Patents have a valuation problem. A patent is a legal instrument, a technical document, and a financial asset, and the price tag a court or acquirer assigns to it has historically depended on which expert witness the retaining party hired. A June 2026 preprint on arXiv proposes replacing that expert judgment with graph-conditioned hierarchical Shapley attribution: a framework that decomposes a patent’s value into per-feature credit assignments rather than producing a single aggregate number. Whether that output is defensible under cross-examination is a separate question from whether it is mathematically coherent.
What the paper proposes
The arXiv preprint (2606.01632) combines two ideas that have each been productive elsewhere but have not, until now, been jointly applied to patent valuation. The first is graph conditioning: representing the relationships between patent claims, prior art, and technical features as a structured graph rather than a flat feature vector. The second is Shapley attribution, borrowed from cooperative game theory, which allocates credit to individual players (here, individual patent features) based on their marginal contribution to the total value.
Shapley values have a strong theoretical pedigree. They are the unique allocation satisfying a set of axioms (efficiency, symmetry, dummy player, additivity) that most economists consider reasonable for fair division problems. The catch is that exact Shapley computation is exponential in the number of players, so practical implementations use approximation. The paper’s title signals a hierarchical approach to that approximation problem, which likely decomposes the computation across subgraphs or clusters of features rather than enumerating all coalitions.
What makes this worth attention is not the novelty of Shapley values themselves, which have been applied to everything from machine learning interpretability to supply chain profit-sharing. It is the application domain. Patent valuation in litigation and M&A contexts has resisted automation precisely because the output has to survive adversarial scrutiny, not just correlate with market prices.
The auditability problem
A Shapley-attributed valuation has mathematical properties that look attractive on paper. Each feature gets a credit score. The scores sum to the total estimated value. The allocation satisfies fairness axioms. For an internal portfolio analysis or a quick screening tool, this may be sufficient.
But patent valuation in practice is not a portfolio-screening exercise. It runs through two high-stakes adversarial processes: litigation damages and M&A due diligence.
In litigation, a damages expert testifies to the reasonable royalty or lost profits attributable to a patent. That testimony is subject to Daubert challenge in US federal courts, where the opposing counsel argues the methodology is unreliable. The expert must explain, in language a judge and jury can follow, how they arrived at the number. A Shapley attribution graph with hierarchical approximation layers is not obviously more explainable than a neural network. The axioms are elegant; the approximation pipeline that actually computed the numbers may not be.
In M&A due diligence, the acquirer’s IP counsel needs to assess whether the patent portfolio justifies the valuation premium. Model-generated valuations can speed up initial screening, but the buyer’s team will still want to understand the basis for any number they use in negotiations. A per-feature decomposition helps more than a single score, but only if the features themselves are legible, the graph structure is documented, and the approximation error is bounded and reported.
Where automated patent valuation stands today
Commercial patent analytics platforms, including Anaqua and PatSnap, already offer automated patent scoring. These tools typically produce portfolio-level metrics (citation counts, family size, forward citation velocity, renewal status) that correlate with economic value but stop short of producing a dollar figure defensible in court.
The gap between a relevance score and a damages estimate is wide. A platform can tell you that a patent is in the top decile of its technology class by citation impact. Translating that into a reasonable royalty rate requires assumptions about the hypothetical negotiation, the entire market value rule, and the specific contribution of the patented feature to the accused product. Those assumptions are where the expert earns their fee.
Patent appraisal is the kind of expert-intensive workflow that automation targets: high labor cost, repetitive structure, and a small number of practitioners who command premium rates. The question is not whether automation reaches patent valuation. It already has, at the screening level. The question is whether it crosses the credibility threshold from internal tool to adversarial evidence.
What IP teams should demand
Any IP team evaluating model-generated patent valuations, whether based on Shapley attribution or another approach, should apply a basic auditability checklist before relying on the output for anything beyond internal prioritization:
Explainability of the graph structure. If the model conditions on a graph of patent features and prior art relationships, the graph must be inspectable. Which nodes represent claims, which represent prior art, and which edges encode what relationships? A black-box graph defeats the purpose of attribution.
Approximation error bounds. Exact Shapley computation is intractable for realistic patent feature sets. Any implementation uses approximation. The approximation error must be quantified and reported, not buried in a supplement.
Sensitivity to graph construction choices. Different graph representations of the same patent will produce different Shapley allocations. The framework should document how sensitive the output is to the choice of graph structure, and the user should understand that the graph is a modeling decision, not a ground truth.
Comparison to expert-appraised benchmarks. If the framework has been validated, the validation should compare model output to actual transaction prices or court-awarded damages, not just to other model outputs. No such validation is reported for this preprint as of June 2026.
Adversarial testing. Before using model output in litigation or negotiation, the producing party should hire opposing counsel to attempt to dismantle the valuation. If the model cannot survive a motivated attacker who understands the methodology, it will not survive a Daubert hearing.
The appeal of Shapley-attributed patent valuation is real. Per-feature credit assignment is more informative than a single score, and the mathematical foundation is stronger than the ad-hoc weighting schemes that commercial platforms currently use. But mathematical fairness and legal defensibility are different properties. The paper’s contribution is a framework for the former. Whether it can be made to satisfy the latter depends on implementation details, documentation standards, and court acceptance that no preprint can deliver.
Frequently Asked Questions
How does Shapley attribution compare to the income, market, and cost approaches patent appraisers already use?
Traditional patent valuation relies on three methods: the income approach (discounted future royalties), the market approach (comparable transaction multiples), and the cost approach (replacement cost of developing the invention). Shapley attribution does not replace any of these. It decomposes the output of one of them into per-feature contributions. A Shapley framework still needs an underlying value estimate to allocate, so it functions as a supplementary layer on top of an existing method rather than a standalone valuation technique.
What happens to Shapley credit when patents in a portfolio share overlapping claims?
Overlapping claims across related patents create a double-counting risk. If two patents each receive Shapley credit for the same technical feature, the portfolio total can exceed the actual economic value. The paper’s graph-conditioned structure could theoretically mitigate this by representing shared features as edges between patent nodes rather than duplicated attributes, but whether the framework handles cross-patent normalization is unconfirmed without the full methodology. Single-patent Shapley attribution ignores a problem that only surfaces at the portfolio level.
Do the Daubert admissibility concerns apply outside US courts?
No. The Daubert standard is specific to US federal courts. Proceedings at the Unified Patent Court in Europe generally do not subject expert testimony to a formal reliability gate. Japanese courts take a judge-led approach to evidence with limited adversarial cross-examination of experts. A Shapley-based valuation could face lower procedural barriers in those jurisdictions, though it would still need to persuade a judge or panel that the underlying methodology is credible.
Have courts accepted other algorithmic valuation methods in litigation?
Real estate automated valuation models went through a comparable adoption arc. US courts accepted AVMs for property tax assessment appeals in several states during the 2010s but rejected them for setting damages in foreclosure and breach-of-contract suits, where judges required a human appraiser’s judgment on specific property conditions. That split (algorithmic methods for screening and tax purposes, human expert testimony for adversarial damages) maps closely onto the trajectory automated patent valuation appears to be following.