| Patent is the crystallization of human intelligence,the achievements of enterprise innovation,technological progress and an important driving force of social development.With the coming of knowledge age,the number of patents and the amount of information are increasing,but the quality and value are uneven.Although there are many traditional patent evaluation methods,each of them has its own shortcomings,and the newly emerging evaluation methods that can handle a large number of patents have become a research hotspot.Many scholars use machine learning algorithms to construct models to optimize the evaluation results,however,because of the single machine learning method and the data source,the patent attribute is not fully considered and the evaluation index system is seldom carded.The purpose of this study is to construct a comprehensive and scientific patent value index system and design a series of machine learning evaluation models.This research transforms the problem of patent value evaluation into the problem of classification of patent value degree.First,on the basis of a multi-level in-depth summary of the theory and practice of patent value evaluation at home and abroad,a complete fourdimensional index system is constructed,the Machine Learning Index System composed of 33 patent attributes is calculated,and the classification labels are made according to the winning situation.Then,we use the XGBoost feature selection algorithm in machine learning to rank the importance of the patent attributes,at the same time,a total of 8 patent valuation models based on naive Bayes,k nearest neighbor,decision tree,random forest,Support vector machine,traditional neural network,artificial neural network and recurrent neural network are designed and constructed.In each model,the patent attribute is added in order of the importance of the features,and the optimal accuracy rate and the optimal feature set of each algorithm are obtained to simplify the evaluation index.According to the result of classification operation of the optimal feature set,the evaluation performance of different machine learning algorithms is analyzed,and the corresponding reduction index system is optimized,the validity and optimization results of each classifier constructed by the above process are verified with larger data set and different classification labels.The experiment shows that the priority of XGBoost features and the optimization process of the priority of features combined with the evaluation accuracy are helpful to improve the evaluation effect,the result of attribute feature selection is better than that of feature full input operation.Secondly,the two neural network models have good evaluation performance,the best performance in the experimental data set,in the test data set next to the random forest.The artificial neural network model has a high accuracy rate of 69.65%,and has a balanced accuracy rate in all kinds of patents,with the best comprehensive performance,the optimal feature subset including 22 indexes such as patent validity and covering four dimensions is obtained.Finally,the most suitable patent value evaluation method,namely the XGBoost-NN model,is obtained by synthesizing the whole process of the experiment.This research constructs and optimizes the patent value evaluation model to improve the accuracy and efficiency,which can be regarded as a scientific and objective patent value evaluation method,in addition to the strong representation of the selected patent data set,the results of this study have some reference significance for the future research of patent value evaluation based on artificial intelligence. |
