Translatable Patent Identification Research Based On Stacking Ensemble Learning

At present,many enterprises in China hold intellectual property rights such as patents and trademarks,and at the same time they are unable to develop rapidly due to liquidity pressure.The transformation of intellectual property rights such as patents attracts "assets" with"intellectual property",which effectively reduces the financing burden of small and mediumsized enterprises and provides new opportunities for small and medium-sized enterprises to develop and grow.The efficient circulation of intellectual property rights accelerates the release of innovation vitality and effectively promotes the flow and allocation of innovation elements represented by intellectual property rights.Based on the relevant patent data,this paper attempts the BERT model,BERT model and TextCNN model,BERT model and TextRNN model to vectorize the patent title data and feature extraction,and finds that the combination of BERT model and TextRNN model achieves good results,so the BERT model and TextRNN model are selected for feature extraction of patent title data.The extracted data and other original numerical features are combined to predict the patent transformation behavior by using KNN,logistic regression,naive Bayes,support vector machine,decision tree,random forest,XGBoost,Stacking ensemble algorithm and Stacking ensemble learning model based on JC indicator adaptive selection of primary learner,respectively,to predict the patent transformation behavior,and find the optimal classification model for predicting whether the patent can be transformed by evaluating the classification effect of each model.It is empirically found that the three models with better performance in the single learner are logistic regression model,KNN model and naive Bayes model,and the AUC value of the Stacking1 ensemble learning model with the above three models as the primary learner is 0.8423,while the three primary learners selected by the JC index are the KNN model,the naive Bayes model and the XGBoost model.The AUC value of the Stacking2 ensemble learning model using these three models as primary learners reaches 0.8631,which is better than other models,which can accurately predict whether patents will be transformed,and provide new ideas for translatable patent identification.