Порівняльний аналіз та вибір архітектури інтерпретованої рекомендаційної системи у фінансовій інженерії

Abstract

This paper presents a comparative analysis of five architectural classes for explainable recommendation systems in AI-driven investment portfolio management, evaluated across criteria of explainability, temporal depth, investment efficiency, resource consumption, and market regime robustness. Prioritizing explainability and temporal depth, the analysis identifies causal AI as the dominant approach, justifying the selection of a hybrid Causal Transformer (CausalFormer) architecture that embeds causal inference into self-attention blocks via an asymmetric temporal mask and causal structure matrix, eliminating dependence on spurious correlations.