Export
Research Overview
Dong Hyun Kim focuses on data mining and machine learning, with specific emphasis on recommender systems and anomaly detection in data streams. His collaborative work addresses challenges such as data sparsity in social recommendations, the integration of side information for product embeddings, and memory-efficient algorithms for real-time outlier detection. The most recent evidence of his research centers on scalable stream processing techniques.
Research Trajectory
The available evidence for Dong Hyun Kim’s research spans from 2016 to 2018, highlighting a trajectory that begins with algorithmic improvements in recommender systems and transitions into the processing of continuous data streams. Throughout this period, Kim has consistently contributed as a second author, indicating a strong collaborative role in developing novel matrix factorization and density-based sampling methodologies.
In 2016, Kim co-authored TRecSo: Enhancing Top-k Recommendation With Social Information, which addressed the pervasive issue of data sparsity in recommender systems. While many contemporary approaches utilized social network data primarily to minimize rating prediction errors (such as RMSE), this work specifically targeted the accuracy of top-k ranking predictions. The proposed method incorporated structural network information by modeling users in dual roles—as trusters and trustees—demonstrating improved ranking performance on real-world datasets.
The following year, Kim continued his collaborative work in recommendation engines, shifting focus toward the representation of items rather than users. In Do "Also-Viewed" Products Help User Rating Prediction? (2017), the research team tackled the challenge of learning high-quality product embeddings. Recognizing that a product's visual appearance alone is insufficient to capture its full functionality and specifications, the authors introduced a matrix co-factorization method. This approach leveraged the implicit side information found in "also-viewed" product networks, allowing the system to capture broader contextual and functional aspects of items to improve user rating predictions.
By 2018, Kim's collaborative focus expanded from static recommendation datasets to the dynamic challenges of data stream mining. He contributed to DILOF: Effective and Memory Efficient Local Outlier Detection in Data Streams, a paper addressing the limitations of applying the well-known Local Outlier Factor (LOF) algorithm to continuous data streams. The research identified two primary bottlenecks in existing stream-based LOF extensions: high memory consumption and the inability to detect long, contiguous sequences of outliers. To resolve these issues, the authors developed a novel density-based sampling algorithm designed to efficiently summarize historical data, thereby reducing memory overhead while maintaining the fidelity required for accurate anomaly detection.
Current Directions
Based on the most recent paper in the available evidence from 2018, Kim’s research interests encompass the challenges of real-time data stream processing and anomaly detection. This direction emphasizes the necessity of memory-efficient algorithms capable of operating in resource-constrained environments. By focusing on techniques like density-based sampling to summarize past data, this work addresses the foundational open problems of maintaining accuracy and scalability when identifying outliers in continuous, high-volume data environments.
Profile updated
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake