Researchers have introduced multi-sink Decision Predicate Graphs (ms-DPGs) to better explain multi-output tree ensembles. These models predict multiple binary labels using shared decision paths, but existing attribution methods often analyze each output independently. ms-DPGs extend decision predicate graphs with a sink for each label, allowing for the definition of Shared Predicate Flow (SPF) to quantify how decision mass is shared across labels. Experiments on nine multi-label benchmarks demonstrated that native multi-output forests exhibit significantly more shared decision structure compared to the Binary Relevance method, even when predictive performance is similar. AI
IMPACT Provides a novel method for interpreting complex multi-output models, potentially improving trust and debugging in AI systems.
RANK_REASON The item describes a new method and its evaluation on benchmarks, fitting the research category. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
- multi-label benchmarks
- multi-output tree ensemble
- multi-sink Decision Predicate Graphs
- native multi-output forests
- Shared Predicate Flow
- TreeSHAP
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →