Invented by Matthew BEERS, Elvir Causevic, Ocean Tomo LLC
The Ocean Tomo LLC invention works as follows
A machine-learning based artificial Intelligence device is disclosed for estimating patent quality such as the patent term or lifetime. A device of this type may be used to receive a set of patent data, and then generate a list binary classifiers. The final binary classifiers are found using a heuristic algorithm such as an artificial neural net (ANN) or genetic algorithm. This is done by iteratively maximising a yield based on a cost function. The device can then receive information about a target patent, and provide an estimate of the quality of that patent based on the final binary classifiers.Background for Patent quality metric based on machine learning
Field of Invention
The present disclosure is a system that includes a CPU, a storage device and a database of patent applications or grants and other data relevant to the estimation of patent quality using machine learning algorithms based on nonlinear models.
Related Art
On the basis of historical data, attempts have been made to evaluate or estimate the expected value or life expectancy of a new patent or patent application. It is difficult to find quantitative metrics that can be tested and reproduced. It is also difficult to combine quantitative factors from a vast universe of patent data in order to estimate a patent’s value, or a patent’s estimated life. This is due to the large number of factors that are patent-related or patent application-related. Finding the optimal combination of factors to produce a patent life/quality profile that maximizes or optimizes patent quality has been a challenging task.
The existing methods of rating patent quality depend on linear combinations of simple variables (e.g. The number of forward citations in combination with the age of the Patent or the traditional linear and statistical mathematics tools based on iterative human driven factors selection processes. By using ‘brute force’ To find the most relevant factors, you must examine every factor in combination with every other factor. Solution space is all possible combinations between factors and coefficients. The only way to arrive at the best solution with brute force is to iterately consider all elements in the solution space. This is called brute-force computing. For a simple problem with A and B and no coefficients the algorithm would have to consider:
Using a brute force approach, every additional factor or combination of factors increases the complexity and processing time exponentially.
A machine-learning-based artificial intelligence device to estimate patent quality is disclosed. This device could include:
The heuristic model may be included in the artificial neural network. The iterative maximizing may include changing the hidden layers of an artificial neural network.
The maximizing can be done iteratively using a genetic algorithms or an artificial neural networks model with a genetic algorithim.
The cost function can be an operating characteristic of a receiver and the yield can be calculated by calculating a region under a curve.
The estimate of patent quality could be an estimation of the lifetime of a patent.
The patent data retriever may be configured to accept a second set patent data, which includes at least one patent application data or patent data for multiple patents.
wherein the device can be configured to test the validity of the final binary classifiers based on the second set patent data.
The device may include an information manager for the user configured to receive information on a patent target and report the estimated patent quality based on the final binary classifiers.
Also envisaged is a method that comprises such a device, in combination with a secondary device communicatively linked to the device via a network. This second device can include:
The following description and the drawings accompanying the disclosure explain further aspects.
A computer system, a network platform, including a server, a processor-readable media, a method and means for implementing a method according to the current disclosure uses a set algorithms based upon training data received from a database containing patent information. This includes granted patents and applications, as well as other relevant data, such aggregate data on patent examination, grant and opposition, abandonment and annuity/maintenance fees. The present disclosure uses a set of binary classifiers that predicts a measure of quality in patents, such as whether an issued patent will continue to be maintained for the duration of the patent. Some other measures of quality include whether or not a patent is licensed, upheld in court and so on. The system can also be used to predict the quality of intangible assets.
The system selects the features by using a heuristic procedure, such as genetic algorithms or simulated annealing. The algorithms take as inputs a list of features identified in information from a number of patents or patent applications. The search is iterated over the data set input. Each feature receives a random or pseudo-random weight. The heuristic determines the optimality of the feature weights at each iteration by evaluating a cost function. Before starting the next cycle, the final step is to modify the feature weights. The algorithm determines the mutation computation. In a genetic algorithm the weights will be mutated pseudo-randomly or randomly, while in simulated annealing, they are modified according an energy transition equation. The heuristic ends after a certain number of iterations, or when the change in the feature weight falls below a threshold. The threshold can be an experience-based parameter that is set by the user. The final factors selected are used to train the binary classifier.
The cost function used by the search procedure is used to optimize an area under the Receiver Operating Characteristics (ROC) Curve (FIG. 1). At each iteration the factors that are currently being considered?those with non-zero weight?are used to “train” a binary classification.
A larger sample may produce a more accurate model. To produce good sampling sets, for example, 100,000 records of patents can be divided into three sets. Sets need not be equal in size. It will be accepted that more or less than 100,000 records can be used. The size of training, validation and test sets does not have to be a strict guideline. It can depend on the population size. The training sets may be different for four million U.S. active patents and 1.5 million EPO active patents, for example. Machine learning can be tailored for country and region. For example, the estimated patent quality/patent value returned for a query may only be based on the data from the patent information of the country or the region where the query was made. The patent value/estimate may also be customized to a specific field of technology, such as mechanical arts, pharmaceuticals or chemical fields. The patent value/estimate returned for a patent in a field of technology or science endeavor X can be based solely on the data obtained from patents/patent application of that field.
Click here to view the patent on Google Patents.