Developing meta-FM machine learning model with application for personalized language learning environment

Abstract

A meta-pretrained factorization machine (FM) framework, termed Meta-FM, is proposed to address the cold-start problem in personalized learning path generation for middle-school foreign language education. The framework further introduces a cross-modal feature augmentation module that integrates categorical metadata (e.g., learning style) with continuous interaction data (e.g., quiz performance) to enrich sparse representations for new learners. The method integrates meta-learning principles with a high-dimensional FM architecture, enabling rapid adaptation to new students from minimal initial data while preserving generalizability across diverse cohorts. The core innovation lies in a three-module system: a meta-pretrained FM backbone that captures feature interactions through latent factors, an adaptive fine-tuning mechanism that updates these factors via few-shot learning, and a cross-modal feature augmentation layer that enriches sparse inputs with structured metadata. Unlike conventional approaches that rely on heuristic rules or extensive per-learner data, the framework dynamically refines recommendations by combining behavioral metrics with cognitive and demographic indicators. The system interfaces with existing diagnostic and ranking modules, replacing one-hot encodings with dense feature representations and stochastic Plackett–Luce ranking for path generation. Implemented with mixed-precision quantization and Transformer-based feature encoding, the method achieves efficient memory usage and robust cold-start performance. Experiments on multilingual student interaction logs demonstrate significant improvements in personalization accuracy and adaptation speed compared with baseline methods. The results highlight the potential of meta-pretrained FMs to bridge the gap between interpretable feature interactions and rapid cold-start adaptation in educational recommendation systems.