Abstract: This paper designed multi-dimensional evaluation indicators of patent value and recognized high-value patents as the prediction target vector to construct the training set and the test set. Then the BP neural network model was used to predict and identify potential high-value patents. At the same time, the MIV algorithm was used to analyze the contribution and influence of various dimension indicators of patent value to the result of the high-value patents prediction. The experimental results show that: (1) The prediction performance of the BP neural network model is relatively good, and the prediction accuracy rate is all over 89%. The BP neural network model that takes the recognized high-value patents through "patent family size" as the predictive target vector has the best performance, while the BP neural network model with the predictive target vector composed of the recognized high-value patents through the combination indicator of "patent family size" and "patent citation frequency" performs relatively poorly. (2) The MIV absolute value can effectively reflect the influence and contribution degree of various patent value indicators on the result of the predication model. The indicator of technical value has the most significant influence on the result of the highly-value patent prediction based on the BP neural network model. From the perspective of the MIV absolute value and the total proportion of every single indicator, the four indicators including the number of patent IPC4 classification, the speed of first citation, the number of claims, and the frequency of patent citations have a more significant influence on the results of high-value patents prediction.

Key words: High-value patent, BP neural network, MIV algorithm, Patent prediction, Machine learning, Patent value evaluation