Multiplicative Update For Fast Optimization Of Information Retrieval Based Neighbor Embedding
Jaakko Peltonen, Ziyuan Lin

Abstract:
Dimensionality reduction of high-dimensional data for visualization has recently been formalized as an information retrieval task where original neighbors of data points are retrieved from the low-dimensional display, and the visualization is optimized to maximize flexible tradeoffs between precision and recall of the retrieval, avoiding misses and false neighbors. The approach has yielded well-performing visualization methods as well as information retrieval interpretations of earlier neighbor embedding methods. However, most of the methods are based on slow gradient search approaches, whereas fast methods are crucial for example in interactive applications. In this paper we propose a fast multiplicative update rule for visualization optimized for information retrieval, and show in experiments it yields equally good results as the previous state of the art gradient based approach but much faster.