is a need for methods to generate the metadata. This paper proposes an automatic method for indexing cooking videos using textual data. The proposed method consists of web crawling for cooking videos and food recipes, the extraction of time-stamped text captions, and caption-recipe alignment using a dynamic programming technique. With information on caption-recipe alignment, automatic indexing is performed for cooking videos. The experimental results show that the proposed approach achieves highly accurate indexing. The user interface allows users to easily move to specific pinpoints with indexed videos.

Discovering Frequent Patterns by Constructing Frequent Pattern Network over Data Streams in E-Marketplaces

The extracting useful information such as itemsets and frequent patterns from the data becomes very important in terms of marketing strategies and maximizing profit in e-marketplaces. Although existing algorithms mining frequent patterns from the data are useful for persistent databases, they have some limitations of data mining from dynamic data arising from the continuous, unbounded and high speed characteristics of data streams. To identify useful frequent patterns in data streams, this paper proposes a frequent pattern network and a new method for discovering frequent patterns through the approximation of frequency counting on the network. The frequent pattern network, whose vertices and edges represent summarized information of transaction data, provides a user-centered environment based on the process of continuously mining frequent patterns because the proposed network is a small and compact data structure, and flexible for minimum support value. The experimental results show that proposed method is more efficient than FP-growth and Apriori methods, and the discussion of memory usage demonstrates the efficiency of the proposed method.

Link Strength-Based Collaborative Filtering for Enhancing Prediction Accuracy

User-based collaborative filtering recommends items to users by analyzing user preferences. Nearest neighbors are identified based on similarity between users and preference prediction of items is performed by using the nearest neighbors. The prediction accuracy depends on how the nearest neighbors are identified among users. In this paper, we propose link strength-based user modeling by applying trust information between users and item ratings to enhance the prediction accuracy. In the proposed user modeling, nearest neighbor candidate is extracted in traditional manner and final nearest neighbor is identified by calculating user ranking with trust information. Trust information between users is presented by link and consists of direct and indirect relation. We evaluate the prediction accuracy on recommended items and experimental results show that the prediction accuracy is improved by applying the proposed method.

Ontology-Driven Visualization System for Semantic Search

Abstract-Technical manuals are very diverse, ranging from manuals on software to manuals on commodities, general instructions and technical manuals that deal with specific domains such as mechanical maintenance. Due to the vast amount of manual information finding the necessary information is quite difficult. In case of electronic maintenance manuals currently used by companies, mechanics should search for the related information to accomplish their tasks. And it is difficult to grasp relationships among contents in manuals. Search process is time-consuming and laborious for mechanics. Many researchers have adopted ontology to solve these problems and semantically represent contents of manuals. However if ontology becomes very large and complex, it is not easy to work with ontology. Visualization has been an effective way to grasp and manipulate ontology. In this research, we model new ontology to represent and retrieve contents of manuals and design the visualization system based on proposed ontology. To model ontology, we analyzed aircraft maintenance process, extracted concepts and defined relationships between concepts. After modeling, we created instances of each class using technical manuals. Our system visualizes related information so that mechanics can intuitively grasp the information. This allows workers to easily get information for given tasks and to reduce their time to search related information. Also, related information can be understood at a time through visualization.

Extracting Relations towards Ontology Extension

Extracting local ontology from domain-specific documents for the purpose of acquiring knowledge or semantic information to extend their ontologies is considered very important. Main components of ontology are concepts and relations between concepts. In this paper, we focus on extracting triples, in which verbs are relations and subjects/objects are concepts, from documents based on natural language. Further, we show that term frequency is the most reliable measure among tf-idf and entropy on evaluating relations extracted from documents, particularly the aircraft maintenance manual.