💻 computer science

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

This paper introduces PatentXAI, a framework that addresses the computational intractability of patent valuation by approximating Shapley values through Markov Blanket-restricted coalitions within a knowledge graph, enabling efficient and accurate hierarchical profit allocation across large patent portfolios.

Original authors: Joy Bose

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joy Bose

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brand-new smartphone. This phone is a masterpiece of engineering, but it's not just one invention; it's a giant puzzle made of 100,000 different patented ideas.

Now, imagine a legal battle. One inventor claims, "I own a tiny piece of this puzzle, and I deserve a huge slice of the phone's profit." The company says, "No way, your piece is small." The judge asks, "How do we fairly split the money among thousands of people who all contributed?"

This is the problem the paper PatentXAI tries to solve. It treats patent valuation like a game of fair sharing, using a mathematical tool called the Shapley Value (think of it as the "Goldilocks" rule for splitting a pie).

Here is how the paper breaks it down, using simple analogies:

1. The Core Idea: The "Fair Share" Game

In a cooperative game, the Shapley value calculates how much each player contributes to the team's total win.

The Problem: If you have 100,000 players (patents), calculating the exact fair share for everyone is like trying to count every grain of sand on a beach. It takes too long and is computationally impossible.
The Solution: The authors created a shortcut. They realized that a patent doesn't need to interact with every other patent in the world to know its value. It only needs to interact with the patents it is directly connected to.

2. The Shortcut: The "Social Circle" (Markov Blanket)

The paper uses a concept from graph theory called a Markov Blanket.

The Analogy: Imagine you are at a massive party with 100,000 people. To figure out your influence on the party's vibe, you don't need to know what the person in the back corner is doing. You only need to know what your immediate social circle is doing: your friends, your friends' friends, and the people your friends are talking to.
How it works: The system builds a "knowledge graph" (a map of who cites whom and who covers which phone feature). For any single patent, it only looks at its "social circle" (the patents covering the same phone parts).
The Result: Instead of checking 100,000 patents, the system only checks about 33% of them (the "blanket"). This makes the math fast enough to run in 10 milliseconds per patent.

3. The Two-Step Hierarchy: The "Company Structure"

The paper suggests a two-level approach to splitting the money, like a corporate hierarchy:

Level 1 (The Departments): First, the total profit of the phone is split among the major "departments" (e.g., the Camera team, the Battery team, the Screen team). This is done using exact math because there are only a few departments.
Level 2 (The Employees): Next, the money allocated to the "Camera Department" is split among the individual camera patents. Here, the system uses the "social circle" shortcut mentioned above to be fast.
The Twist: The system also weighs patents based on their "importance" (like how famous or central they are in the citation network). A patent that is a "hub" of connections gets a slightly larger slice of the department's budget.

4. Testing the System: The "Simulation Lab"

The authors tested their method using computer simulations (since they didn't have real-world court data yet).

The Test: They created fake patent portfolios ranging from 12 to 100 patents.
The Findings:
- Speed: It was incredibly fast.
- Accuracy: The "shortcut" was very close to the "perfect" (but slow) calculation. The difference was tiny (less than 0.1).
- The "Dense" Test: They tested a scenario where 80% of patents all covered the exact same feature (like a crowded room where everyone is shouting the same thing). Surprisingly, the system got more accurate in this crowded scenario because the "social circles" merged into one big, clear group.

5. The Big Caveat: "We Have the Calculator, Not the Ingredients"

This is the most important part of the paper. The authors are very honest about what they haven't solved yet.

The Missing Piece: The Shapley math needs a starting number: "How much money does a specific group of patents actually generate?" (This is called the characteristic function).
The Reality: In machine learning, you can easily ask a computer model, "What happens if we remove this feature?" But in the real world, you can't easily ask, "How much money would this phone make if we removed this one patent?"
The Paper's Stance: The authors say, "We built a perfect, fast calculator for splitting the pie, but we haven't figured out how to measure the size of the pie yet." They suggest using real-world data (like court cases and market prices) to estimate that pie size in the future.

Summary

PatentXAI is a new, super-fast way to fairly split profits among thousands of patent owners by only looking at their "social circles" rather than the whole world. It works like a two-step manager: first splitting money between departments, then between individual patents.

However, the paper admits it is currently a theoretical framework. It provides the engine and the steering wheel, but the driver (the real-world data to determine actual patent value) still needs to be built and tested.

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

1. Problem Statement

The paper addresses the long-standing unsolved problem in intellectual property economics of estimating the economic contribution of a single patent within a product embodying tens of thousands of patents. Current legal frameworks (e.g., Georgia-Pacific, Nash bargaining, incremental value analysis) struggle to combine graph-conditioned computation with hierarchical allocation while satisfying axiomatic fairness.

A critical distinction is made between the computational problem and the economic estimation problem. While machine learning models can directly compute a characteristic function $v(S)$ (revenue of a feature subset), patent valuation cannot observe $v(S)$ directly. The authors explicitly state that estimating $v(S)$ from real data remains an open problem; this paper focuses on the computational contribution: providing an efficient, axiomatic framework for Shapley value computation given a well-defined $v(S)$ .

2. Methodology: The PatentXAI Framework

The authors propose PatentXAI, a four-layer framework that treats patent valuation as an Explainable AI (XAI) feature attribution problem. The framework relies on the Shapley value, which satisfies efficiency, symmetry, dummy, and additivity axioms.

Layer 1: Comprehensive Patent Knowledge Graph

The technology landscape is modeled as a directed property graph containing:

Nodes: Patents, Features, Components, and Standards.
Edges: COVERS (Patent to Feature), IMPLEMENTED_IN (Feature to Component), CITES (Patent to Patent, forming a Comprehensive Patent Citation network), and DECLARED_ESSENTIAL_TO (Patent to Standard).
Centrality: Patents are assigned composite centrality weights ( $w_i$ ) based on PageRank, HITS authority, and normalized in-degree within the citation network.

Layer 2: Economic Impact Estimation

Components are assigned economic weights using methods such as conjoint analysis, hedonic pricing, or engineering cost analysis. For Standard-Essential Patents (SEPs), these weights are anchored to FRAND royalty rates and court determinations.

Layer 3: Markov Blanket Coalition Restriction

To make Shapley computation tractable for large portfolios ( $n=100+$ ), the authors restrict the coalition for each patent $i$ to its Markov Blanket ($MB(i)$).

Definition: $MB(i)$ includes all components covered by patent $i$ and all other patents covering at least one of those same components.
Theoretical Basis: This restriction is grounded in the C-SVE Conditional Independence Theorem (Li et al., 2020), which posits that under local conditional independence, a neighborhood-restricted Shapley value equals the global value almost surely.
Computation: Shapley values are approximated via Monte Carlo sampling over permutations of the Markov Blanket rather than the full portfolio.

Layer 4: Hierarchical Profit Allocation

Allocation proceeds in two steps to handle scale:

Component-Level: Exact Shapley values are computed for macro-components (typically 4–8) to distribute total product profit among them, capturing synergies.
Patent-Level: Each component's allocated budget is distributed among its covering patents. Crucially, the centrality weight $w_i$ is folded into the local characteristic function $v_{local}(S)$ (Equation 2) rather than applied as a post-hoc multiplier. This creates a "Priority-Aware Shapley Value" that respects the efficiency axiom and shifts allocation toward the cooperative game core.

3. Key Contributions

The paper outlines six specific contributions:

Formal Mapping: Establishes the structural identity between XAI feature attribution and patent valuation, proving hierarchical Shapley satisfies axioms under graph-conditioned approximation.
Markov Blanket Restriction: Introduces a tractable coalition restriction grounded in C-SVE, validated by scaling experiments.
Dense-Component Resilience: Demonstrates that the approximation does not degrade catastrophically in dense SEP-like sub-portfolios; instead, it expands to cover dense clusters, improving accuracy in homogeneous settings.
Literature Integration: Explicitly addresses the Ma et al. (2020) finding on the disassociation between Shapley rankings and Markov Blanket membership, clarifying that PatentXAI is designed for budget-constrained apportionment, not causal feature selection. It also distinguishes itself from Dehez and Poukens (2014) by using graph-conditioned restriction and hierarchical allocation.
Four-Layer Architecture: Integrates CPC graphs, centrality-weighted characteristic functions, SHAP-Tree-MCDA mathematics, and D-Shap dynamic updates for evolving portfolios.
Empirical Roadmap: Provides a concrete plan for validation using public datasets (ETSI, USPTO, Lens.org) and landmark litigation data.

4. Experimental Results

Experiments were conducted on synthetic Pareto-distributed coverage graphs (simulating real SEP structures) using a single Intel i7 core.

Scaling (n=12 to n=100):
- At $n=100$ , the median Markov Blanket size is 32.9% of $n$ , and the 90th percentile is 55.2%.
- Runtime is approximately 10 milliseconds per patent.
- The difference against a high-sample Monte Carlo reference (5,000 samples) is 0.062 ± 0.003 at $n=100$ .
- The approximation error does not systematically increase with $n$ .
Dense-Component Experiment:
- In a scenario where 80% of patents share a single component (mimicking a dense SEP sub-portfolio), the Markov Blanket expands to cover 80% of the patents.
- Counterintuitively, the error decreases to 0.039 in this dense setting because the pooled coalition becomes structurally homogeneous, making the within-blanket sampling more accurate.
Toy End-to-End Pipeline:
- A 10-patent, 4-component example demonstrated that the framework satisfies the efficiency axiom, where the sum of individual patent valuations exactly equals the total product profit ($200).

5. Significance and Claims

The paper claims that PatentXAI offers a principled, axiomatic foundation for patent valuation that has been missing in legal and economic literature. By treating valuation as an XAI problem, it leverages established mathematical tools (Shapley values, Markov Blankets) to solve the computational intractability of portfolio-scale allocation.

Modest Claims and Limitations:

v(S) Estimation: The authors explicitly state that estimating the characteristic function $v(S)$ from real data is the "primary open problem." The framework assumes $v(S)$ is given; the economic accuracy of the final valuation depends entirely on the quality of this input.
Validation: No part of PatentXAI has been run on real patent data yet. The current results are based on synthetic data and theoretical proofs.
Faithfulness Assumption: The theoretical guarantee of the Markov Blanket restriction relies on a faithfulness assumption (that the graph structure encodes conditional independencies in the profit function). The authors acknowledge this may not hold in all real-world scenarios and treat the restriction as a computationally motivated heuristic with empirically validated accuracy.
Legal Acceptance: While the framework addresses the "disassociation" issue (where redundant patents might receive high Shapley values), the authors note it is an empirical question whether courts will accept centrality-weighted rankings over traditional legal precedents.

In conclusion, PatentXAI is presented not as a fully validated end-to-end valuation system, but as an efficient and axiomatic computation framework that bridges the gap between cooperative game theory and the practical constraints of patent portfolio analysis, with a clear roadmap for future empirical validation.