Multi-Variant Recommendation Module for Improving the Activities in the Academic Staff Performance Appraisal System

Abstract

This paper presents the module and software implementation of an intelligent recommendation system designed to enhance the professional performance of university academic staff. Unlike traditional appraisal systems, which primarily function as historical data archives or offer recommendations for limited aspects of research activity, the proposed solution focuses on generating recommendations for future activities using a diverse set of tools. The system's core is built upon a hybrid architecture combining a knowledge graph for contextual modeling, Graph Neural Networks (GNN) for predicting potential activities, and Large Language Models (LLMs) for creating various performance improvement scenarios. A distinctive feature of the approach is the provision of alternative recommendation paths, which can be customized by modifying instructions within the prompt template, and an interactive refinement mechanism that allows users to clarify and expand the details of suggested recommendations. The module was deployed within a real university academic staff performance appraisal system. The results of the pilot operation indicated that users positively evaluated the generated personalized recommendations, leading to a noticeable improvement in annual performance indicators compared to previous periods.