The staff examined their method on each public benchmark datasets and actual electrical load forecasting information. Their mannequin constantly outperformed different state-of-the-art GAMs, particularly in difficult situations like excessive climate occasions. What’s notably encouraging is how the mannequin’s accuracy improves when consultants add their area information. On excessive excessive temperature days, incorporating skilled information decreased prediction errors by over 34%.
This work represents an necessary step towards making AI techniques that may really collaborate with human consultants. The staff has already built-in their method into Alibaba’s eForecaster platform, the place it’s being utilized by energy corporations throughout a number of Chinese language provinces.The implications lengthen past simply electrical load forecasting. Any area the place you want each correct predictions and human oversight may benefit from this method — from monetary forecasting to medical analysis.
The way forward for AI isn’t nearly constructing extra highly effective black packing containers. It’s about creating techniques that may work alongside human experience in an interpretable and collaborative approach. The analysis staff’s work demonstrates that we don’t have to decide on between mannequin accuracy and interpretability. By thoughtfully combining mathematical innovation with human-centered design, we will create AI techniques which are each exact and sensible.The staff’s paper supplies full technical particulars of their method, together with the mathematical foundations and experimental outcomes. Their code has been open-sourced to assist different researchers construct on this work.
This work was performed by YANG Linxiao, REN Rui, GU Xinyue, and SUN Liang at Alibaba DAMO Academy, printed in KDD ’23
