Technology Commercialization Activation Model Using Imagification of Variables

Abstract

Various institutions such as universities and corporations strive to commercialize technologies produced through R&D investment. The ideal way to commercialize technology is to transfer it, recognizing the value of the developed technology. Technology transfer is the transfer of technology from R&D entities, such as universities, research institutes, and companies, to others, with the advantage of spreading research results and maximizing cost efficiency. In other words, if enough technology is transferred, it can be commercialized. Although many institutions have various support measures to assist in transferring technology, there is no substitution for quantitative, objective methods. To solve this problem, this paper proposes a technology transfer prediction model based on the information found in patents. However, it is not realistic to include the information from all patents in the quantitative, objective method, so patterns related to technology transfer must be identified to select the appropriate patents that can be used in the predictive model. In addition, a method is needed to address the insufficient training data for the model. Training data are limited because some technology transfer information is not disclosed, and there is little technology transferred in new technology fields. The technology transfer prediction model proposed in this paper searches for hidden patterns related to technology transfer by imaging the patent information, which can also be applied to image analysis models. Furthermore, augmenting the data can solve the problem of the lack of learning data for technology transfer. To examine whether the proposed model can be used in real industries, we collected patents related to artificial intelligence technology registered in the United States and conducted experiments. The experimental results show that the models trained by imaging patent information performed excellently. Moreover, it was shown that the data augmentation technique can be used when there are insufficient data for technology transfer.

1. Introduction

Technology commercialization is the act of creating economic benefit by directly or indirectly transferring technology developed through R&D investment and using it for business activities [1,2,3,4]. In other words, technology commercialization connects R&D investment with added value creation, and it is an important management strategy carried out to spread research results and strengthen national and corporate competitiveness. Therefore, many countries strive to commercialize technology [4]. They also support promoting technology commercialization to address major factors that could lead to market, system, or coordination failure [4,5,6]. However, despite national support, many difficulties persist, such as the valley of death and the Darwinian sea, preventing most technologies from being commercialized [2]. Moreover, it is challenging for an R&D institution to commercialize its research results because of the researchers’ perception of technology commercialization and problems with performance management.

Technology transfer is intended to increase the success rate of technology commercialization by spreading the R&D results to consumers and industries [7,8,9]. Forms of technology transfer include licensing, mergers and acquisitions (M&A), and the transfer of intellectual property rights [10]. Of these, the transfer of intellectual property rights, that is, the transfer of patents to others, is an important form of technology transfer that promotes commercialization [11,12]. Transferring patents has the advantage of spreading research results and maximizing cost efficiency because the value of already-developed R&D results is recognized and transferred to others. Because of these advantages, many countries and institutions prefer to commercialize technology by transferring it rather than connecting R&D directly to commercialization. In other words, technology transfer has several advantages, such as the prevention of overlapping research, reduction of the R&D cost, and the possibility of professional commercialization, and it is the ideal way to commercialize technology [13]. A significant amount of technology must be transferred for commercialization, so some governments establish and operate support systems to increase the technology transfer success rate. Nevertheless, the success rate is still not high because of existing alternative technologies and market uncertainty. In addition, much time and money are consumed analyzing the value of patents to acquire, and the analysts’ judgments are subjective [14]. Solving this problem requires a data-driven, quantitative, objective technology transfer analysis model [15,16].

Existing quantitative technology transfer analysis models use various patent data information [17,18,19,20,21,22]. However, in reality, quantitative models cannot use all the information in patents. Therefore, it is necessary to search effectively for patterns related to technology transfer in limited patent information and reflect them in the model. In addition, technology transfer is not open to the public for reasons such as security, and there are industries where technology is less often transferred. In such cases, there are insufficient training data to build a technology transfer prediction model. If the training data are insufficient, the model is less likely to be accurate. Therefore, constructing a technology transfer analysis model requires incorporating patent information effectively and addressing the lack of learning data. In this paper, we propose a predictive model that can effectively reflect patterns related to technology transfer by imaging the patent information and addressing the lack of training data. The proposed model uses both numerical and textual information from the patents and uses the DeepInsight [23] algorithm to image the patent information. Imaging the patent information makes it possible to search for characteristics related to technology transfer and to apply a specialized model for image analysis. Moreover, data augmentation is proposed when there are insufficient data for model training. To examine whether the model proposed in this paper can be used in actual industrial fields, we collected patents related to artificial intelligence (AI) technology registered with the United States Patent and Trademark Office (USPTO) to conduct experiments.

This paper is structured as follows. First, in Section 2, related theories and previous studies are reviewed. Section 3 describes the proposed model in detail. Section 4 shows the experimental results from applying the proposed model with actual data. Section 5 discusses the proposed model and experimental results. Finally, Section 6 discusses the limitations of the proposed model, conclusions, and future research tasks.

2. Related Work

2.1. Technology Commercialization

Technology commercialization is the creation of economic benefits by directly or indirectly using R&D output for business [1,2,3]. Because technology commercialization has advantages, such as spreading research results and improving corporate and national competitiveness, many countries operate support measures [4,5,6]. However, despite national support, various difficulties have prevented technology from being sufficiently commercialized. Various studies have proposed methods to address these difficulties and commercialize technology [24,25,26,27,28,29]. Chen analyzed the performance of venture companies using a regression model and identified factors affecting technology commercialization [24]. Jung et al. used a decision-tree model to analyze the success and failure factors of public technology commercialization [25]. Kim and Cho analyzed the major factors contributing to the success of commercializing technology transferred from government-funded research institutes to small and medium enterprises using logistic regression and a decision-tree model [26]. Chen, Jung et al., and Kim and Cho analyzed the success and failure factors of technology commercialization using a quantitative model, but they focused only on specific institutions. Furthermore, they had little data to analyze. Kirchberger and Pohl analyzed previous studies on technology commercialization and searched for the success factors based on the type of institution [27]. Anokhin et al. analyzed the framework for commercializing misfit technology and suggested alternatives for the various cases [28]. Carayannis et al. compared and analyzed the characteristics of the technology commercialization process of US and Russian universities [29]. Studies analyzing the framework of technology commercialization suggested alternatives to the process, from R&D to commercialization. In reality, performing everything from R&D to technology commercialization in one institution is costly and highly likely to fail. Therefore, it is necessary to revitalize technology commercialization via technology transfer using research results that have already been developed.

2.2. Technology Transfer Using Patent Analysis

Technology transfer is intended to increase the success rate of technology commercialization by disseminating the research results to consumers and industries [7,8,9]. There are various types of technology transfer, but the transfer of a patent is a common form. Patent-based technology transfer can maximize the cost efficiency and dissemination of research results by transferring R&D results that have already been developed. A technology transfer study using patent analysis was previously conducted [17,18,19,20,21,22]. Patent analysis is primarily divided into qualitative and quantitative techniques [30,31]. In the qualitative technique, the patents are analyzed by experts, whose subjectivity can be seen in the results, and it requires a significant amount of time and money [14]. Solving these problems requires technology transfer research using quantitative, objective techniques [15,16].

Such techniques use various information from patent data [17,18,19,20,21,22]. Trappey et al. extracted quantitative indicators through principal component analysis and the Kaiser-Meier-Olkien test to evaluate the quality of technology-transfer patents. In addition, they constructed a patent-quality evaluation model with a neural network [17]. Park proposed a model to analyze technology transferability using patent citation information [18]. Park et al. analyzed the potential functions of patents through text-mining techniques and proposed a model to examine the possibility of technology transfer in various industries [19]. Choi et al. used patents to analyze factors affecting technology transfer [20]. Furthermore, technology transfer characteristics were analyzed by applying social network analysis (SNA) to each institution. These prior studies searched for technology transfer fields and influence factors based on patent analysis. In fact, R&D institutions or technology consumers must focus on things such as marketing by searching for patents with high transfer potential. Therefore, a quantitative model is needed to identify potential technology transfer patents in advance. Jun et al. applied text-mining techniques to structure patents and predicted the possibility of technology transfer through SNA [21]. Lee et al. applied textual information and quantitative indicators to an ensemble model to improve the prediction of technology transfer patents [22]. Previous studies for predicting technology transfer patents have built quantitative models using various patent information, but the patent information used to build the actual model was limited. Furthermore, due to the characteristics of technology transfer, the amount of patent training data was insufficient. This makes it difficult to capture small technology-transfer-related variations in the data, and predictive performance may be poor in certain industries due to the lack of training data. The model proposed in this paper converts patent information into images and effectively searches for information related to technology transfer. Moreover, data augmentation allows it to be used when the training data are sparse.

2.3. DeepInsight: Imagification of Variables

Analysis techniques such as statistics and machine learning are usually applied to tabular or similar data composed of numbers because it can be challenging to understand the information contained in non-image data without such techniques. When it is difficult to capture minor variations in such tabular data, they are often converted into images for analysis [23,32]. This conversion into image data allows the inherent information to be more effectively identified using a model specialized for image analysis, such as a convolutional neural network (CNN). Due to these advantages, non-image data are converted into image data in various fields such as genetic analysis and defect measurement [23,32,33,34,35,36].

DeepInsight, proposed by Sharma et al., is an algorithm developed to capture small variations in non-image data and easily discover differences between data [23]. DeepInsight constructs an image by placing similar elements or features together and dissimilar elements farther apart. To transform image data, DeepInsight first defines the location of features using a dimension reduction technique such as kernel principal component analysis or t-distributed stochastic neighbor embedding (tSNE). This study uses tSNE to reduce the dimensions when DeepInsight is used. Equation (1) is the tSNE cost function [37].

$$C=KL(P||Q)={{\displaystyle \sum }}_i{{\displaystyle \sum }}_j{p}_{ij}\log \frac{p_{ij}}{q_{ij}},$$

(1)

where $p$ is the probability that the $j$-th neighbor is selected when given the $i$-th object existing in a high-dimensional space. Moreover, $q$ is the probability that the $j$-th neighboring object will be selected, given the $i$-th object embedded in the low-dimensional space. Next, DeepInsight uses the convex hull algorithm on the location information of the features to find the smallest rectangle containing all objects, rotate them, and map them to pixels. DeepInsight converts various data such as text and genetic analysis into image data. DeepInsight was used in this study with patent information to discover hidden information related to technology transfer.

3. Proposed Model

In this study, we propose a technology commercialization activation model based on technology transfer prediction. The proposed model uses both the numerical and textual information in patents. Moreover, we used DeepInsight to image patent information to discover hidden patterns related to technology transfer and to improve the predictive model performance. Data augmentation was applied so that the model can be used in industries where technology transfer information is not disclosed and technology transfer is rare. The flow of proposed model is shown in Figure 1.

3.1. Data Description and Preprocessing

In this study, we propose a technology commercialization activation model based on technology transfer prediction. To examine whether the proposed model can actually be used in the industrial field, patents were collected from the Database, and experiments were carried out. A total of 15,193 patents related to AI technology and registered with the USPTO were collected. The patents were filed between 1989 and 2016. Among these, 4137 cases of technology transfer occurred. Technology-transferred patents are cases transferred to another person. The transferred patent was checked on keywert, a paid patent information site [38].

Table 1. Experimental data.

Technological Field	DB	Application Period	Status	Number of Patents (Number Transferred)
Artificial Intelligence (AI)	USPTO	1989–2016	Registered	15,193 (4137)

summarizes the characteristics of the experimental data.

The technology transfer prediction model proposed in this paper was built by learning patent data. Therefore, both numerical and textual information from the patents were used. The numerical information used to build the model includes the indicators shown in

Table 2. Numerical information.

Indicator	Description	Measurable Value Information
app_to_regi (days)	Time required from application to registration	Utility value
all_claim_num	Number of claims	Rights
applicant_num	Number of applicants	Utility
inventor_num	Number of inventors	Sustainable development
current_owner_num	Number of current rights owners	Market impact
IPC_num	Number of IPC codes	Technology scalability
b_citation_num	Number of backward citations	Technology impact
f_citation_num	Number of forward citations	Technology impact
fam_nation_num	Number of family nations	Market impact
fam_doc_num	Number of family patents	Market impact
alone_app_yn	Sole application status (dummy)	Utility value
stand_patent_yn	Standard patent status (dummy)	Technology impact
lit_yn	Litigation status (dummy)	Utility value

.

As shown in Table 2, indicators that measure various values were extracted and used to build the model. app_to_regi is the number of days elapsed from filing a patent application to registration. In general, applicants request priority examination to grant rights quickly for patents with high utility value. Thus, when the app_to_regi value is low, the utility value is high. all_claim_num is the number of claims, indicating the scope of the rights. Therefore, all_claim_num can be used as a measure of rights. applicant_num represents the number of applicants. Because it indicates whether multiple applicants are involved in the corresponding patent application, it can be used to measure the utility value. inventor_num represents the number of inventors who participated in a patent, so it can measure the sustainability of the development. current_owner_num represents the number of current patent holders, so it is a measure of the influence of the patent in the market. IPC_num is the number of technical field codes associated with the corresponding patent. A higher value means that the patent can be extended to various technologies. b_citation_num and f_citation_num are the numbers of backward and forward citations, respectively. In the backward citation, another patent cited the patent, and in the forward citation, the patent was cited by another patent. The citation degree of a patent can be used to measure the impact of a technology. fam_nation_num and fam_doc_num indicate the number of countries and patents included in family applications. Because a family application is intended to exercise rights in various countries, these indicators can measure the market influence. alone_app_yn indicates whether a patent application is a sole application. A sole patent is less encumbered than a patent in which technology transfer and an overseas application are filed jointly, making it a measure of the utility value. stand_patent_yn indicates whether it is a standard patent. A standard patent plays a key role in the relevant technical field, so it is possible to measure the technological influence. Finally, lit_yn indicates whether or not patent litigation has been filed. Patents that have been litigated are likely relatively important technologies, as they have caused disputes in their field. Therefore, it is possible to measure the utility value based on whether a lawsuit has been filed. lit_yn identified through keywert [38].

In addition to the quantitative information mentioned above, a patent contains detailed information about the developed technology in text form, such as the abstract and title. Patents have different index values depending on the elementary technology field to which they belong. For this reason, the elementary technology must be identified for each patent and reflected in the model training. Elementary technology can be identified through the text information and clustering [39]. In this paper, the abstract section of the patent was used to identify the elementary technology. However, the corpus size is minimal because the text information extracted from the patent data is limited to a specific technology. To derive effective clustering results, patents must be embedded based on a sufficiently large corpus. Therefore, this study used bidirectional encoder representations from transformers (BERT) pre-trained with a large corpus.

Table 3. Information about the pre-trained BERT.

Model	Description	Number of Dimensions	Size
all-MINIlm-L12-v2	Trained on a large, diverse dataset of over one billion training pairs	384	120 MB

shows the characteristics of the BERT model used for patent embedding [40].

As shown in Table 3, the pre-trained BERT learns various huge text datasets and embeds a set of documents to be analyzed as real numbers in a 384-dimensional space. In this study, 15,193 patents were embedded in 384 dimensions for clustering. In addition, by applying K-means clustering to the embedded patents, multiple clusters were formed, and the elementary technology was identified. When clustering was applied, the optimal number of clusters was selected by comparing values derived from various measures.

Table 4. Measures for deriving the optimal number of clusters.

Measure	Description
CH (Calinski-Harabasz index)	Calculated as the ratio of the sum of dispersion between clusters and the dispersion within clusters for all clusters. The higher the value, the better the performance [41].
DB (Davies-Bouldin index)	Calculated as the average similarity between each cluster and its most similar cluster, where the similarity is the ratio of the distance within the cluster to the distance between clusters. The lower the value, the better the performance [42].
SS (Silhouette score)	A measure of how similar an object is to its own cluster compared to other clusters. The higher the value, the better the performance [43].
SSE (Sum of squared errors)	The sum of the squared differences between each observation and its group’s mean. SSE is used with elbow plots to find the optimal number of clusters [44].

summarizes the measures for deriving the optimal number of clusters [41,42,43,44].

Clusters define and label the elementary technology using high-frequency words. The elementary technology label assigned to each patent is converted into a dummy variable, merged with the quantitative information, and used for the technology commercialization activation model learning.

3.2. Technology Commercialization Activation Model

Patents contain detailed, diverse information about technologies that have been developed, but they require much analysis time and know-how to use them to commercialize technology. In addition, when quantitative analysis is used for technology commercialization, the performance is poor due to limitations in learnable variables and an insufficient number of technology transfers. The performance can be expected to improve if hidden patterns related to technology transfer are searched for effectively within the range of available patent information and reflected in the predictive model construction. Moreover, the lack of technology transfer learning data must also be addressed. The technology commercialization activation model proposed in this paper explores technology transfer patterns inherent in the data by imaging the preprocessed variables and reflecting them to improve the predictive model performance. The imaged data are augmented to solve the lack of learning data.

Converting non-image data into image data makes it possible to search for potential information that has not yet been identified. It can also be used with models that analyze image data, such as CNNs. Therefore, non-image data is imaged and used for analysis in various fields such as gene and text analysis. In this paper, DeepInsight, which converts non-image data into image data, was applied to the patent information. Figure 2 shows the schematic of converting the patent information into image data.

As shown in Figure 2, DeepInsight was applied to the preprocessed patent information, converting it into image data. The non-image data from the patent was normalized using Equation (2) and used with DeepInsight.

$$\frac{x-x_{min}}{x_{max}-x_{min}},$$

(2)

where $x_{min}$ is the minimum value of the data and $x_{max}$ is the maximum value. Normalized values were projected into a two-dimensional space via tSNE, which uses a convex hull algorithm for projected spatial information to find the smallest rectangle that encloses all the data, rotate it, and map it to a pixel. The number of converted images is equal to the number of patents.

Table 5. Imagification parameters.

Parameter	Value
Dimensionality reduction method	tSNE
Pixel size	120 × 120
Channel	3 (RGB)

summarizes the parameters used for data image transformation.

The comparison models are logistic regression (LR), K-nearest neighbor (KNN), decision tree (DT), random forest (RF), AdaBoost (AB), and a CNN. The comparison models learn non-image and image data, respectively, and compare the technology transfer prediction performance. Additionally, data augmentation was performed to prepare for situations with only a small amount of technology transfer learning data, and the learning and prediction results are also presented. Data augmentation uses the synthetic minority oversampling technique (SMOTE) and random oversampling (RS), which are primarily used in label imbalance problems. Furthermore, accuracy, precision, recall, and the F1-score were used to evaluate the model prediction performance in this paper.

4. Experimental Results

Various models were used to evaluate the technology transfer prediction performance of the technology commercialization activation model proposed in this paper, and the results were compared. To evaluate the technology transfer prediction performance of the proposed technology commercialization activation model, actual registered patents were collected, and experiments were conducted. A total of 15,193 patents related to AI technology were collected. Of these, 12,154 were used as the training data, and the remaining 3039 were used for the test data.

Table 6. Characteristics of the training and test datasets.

Data		Number of Patents (Size: Height × Width × Channel)	Number of Labels (0:Non-Transferred, 1:Transferred)
Non-image data	Training data	12,154	0:8845 1:3309
	Test data	3039	0:2211 1:828
Image data	Training data	12,154 (120 × 120 × 3)	0:8845 1:3309
	Test data	3039 (120 × 120 × 3)	0:2211 1:828

summarizes the characteristics of the training and test datasets.

Table 6 shows there were 8845 (72.77%) non-transferred and 3309 (27.23%) transferred patents in the training dataset. In addition, the test dataset contained 2211 (72.75%) non-transferred and 828 (27.25%) transferred patents. The image data had a size of 120 × 120, consisting of three channels (RGB).

Text information was extracted from the patent abstract, and the documents were embedded in 384 dimensions using the pre-trained BERT. Next, the elementary technology was identified by applying K-means clustering to the embedded patent. The optimal K value was then selected by comparing the values derived for each measure, as shown in

Table 7. Result of optimal number of clusters.

K	Measure
	CH	DB	SS	SSE
2	782.671	4.315	0.047	183,495.098
3	676.904	3.042	0.058	177,159.764
4	635.130	3.258	0.063	171,663.809
5	580.094	3.149	0.064	167,377.781
6	516.671	3.587	0.044	164,899.573
7	473.446	3.640	0.043	162,545.410
8	438.377	3.609	0.045	160,513.242
9	409.199	3.638	0.046	158,727.356

.

The areas colored in yellow in Table 7 show the optimal values derived for each measure. For CH and DB, it was found that K values of two and three, respectively, were appropriate. For SS and SSE, a K value of five was found to be appropriate. Based on these results, the collected patent data were grouped into five elementary technology clusters in this study.

Table 8. Elementary technologies.

Cluster	Top-Frequency Words	Elementary Technology
0	Signal, Sound, Speech	Auditory intelligence
1	Image, Detect, Motion	Visual intelligence
2	Language, Translate, Generate	Language intelligence
3	Device, Control, Circuit	AI semiconductor
4	Assist, Person, Digit	Intelligent agent

defines the high-frequency words and elementary technology by cluster.

As shown in Table 8, the basic technology labels identified in patents were converted into dummy variables and merged with preprocessed quantitative information to be used with DeepInsight. Figure 3 shows an example of a converted patent information image.

Figure 3 shows the difference in the patent information image based on whether or not the technology was transferred. The technology transfer prediction performance was compared by training the models on the derived patent information images using the parameters shown in

Table 9. Parameters for comparison models.

Model	Parameters
Logistic regression (LR)	L2 Penalty
K-Nearest neighbor (KNN)	K = 3, Distance = Minkowski
Decision tree (DT)	Criterion = Gini
Random forest (RF)	Criterion = Gini, # of features = sqrt
AdaBoost (AB)	Criterion = Gini, # of features = sqrt
Convolutional neural network (CNN)	Layers = 3, activation function = relu/softmax, Optimizer = Adam, Epochs = 50

.

We trained the models shown in Table 9 using non-image and image data and compared the technology transfer prediction performance.

Table 10. Prediction performance by data type.

Data Type	Measure	Model
		LR	KNN	DT	RF	AB	CNN
Non-image data	Accuracy	0.73	0.66	0.66	0.76	0.74	-
	Precision	0.63	0.53	0.57	0.7	0.75
	Recall	0.52	0.52	0.58	0.61	0.54
	F1-score	0.47	0.52	0.57	0.62	0.5
Image data	Accuracy	0.73	0.7	0.66	0.77	0.74	0.75
	Precision	0.63	0.61	0.58	0.73	0.75	0.68
	Recall	0.52	0.59	0.58	0.61	0.54	0.62
	F1-score	0.47	0.59	0.58	0.62	0.5	0.63

summarizes their prediction performance by data type.

The values colored in yellow in Table 10 indicate the best performance for each indicator. In general, the predictive performance of the models trained with image data was better than that of the models trained with non-image data. RF performed best, and its accuracy and precision improved when trained with image data. In addition, all measurement indices for KNN were better when it was trained with image data than when it was trained with non-image data. CNN, which is specialized in image data analysis, had an accuracy of 0.75, the second-highest prediction performance after RF, and its recall and F1-score were the best of all the models.

A significant proportion of the previously learned dataset was labeled as transferred technology, at a ratio of 7:3. In general, technology transfer in industry is infrequent. Therefore, it is necessary to prepare for cases where the labels are imbalanced. In this study, the image data were augmented using the oversampling techniques SMOTE and RS, and the training and prediction performance of the models were compared.

Table 11. Parameters of comparison models.

Training Data Label	Original Proportion in Dataset (%)	Proportion after Data Augmentation (%)
0: Non-transferred	8845 (72.77%)	8845 (50%)
1: Transferred	3309 (27.23%)	8845 (50%)

summarizes the data augmented by oversampling.

As shown in Table 11, data augmentation adjusted the ratio of the technology transfer label from 7:3 to 5:5. In other words, a total of 12,154 training data elements were augmented to 17,690.

Table 12. Comparison of prediction performance by data augmentation technique.

Technique	Measure	Model
		LR	KNN	DT	RF	AB
SMOTE	Accuracy	0.61	0.63	0.65	0.73	0.67
	Precision	0.58	0.59	0.58	0.65	0.59
	Recall	0.6	0.6	0.59	0.63	0.59
	F1-score	0.57	0.58	0.59	0.63	0.59
RS	Accuracy	0.62	0.63	0.66	0.75	0.66
	Precision	0.58	0.59	0.57	0.68	0.6
	Recall	0.6	0.6	0.57	0.63	0.61
	F1-score	0.58	0.58	0.57	0.64	0.6

summarizes the prediction performance of each model based on the data augmentation technique.

Table 12 shows that RF outperformed the other models. The accuracy of the models was lower than when trained on the data before data augmentation. However, the F1-scores were higher in most models than when trained with the original training dataset. The CNN prediction performance was evaluated after learning with data augmented by the commonly used Keras ImageDataGenerator [45].

Table 13. Parameters of ImageDataGenerator and CNN prediction performance.

Image Data Generator Parameters	CNN
	Measure	Value
Horizontal flip = True	Accuracy	0.76
Vertical flip = True	Precision	0.7
Rotation range = 0.45	Recall	0.6
Zoom range = 0.2	F1-score	0.61

summarizes the parameters used in ImageDataGenerator and the prediction performance of CNNs trained on the augmented data.

Table 13 shows that the CNN trained using the augmented image data had better accuracy than the original model. In addition, the precision, recall, and F1-scores were higher than those of the other models trained on the augmented data. Thus, when sufficient training data cannot be obtained or the technology transfer label is highly imbalanced, training a prediction model using data augmentation may be a better method.

5. Discussion

In this study, we used DeepInsight to image variables, searched for patterns related to technology transfer in limited patent information, and reflected them in building a predictive model. In addition, a data augmentation technique that can be used when there are insufficient training data was applied to the model. Experiments were conducted by collecting patents registered with the USPTO. The results showed that the models trained by converting patent information into images generally outperformed those trained on non-image data. Among them, the RF model accuracy and CNN recall and F1-score were the highest values. In the data augmentation experiment based on oversampling, the RS method performed better than SMOTE, and the CNN accuracy when the model was trained through ImageDataGenerator was better than that of the original model. Unlike the oversampling technique, which generates new data, CNN is considered to have improved prediction performance because it uses data augmented by converting existing data by such actions as rotation and horizontal flip.

In this study, only patents registered in the USPTO were used for modeling. A patent is a territorial system in which rights are granted only in the country in which it is filed. Therefore, the characteristics of patents may differ from country to country. It will be necessary to collect and model patents from various countries and to compare and analyze the results.

6. Conclusions

Technology commercialization is an important management strategy to strengthen the competitiveness of countries and companies, which various countries are trying to revitalize. Despite their efforts, various challenges make technology commercialization difficult to revitalize. Among the various forms of technology commercialization, technology transfer, in which patents are transferred to others, is the ideal commercialization model. Many institutions strive to promote technology commercialization through technology transfer. However, analyzing the target patents for technology transfer consumes substantial time and money, and the existing quantitative models have poor prediction performance because of limitations in patent information and a lack of training data. In this study, patent information was converted into image data using DeepInsight, and patterns related to technology transfer were explored. In addition, a data augmentation technique that can be used when the data for technology transfer learning are insufficient was applied to the model. The proposed model was tested with an actual patent and showed that it can be used in industry. The model proposed in this paper can be used for predicting technology transfer related to AI. Using the proposed model, it is believed that it will be possible to identify patents with high potential for technology transfer in the AI field in advance and establish various technology commercialization strategies.

The model proposed in this paper can help commercialize technology by predicting technology transfers in advance. However, the proposed model has the following limitations. First, the model was built focusing on only one type of technology transfer, namely, transferring patents. In actuality, technology is transferred in various ways, such as licensing and M&A, but all these methods were not reflected in the model. Next, some of the original image information, such as drawings in the patents, was not used. Patents include images such as drawings to explain the technology developed and to secure rights. The original images from these patents should be combined with the existing quantitative information. In future research, more effective quantitative technology commercialization revitalization methods should be studied by addressing the above-mentioned limitations.