As technological complexity continues to increase, it is imperative to analyze and identify complementary relationships among patents in order to facilitate the recombination of multi domain technical knowledge and foster innovation. Hence, this research proposes to identify patent relationships from the perspective of complementarity according to IPC classification numbers. Specifically, representation learning is performed on the external structure and textual content of patents, and on this basis, a convolutional neural network with attention mechanism is utilized to mine the complementary relationships between patents. It is mentioned that a complementary patent dataset is generated based on the IPC classification numbers of the patents for model training. Empirical analysis in the field of new energy vehicles demonstrates that this approach can effectively identify potential patent complementarities, which can facilitate breakthroughs in key core technologies for enterprises and countries.
The complementarity of patents in technology
refers to the degree to which two patent subjects
with the same broad field of technology focus on
different narrow domain technologies [
Existing research related to technology mining and analysis is usually based on the similarity between patents, ignoring patent complementarity. It is essential to further utilize the distinctions among various technology components, then form a standardized dataset comprising complementary patents. In addition, the present research relies primarily on patent classification or citation information, which may not accurately reflect the patent complementarity in regards to technical content. Therefore, it is necessary to incorporate their textual content in order to delve deeper into their complementarity. Moreover, certain studies assess patent complementarity according to the co-occurrence of patent classifications, which may not exactly indicate the degree of complementarity between patents. Hence, there is a need to integrate deep learning techniques to establish a quantitative approach to identify complementary patents.
To address the above issues, this paper first constructs a complementary patent dataset for subsequent model training based on the IPC classification numbers arranged in a hierarchy. Secondly, drawing on state-of-the-art representation learning techniques, we characterize the external structure and textual content of patents, and comprehensively generate the semantic representations. In the end, based on the above two steps, we propose a complementary patent identification approach, which employs deep learning-based model training to evaluate the degree of complementarity between distinct patents.
In order to fully exploit the complementary relationships between patents, we firstly gather the complete patent records in the field of new energy vehicles. Then we perform patent semantic representation using external structures and textual contents with a deep representation learning method. Finally, we propose a complementary patent identification method based on deep learning techniques. In this method, a dataset of complementary relationships annotated via IPC classification numbers is constructed for model training. The proposed approach is comprised of three main components, as illustrated in Figure 1. 2.1.
To demonstrate the application of the proposed approach, we have selected the field of new energy vehicles for our study. This innovative area encompasses a wide range of industry chains, multiple actors and numerous links. Our study collected full-text patent data from USPTO for new energy vehicles published in 2022. We narrowed our search to the title, abstract, and claim of patents, utilizing keywords new energy vehicle (automobile), hybrid vehicle, electric vehicle, plug-in electric vehicle and fuel vehicle. It is worth noting that all of these terms were established by the Ministry of Industry and Information Technology of the People's Republic of China in 2009. Furthermore, we limited our search to invention patents and ultimately retrieved 8267 patents.
Regarding the establishment of dataset, we
utilize the IPC classification numbers to
determine the technical fields of patents. If two
patents are focused on different technical features
but both fall into the same technical category, we
categorize them as complementary patents [
2.2.
Initially, the multi-layer association
relationships between patents are utilized to
construct a patent heterogeneous network denoted
as = ( , ℰ, , ℛ), where the nodes in contain
multiple external features, ℰ represents the
connections between patents and their
corresponding features, involves various node
types in , while ℛ denotes the linkage types,
such as cite/cited, invent/invented, include/belong,
and others. Afterwards, we employ the
CompGCN graph neural network model [
Finally, the central node's final embedding, patent , is obtained by aggregating Equation 2. Finally, we obtain either a label with 1 or 0 to indicate the existence or absence of a complementary relationship. 2.3.
To begin with, sentence vectors are learned
through the ESimCSE sentence embedding model
[
Afterwards, we utilize attention mechanism to weigh the sentences and calculate the correlation between each dimension in the sentence vectors. Equation 1 demonstrates the specific calculation, where ( , ) measures the matching degree of the th pair of text features between the two sentence vectors ( , ) in the patent.
Furthermore, represents the unique heat vector representation of the th pair of features, with only the th element being equal to 1 and the rest being 0. The ReLU (Rectified Linear Unit) activation function is represented as (∙).
( , ) = ( ( ⊙ ⊙ ) + ) + (1) 2.4.
By
treating
complementary
patent
identification method as a classification problem,
we utilize the semantic vector of patents and
as input data and employ the CBAM module [
Afterward, the characteristics are enhanced by a multi-layer neural network. Subsequently, we apply a shifted sigmoid function to the combined features to determine the conclusive term, as , = ( − , ) = 1 +
1 ( , − ) (2) 2.5.
To evaluate the performance of this approach in identifying complementary patents, a quantitative assessment can be conducted by utilizing both predicted outcomes and actual complementary
relationships in the test set, through metrics such as precision, recall, and f1 score.
To affirm the effectiveness of the proposed approach, a comparison with different methods that utilize either the structural or textual dimension was conducted. As depicted in Table 1, all the three methods yielded precision above 85%, demonstrating their ability to accurately identify complementary relationships with a high degree of matching and predict effectively. Out of the three methods, the combined structure and text method produced the greatest precision rate of more than 90%. Notably, the textual dimension method had the highest recall rate, indicating its proficiency complementary in
accurately relationships with identifying a limited degree of matching. In particular, our proposed method attained the highest F1 score, indicating its superiority in identifying complementary relationships between patents.
90.3
73.1 76.9 75.8
82.4 carried out a specialized examination of the ten patent pairs exhibiting the maximum complementary relationship probabilities. Table 2 presents the specifics, including patent numbers and complementary probabilities of these ten patent pairs.
High probability indicates the likelihood of room for cooperation, so we take the first patent pair as an example to carry out the analysis. The first patent, US11431046, describes an energy storage device that can function as an electrochemical battery with both positive and negative electrodes. Conversely, patent US11522184 presents a technique for preparing positive active material. This technique can be utilized as one of the pathways to enhance the energy storage efficiency of the former patent, and thereby a potential direction for collaboration. Thus, this approach is proved to be valuable in assessing the possibility of a patent complementary relationship as well as determining technological compatibility, which can help enterprises in making informed decisions.
Our paper aims to examine the correlation between patents in terms of complementarity and proposes an approach for complementary patent identification utilizing deep learning. This approach solely utilizes fundamental patent data as input with minimal pre-processing and eliminates the need for costly and labor-intensive feature engineering. We confirmed the effectiveness of this method in identifying complementary relations for new energy vehicle patents by conducting ablation experiments. Our forthcoming research will involve adding patent image data to structural and textual features to boost patent retrieval and matching tasks.
This work is supported by the National Natural Science Foundation of China (No. 71974095) and the Postgraduate Research & Practice Innovation Program of Jiangsu Province (No. SJCX22_0152).