New crEG graph algorithm speeds multimedia retrieval, improves exploration search

As multimedia datasets grow, approximate nearest neighbor search (ANNS) becomes critical for fast retrieval. Graph-based methods offer the best precision-speed trade-off, but construction can be slow and memory-hungry. In a new paper published at ICMR 2024, researchers Nico Hezel, Kai Uwe Barthel, Konstantin Schall, and Klaus Jung introduce the continuous refining Exploration Graph (crEG), a novel approach that constructs a compact exploration graph rapidly while achieving state-of-the-art search performance. The algorithm is specifically designed for undirected graphs with even degrees and guarantees graph connectivity at all times, a property that makes it particularly useful for exploratory search, where the query itself is part of the database.

crEG also includes an optional edge optimization algorithm that further improves retrieval effectiveness without sacrificing construction speed. Experiments show that high ANNS efficiency does not automatically translate to good exploratory search performance, highlighting a gap the authors address directly. This breakthrough enables faster, more memory-efficient multimedia retrieval, opening doors for better recommendation systems, content exploration tools, and large-scale similarity searches where traditional methods struggle.