Taehee Lee

Building an operational product ontology system

A base of clearly defined product information is a key foundation for an e-commerce system. The manipulation and exchange of semantically enriched and precise product information can enhance the quality of an e-commerce system and offer a high level of interoperability with other systems. Product information consists of product attributes and the relationships between products. Product categorization (or classification) is one type of such relationships. Ontology can play an important role in the formalization of product information. Although the idea of utilizing ontology for e-Catalogs has been raised before, we are yet to find an operational implementation of applying ontology in the domain. In this paper, we report on our recent effort to build an operational product ontology system for a government procurement service. The system is designed to serve as a product ontology knowledge base; not only for the design and construction of product databases but also for search and discovery of products. Especially, the keyword-based searching over product ontology database demands different techniques from those over conventional document databases or relational databases, and should be designed to reflect particular characteristics of product ontology. We also introduce some other issues that we have experienced in the project, and those issues include product ontology modeling, ontology construction and maintenance, and visualization. Our work presented herein may serve as a reference model for similar projects in the future.

An Ontology-Based Product Recommender System for B2B Marketplaces

An ontology-based product-recommender system can help catalog administrators in B2B marketplaces maintain up-to-date product databases by acquiring mapping information between the new product data and existing data. The proposed approach is keyword-based and independent of the underlying physical structure of product ontology. With a Bayesian belief network as its basis, the ranking algorithm utilizes semantics embedded within relationships defined in ontology to probabilistically determine the ranking scores. The methodology is implemented on a practical ontology system powerful enough to assist users in B2B marketplaces. Its effectiveness is demonstrated in comparison to the conventional search engines.

Practical issues for building a product ontology system

A base of clearly defined product information is a key foundation for an e-commerce system. The manipulation and exchange of semantically enriched and precise product information can enhance the quality of an e-commerce system and offer a high level of interoperability with other systems. Product information consists of product attributes and the relationships between products. Product categorization (or classification) is one type of such relationships. Ontology can play an important role in the formalization of product information. Although the idea of utilizing ontology for e-catalogs has been raised before, we are yet to find an operational implementation of applying ontology in the domain. In this paper, we report on our recent effort to build an operational product ontology system for the government procurement service. The system is designed to serve as a product ontology knowledge base; not only for the design and construction of product databases but also for search and discovery of products. We introduce some of the issues that we have experienced in the project, so that our work may serve as a reference model for similar projects in the future.

Modified naïve bayes classifier for e-catalog classification

As the wide use of online business transactions, the volume of product information that needs to be managed in a system has become drastically large, and the classification task of such data has become highly complex. The heterogeneity among competing standard classification schemes makes the problem only harder. However, the classification task is an indispensable part for successful e-commerce applications. In this paper, we present an automated approach for e-catalog classification. We extend the Naive Bayes Classifier to make use of the structural characteristics of e-catalogs. We show how we can improve the accuracy of classification when appropriate characteristics of e-catalogs are utilized. Effectiveness of the proposed methods is validated through experiments.

BestChoice: a decision support system for supplier selection in e-marketplaces

A growing number of companies are outsourcing their purchasing processes to independent purchasing agencies. These agencies now have to process an ever increasing number of purchase requests each day. The conventional methods of selecting the right suppliers for the purchase requests incur heavy human and time costs. We have designed and implemented BestChoice, a decision support system for supplier selection. It allows the evaluator to create rules for supplier evaluation based on the Multi Attribute Utility Theory, a theory for evaluating the utility of alternatives. BestChoice provides rule structures that can be saved and reused for similar selection cases. The architecture and selection rules of BestChoice are presented. Performance of BestChoice at one of the largest procurement agencies is analyzed.

Exploiting Attribute-Wise Distribution of Keywords and Category Dependent Attributes for E-Catalog Classification

E-catalogs are semi-structured documents that consist of multiple attributes and values. Although the conventional text classification techniques are applicable to the e-catalog classification as well, they cannot use the attribute information effectively to improve the classification accuracy. In this paper, we propose an e-catalog classification algorithm by extending Naive Bayesian Classifier to use the attribute information. Specifically, we focus on exploiting two e-catalog specific characteristics: the attribute-wise keyword distribution and the category dependent attributes. Experiments on real data validate the proposed method.

A pragmatic approach to model and exploit the semantics of product information

Recently researchers have tried to apply ontology to the product information domain. From a practical point of view, a key problem to streamline this trend is how to make a product ontology database operational. Technical solutions should consider the characteristics that a pragmatic product ontology database contains; first, the database size is quite huge, and second, ontological manipulation and utilization should be realistically feasible. We recently engaged in a project to build an operational product ontology system. The system is designed to serve as a product ontology knowledge base, not only for the design and construction of product databases but also for the search and discovery of products. From the insights gained through this project, we believe that ontological modeling and its implementation on an operational database, as well as the building applications which exploit ontological benefits, are the most important facets towards the successful deployment of a practical product ontology system. As such, searching techniques should take into account the features of an underlying ontological model especially with product searching being one of the most popular applications within product information systems. In this paper, we present these two issues; product ontology modeling and searching techniques. Although our work presented herein may not be the only way to build an operational product ontology database, it may serve as an important reference model for similar projects in future.

Right-lateral strike-slip movement of the South Korea Plateau associated with the opening of the East Sea (Sea of Japan)

Multi-channel seismic profiles and swath bathymetric data were used to investigate the tectonic evolution of the South Korea Plateau (SKP) associated with the opening of the East Sea (Japan Sea). The SKP is a deformed fragment of continental crust with numerous horsts and sediment-filled grabens. Three sedimentary units in the plateau were identified, which consist of the lower sequence (Unit I) interpreted as syn-rifting deposition during the early to middle Miocene, and the middle and upper sequences (Units II and III) considered as post-rifting deposition since the late Miocene. The fault system in the SKP includes the South Korea Plateau Fault (SKPF) trending NNW-SSE and smaller en echelon normal faults oriented NE-SW. We interpreted the information to postulate that the formation of the SKPF is the result of divergent right-lateral strike-slip movement in the SKP. This study suggests that the dextral movement of the SKP was induced by WSW-ward propagation of the spreading center located in the Japan Basin from the early to middle Miocene times.

Identifying relevant constraints for semantic query optimization

Abstract Semantic query optimization is the process of utilizing information implied by integrity constraints to reformulate the query into one that generates the same set of answers in a more efficient way. The difficulties of identifying relevant integrity constraints for a given query have been well recognized as have the various solutions. However, most of the previous works consider the query consisting of join(s) of base relations and the integrity constraints on base relations only. We generalize these restrictions and propose a method of identifying relevant integrity constraints for queries involving any combinations of joins and unions of base and defined relations. Our method utilizes a query graph that can be constructed dynamically during the query processing time, and, as a consequence, does not rely on heavy preprocessing or normalization. The method is extended to include the use of heuristics for generating a subset of answers.

Using Relational Database Constraints to Design Materialized Views in Data Warehouses

Queries to data warehouses often involve hundreds of complex aggregations over large volumes of data, and so it is infeasible to compute these queries by scanning the data sources each time. Data warehouses therefore build a large number of materialized views to increase system performance. However, materialized views need to be immediately updated when its sources are changed, leading to a possible decrease in system performance. The goal of the materialized view selection problem is to select an appropriate set of views that minimize total query response time as well as the view maintenance cost. In this paper, we develop a solution for identifying the candidate view space of materialization. In particular, we present algorithms for generating a union-view and a partial-view, both of which are good candidates for materialization. The proposed candidate view space guarantees to find a polynomial bounded set of optimal views, and any selection algorithm from previous research, e.g. greedy algorithm, can be ran on the candidate view space to find the optimal materialized views.