LogitMat : Zeroshot Learning Algorithm for Recommender Systems without Transfer Learning or Pretrained Models

TL;DR

LogitMat introduces a novel zeroshot learning algorithm for recommender systems that leverages Zipf's Law and logistic regression, eliminating the need for transfer learning or pretrained models, and demonstrating competitive performance.

Contribution

This paper presents LogitMat, a new zeroshot learning method for recommender systems that does not rely on transfer learning or pretrained models, utilizing Zipf's Law and logistic regression.

Findings

LogitMat is fast and robust in experiments.

It achieves competitive results without transfer learning.

The method effectively addresses the cold-start problem.

Abstract

Recommender system is adored in the internet industry as one of the most profitable technologies. Unlike other sectors such as fraud detection in the Fintech industry, recommender system is both deep and broad. In recent years, many researchers start to focus on the cold-start problem of recommender systems. In spite of the large volume of research literature, the majority of the research utilizes transfer learning / meta learning and pretrained model to solve the problem. Although the researchers claim the effectiveness of the approaches, everyone of them does rely on extra input data from other sources. In 2021 and 2022, several zeroshot learning algorithm for recommender system such as ZeroMat, DotMat, PoissonMat and PowerMat were invented. They are the first batch of the algorithms that rely on no transfer learning or pretrained models to tackle the problem. In this paper, we follow this line and invent a new zeroshot learning algorithm named LogitMat. We take advantage of the Zipf Law property of the user item rating values and logistic regression model to tackle the cold-start problem and generate competitive results with other competing techniques. We prove in experiments that our algorithm is fast, robust and effective.