Analysis of the Impact of Machine Learning Models on Intellectual Property Protection from a Jurisprudential Perspective

Machine learning is a discipline that studies how to make computers simulate or implement human learning behaviors. And machine learning model training is an important part of machine learning, which refers to the given training data and corresponding labels, so that the computer learns the relationship between inputs and outputs, and predicts or classifies according to the learned patterns. Machine learning models mainly include features, labels, algorithms and other basic elements. Machine learning model plays an important role in the development of modern technology. It is the core of the field of artificial intelligence, through the learning and analysis of data, can simulate the human thinking and decision-making process, and has an important impact on the protection of intellectual property rights [1–4].

Intellectual property rights (IPR) refers to “the exclusive rights of the right holder to the fruits of his or her intellectual work”, which are generally valid only for a limited period of time. Various intellectual creations such as inventions, literary and artistic works, as well as logos, names, images, and designs used in commerce can be considered as intellectual property owned by a person or organization. Intellectual property is the ownership of the fruits of intellectual labor, and intellectual property protection is the law and system to protect and maintain intellectual property rights, aimed at encouraging innovation and promoting social and economic development, and with the development of science and technology, artificial intelligence has brought new challenges to intellectual property protection [5–8].

Literature [9] lies in examining deep neural network IP security techniques. A taxonomy in favor of deep neural network IP protection techniques and the challenges faced by these methods and their ability to withstand attacks are proposed in terms of size and methods. Threats to deep neural network IP security techniques are also explored in terms of model modifications and other aspects. Literature [10] provides a survey of current DNN IP. It not only outlines the deployment models of DNN models and the problems of DNN IP protection, but also categorizes and introduces the protection strategies, compares the protection methods, and discusses the topics related to the future research of DNN IP protection. Literature [11] proposes a method to protect the IP of machine learning models by fingerprinting classification boundaries. By describing the difference between fingerprinting and watermarking, it is emphasized that the model owner's intellectual property can be protected robustly by using the fingerprint of the model classification boundary. Literature [12] examines the complex relationship between IP and the AI domain. Research including the utility of patent law for generating inventions in AI systems and analyzing the role of trademarks in digital drives will be explored, providing insights into the mechanisms needed to preserve IP principles while fostering innovation. Literature [13] provides a decomposition of the machine learning process and categorizes this decomposition into steps such as, problem definition, data collection, and model selection. It is emphasized that most machine learning jumps back and forth between these steps and rotates through model construction and improvement rather than a straight linear path. Literature [14] aims to provide an overview of Artificial Intelligence in general and machine learning in particular. Introducing, the application of machine learning in the field of intellectual property has sparked a lot of debate in legal academia, but it remains ambiguous in many technical aspects, thus leading to a confusing situation from time to time.

Literature [15] emphasizes the importance of balancing intellectual property rights and ethical behavior in the context of artificial intelligence. Based on a comprehensive historical review, the creation and application of ethical AI is reflected upon while protecting intellectual property, and the concepts of intellectual property and ethical behavior in the context of AI are elaborated. Literature [16] proposes an obfuscation framework for the protection of neural networks, which guarantees that only users who are authorized by a trusted hardware device can run DL applications using the published model. An experimental evaluation shows that the accuracy of any deep learning model that is not authorized to use this obfuscation is degraded, as well as the robustness of the HPNN framework to model fine-tuning type attacks. Literature [17] explores the impact that Orland ML has on intellectual property laws. The use of ownership and authorship of AI and machine-generated inventions in patent law is discussed based on methods such as literature review, with a view to contributing to the discussion on AI, machine learning and intellectual property law. Literature [18] proposes methods of intellectual property protection based on the application of machine learning and IoT in green manufacturing molding technology and measures to stop the illegal theft of the technology. Experiments and case studies verified that machine learning and IoT not only improve the manufacturing capability of green manufacturing, but also provide a basis for intellectual property protection. Literature [19] describes model watermarking for protecting the intellectual property of machine learning participants. It is emphasized that various watermarking solutions have been proposed in recent years to cope with model extraction, theft or misuse. These techniques are discussed and their use is suggested.

The research firstly elaborates on the application of machine learning model in intellectual property protection, and in order to make the comprehensive evaluation of intellectual property scientific, the evaluation index system of legal intellectual property is constructed from five aspects, namely, the quantity of intellectual property, the quality of intellectual property, the development of intellectual property, the implementation of intellectual property and the management of intellectual property. Then the GA-BP neural network model is designed, specifically, after selecting and processing the training samples, the BP neural network construction is carried out, considering the advantages and disadvantages of the BP neural network, and the disadvantages of the BP neural network are optimized by using the GA genetic algorithm. Finally, the GA-BP neural network prediction model proposed in this paper is tested for performance and specific empirical tests to explore the impact of machine learning models on intellectual property protection.

2

Proposal for a machine learning model for intellectual property protection

2.1

Application of Machine Learning Models to Intellectual Property Protection

Machine learning algorithms have been widely used in the computer science and statistics literature, where econometrics focuses on reducing estimator bias and machine learning algorithms focus on minimizing prediction error for test sets, machine learning's success in intelligent tasks is largely due to the fact that it is able to find complex structures that have not been pre-specified, and can manage to fit the data with a complex and flexible functional form, rather than simply overfitting, so it can find functions that still perform well out of sample [20]. Such algorithms are becoming more and more popular among scholars as the technology continues to evolve, which provides them with a new toolbox that can be used for purely predictive tasks [21].

Machine learning algorithms based on the analysis of historical data to make predictions about future outcomes can be used to estimate demand, forecast price changes, predict customer behavior and preferences, assess risk and predict endogenous or exogenous shocks that may affect the market environment. All of this information is valuable for improving decision-making, enabling companies to plan their business strategies more effectively and develop innovative and customized services. Machine learning can also be used to optimize business processes, allowing companies to respond effectively to market conditions by reducing production and transaction costs, segmenting consumers, or setting optimal prices to gain a competitive advantage. The particular ability of algorithms to optimize processes is a result of their automated nature and powerful computational capabilities, which allow them to handle large data sets, respond quickly, and do so at a lower cost than when the same task is performed by a human.

2.2

Establishment of evaluation index system for intellectual property protection

The construction of an intellectual property evaluation index system from the perspective of jurisprudence should fully consider the practical application of intellectual property in jurisprudence. For law, intellectual property rights mainly include intellectual property rights, trademark rights, copyrights, trade secrets and so on. This paper follows the principle of constructing the evaluation index system, and constructs the evaluation index system of intellectual property rights from the aspects of intellectual property rights quantity, intellectual property rights quality, intellectual property rights development, intellectual property rights implementation and intellectual property rights management on the basis of sufficient literature research and actual investigation. The intellectual property evaluation index system is shown in Table 1.

Table 1.

Intellectual property evaluation index system

	Primary indicator	Secondary indicator
Intellectual property evaluation index system	Quantity indicator(C1)	Patent application(C11)
		Application rate of patent application(C12)
		Publication number(C13)
	Quality indicator(C2)	Patent authorization(C21)
		Patent licensing rate(C22)
		International patent application(C23)
		Copyright registration number(C24)
	Development indicator(C3)	Total number of professionals(C31)
	Development indicator(C3)	The professional personnel proportion of employees(C32)
	Implementation index(C4)	Patent implementation(C41)
	Implementation index(C4)	Patent yield(C42)
	Management index(C5)	The number of patent infringement cases(C51)
		Number of copyright infringement cases(C52)
		The number of cases of intellectual property(C53)
		Case rate of intellectual property(C54)

2.3

Proposed GA-BP Neural Network Modeling

2.3.1

Selection and processing of training samples

Neural networks, because of their highly nonlinear input-output relationship, generally speaking, the preprocessing of learning data is not necessary. However, in view of better improving the learning efficiency, accelerating the convergence speed, improving the ability of neural networks to quickly and accurately process the learning data, and thus improving the evaluation of the longitude, it is essential to take certain learning data preprocessing.

Normalization processing In general, the normalization of data is one of the commonly used ways of data preprocessing, through the normalization process, all the data can be converted to a number between (0,1), which can greatly reduce the range of data recognition, the prediction success rate has been greatly improved.

The normalization method is as follows: 1 $X_{i j} = \frac{X_{i j} - M i n (X_{i j})}{M a x (X_{i j}) - M i n (X_{i j})} (i = 1, 2, 3; j = 1, 2, \dots, 7)$

In order to avoid bias in the model fit due to the different magnitudes of the indicators, the data need to be processed dimensionless. The processing method is as follows: 2 ${X^{'}}_{i j} = \frac{X_{i j} - \bar{X_{j}}}{S_{j}^{2}}$ where $\bar{X_{j}}$ is the mean of the variable in column j. $\bar{X_{j}} = \frac{1}{n} \sum_{i = 1}^{n} X_{i j}$ .

$S_{j}^{2}$ is the sample variance of the variable in column j, $S_{j}^{2} = \frac{1}{n - 1} \sum_{i = 1}^{n} {(X_{i j} - X_{j})}^{2}$ .

Principal component analysis through the above research, we have selected a large number of evaluation indicators of intellectual property value, however, in this part of the indicators have a part of the indicators have linear correlation between the indicators, and a part of the indicators have the repetition of information between the indicators, and the existence of these shortcomings may affect the effectiveness of the output results in the neural network, but if you simply remove the data represented by these indicators, you can not very well show the comprehensiveness of the evaluation index system [22]. In the international popular method to deal with the problem of linear correlation and repeatability, the principal component analysis method is generally adopted. Through the principal component analysis of the indicator data, the above indicators are divided into several comprehensive indicators that are no longer relevant, and the processed indicators will provide a comprehensive mapping of the content of the indicator system, while solving the problem of repetitive data in the original data. In addition, in the neural network analysis, in order to achieve better simulation results, the general principle in the selection of indicators is to choose as few indicators as possible, which on the one hand reduces the data storage space of the model, and on the other hand improves the effectiveness of the program. Therefore, before training the BP model, the selected indicators should be processed using the principal component analysis method to improve their output.

In addition, the process of principal component analysis can be carried out in the selection of the number of comprehensive indicators, the selection criteria are usually based on the principal component factors corresponding to the eigenvalue greater than 1, and the general cumulative variance of the contribution rate is required to be more than 75%. The principal component data can be calculated by substituting the original indicator data into the principal component expression, as reflected in the SPSS processing matrix of principal component coefficients.

2.3.2

BP Neural Network Construction

1)

Determine the number of BP neural network layers

BP neural network is a multilayer network composed of input layer, hidden layer and output layer, generally it has one input layer and one output layer, respectively, for receiving and outputting information, and the number of layers of the hidden layer is different with the different characteristics of the problem, and the number of layers is also different. When the complexity of the information between the data is higher, and it is difficult to learn the generalization of the features within the data, the BP neural network with a larger number of hidden layers can be processed by a stronger information processing ability. However, with the increase in the number of hidden layers, the more complex the structure of the BP neural network, the time required for processing increases accordingly, and the convergence speed is slower, which affects the efficiency of processing problems [23]. For general complex problems, with a hidden layer, that is, the three-layer network structure of the BP neural network has the learning ability is sufficient to achieve any m -dimensional data to n - dimensional data mapping, for this paper has been sufficient to achieve the input benefit index value to get the corresponding project portfolio benefits, so as to realize the project portfolio benefit prediction. Therefore, this paper adopts the BP neural network with three-layer network structure to predict the benefits of the project portfolio.

2)

Determine the BP neural network structure

Since each layer of the BP neural network consists of neurons, after determining the number of layers of the BP neural network, the number of neurons that each layer has should be determined to determine the BP neural network structure. Because the number of neurons in a BP neural network affects the performance of the BP neural network, it is very important to reasonably determine the number of neurons in each layer of the topology, otherwise the performance of the BP neural network prediction will be affected. Among them, the number of neurons in the input and output layers is usually set according to the actual situation.

(1)

Number of neurons in the input layer

The number of neurons in the input layer of the BP neural network is usually set according to the size of the sample data of the problem being studied. Generally speaking, the number of neurons in the input layer is as many as the number of indicators in the constructed indicator system. Therefore, the number of neurons in the input layer set in this paper is equal to the number of benefit indicators, and the number of neurons in the input layer of this study is 30 according to the benefit indicator system constructed in Chapter III.

(2)

Number of neurons in the hidden layer

The BP neural network's hidden layer is the primary component in mapping the relationship between input and output, and its neuron nodes have a significant impact on its operation. The number of neurons in the hidden layer should not be too many or too few, and must be set within a reasonable range. If the number of hidden layer neurons is too small, the BP neural network will produce obvious errors, resulting in poor fitting performance or failure to converge. However, when the number of hidden layer neurons is too many, overfitting is likely to occur, which decreases the generalization ability of the BP neural network and thus affects its prediction accuracy. Therefore, choosing an appropriate number of hidden layer neurons is crucial for BP neural networks.

Currently, there is no uniform method to determine the number of hidden layer neurons, and the main method is to determine the number of hidden layer neurons through the “trial and error method”. This method first calculates the range of hidden layer neurons according to the empirical formula, then sets the number of hidden layer neurons within this range, and after training, finally selects the number of hidden layer neurons with the smallest training error of the BP neural network. The commonly used empirical formula, as shown in equation (3): 3 $S = \sqrt{M + N} + A$ where S, M, and N represent the number of nodes in the hidden, input, and output layers, respectively, and A is a constant in the interval (0,10).

(3)

Number of neurons in the output layer

The number of neurons in the output layer depends on the output that the model expects to get. Since this paper expects to get the value of project portfolio benefits, the number of neurons in the output layer of the model constructed in this paper is set to 1.

3)

Setting the activation function

BP neural network can deal with a variety of complex problems, and is widely used in various fields, the main reason is that its activation function is to change the linear characteristics of the BP neural network to realize the nonlinear mapping of the conversion function. The activation function has a variety of functional forms, which can be set according to the needs of the model, and it is mainly used to activate the received information at the neuron to realize the nonlinear transformation. The following three activation functions are commonly used:

Type S activation function, also known as the Sigmoid function. This function is a mapping of numbers taking the value (–∞,∞) to between (0,1). Its expression: 4 $f (x) = \frac{1}{1 + e^{- x}}$

Tanh function, whose definition domain is (-∞,+∞), the value domain is (0,1), as the definition domain value becomes larger, the function image constantly converges to 1, and as the definition domain shrinks gradually converges to 0. The function image becomes faster when it is close to the center of the definition domain 0, but disperses to the ends of the relatively smooth. Its expression is: 5 $f (x) = \frac{e^{x} - e^{- x}}{e^{x} + e^{- x}}$

Both the Sigmoid function and the Tanh function may produce the phenomenon of vanishing gradient, when the function value is close to saturation, because the curve slows down, resulting in the derivative close to 0. The final value obtained in the mathematical equation after combining through a number of chains may be interfered with by the too-small-value property of the derivative, producing the phenomenon of vanishing gradient.

The ReLu activation function, known as the modified linear unit, is a segmented linear function with the expression: 6 $f (x) = {\begin{array}{l} 1, & x > 0 \\ 0, & x \leq 0 \end{array}$

The ReLu activation function compensates for the gradient vanishing problem of the sigmoid function as well as the tanh function, and the computational complexity is very low without the power exponential operation which is more voluminous, which makes it suitable for backpropagation algorithms. In addition, because of the simplicity of the operation, it is also easy to learn and optimize. Therefore, in this paper, the ReLU activation function is chosen as the activation function of the BP neural network.

2.3.3

GA-BP neural network model design

In this paper, the model based on GA-BP neural network is shown in Figure 1, in which the input on the left is the evaluation value of each sample event under five indicators determined in Section 2.2, and the output is the purpose. To simplify the idea of the model, there are three processes including “input-processing-output”, and the BP neural network model optimized by genetic algorithm is used to process the input and output values. In this paper, a typical three-layer neural network (containing only one hidden layer) is used.

1)

BP end design

(1)

Determination of input, hidden and output layers

After modeling machine learning from the perspective of jurisprudence, the number of nodes in the input and output layers can be determined. In the machine learning model studied in this paper, the indicators affecting the protection of intellectual property rights are entered into the BP end as input values, and the number of nodes in the input layer is derived as 9 according to the results of the construction of the intellectual property evaluation index system.The purpose of the study in this paper is to obtain the level of protection of intellectual property rights, including the extremely dangerous level, the dangerous level, the warning level, and the slight level, so the number of nodes in the output layer is 4, and the expected output modes of the warning model are respectively (1000), (0100), (0010), and (0001), which represent the protection level of intellectual property as slight, warning, dangerous, and extremely dangerous.

The determination of the number of nodes in the hidden layer is also important, and it has a very large impact on the learning ability and model accuracy of the established neural network. If the number of hidden layer nodes is too small, system errors will be reduced, but the learning and training ability of the BP neural network model will be reduced. Too many hidden layer nodes not only increase the complexity of the network structure, and the training speed will become slower, making the BP neural network in the learning process is more prone to overfitting into the local extreme value. In this paper, we first determine the range of the number of hidden layer neurons according to the formula, and then find the optimal value. 7 $p = \sqrt{m + n} + a$

Where: m = 9 is the number of nodes in the input layer, n = 4 is the number of nodes in the output layer, and a is a constant between 1 and 10, which determines the number of nodes in the hidden layer of this paper to be between 413.

(2)

Number of hidden layers

Any continuous function in a closed interval can be approximated by a BP neural network containing one hidden layer, i.e., a three-layer BP neural network model can solve any nonlinear problem. Moreover, an excessive number of hidden layers is not all good, it cannot increase the correctness of the prediction, for example, it will make the model complex, the performance will become worse, and the learning time of the neural network will increase. The BP neural network in this paper uses the basic three-layer structure, i.e., it contains only one hidden layer, which can already learn the sample data well.

(3)

Other network structure parameters

a)

Determination of initial weights for BP neural network

For BP neural network, at the beginning of training, the determination of the connection weights between layers is randomly generated by the computer, generally between 0 and 1, which is not the optimal weights, so the learning time of the network is prolonged, and the training speed is slow. Therefore, this paper introduces a genetic algorithm to optimize this shortcoming of BP neural networks, which can find the optimal initial weights of the BP neural network and improve the convergence speed of the model.

b)

Setting of learning rate and training times of BP neural network

The learning rate has a direct impact on the training process of the BP neural network, which is reflected in the correction of the weights. Too high a learning rate will make the system unstable. Too low a learning rate will increase the learning time. In this paper, we adopt an adaptive approach to automatically select the optimal learning rate within the value range of 0.01~0.8 learning rate.

In this paper, we set the learning rate of the BP neural network model to 1000, and the model will terminate the training regardless of whether the output results are satisfactory or not.

c)

Setting of BP neural network training error

Generally speaking, the more accurate the results obtained from the research problem, the smaller the training error is designed. In this paper, the research problem is the warning level of public opinion on the network of public emergencies, and this output result is to be very precise, and no error is allowed, therefore, the training error of the BP neural network model in this paper is set to 0.00001.

2)

Genetic algorithm end design

(1)

Genetic algorithm initialization. The initialization of the genetic algorithm is mainly to determine the initial population, the size of the initial population is the number of individuals in the group, the greater the number of individuals means the larger the population size. Too large a population size will extend the learning and training time. Too small a population size will make it more difficult to find the optimal solution. In this paper, the initial population is randomly selected within the range of 20 to 100.

(2)

Selection operation. In this paper, the selection operation is carried out through the fitness function. The better an individual is, the higher their fitness value is, and the higher their probability of being selected. Therefore, the fitness function in this paper is: the sum of the absolute value of the error between the actual output value and the expected output value of the intellectual property protection level.

(3)

Crossover operation. Generally, when the crossover operation is performed, the probability is set between 0.4 and 0.9. To ensure the stability of the GA-BP neural network model in performing the intellectual property protection warning, the crossover probability in this study is set to 0.4.

(4)

Mutation operation. The mutation probability directly affects whether the BP neural network model can quickly and accurately reach the optimal solution. However, the mutation probability is too large, which will reduce model convergence. Generally, the mutation probability is between 0.001 and 0.01. This paper sets the mutation probability to 0.001.

3

Model testing and empirical applications

3.1

Model Performance Testing

3.1.1

Determination of GA parameters

1)

Individual coding

In this paper, binary encoding is used to encode the initial weights and thresholds, including the connection weights from the input layer to the hidden layer ϖ_ij, the hidden layer threshold θ_j, the connection weights from the hidden layer to the output layer ϖ_jk, and the output layer threshold γ_k. In this model, the coding length of the individual is related to the network structure, which is determined according to the number of nodes in the input layer, the hidden layer, and the output layer, and the BP neural network in this paper is a three-layer of 15:7:1 network structure, thus the individual coding length is 120.

2)

Determination of population size and number of iterations

Reasonable or not the population size is related to the convergence speed of the model, this paper, after many tests, to determine the number of populations of 10 when the global search ability is the best. The number of iterations and the trend of the fitness function curve is related to the number of iterations, that is, as the number of iterations increases, the fitness function curve becomes more and more flat, the number of iterations of the critical value of the curve is considered as the optimal value, and the number of iterations of this paper is set to 30, at this time, the prediction error is basically stable.

3)

Determination of the fitness function

The fitness function is the basis for genetic algorithm optimization, which is used to measure the probability that each individual in the population may be close to the optimal solution in the continuous optimization calculation. A high fitness is a high probability. In this paper, the difference between the predicted output and the desired output of the GA-BP neural network is used as the fitness function, i.e., the smaller the error is, the higher the probability that the individual is selected.

4)

Genetic operator operation

The selection algorithm is the key to determine the direction of the population, usually with a high probability of retaining the good individuals of the parent generation to the offspring, so as to make the population evolve in the direction of the best fitness value. In general terms, that is, there are a number of alternatives, and each program has its own potential score, but in the selection is not entirely in accordance with the high and low scores, but a certain probability of acceptance, high score acceptance probability is high, the lower score acceptance probability is also low.

3.1.2

GA-BP neural network model evaluation results

1)

Validation of GA-BP neural network model

As with the BP neural network, the evaluation accuracy of the GA-BP neural network model is verified, and the results of the comparison between the true value and the predicted value and the percentage error between the true value and the predicted value of the 2000 test samples trained using the GA-BP model are indicated, respectively. The comparison of predicted and true values of the neural network test samples and the percentage error of the neural network test samples are shown in Figures 2 and 3. The distribution of true values and network predicted outputs in the figure is basically the same, and the percentage error is concentrated in 0~0.12, which is small, proving that the model simulation effect is good.

2)

Analysis of GA-BP neural network model evaluation results

The absolute error distribution of the GA-BP neural network test samples and the relative error distribution of the neural network test samples are shown in Figure 4 and Figure 5. The distribution of absolute error and relative error of test samples obtained by simulation using GA-BP neural network can be seen that the absolute error is the smallest 0.0023, and the absolute error of 2000 samples is less than 4. 99.9% of the test samples have relative error less than 0.1.

3.2

Model application

3.2.1

Data collection

This paper obtains intellectual property data through the State Intellectual Property Office and IncoPat database. Firstly, using the intellectual property search of the State Intellectual Property Office to collect relevant legal intellectual property rights, and after filtering out invalid intellectual property rights as well as foreign intellectual property rights, the remaining 71 intellectual property rights are selected, and December 1, 2023 is taken as the base date for assessment. The secondary indicators filtered by the IPR assessment index system constructed in Chapter 2 are used as input vectors, and the relevant information of the selected IPR samples is searched in Incopat database, and the calculated sample values are inputted into the previously constructed and trained network model as input values, and then the model is used to make predictions and output the degree of shared value. Due to the differences in the quantification criteria of the input metrics and the significant differences in the ranges of some of the data, this may lead to slower convergence and growth in the training time of the neural network model, and it becomes unfavorable to proceed directly with the model training. Therefore, in order to solve this problem, the input samples need to be normalized and the output values of the model need to be back-normalized to ensure the consistency and reasonableness of the data. The collected IPR information is shown in Table 2. Observing the 46 collected IPR information, combined with the weights of the indicators, it can be seen that the quality indicator, development indicator and implementation indicator are the three factors that have the greatest influence on the degree of value of IPR sharing.

Table 2.

Patent information

	C11	C12	C13	C21	C22	C23	C24	C31	C32	C41	C42	C51	C52	C53	C54
CN117164526A	6	1	85	4	4	7	2	19	1	4	7	1	12	6	0
CN117164580A	2	1	32	4	1	11	3	15	0	4	5	0	8	2	1
CN117164581A	0	2	42	4	5	13	2	18	1	4	9	5	8	0	2
CN117003737A	0	1	31	1	11	8	0	19	1	5	10	3	8	1	1
CN116813646A	0	1	149	4	4	11	1	18	4	5	6	3	8	0	1
CN115043817A	4	2	127	3	3	11	3	15	2	5	7	4	12	4	2
CN116332959A	1	0	158	3	3	13	2	18	5	3	7	3	9	1	0
CN116253802A	1	0	88	5	8	11	3	18	1	4	8	3	7	1	0
CN116196434A	4	1	107	3	7	15	3	20	4	6	6	4	4	4	0
CN116583518A	0	2	51	5	0	10	2	19	4	5	8	8	7	0	2
CN116323590A	1	0	217	14	2	15	4	18	2	6	10	8	10	1	0
CN115611865A	1	1	166	15	14	12	2	16	1	4	6	5	10	1	0
CN115768774A	3	1	39	1	2	9	1	15	6	6	6	1	5	3	1
CN115215844A	0	1	32	1	6	11	6	17	1	3	7	3	8	0	1
CN115215869A	1	2	42	2	7	10	2	14	3	3	6	7	9	1	2
CN115109078A	1	4	36	3	2	11	4	13	1	4	7	9	12	1	4
CN115043842A	0	5	58	3	20	12	0	16	0	2	8	12	11	0	5
CN114907324A	4	2	144	12	10	12	1	19	4	3	7	3	4	4	2
CN114835703A	1	3	44	2	5	10	1	19	1	5	4	10	8	1	3
CN114835719A	0	4	60	0	14	7	2	17	2	3	12	10	6	0	4
CN115175908A	0	1	178	5	3	10	3	16	3	4	8	8	8	0	1
CN114685426A	0	1	35	3	15	11	2	13	3	6	8	5	7	3	1
CN114671866A	1	1	156	7	2	14	4	15	2	5	6	3	5	1	1
CN114539221A	1	0	36	0	6	10	1	16	5	2	9	11	10	1	0
CN114269741A	0	1	222	7	4	13	0	17	5	5	8	10	11	0	1
CN114380805A	0	1	79	8	15	12	2	15	3	4	9	7	8	0	1
CN114369167A	1	2	22	1	3	6	3	16	2	6	8	2	6	1	2
CN114369166A	3	1	17	3	3	9	0	16	1	4	6	7	4	3	1
CN113929784A	1	0	14	1	4	9	0	16	0	3	6	7	6	1	0
CN113801113A	6	2	130	10	6	8	1	19	1	5	10	8	10	6	2
CN113666923A	0	3	85	3	6	12	2	12	4	5	7	11	9	0	3
CN113666943A	1	1	27	0	16	11	1	15	3	3	8	5	7	1	1
CN113461817A	16	1	40	5	2	13	2	14	1	3	10	6	11	16	1
CN113105448A	2	4	67	6	7	10	1	16	2	7	7	10	8	2	4
CN112694475A	4	4	169	4	2	11	1	14	4	2	8	10	10	4	4
CN112824410A	1	2	70	1	16	8	3	14	0	8	7	9	7	1	2
CN112778372A	2	3	99	9	11	13	1	16	4	3	10	3	9	2	3
CN112707892A	6	4	89	3	20	11	1	15	2	2	10	9	10	6	4
CN112645929A	4	2	25	4	8	11	2	13	0	7	8	4	8	4	2
CN112209896A	3	2	51	5	6	9	1	17	3	8	9	11	11	3	2
CN111718388A	10	0	21	12	2	11	2	15	3	3	8	1	7	10	0
CN111423422A	0	1	74	0	21	11	3	12	1	4	11	11	6	0	1
CN107709304A	0	1	0	20	3	17	0	9	2	7	9	2	10	0	1
CN107250149A	5	1	4	30	17	11	1	14	2	3	11	14	10	5	1
CN106632564A	7	8	24	1	7	11	2	9	2	9	9	9	5	7	8
CN106316963A	3	2	33	2	2	11	2	10	2	9	7	4	9	3	2

3.2.2

GA-BP Neural Network Model Evaluation of Intellectual Property Rights

The organized 71 IPR data are used as neurons in the input layer, the input layer neurons are set to 13, the output neuron is 1. According to the results of the model training in the previous section, the hidden layer neurons of the model are set to 25, the trainlm training function is selected, the maximum number of iterations is set to 1,000, the error target of the model is set to 10⁻⁴, and the learning rate is determined to be 1e⁻³. The results are obtained and then inverse normalization is performed to obtain the predicted value. Since the degree of shared value is an integer, the predicted value is rounded up and compared with the real value, and the IPR prediction results are shown in Table 3. The comparison results show that the use of GA-BP neural network model for intellectual property value degree can be more accurate. The average error between the output value and the real value based on the GA-BP neural network model is 0.034, and the accuracy rate after rounding reaches 93%, so it can be considered that the model is effective in estimating the degree of value of intellectual property rights. 71 intellectual property rights have an error of less than 10% between the actual value and the appraisal value, and only 7 intellectual property rights have an error in appraisal after rounding, and the absolute value is within 1, and the error is less than 1, and the absolute value is less than 1. After rounding, only 7 IPRs have errors in the evaluation results, and the absolute values are all within 1, which is a small error, indicating that the predicted values of the GA-BP neural network model have credibility.

Table 3.

Patent prediction results

Patent number	True value	Predictive value	True error rate	The predicted value is rounded up	error
CN116323590A	9	9.3	3%	9	0
CN116323590A	9	9.12	1%	9	0
CN116583518A	10	9.88	-1%	10	0
CN114835703A	8	8.21	3%	8	0
CN115109078A	7	7.05	1%	7	0
CN115043842A	10	9.3	-7%	9	1
CN107250149A	10	10.12	1%	10	0
CN114835719A	8	8.87	11%	9	-1
CN106232602A	10	9.59	-4%	10	0
CN106316964A	9	8.95	-1%	9	0
CN111423422A	10	9	-10%	9	1
CN107709304A	9	9.06	1%	9	0
CN112645929A	8	8.07	1%	8	0
CN111718388A	6	6.05	1%	6	0
CN112209896A	9	9.21	2%	9	0
CN114539221A	9	9.2	2%	9	0
CN114269741A	9	9.01	0%	9	0
CN114380805A	9	9.24	3%	9	0
CN113929784A	9	9.12	1%	9	0
CN113801113A	9	8.97	0%	9	0
CN114907324A	7	7.34	5%	7	0
CN113105448A	7	7.52	7%	8	-1
CN112778372A	7	8.42	20%	8	-1
CN114369167A	6	6.12	2%	6	0
CN113666923A	9	9.09	1%	9	0
CN106316963A	8	8.06	1%	8	0
CN112694475A	8	8.57	7%	9	-1
CN112824410A	9	9.28	3%	9	0
CN114685426A	9	9.42	5%	9	0
CN112707892A	8	7.94	-1%	8	0
CN104817559A	9	8.99	0%	9	0
CN114014804A	8	7.97	0%	8	0
CN113666943A	7	6.88	-2%	7	0
CN109096171A	9	9.33	4%	9	0
CN106632564A	9	9.3	3%	9	0
CN102190587A	8	8.05	1%	8	0
CN102675018A	8	8.19	2%	8	0
CN102516154A	10	9.6	-4%	10	0
CN103570613A	9	9.11	1%	9	0
CN103387536A	8	8.07	1%	8	0
CN114671866A	9	9.08	1%	9	0
CN101676266A	6	5.56	-7%	6	0
CN103301067A	10	9.51	-5%	10	0
CN102803221A	9	9.31	3%	9	0
CN101906436A	9	9.07	1%	9	0
CN104109149A	9	9.02	0%	9	0
CN103965114A	8	8.29	4%	8	0
CN102690803A	8	7.98	0%	8	0
CN104955811A	10	9.39	-6%	9	1
CN102190616A	9	9.07	1%	9	0
CN115768774A	10	9.14	-9%	9	1
CN115175908A	6	5.63	-6%	6	0
CN113461817A	9	9.07	1%	9	0
CN103301066A	9	9.02	0%	9	0
CN102803220A	9	9.18	2%	9	0
CN115043817A	10	9.04	-10%	9	1

4

Conclusion

This study innovatively proposes the application of GA-BP neural network modeling for the assessment of intellectual property rights protection and draws the following conclusions:

The effectiveness of the developed prediction model was validated through experimental training and testing in this study. In the validation analysis of the GA-BP neural network model, the distribution of the true values and the network predicted outputs are basically the same, and the percentage errors are concentrated in the range of 0 to 0.12.

After using the GA-BP neural network model to evaluate the intellectual property rights, it was found that the average result error between the output value and the real value of the GA-BP neural network model was 0.034, and the accuracy rate after rounding was 93%, which was able to effectively evaluate the protection status of the intellectual property rights.

Language:: English

Publication timeframe:: 1 times per year
Journal Subjects:: Life Sciences, Life Sciences, other, Mathematics, Applied Mathematics, General Mathematics, Physics, Physics, other

Journal RSS Feed

Analysis of the Impact of Machine Learning Models on Intellectual Property Protection from a Jurisprudential Perspective

Zhiqiang Song

Published Online: Mar 19, 2025

Received: Oct 14, 2024

Accepted: Feb 06, 2025

DOI: https://doi.org/10.2478/amns-2025-0433

KeywordsGA-BP neural networks, Machine learning models, Intellectual property, Protection, Jurisprudence

© 2025 Zhiqiang Song, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Keywords
GA-BP neural networks, Machine learning models, Intellectual property, Protection, Jurisprudence