Patent value assessment

The challenge of quantifying patent value has necessitated diverse measurement approaches, each targeting different outcome variables that serve as proxies for intellectual property value. Forward citation counts represent the most established target variable [1,18,19], though they primarily capture technological rather than commercial significance [20]. Alternative approaches target patent maintenance decisions [3], network centrality measures [2], and quality scores [14]. These diverse targets reflect fundamental trade-offs: citation-based outcomes may reflect academic rather than commercial interest, maintenance decisions depend on firm-specific strategies, and network centrality emphasizes positional importance rather than market potential [5]. Patent licensing outcomes, in contrast, represent explicit market transactions in which parties negotiate monetary value for intellectual property rights, providing direct indicators of commercial value that distinguish them from these indirect proxies.

Temporal dynamics in patent analytics

The temporal dimension has gained increasing recognition in patent value assessment. Survival analysis of patent maintenance reveals that value indicators evolve over time, with family size emerging as the strongest early predictor while forward citations show delayed effects [3]. Recent advances in machine learning-based survival analysis have enabled more accurate predictions of patent lifespan by incorporating multidimensional characteristics including technical, legal, market, patentee features, and textual semantic features [21]. Citation timing analysis demonstrates that patents with shorter technology cycle times, larger patent families, and higher claim counts receive citations more quickly [4].

However, licensing timing differs fundamentally from citation or maintenance patterns. Research on university technology licensing reveals complex strategic considerations, showing that licensees with strong technological capabilities tend to pursue early licensing before patent rights are fully established, though this tendency is moderated by appropriation concerns when significant technological overlap exists [7]. Survival analysis using Cox proportional hazards models has been applied to healthcare patent licensing prediction, integrating topic modeling with expert judgment [15]. Unlike citations that reflect technological building upon prior work or maintenance decisions at predetermined intervals, licensing events exhibit heterogeneous timing patterns influenced by technology readiness levels, market development cycles, commercial viability assessments, and competitive dynamics emerging at various points during a patent’s lifetime. This distinct temporal complexity suggests that licensing prediction requires specialized modeling approaches that can capture the unique timing characteristics of commercial technology transfer events.

Patent licensing prediction models

The PatVal-EU study [22] provided empirical evidence showing that about 11% of patents were licensed, with an additional 7% of patent owners willing but unable to license due to transaction costs, demonstrating that factors affecting willingness to license differ from those affecting actual licensing probability. Predictive modeling approaches have evolved from early work using neural networks for patent transferability [13] to sophisticated machine learning approaches integrating patent-based features [11] and Graph Convolutional Networks modeling licensing relationships [12]. Recent research has emphasized patent scope measures, with citation network analysis revealing that dynamic characteristics, particularly inward degree centrality, are significant factors in transfer prediction [23]. Transformer-based approaches using BERT embeddings [14] and domain adaptation techniques [24] have further advanced the field. However, these approaches treat patent transfer as a classification problem without accounting for censoring effects, where non-transferred patents during observation periods may eventually transfer in the future, potentially leading to biased predictions that conflate truly non-transferable patents with those whose transfer events have not yet been observed.

Deep survival analysis

Deep survival models, developed primarily in medical and social science domains, extend classical survival analysis with neural networks, enabling analysis of high-dimensional feature spaces and nonlinear relationships while properly handling right-censored data. Established frameworks include DeepSurv [8], which implements neural Cox proportional hazards models, and DeepHit [25], which models survival distributions through discrete-time competing risks frameworks. Despite their success in medical and social applications, deep survival models have seen limited adoption in patent analytics. Marusaki et al. [16] applied them to patent maintenance duration prediction, and Rhee et al. [17] recently demonstrated DeepSurv’s superiority over Random Survival Forests and classical Cox models on Korean patent transaction data, establishing neural survival models as viable for patent transaction timing prediction.

Standard survival models assume a single homogeneous population where all subjects will eventually experience the event of interest. Classical cure models [9] address population heterogeneity by explicitly modeling both susceptible and non-susceptible subgroups, with recent deep learning extensions demonstrating feasibility for recidivism prediction [10]. In patent licensing contexts, this framework may be particularly relevant as many patents could be inherently unlicensable due to technical obsolescence, narrow claim scope, or market misalignment. However, the relative merits of standard deep survival models versus cure rate extensions for patent licensing prediction remain unexplored. Moreover, prior work has not systematically compared survival modeling against static binary classification approaches to quantify the value of explicit temporal modeling. This study addresses these gaps by providing, to the best of our knowledge, the first systematic comparison of these approaches—including cure rate models and static classification baselines—on large-scale patent licensing data characterized by extreme class imbalance and heavy censoring.

Data acquisition

We retrieved the complete bibliographic, textual and citation record of every utility patent published in the United States between 2000 and 2021 from the Google Patents Public Data tables [26]. Licensing data was obtained from the USPTO Patent Assignment Dataset [27]. Because only two licensing events were found in this dataset after 2018, we restricted the licensed sample to USPTO events dated up to the end of 2017 (Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Annual patent volume and licensing ratio trends from 2000–2021.

The dual-axis chart shows the total number of patents (blue bars, left y-axis) and corresponding license ratio percentage (orange line, right y-axis) over a 21-year period.

https://doi.org/10.1371/journal.pone.0355826.g001

Patent volume shows consistent growth from approximately 158,000 patents in 2000 to peak levels exceeding 355,000 patents in 2019–2020. The apparent decline to around 251,000 patents in 2021 reflects data truncation due to incomplete observation of the full year 2021 (the observation period ended on 28 September 2021). Similarly, the license ratio shows a decreasing trend from 1.5% in 2000 to approximately 0.0% by 2018–2021. This declining pattern in license ratios for more recent years reflects data censoring effects, as there is typically a time lag between patent publication and subsequent licensing activities. Both the reduced patent count in 2021 and the consistently declining licensing ratios in recent years are artifacts of the observation period ending partway through 2021, rather than representing actual decreases in patenting activity or licensing rates.

For each normalized publication number, the dataset contains:

• Application date, publication date, document type and claim count;
• The full abstract;
• The ordered concatenation of all claim sentences;
• The complete set of current Cooperative Patent Classification (CPC) group identifiers;
• Disambiguated assignee names (organizations or inventors) recorded at the time of patent grant, which remain fixed and do not reflect subsequent ownership transfers;
• USPTO-internal backward citations; and
• Voluntarily filed license information (licensee name and license date).

For USPTO licensing data, we use the execution date of the licensing agreement as the licensing event time. Survival time is calculated as the number of days from the patent application date to the execution date. In cases where the execution date preceded the application date (1,468 cases, 2.65% of licensed patents), we set the survival time to zero during model training to accommodate the non-negative survival time requirement. Among licensed patents, the median time-to-licensing was 1.6 years (IQR: 0.3–4.7 years), with 42.2% licensed within 1 year, 64.3% within 3 years, and 76.6% within 5 years of application. This concentration of early licensing events is consistent with prior findings by Graham et al. [27], who observed that many ownership changes occur within 3 years of grant.

The USPTO Patent Assignment Dataset captures licensing transactions through voluntary disclosure filings submitted by patent holders and licensees. Parties are not legally required to record licensing agreements with the USPTO, although they may choose to do so to establish public notice of their rights or for other strategic reasons. Consequently, the dataset represents a subset of actual licensing activity, with many licensing transactions remaining unrecorded. This voluntary nature means that patents labeled as unlicensed in our dataset may include cases where licensing occurred but was not disclosed, creating an incomplete but valuable record of observable market transactions in intellectual property rights.

When multiple license events were recorded for a single patent, we used the earliest license date as the event time for survival analysis. This approach treats the first observed licensing transaction as the primary event of interest, consistent with the economic interpretation that initial licensing represents the patent’s entry into commercial use. This methodology ensures that each patent contributes exactly one survival time to the analysis, maintaining the standard assumptions of survival modeling while capturing the most temporally proximate licensing event to patent publication. All underlying data are publicly available and do not contain personally identifiable information.

We employ a temporal split strategy to ensure a realistic evaluation of the out-of-sample. Given the large sample size (n = 2,903,422 for training; n = 299,382 for validation), cross-validation was not necessary for the estimation of stable parameters. Our approach mirrors real-world deployment scenarios in which models trained on data up to a certain point are used to make predictions about subsequent periods.

Patents were ordered chronologically and partitioned as follows.

• Training period (2000–2014). All patents published between 2000 and 2014 (n = 2,903,422; licensed = 43,502, 1.50%) constituted the development pool. This dataset was used for weight updates.
• Validation period (2015). All patents published in 2015 (n = 299,382; licensed = 4,070, 1.36%) constituted the validation set. This dataset was crucial for parameter tuning and model selection.
• Test period (2016–2017). All patents published between 2016 and 2017 (n = 624,129; licensed = 7,872, 1.26%) were held out ex ante and used for USPTO-based evaluation.

This strictly time-stratified protocol mirrors real-world deployment, preventing information leakage from future documents into model training.

Model architectures

We evaluated four modeling approaches representing key methodological variations in survival analysis and patent prediction: (1) classical Cox regression baseline, (2) deep neural network for static binary classification, (3) DeepSurv implementing neural Cox proportional hazards, and (4) Cox cure rate model.

These four models are designed as a controlled, ablation-style comparison. The three survival models—the classical Cox baseline, DeepSurv, and the Cox cure rate model—share a common continuous-time, Cox-based proportional-hazards foundation and differ only in individual modeling components, namely feature representation (hand-crafted versus learned) and the presence of an explicit cure fraction; the deep neural network is included as a static binary-classification reference, so that its contrast with the survival models isolates the value of temporal modeling. Restricting the comparison to this shared foundation is what allows performance differences to be attributed to specific components: survival paradigms built on fundamentally different assumptions—such as DeepHit [25], which uses a discrete-time competing-risks formulation, and Random Survival Forests [28], which is tree-based and does not assume proportional hazards—would change the modeling foundation itself and confound this attribution, and are therefore outside the present scope and deferred to future work.

By design, each approach is also paired with the feature construction under which it is conventionally deployed: the classical Cox model uses interpretable hand-crafted features, whereas the deep models use learned embedding representations. This is an intentional choice that represents each paradigm in its standard, realistic form rather than forcing an artificial common input; in particular, supplying high-dimensional dense text embeddings directly to a linear Cox model would not reflect how such a model is used in practice. As a result, the comparison between the classical Cox baseline and the deep models is a contrast between entire modeling pipelines (method together with its typical feature representation), not between architectures in isolation, whereas the three deep models (DNN, DeepSurv, and Cox Cure) share an identical feature encoder, so comparisons among them isolate the effect of the temporal and cure-modeling components.

The classical Cox baseline implements standard Cox proportional hazards with linear predictor where X represents a 999-dimensional hand-crafted feature vector comprising: scaled claim count (1), scaled backward citation count (1), 768-dimensional text embeddings from BERT-for-Patents [29], binary indicators for all 129 CPC codes, and binary indicators for the top 100 most frequent assignees. These features are processed directly without neural encoding.

Neural models (DNN, DeepSurv, Cox Cure) employ identical feature encoders processing 1042-dimensional input vectors comprising: scaled claim count (1), scaled backward citation count (1), 768-dimensional text embeddings from BERT-for-Patents [29], 16-dimensional mean-pooled CPC classification embeddings (vocabulary size 129), and 256-dimensional mean-pooled patent assignee embeddings (vocabulary size 207,499). The shared encoder consists of fully connected layers () with ReLU activations and dropout (0.30) after each layer, feeding into model-specific prediction heads.

The DNN model uses a prediction head consisting of fully connected layers () with ReLU activations, dropout (0.30), and sigmoid output. Three separate instances are trained independently for 1-year, 3-year, and 5-year prediction horizons.

DeepSurv implements neural Cox proportional hazards with an identical prediction head architecture but linear output representing log hazard ratio , with a single model predicting across all time horizons.

The Cox cure rate model extends DeepSurv with two parallel prediction heads from the shared encoder (Fig 2). The licensability predictor uses layers () with ReLU activations, dropout (0.30), and sigmoid output, producing probability that a patent is ever licensable. The Cox linear predictor uses identical architecture to DeepSurv, producing log hazard ratio for licensable patents. An additional learnable baseline hazard network () with ReLU activations and Softplus output models h₀(t).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Semi-parametric Cox cure rate model architecture.

The model processes patent features through a shared encoder ( dimensions) that feeds into three specialized components: a licensability predictor outputting probability p, a Cox linear predictor outputting log hazard ratio , and a learnable baseline hazard function h₀(t). The final licensing probability combines these components as , where is the survival function with Cox proportional hazards framework.

https://doi.org/10.1371/journal.pone.0355826.g002

The final licensing probability at time t combines both components:

(1)

where the survival function follows the Cox proportional hazards framework:

(2)

and is the cumulative baseline hazard.

Training procedure

Neural models employed Adam optimizer (learning rate 10⁻⁴, weight decay 10⁻⁴, batch size 256) with early stopping (patience 10 epochs) and maximum 50 training epochs. Cox Cure required additional optimization techniques for training stability: learning rate scheduling (ReduceLROnPlateau with factor 0.5, patience 3), gradient clipping (maximum norm 5.0), and 3-epoch warm-up with linear learning rate scaling from to full rate. DNN and DeepSurv trained stably without these techniques. Classical Cox baseline used LBFGS optimization with L2 regularization (, tolerance 10⁻⁹, maximum 100 iterations).

DNN employed binary cross-entropy loss with class weights inversely proportional to class frequencies:

(3)

where addresses extreme class imbalance.

DeepSurv implemented the Cox negative log-partial likelihood:

(4)

where denotes the risk set at time .

Cox Cure combined licensability classification with Cox survival modeling. For observed events ():

(5)

where . For censored observations ():

(6)

The total log-likelihood is:

(7)

The cumulative baseline hazard was computed via trapezoidal rule with 20 grid points during training and 50 during validation.

Classical Cox baseline optimized the standard Cox partial likelihood using hand-crafted features as implemented in scikit-survival [30].

Experimental setup

We evaluated out-of-sample performance on the test set (2016–2017) using four modeling approaches to address fundamental questions about patent licensing prediction. The comparison evaluates whether deep learning provides value beyond classical approaches, whether temporal modeling improves over static classification, and whether explicit cure rate modeling enhances performance when many patents may never be licensed.

Our evaluation employed Recall@K as the primary metric, measuring the proportion of actual licensing events captured within the top K% of patents ranked by predicted cumulative licensing probability. We evaluated Recall@K at 1, 3, and 5 years post-publication with retrieval thresholds of 1%, 5%, 10%, 20%, and 30% to provide comprehensive coverage of practical use cases. Supplementary metrics included Lift@K (ratio of precision to base rate, quantifying business value) and concordance index (C-index, measuring rank-ordering capability for survival models).

We focus on recall-based metrics rather than precision-based measures because patent licensing data in the USPTO Patent Assignment Dataset reflects voluntary disclosures rather than mandatory reporting requirements. Patents labeled as unlicensed may include cases where licensing occurred but was not disclosed or captured in available records, creating false negatives that make precision calculations unreliable. Recall-based metrics remain valid as they measure the proportion of known licensing events successfully identified, providing actionable guidance for patent portfolio screening despite incomplete reporting.

For patents with multiple recorded licensing events, we used the earliest license date as the survival event time, treating the first observed licensing transaction as the primary event of commercial interest.

Statistical significance of performance differences was assessed using patent-level bootstrap resampling (B = 1,000 resamples, seed = 42). For each resample, a bootstrap sample of n = 624,129 patents was drawn with replacement from the test set using a single index shared across all models, ensuring paired comparisons. Within each resample, Recall@10% was computed for every model at the 1-, 3-, and 5-year horizons, and each pairwise model difference was averaged across the three horizons to yield one horizon-averaged difference per resample; the resulting B horizon-averaged differences form the bootstrap distribution for each comparison. Two-sided bootstrap p-values were computed by recentring this distribution at zero under the null hypothesis and measuring the proportion of recentred values at least as extreme as the observed mean difference, bounded below by the resolution 1/B = 0.001. Ninety-five percent confidence intervals were obtained by the percentile method.