Web Image Search Re-Ranking With Click-Based Similarity and Typicality

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• The intent gap, distinct from the semantic gap, hinders image retrieval by misaligning user queries with their true information needs.
• Conventional re-ranking methods often neglect valuable user implicit feedback, such as click-through data.
• Improving image search performance requires addressing both image similarity and relevance typicality effectively.

Purpose of the Study:

• To propose a novel spectral clustering re-ranking approach that leverages user click-through data to overcome the intent gap.
• To develop a click-based multi-feature similarity learning algorithm for more accurate image similarity measurement.
• To enhance image search result relevance and typicality by incorporating implicit user feedback.

Main Methods:

• Utilized image click-through data as implicit user feedback to infer user intent.
• Developed a click-based multi-feature similarity learning algorithm using metric learning and multiple kernel learning.
• Applied spectral clustering based on the learned similarity to group images and re-rank them using click-based typicality measures.

Main Results:

• The proposed spectral clustering re-ranking approach significantly improves initial image search results.
• The method effectively incorporates click-based similarity and typicality, outperforming several existing re-ranking approaches.
• Experiments on real-world datasets demonstrate the robustness and effectiveness of the click-enhanced re-ranking strategy.

Conclusions:

• Click-through data is a crucial factor for bridging the user intent gap in image retrieval.
• The proposed spectral clustering re-ranking method offers a significant advancement in image search performance.
• Integrating implicit user feedback into similarity and typicality measures leads to more relevant and satisfying search results.