Year: 2025
Type: conference-paper
Abstract: In recent years, the machine learning community has engaged on training increasingly powerful large language models (LLMs) for text classification tasks. These LLMs demand huge computational capabilities, requiring billions of parameters and enormous amounts of labeled data, which can make them expensive to use, optimize, and deploy in practice. However, recent studies on compressor-based distance metrics for text classification have been proposed, aiming to achieve competitive accuracy when compared with these LLM models. Compression-based models are usually based on k-nearest neighbors (KNN) and demand very few training parameters. In this paper, we propose a compression-based method that uses a Burkhard-Keller tree (BKTree) for similarity search, accelerating the KNN on compressed data. When compared with the state-of-the-art, the proposed method achieved speedups of 20x, 25x, 7x, 6.6x, 8x, 10x, 1.4x, 1.9x, 11x, and 12x for the Brotli, FSST, LZ4, LZAV, LZF, QuickLZ, Shoco, Smaz, Snappy, and ZLib compression algorithms, respectively. The proposed method is able to achieve similar prediction performance results while reducing the average runtime and computational resources, demonstrating its advantage in saving computational resources.
Authors: Bruno Vargas de Souza, Pedro Garcia Freitas
Institution: Universidade de Brasília
Language: English
Subfield: Theoretical Computer Science
Field: Mathematics
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7108320567