Jaekyung Yang

Optimization-based feature selection with adaptive instance sampling

Preprocessing the data to filter out redundant and irrelevant features is one of the most important steps in the data mining process. Careful feature selection may improve both the computational time of inducing subsequent models and the quality of those models. Using fewer features often leads to simpler and easier to interpret models, and selecting important feature can lead to important insights into the application. The feature selection problem is inherently a combinatorial optimization problem. This paper builds on a metaheuristic called the nested partitions method that has been shown to be particularly effective for the feature selection problem. Specifically, we focus on the scalability of the method and show that its performance is vastly improved by incorporating random sampling of instances. Furthermore, we develop an adaptive variant of the algorithm that dynamically determines the required sample rate. The adaptive algorithm is shown to perform very well when applied to a set of standard test problems.

Intelligent Partitioning for Feature Selection

This paper develops a new optimization-based feature-selection framework for knowledge discovery in databases. Algorithms following this new framework have attractive theoretical properties such as proven convergence to an optimal set of relevant features and the ability for deriving rigorous statements regarding the quality of the set that is found. Within this framework both wrapper and filter algorithms are derived, and numerical experiments show the new methodology to perform well with respect to accuracy and simplicity of the set of features found to be relevant.

An optimization approach to partitional data clustering

Scalability of clustering algorithms is a critical issue facing the data mining community. One method to handle this issue is to use only a subset of all instances. This paper develops an optimization-based approach to the partitional clustering problem using an algorithm specifically designed for noisy performance, which is a problem that arises when using a subset of instances. Numerical results show that computation time can be dramatically reduced by using a partial set of instances without sacrificing solution quality. In addition, these results are more persuasive as the size of the problem is larger.

Two-phase edge outlier detection method for technology opportunity discovery

This article introduces a method for identifying potential opportunities of innovation arising from the convergence of different technological areas, based on the presence of edge outliers in a patent citation network. Edge outliers are detected via the assessment of their centrality; pairs of patents connected by edge outliers are then analyzed for technological relatedness and past involvement in technological convergence. The pairs with the highest potential for future convergence are finally selected and their keywords combined to suggest new directions of innovation. We illustrate our method on a data set of US patents in the field of digital information and security.

Design and Implementation of Robotic Parcel-Sorting Systems

With the recent increase in e-business, post offices or door-to-door parcel delivery companies have witnessed a corresponding increase in the number of parcels handled. Most of these parcels are still being handled manually, which results in an increased workload for employees and lowers efficiency. Korea Post has considered a smart-robot manipulation system to automate parcel-sorting procedures that remain manual. This study examined parcel-sorting methods using a robotic system with the aim of reducing the workloads of deliverymen and enhancing parcel-sorting efficiency. We developed a methodology for the design of a robotic parcel-sorting system, and the system was then tested and implemented in accordance with this methodology. The results of this study are expected to reduce parcel workloads of deliverymen and improve parcel-sorting efficiency.

Finding the time lag effect of the R&D activity for a government research program of Korea

This study examines the relationship between R&D investment and subsequent outputs of the research activity. Usually, there is some time difference between the production of research outputs, such as academic papers and application or registration of patents, and the investment of R&D expenditure. The time lag for producing this kind of research outputs should be considered to evaluate the performance of research activity exactly. The purpose of this study is to identify time lag effect between the times of input and output of a R&D activity and to derive the degree of time lag using the data set of a long term R&D program supported by Korean government. A modified Almon model is suggested to identify the time lag effect between input and output of research activities performed by this program. Time-series cross-section data from 16 research centers between 2001 and 2009 are used to find time lag effect.

Scalable clustering with adaptive instance sampling

Most of the clustering algorithms are affected by the number of attributes and instances with respect to the computation time. Thus, the data mining community has made efforts to enable induction of the clustering efficient. Hence, scalability is naturally a critical issue that the data mining community faces. A method to handle this issue is to use a subset of all instances. This paper suggests an algorithm that enables to perform clustering efficiently. This is done by using nested partitions method for solving the noisy performance problems, which arises when using a subset of instances and adjusting the sample rate properly at each iteration. This Adaptive NPCLUSTER algorithm had better similarity in small dataset and had worse similarity in large dataset than NPCLUSTER, but it had shorter computation time than NPCLUSTER.

Scalable optimization-based feature selection with application to recommender systems

Along with development of a variety of data mining techniques, numerous feature selection methods have been introduced to reduce dimensionality. This may improve scalability and make interpreting learning models easier. In this dissertation a new optimization based feature selection method using the nested partition (NP) approach is presented, including both basic analysis of the NP framework and numerical results on various experiment problems. The numerical results show how the optimal structure of the NP makes contributions on a feature selection process. Further, it is addressed how the new intelligent partitioning method obtains very high quality partition efficiently. The feature selection method is implemented as both a filter and a wrapper. In addition, the scalability of the algorithm, which is the most significant issue in mining large databases, is also dealt with according to the instance dimension, the feature dimension, and new features adaptation. However, since the NP naturally handle the feature dimension effectively, the dissertation mostly focuses on scalability with respect to the instance dimension. In this research problem, two systematic approaches to improve scalability of instance dimension are presented, which both utilize random sampling. Through this study, a predicted best solution for the size of instance samples is presented using a two-stage version of the NP that also incorporates statistical selection, and a heuristic solution is as well presented in a new adaptive version of the algorithm. Numerical results report that those two approaches are effective for scalability improvement, and perform better than the generic NP method that uses a static sampling approach. In order to have the NP feature selection method flexible for handling mixed type of features, feature quality evaluators are introduced to determine the order of partitioning with experiment results reporting which one performs better based on a data domain. Finally as a case study, a recommender system that can be effectively used in B2B (business to business) e-business systems is provided using classification, association rules and the new NP-based feature selection method.