A Framework for Graph-Conditioned Hierarchical Shapley Attribution in Patent Valuation

Researchers introduce PatentXAI, a framework using Shapley values and graph-conditioned Markov Blankets to fairly allocate patent valuations within complex products containing thousands of patents. The method scales computationally by restricting coalition analysis to relevant patent subsets, achieving sub-100 millisecond processing times while maintaining accuracy within 6.2% of Monte Carlo benchmarks.

PatentXAI addresses a fundamental challenge in intellectual property economics: determining individual patent contributions to product value when thousands of patents interlock in modern products. The framework leverages Shapley value theory from cooperative game theory, translating the patent valuation problem into explainable AI by modeling revenue functions across patent subsets. This bridges academic game theory with practical IP management.

The technical innovation centers on computational tractability. Computing exact Shapley values grows exponentially with coalition size, making it infeasible for large patent portfolios. The authors constrain each patent's coalitions to its Markov Blanket—the minimal set of patents whose removal makes that patent conditionally independent—within knowledge graphs grounded in formal conditional independence theorems. Testing on graphs up to 100 patents demonstrates median Markov Blanket sizes of roughly 33% of total patents, with 90th-percentile sizes around 55%, reducing computational burden dramatically while maintaining accuracy.

The hierarchical profit allocation approach proves particularly relevant for complex products. Exact Shapley distributes profits among major components, then centrality-weighted Shapley allocates component budgets among constituent patents. Dense-component experiments show the framework correctly expands Markov Blankets when patents cluster heavily, improving accuracy to within 3.9% of benchmarks for homogeneous portfolios.

The framework's practical deployment hinges on accurately estimating revenue functions from real data—explicitly identified as the primary unsolved problem. The authors outline validation roadmaps using public ETSI, USPTO, and Lens.org datasets, indicating awareness that computational elegance requires empirical grounding. This positions PatentXAI as a methodological foundation awaiting real-world calibration.

• →Shapley value-based framework enables fair patent contribution allocation in products containing thousands of patents
• →Markov Blanket restriction reduces computational complexity from exponential to tractable scales, processing 100 patents in 10 milliseconds
• →Hierarchical allocation separates macro-component distribution from micro-patent budgeting within components
• →Accuracy within 6.2% of Monte Carlo references at n=100, improving to 3.9% accuracy for dense patent clusters
• →Revenue function estimation from real data remains the critical unsolved problem blocking practical deployment