Sentiment Text Analysis for Providing Individualized Feed of Recommendations Using Reinforcement Learning

We present an implementation of a new method of producing personalized text recommendations to users based on their sentiment. This method aims to identify patterns in user behavior based on a reward that results from each actual application in which the algorithm is used. The system consists of a semantic analysis algorithm that combines clustering and word processing, a level of neural networks designed to predict user reward for each text, and a level of algorithms that attempt to explore the problem area to adapt to changes in user behavior. This research work combines many known methods as well as variations thereof into a single system which in addition has the technical specifications to be used in typical applications like newsfeeds.

I. Introduction

Personalized feeds [1] are a very important feature of many applications in machine learning [2] and are intended to offer the user an experience that will be related to their interests and goals. This work focuses on text sentences, which means that in order to be implemented, there needs to be a system that can understand natural language mainly through the extraction of emotions (sentiment analysis). A key feature of such a system is its online character [3]. This means that the system must be constantly trained but also be able to generate flows from the very beginning. It must also try to suggest different content, which leads to the need for a proper balance between exploration and exploitation.