CoNN: Collaborative Neural Network for Personalized Representation Learning with Application to Scalable Task Classification

Abstract

Inspired by the personalized recommendation system, the Collaborative Neural Network (CoNN) is proposed to solve the problem of insufficient personalization of the existing representation learning methods. According to the input sample, CoNN generates personalized feature extractor (PFE), which is used for personalized representation learning. Specifically, we present two variants of CoNN: Unconditional Collaborative Neural Network (U-CoNN) and Conditional Collaborative Neural Network (C-CoNN). They are suitable for single-task and multi-task personalized representation learning, respectively. So based on CoNN we can solve the problem of personalized representation learning in the Scalable Task scenario. To evaluate CoNN, we experimented with two datasets, MNIST and Fashion-MNIST, which are commonly used in image classification problems. The results showed that U-CoNN improved the classification accuracy of 4.26% in a single task classification, and C-CoNN improved the classification accuracy by 1.25% in multiple task classification.