LG AI Research Unveils SciNO Framework for Causal Inference in High-Dimensional Data

Recently, the buzzword in AI academia has gone beyond growing model performance to securing reasoning ability and reliability. In particular, Causal Discovery research, which uncovers causal relationships between variables, is drawing attention as a practical technology for solving complex real-world problems.

Existing order-based Causal Discovery methods had limitations in that memory usage surges in high-dimensional environments, and for MLP-based models, the stability of derivative estimation deteriorates. To address this, LG AI Research proposed the SciNO (Score-informed Neural Operator) framework.

SciNO adopts the theoretical perspective of the Hilbert Diffusion Model (HDM), treating data not as simple vectors but as continuous functions. Through this, it is designed so that the Neural Operator can stably approximate not only the score function but also its higher-order derivatives.

The key features and achievements are as follows:

• Through an FNO (Fourier Neural Operator)-based architecture, it processes signals in the frequency domain to effectively learn derivative information.
• It introduces the LTE (Learnable Time Encoding) module to provide learnable embeddings for continuous time variables.
• It demonstrated its performance by achieving an average Order Divergence reduction of 42.7% on synthetic datasets and 31.5% on real datasets compared to the existing model DiffAN.

Additionally, LG AI Research also presented a methodology that integrates semantic information into the causal order inference process through a probabilistic control algorithm that combines SciNO with autoregressive models such as LLMs.