Boon-Siew Seah

3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Boon-Siew Seah focuses on multimedia information retrieval, social image search, and metadata analysis. Early work addressed the challenge of summarizing image search results by combining visual features and user-generated tags to create semantically coherent clusters. More recent research investigates the concurrent ranking of noisy social metadata to improve the discoverability and organization of visual content.

Research Trajectory

The available evidence outlines a focused research trajectory centered on improving how users interact with and retrieve social multimedia. Beginning in 2014, Seah’s work addressed the limitations of traditional tag-based social image search engines, which typically present results as flat, ranked lists that can be difficult for users to process intuitively.

As first author, Seah developed novel algorithms to summarize these search results by exploiting both visual features and user-generated tags. This approach was introduced in Summarizing social image search results, which demonstrated how to partition search outputs into visually and semantically coherent clusters while maximizing the coverage of the original results. This foundational work was expanded upon with the introduction of the Prism algorithm in PRISM: concept-preserving social image search results summarization. Prism advanced the summarization process by constructing a visual similarity graph—where nodes represent images and edges denote visual similarities—to ensure that the resulting summaries preserved the core concepts of the initial search query.

By 2018, Seah’s focus shifted toward addressing the underlying data quality issues inherent in social multimedia platforms like Flickr. Because user-generated metadata is often noisy—containing irrelevant tags, spam, or off-topic chatter—relying on raw metadata for search and recommendation is suboptimal. In Killing Two Birds With One Stone: Concurrent Ranking of Tags and Comments of Social Images, Seah led the development of a lightweight, visual signature-based model designed to filter and rank this metadata. Rather than treating tags and comments in isolation, this model concurrently generated ranked lists of both by evaluating their joint relevance to the image's visual features, the user comments, and the user tags.

Current Directions

Based on the provided corpus, the most recent paper in the evidence is from 2018. This work indicates an interest in the intersection of computer vision and natural language processing within social media ecosystems. The primary open problems addressed in this recent research involve the automated curation of noisy, crowdsourced metadata. By developing lightweight models that jointly evaluate visual signatures alongside textual comments and tags, this direction aims to improve downstream applications such as tag recommendation and image retrieval in highly unstructured social environments.

Profile updated

See something wrong? Report a mistake