大規模線性分類資料低階多項式映射中雜湊函數之應用

Abstract

Nonlinear mappings have long been used in data classification to handle linearly inseparable problems. Low-degree polynomial mappings are a widely used one among them, which enjoys less time and space consumption and may sometimes achieve accuracy close to that of using highly nonlinear kernels. However, the explicit form of polynomially mapped data for large data sets can also meet memory or computational difficulties. To solve this, hash functions like murmur and fnv hash are used in some packages like vowpal wabbit to have flexible memory usage. In this thesis, we propose a new hash function which is faster and could achieve the same performance. The results are validated in experiments on many datasets.