Independence Program Recommendation For Autism Spectrum Disorder Using Hybrid Collaborative Filtering Model

Abstract

Heterogeneity in the characteristics of children with Autism Spectrum Disorder (ASD) requires precise personalization of self-help skill programs. However, conventional Collaborative Filtering tends to suffer from popularity bias and struggles to capture clinical profile similarities between individuals, while patient-similarity approaches alone underuse collective success patterns, requiring a hybrid architecture. This study develops and evaluates a hybrid system integrating a Patient Similarity Model (PSM) based on cosine similarity and Collaborative Filtering based on Alternating Least Squares (ALS). The experiment used functional profile data from 50 students at a Special Education School (SLB), covering four core clinical aspects: motor, sensory, attention, and communication, along with the success history of 8 independence targets and 7 teaching strategies from teachers. An inverse frequency penalty mechanism was applied to address data popularity bias. Leave-One-Out Cross-Validation showed the hybrid model performed superiorly, achieving Hit Rate@3 of 0.8400, NDCG@3 of 0.8079, and MRR@3 of 0.7967, retaining PSM-Only's coverage advantage while empirically outperforming CF-Only (Hit Rate@3: 0.6200). The system also features a conditional rule engine based on pedagogical keyword mapping for accountable strategy recommendations. Key innovations include converting observational text into structured data, eliminating popularity bias, and producing explainable intervention outputs to support Individualized Learning Program (ILP) development.