A method of storing vector data in compressed form using clustering

Abstract

The development of the machine learning algorithms for information search in recent years made it possible to represent text and multimodal documents in the form of vectors. These vector representations (embeddings) preserve the semantic content of documents and allow the search to be performed as the calculation of distance between vectors. Compressing embeddings can reduce the amount of memory they occupy and improve computational efficiency. The article discusses existing methods for compressing vector representations without loss of accuracy and with loss of accuracy. A method is proposed to reduce error by clustering vector representations using lossy compression. The essence of the method is in performing the preliminary clustering of vector representations, saving the centers of each cluster, and saving the coordinate value of each vector representation relative to the center of its cluster. Then, the centers of each cluster are compressed without loss of accuracy, and the resulting shifted vector representations are compressed with loss of accuracy. To restore the original vector representations, the coordinates of the center of the corresponding cluster are added to the coordinates of the displaced representation. The proposed method was tested on the fashion-mnist-784-euclidean and NYT-256-angular datasets. A comparison has been made of compressed vector representations with loss of accuracy by reducing the bit depth with vector representations compressed using the proposed method. With a slight (around 10 %) increase in the size of the compressed data, the absolute value of the error from loss of accuracy decreased by four and two times, respectively, for the tested sets. The developed method can be applied in tasks where it is necessary to store and process vector representations of multimodal documents, for example, in the development of search engines.