Discovering Sustainable Business Partnerships through a Deep Learning Approach to Maximize Potential Value

Abstract

Discovering sustainable business partnerships is crucial for small and medium-sized companies, where they can realize potential value through operational resources and abilities. Prior studies have mostly focused on predicting and developing new business partners using various machine learning techniques or social network analyses. However, effectively estimating potential benefits from business partnerships is much more valuable to companies. Therefore, this study proposes a method which combines deep learning and network analyses to estimate the potential value of business partnerships for companies. To demonstrate the effectiveness of the proposed method, we expand business partnerships between companies and assess potential value derived from the parenthesis using business transaction data collected from the Republic of Korea. The results suggest that companies can gain more potential value from extended networks when compared to previous ones. Furthermore, potential value results show clear distinctions between industries. Our findings provide evidence that small and medium-sized companies can experience significant benefits by establishing adequate business partnerships.

1. Introduction

The growth of small and medium-sized companies is of paramount importance in driving economic progress [1]. Business partnerships play key roles in the growth of these companies, and previous studies have explored how companies develop and discover business partners (BPs) to create value [2,3,4]. For example, companies benefit from business partnerships that are associated with their products and are geographically proximate, which improves product maturity and reduces supply costs [5,6]. By discovering new BPs that have large operational resources and abilities, small and medium-sized companies can extend additional business [7,8]. Most small and medium-sized companies aim to significantly maintain and explore business partnerships. They strengthen regional business partnerships through regional business transaction clusters [9].

Recently, machine learning (ML) was utilized to develop and discover potential BPs. ML has shown outstanding performances in predicting potential BPs. Methods such as support vector machine (SVM), case-based reasoning (CBR), decision tree (DT), and neural networks (NNs) have been used to forecast BPs [3,10,11].

Table 1. Prior BP prediction studies based on ML techniques.

References	Methods	Highlights
[12]	SVM	Utilizing reciprocal relationships between companies for seller selection
[10]	CBR, NN	Combining CBR and NN to predict potential suppliers
[3]	SVM	Hierarchical SVM combined with DT for supplier sections
[11]	DEA, DT, NN	Combining DEA and DT or NN to predict new suppliers

summarizes some examples of previous studies related to BP predictions based on ML. Prior studies that have used different ML techniques have however showed limited results in predicting potential BPs.

Social network analysis (SNA) has been utilized to explore BPs. SNA provides the potential to understand a business transaction network through the structural characteristics of business partnerships. Autry and Golicic (2010) [13] used SNA to investigate the relationship strength of buyer–supplier relationships in the construction industry over time. In [14], authors identified a significant impact of network measures on forming new manufacturing joint ventures by exploring ego networks in automotive companies. Kao et al. (2017) [15] examined various network measures which were closely associated with both firm-level and chain-level productive efficiency in a supply network. Shao et al. (2018) [16] proposed a data analytics approach and developed a mathematical model where multiple network centrality measures were used to explore a supplier’s structural position in nexus supplier networks.

While SNA or ML techniques have been used to predict or explore business partnerships, little attention has been paid to associating business partnerships with potential company benefits. Sasaki and Sakata (2021) [17] used various network centrality measures, combined with company attributes and a Bayesian network, to enhance explanatory power when predicting potential business partners in a supply chain network.

However, analyzing potential value from existing business partnerships and new business partnerships is very important in improving company finances, but previous studies have focused on predicting potential business partners. Therefore, in this study, we propose a method that combines deep learning and network analysis approaches to estimate potential value. The main differences between our research and previous studies are listed below.

• The deep learning-based recommendation method [18] is used to extend business networks by discovering new business partnerships. This method successfully discovers new business partnerships compared to existing ML-based BP prediction models.
• We propose a method that calculates potential value using network measures and financial data. Using two methods, we evaluate potential gain and loss of companies from an extended business network.
• To verify the effectiveness of our method, we use real-world business transaction data collected from the Republic of Korea. The results demonstrate that companies can gain more potential value from extended networks when compared with previous ones. Furthermore, potential value results show clear distinctions between industries.

Our research questions are:

• When a small and medium-sized company either maintains existing business partners or establishes new business partnerships with other companies, is it possible to estimate the potential benefits a small and medium-sized company can achieve from these partnerships? If so, how can these benefits be estimated?
• Are there any distinctions in benefits from these business partnerships at the industry level?

To answer these research questions, this study adapts a deep learning-based model proposed by Lee and Kim (2022) [18] that provides a recommendation system for possible business partnerships. Our study proposes a novel method to estimate the potential value of business partnerships for a small and medium-sized company by implementing network measures and analyzing company financial data. The proposed method is expected to facilitate small and medium-sized companies identify promising business partners that can provide more value.

The rest of this study is organized as follows. First, prior studies related to this research are reviewed in Section 2. Subsequently, our methodology that estimates potential value is presented in Section 3. In Section 4, we describe experimental results with company financial data. In Section 5, estimated added value is analyzed at the industry level. Finally, we conclude with a summary and implications from this study in Section 6.

2. Literature Review

We reviewed prior studies that proposed a model of predicting business partnerships using such methods as analytic hierarchy process (AHP), Bayesian network (BN), and machine learning (ML), and investigated relationships in business transaction networks through social network analysis (SNA).

Table 2. Common methods used in previous studies to develop business partnerships.

Applied Methods	Illustrative Examples	References
Analytic Hierarchy Process	Select a global supplier for manufacturing companies	[19]
	Select a supplier for automobile companies in India	[20]
	Evaluate and select appropriate building material suppliers	[21]
Bayesian Network	Select a supplier using Monte Carlo Markov Chain simulation	[22]
	Select suppliers in a large-scale automobile manufacturer in Turkey	[23]
	Select suppliers and develop multitier sustainable supply chain management	[24]
Social Network Analysis	Analyze business partnerships in supply chain networks	[25]
	Select suppliers in an automotive supply chain network	[26]
	Identify hidden critical suppliers in a Honda supply chain network	[16]
	Select suppliers in a supply chain network in Dubai city	[27]
	Select business partners in a supply chain network in Tohoku city	[17]
Machine Learning	Select suppliers for manufacturing companies in Beijing	[3]
	Suggest sellers for manufacturing companies in Tokyo	[12]
	Select business partners in the supply chain network	[28]
	Recommend business partners for companies in the Republic of Korea	[18]

summarizes these studies. While AHP and BN have been used to evaluate and select business partners [19,20,21,22,23,24], our study focuses primarily on SNA and ML as they are relevant and appropriate for discovering business partnerships and exploring potential value through the structural positions of companies.

Previous studies have used SNA within the context of supply chain networks to explore business partnerships [25,26]. Various network centrality measures such as in-degree and out-degree centralities, closeness centrality, and betweenness centrality have been used to analyze operational load levels for companies, contractual relationships, and potential effects on daily operations for upstream and downstream companies. Some studies have attempted to identify suppliers using network centrality measures combined with other techniques, such as data envelopment analysis [16] and graph mining [27]. Network centrality measures were found to be beneficial for companies in capturing different aspects of structural importance of a supply chain network. Sasaki and Sakata (2021) used network centralities, with a Bayesian network, to effectively select BPs in a supply chain network. Their findings implied that SNA could be a highly beneficial tool to adequately describe transactions between companies in the network.

From another perspective, SNA is used to analyze potential value. A study used centrality and density measures to analyze return on asset (ROA) and found that higher levels of centrality and density values had a positive influence on ROA [29]. Another study also found that higher levels of network centrality improved supplier performance in a supply network in terms of ROA [30]. Arora and Brinturp (2021) [31] used network centralities to investigate the effects of net profit and productivity in complex supply networks. Furthermore, network centralities were used to analyze correlations between degree centrality and financial performance [32], as well as operational profits [33]. However, since these studies relied on existing networks, they were unlikely to consider new potential value derived from new business partnerships.

ML techniques, such as support vector machine (SVM), decision tree (DT), and neural networks (NN), have also been used in many studies to develop and forecast new business partnerships. Previous studies proposed an SVM-based BP selection model using reciprocal relationships to investigate business transaction characteristics between companies and help companies find new BPs [12]. A hierarchical SVM combined with DT was used to efficiently select BPs using various features, such as quality, price, technical support, and packaging ability [3]. Liu and Ran (2020) [28] adopted NN to extend supply chain partnerships in a green supply chain network and effectively select BPs. Lee and Kim (2022) [18] proposed a deep learning-based model to extend business partner networks using multiple business company characteristics in business transaction networks and recommend tailored business partnerships.

3. Methodology

The overarching structure of the proposed framework is presented in Figure 1. This study adapted the deep learning-based model called “DBR model”, which was proposed in a previous study [18]. As shown in the red arrows, The DBR model was used to extend BP networks through existing transaction networks using multiple business characteristics, such as industry, products, locations, and transactions. As shown in Figure 1, $n$ small and medium-sized companies are defined as set $B=\{b_1,b_2,\dots ,b_n\}$ where $b_i$ and $b_j$, $b_i\ne b_j$, are defined as $i^{th}$ and $j^{th}$ companies. The network in this research is a directed network that is composed of a pair $\left(B,B^{\prime }\right)$; $B^{\prime }$ also contains a set of the collection of all pairs $\left(b_i,b_j\right)$. Each pair $\left(b_i,b_j\right)\in B$ (or $B^{\prime })$, is defined as an edge and company $b_i$ is directly associated with company $b_j$ as a business partner (BP).

The proposed model initially uses characteristics, such as product property, industry, geographical distance, and relative transaction volume for companies $b_i$ and $b_j$, and these characteristics are embedded in an input vector. Then, this input vector is used to generate a set of probabilities, $Q$; $q_{i,j}\in Q$, where $q_{i,j}$ refers to a level of business partnership between $i^{th}$ and $j^{th}$ companies. $q_{i,j}=0$ refers to a “no business partnership” between two companies $b_i$ and $b_j$, while $q_{i,j}~1$ indicates that company $b_j$ is more likely to be suitable as a BP to company $b_i$.

Computation of Potential Values

To compute potential values obtained through business partnerships, this study uses the out-degree and closeness centralities [34] of each company after the initial business transaction network is extended through the DBR model. The out-degree centrality value of company $b_i$ is denoted as $O\left(b_i\right)$ and formalized as in Equation (1).

$$O\left(b_i\right)=\sum _{j=1,i\ne j}^n(b_i,b_j)$$

(1)

Closeness centrality for company $b_i$ is defined as $L\left(b_i\right)$ and calculated as in Equation (2), where $d_{i,j}$ is the distance for a pair $\left(b_i,b_j\right)$. Closeness centrality refers to the length of the shortest path. Out-degree and closeness centralities are ranged from 0 to 1.

$$L\left(b_i\right)=\frac{1}{\sum _{j=1,i\ne j}^n{d}_{i,j}}$$

(2)

This study uses selling expenses and cash flows generated from operations and extracted from a company’s financial data. For company $b_i$, an estimated loss is defined as $C\left(b_i\right)$ and computed as in Equation (3), where $e_i$ and $s_i$ are the number of employees and selling expenses, respectively.

$$C\left(b_i\right)=\frac{\sum _{j=1}^n\mathit{log}\left(\frac{s_i}{e_i}\right)\times (1-q_{i,j})}{O\left(b_i\right)}$$

(3)

An estimated loss refers to a relative cost that may be accrued whenever a company $b_i$ establishes a business partnership with company $b_j$. In addition, a level of business partnership between i^th and j^th companies, $q_{i,j}$, is considered in this estimation. Each probability value of $q_{i,j}$ is computed using the deep learning estimation and the estimated loss is adjusted by an out-degree centrality value, $O\left(b_i\right)$, of company $b_i$. An estimated gain for a company $b_i$ is defined as $G\left(b_i\right)$ and formalized as in Equation (4), where $z_j$ is a cash flow generated from operations.

$$G\left(b_i\right)=\frac{\sum _{j=1}^n\mathit{log}\frac{z_j}{e_j}\times L\left(b_i\right)}{O\left(b_i\right)}$$

(4)

An estimated gain refers to a relative benefit that may be realized whenever a company $b_i$ establishes a business partnership with company $b_j$. We focus on a company’s financial data related to its operating activities over a period. To estimate the value, closeness centrality, $L\left(b_i\right)$, for a company $b_i$ is used to incorporate the possible benefits of directly reaching out to more companies in the network. In a similar fashion to that of estimated loss, the value of estimated gain is adjusted by an out-degree centrality value, $O\left(b_i\right)$. For both Equations (3) and (4), the number of employees for a company $b_i$ is used to standardize values across companies. In addition, the $log$ function is used to reduce deviations from the value of selling expenses and cash flows generated from operations.

4. Experiments

4.1. Dataset

Since it is difficult to collect data that include business transactions and financial information, both of which possess unique characteristics, we obtained data by making a request to KoDATA, a South Korean company [http://www.kodata.co.kr (accessed on 14 September 2023)], for research purposes.

The dataset contains 186,399 small and medium-sized companies, including company’s general information and financial data. Over 99% of companies have less than 500 employees. The companies belong to four industries, such as manufacturing (45.4%), wholesale and retail (29.0%), construction (17.7%), and information and communication (7.9%) industries.

Table 3. The number and average sales of companies by industry.

Industry Average Sales No. of Companies		Number of Employees
		≤500	≤1000	≤1500	≤2000	>2000
Manufacturing		$9073 84,334	$413,229 226	$938,159 53	$1.424 M 26	$7.586 M 57
Wholesale and retail		$10,216 53,994	$569,974 49	$2391,317 13	$2.816 M 7	$3.833 M 14
Construction		$6849 32,886	$446,970 16	$164,331 1	$1.063 M 3	$3.740 M 5
Information and communication		$4439 14,644	$245,179 37	$335,593 12	$0.933 M 4	$3.125 M 18
Total	$30.055 M 186,399	$30,579 185,858	$1.375 M 328	$3.827 M 79	$6.236 M 40	$18.285 M 94

Note: Average sales: 1 USD = 1100 KRW and Unit: 1000 USD.

shows the average sales and total number of companies for the four industries.

Table 4. Business transaction volumes between each industry.

	Buyers	Manufacturing	Wholesale and Retail	Construction	Information and Communication
Sellers
Manufacturing		$3044 M	$2046 M	$15 M	$59 M
Wholesale and retail		$749 M	$2468 M	$159 M	$13,918 M
Construction		$1560 M	$6671 M	$79 M	$454 M
Information and communication		$279 M	$7396 M	$296 M	$15,134 M

Note: Average sales: 1 USD = 1100 KRW and Unit: 1000 USD.

presents the business transaction volumes of each industry with the other industries. The “information and communication” industry shows much higher transaction volumes within and between the “wholesale and retail” industry than the other industries. Python 3.9 and Ucinet 6.765 software were used to assess the DBR model and SNA. Gephi 0.9.5 software was also used to visualize BP networks.

4.2. Data Analysis

The current dataset did not provide information on the entire BPs for each company. Since Lee and Kim (2022) [18] showed evidence that the DBR model precisely estimated the “expected” business transaction network, we adapted this model to revise the business transaction network before we conducted further data analyses. Figure 2 presents business transaction networks across the country before and after running the model.

With the revised network, out-degree centrality and closeness centrality are computed using Ucinet. Out-degree centrality shows a frequency of information flow with other companies in the network, while closeness centrality becomes higher when a company reaches out more directly to other companies in the network [35]. The higher the closeness centrality that a company holds, the more chances it can efficiently establish business partnerships in the network. Then, to compute potential values for business partnerships, we use selling expenses and cash flows generated from operations, extracted from the dataset, to estimate the loss and gain that a company can most likely generate from business partnerships in the network. The related formulas used to compute these measures are presented in the previous section (“Computation of potential values”).

Direct and indirect links with BPs in the network are also considered when assessing relevant loss and gain. A direct link refers to an immediate link from a company (called “source”) to another company (called “target”) in the network, while an indirect link presents links where the “source” company extensively reaches other companies through the “target” company in the network. Figure 3 illustrates these links under consideration. Black and red circles refer to “source” and “target”, respectively. Black and blue lines are direct and indirect links, respectively.

Segregated by industry, average and standard deviation values for loss and gain by industry, as estimated from both direct and indirect links in the network, are presented in

Table 5. Average estimated loss and gain values for direct and indirect links by industry.

Industries	Estimated Loss		Estimated Gain
	Mean	Std	Mean	Std
Manufacturing
Direct link	3.108	2.042	1.996	1.368
Direct & indirect links	1.198	1.981	2.718	1.979
Wholesale and Retail
Direct link	3.050	2.367	2.329	1.476
Direct & indirect links	1.126	1.912	3.186	1.969
Construction
Direct link	3.253	2.034	1.624	1.563
Direct & indirect links	2.098	2.283	3.269	1.733
Information and communication
Direct link	2.846	2.210	2.851	1.593
Direct & Indirect links	0.544	1.304	4.474	1.979

.

In further analysis, ANOVA was conducted to determine if estimated loss and gains were significantly different among industries, as shown in

Table 6. ANOVA results for estimated losses and gains.

	Sum of Squares	DF	Mean Squares	F-Scores
Direct link Between groups Within groups Total	13,593.93 1,398,982.54 1,412,576.47	3 2585 2588	4531.31 541.19	8.37 *
Direct & indirect links Between groups Within groups Total	16,285.06 902,007.28 918,292.34	3 2585 2588	5428.35 348.94	15.56 *

Note: * indicated $p<0.001$.

. The findings show that estimated values from direct and indirect links significantly differ among the four industries: direct link [$F_{3, 2585}=8.37, p<0.001$] and both direct and indirect links [$F_{3, 2585}=15.56, p<0.001$].

Games–Howell post hoc tests are also performed to further investigate differences between industries.

Table 7. Multiple comparisons between industries using Games–Howell post hoc tests.

Networks	Industries	Target Industries	Mean Differences	Std. Errors	Sig. Values
Direct link	Manufacturing	Wholesale and retail	2.96	1.50	0.201
		Construction	−3.39	−1.58	0.142
		Information and communication	−7.11 *	2.20	0.008
	Construction	Wholesale and retail	6.35 *	2.05	0.011
		Information and communication	−3.72	2.61	0.484
	Wholesale and retail	Information and communication	−10.07 *	2.56	<0.001
Direct & Indirect links	Manufacturing	Wholesale and retail	−2.25	1.10	0.171
		construction	−3.62 *	1.34	0.036
		Information and communication	−8.38 *	1.92	<0.001
	Construction	Wholesale and retail	5.87 *	1.63	0.002
		Information and communication	−4.76	2.27	0.157
	Wholesale and retail	Information and communication	−10.63 *	2.13	<0.001

Note: * indicated $p<0.001$.

presents multiple comparisons between industries.

Using estimated loss and gains, we further explore the direct and indirect links at the company level to investigate how companies use existing business partnerships in the search for new business partnerships. Figure 4 presents the number of small and medium-sized companies with differences shown between estimated loss and gains. Figure 4 contains two histograms; (a) one from the direct link, and (b) the other from both direct and indirect links. In Figure 4, “black” bars refer to the number of companies that have negative values in differences between estimated gain and loss, while “red” bars present the number of companies with positive values with respect to these differences. In addition, percentages indicate the proportion of companies that belong to each group.

The “information and communication” industry stands out for its relatively low realized losses and high realized gains when compared to the other industries. The findings show that more companies could realize potential gains when they consider extended business partnerships (53.63%) in the network rather than when they count only immediate partnerships (28.94%).

Two different planes for each industry are shown in Figure 5. Both present contour heat maps with estimated loss and gains as its axes; upper plane for direct links and lower plane for both direct and indirect links. Values in lower planes extend over the areas more than values in upper planes since the lower plane presents estimated values from both direct and indirect links. The results indicate that the “manufacturing and construction” industry exhibit similar patterns, while the “information and communication” industry particularly displays a distinctive pattern with higher estimated gains in the lower plane.

The dynamics of estimated loss and gain for the four industries in the network are also investigated.

Table 8. Proportion of companies in terms of changes in estimated loss and gain by industry.

Industries	No Change (−Margin)	− to +Margin	+ to −Margin	No Change (+Margin)
Manufacturing	46.06%	27.47%	2.85%	23.61%
Wholesale and retail	38.31%	28.14%	2.37%	31.19%
Construction	56.08%	23.37%	2.34%	18.22%
Information and communication	21.93%	33.16%	1.07%	43.89%

Note: margin = estimated gain—estimated loss; “−” = negative & “+” = positive; each column shows.

summarizes the influence of indirect links on expected value added, where margin is calculated “estimated gain—estimated loss.” In response to indirect links, a company’s value responds in one of four ways: (1) it maintains a positive margin, (2) it moves from a positive margin to a negative margin, (3) it moves from a negative margin to a positive margin, and (4) it maintains a negative margin. These findings show that the dynamics composition is quite different across industries. In particular, when indirect links are included, the “information and communication” industry contains a much higher number of companies with positive margins when compared to other industries, while the “construction” industry contains a larger number of companies with negative margins when compared to the other industries.

This study also explores location-based information as shown in Figure 6; the number of small and medium-sized companies, located at four major regions, with differences between estimated gain and loss. We split the country into four major regions: Seoul Metroplex, Central region, South Eastern region, and South Western region. Figure 6 shows the distribution of companies for each region, when considering direct and indirect links. We found that after including indirect links, large portions of estimated losses are recovered and most companies across regions experience benefits.

5. Discussions and Implications

This study investigated a BP network using a deep learning model adapted from Lee and Kim (2022) [18] and social network analysis to explore the potential value of business partnerships in a network. Theoretical implications from our results in Figure 2 show that business transaction networks, revised by the proposed method, are extended particularly for companies with multiple existing business partnerships. This suggests that the contributions from this study are similar to our previous study [25]. Furthermore, the proposed method was successful in computing the potential value for each SME from both original and extended business transaction networks. In contrast to a previous study [25], our study demonstrated that the proposed method could help SMEs make sophisticated decisions for business partnerships.

ANOVA and post hoc test results illustrated that companies belonging to each industry may realize distinctive value when they extensively engage in business partnerships. The findings also demonstrate that the “information and communication” industry holds quite different patterns in estimated value from other industries. Our findings suggest the following managerial implications. When companies struggle from limited resources or value extracted from business partnerships within the same industry, they may achieve possible benefits by extending business partnerships across other industries.

In addition, the findings in Figure 4 show that the proportion of small and medium-sized companies with positive values increased from 28.94% to 53.63% when second-hand business partnerships were considered, which implies that many companies can experience potential benefits when they not only establish immediate business partnerships, but also explore business partnerships through extended connections in the network.

The results from Figure 5 show that when second-hand business partnerships are considered, “manufacturing” and “construction” industries tend to display similar patterns, while patterns in both “wholesale and retail” and “information and communication” industries are distinctive. These findings suggest that companies in “manufacturing” and “construction” industries may consider expanding their business partnerships across other industries, to access new markets, enhance their product offerings, and/or accelerate the pace of innovation.

The results from Figure 6 show that when second-hand business partnerships are considered, “manufacturing and construction” industries tend to display similar patterns, while patterns in both “wholesale and retail” and “information and communication” industries emerge distinctively. Companies in the “information and communication” industry should consider expanding their business partnerships to derive more gains. This can include business partnerships with other industries. By leveraging these business partnerships, companies can tap into new markets, enhance their product offerings, and accelerate the pace of innovation.

Our findings in Table 8 confirm the results in Figure 5. We found that the “information and communication” industry revealed much higher proportions of positive value (33.16% and 43.89%) than the other industries. These findings imply that companies may achieve more financial benefits when they develop “technology”-related business partnerships than “traditional product”-related business partnerships. In addition, our results show that the “construction” industry contains higher negative values (56.08%) than the other industries, even after second-hand business partnerships are included. These findings suggest that companies in the “construction” industry may not realize much benefit through business partnerships.

The results in Figure 6 show that benefits from second-hand business partnerships have accrued in many companies regardless of location. However, some companies located in south-east and central regions still suffered from losses. These findings also illustrate that most business partnerships, particularly in the “information and communication” industry, have focused on the “Seoul metroplex” region. Our findings suggest that the regional government may consider how to effectively allocate resources or adjust their policies and regulations to facilitate business partnerships that are adequately balanced across regions.

6. Conclusions

This study suggests a new approach to estimate the potential value of business partnerships for companies by implementing network measures and analyzing company financial data. The results in Lee and Kim (2022) [18], deep learning-based BP recommendation model.

This approach facilitates the identification and development of new business partnerships for small and medium-sized companies. To test the proposed method, we used secondary business transaction data in the Republic of Korea. The results suggest that the proposed method may be effective in helping small and medium-sized companies realize potential benefits by establishing and extending business partnerships. The findings also suggest that companies should prudently form business partnerships from the same or different industries to maximize benefits.

In addition, this study provides insights for small and medium-sized companies in terms of how to effectively manage business partnerships. As the potential benefits of business partnerships differed across industries, one key insight is that a company needs to carefully evaluate potential benefits before establishing new business partnerships. To evaluate these benefits, a company may strategically analyze financial data as well as the extended network of potential business partners. In addition, a company may consider establishing adequate business partnerships to sustain its competitiveness and improve the chances of success in the marketplace. Furthermore, our study suggests that the government may introduce different regulations or policies over different industries to enable the growth of small and medium-sized companies.

However, this study contains some limitations. First, as business transaction data used in this study included financial data from the COVID-19 pandemic, the estimated values might be biased. Further research is required to explore longitudinal data to reduce possible biases. Second, the results might be limited since the data were collected only in the Republic of Korea. To generalize the findings, an extended set of data may be needed to investigate this fully. Finally, even though this study estimated values from both direct and indirect links in the business transaction network, the results may be limited for the complete network. Further research is required to realize full potential benefits for companies. Furthermore, as the probability derived from our deep learning-based method was used to estimate potential value for a company, outcomes may depend on the performance of the deep learning-based method. Efforts to enhance performance are necessary for future work.