Abstract
System design has been facing the challenges of incorporating complex dependencies between individual entities into design formulations. For example, while the decision-based design framework successfully integrated customer preference modeling into optimal design, the problem was formulated from a single entity’s perspective, and the competition between multiple enterprises was not considered in the formulation. Network science has offered several solutions for studying interdependencies in various system contexts. However, efforts have primarily focused on analysis (i.e., the forward problem). The inverse problem still remains: How can we achieve the desired system-level performance by promoting the formation of targeted relations among local entities? In this study, we answer this question by developing a network-based design framework. This framework uses network representations to characterize and capture dependencies and relations between individual entities in complex systems and integrate these representations into design formulations to find optimal decisions for the desired performance of a system. To demonstrate its utility, we applied this framework to the design for market systems with a case study on vacuum cleaners. The objective is to increase the sales of a vacuum cleaner or its market share by optimizing its design attributes, such as suction power and weight, with the consideration of market competition relations, such as inter-brand triadic competition involving three products from different brands. We solve this problem by integrating an exponential random graph model (ERGM) with a genetic algorithm. The results indicate that the new designs, which consider market competition, can effectively increase the purchase frequency of specific vacuum cleaner models and the proposed network-based design method outperforms traditional design optimization.
1 Background and Introduction
In today’s rapidly changing technological landscape, the increasing scale of autonomous agents in a system and their complex interactions are posing challenges for designers to effectively model system behaviors and optimize system performance [1]. To find effective solutions to these challenges, the literature has explored various methodologies. One notable work is the decision-based design framework, which models design as a decision-making process seeking to maximize the value of a designed artifact [2,3] by integrating customer preferences. Other significant contributions include the application of Taguchi methods for robust system design [4], game-theoretic approaches to system optimization [5], and the utilization of Monte Carlo simulation for system modeling and design optimization [6]. However, one limitation in many of these works is the lack of effective ways to model dependencies between individual entities.
The recent development in network science has made it a suitable theoretical foundation to address this limitation, thus supporting research on various complex relationships between entities [7,8]. Some representative network-based design studies include (1) mechanical system representations in networks [9–11], (2) social network analysis in design, involving collaboration within virtual design communities [12–15] and co-evolution between design teams and product development [16–20], (3) ecological network-inspired design of engineered products, systems, and industries [21–25], (4) network models for design ideation [26,27], and (5) network-based customer preference modeling [28–32].
Furthermore, there has been growing interest in understanding the influence of local-level networks (as defined in Fig. 1) on system-level functionalities. Research efforts in this area include the use of network motif theory to study system resilience and reliability by assessing specific subsystem structures [33–37]. Other studies investigate how local-level network topologies (e.g., closed or open triangles) influence customer behaviors based on exponential random graph models (ERGMs) [38,39]. More recently, with the increasing popularity of graph neural network (GNN) models, their applications in complex systems engineering have been actively explored [40–44].
Despite the rapid development and wide application of network theories and models in the literature on systems engineering and design, particularly the emerging interest in studying the role of local networks in system design-related issues, the current focus remains on modeling and analyzing global- and local-level system information and their correlations to system performance. There is a scarcity of research on how the results obtained from network modeling and analysis, particularly the insights into significant local dependencies and global–local interactions, can be used to inform design decisions. The latter is an inverse problem. To solve this problem, there are multiple challenges. For example, (1) how do we identify essential local-level dependencies? (2) How do we formulate a system design problem considering those local-level dependencies? and (3) How to search for the optimal design in a network-transformed design space?
To address these challenges and advance the current state-of-the-art, we introduce a network-based system design framework that considers local dependencies. This framework includes: (1) mining significant local networks using network motif theory, (2) transforming the original system design objective using the identified local networks, and (3) searching optimal design attributes for desired system performance combining the genetic algorithm (GA) and an ERGM model. To demonstrate the utility of this framework, we conducted a case study using customer survey datasets on US household vacuum cleaners collected in our previous work [45]. Based on the proposed network-based design framework, this study makes two unique contributions: (1) we propose a novel network-based product design representation that incorporates local triadic competitions of individual products and (2) we develop a new algorithm for calculating the probability that target local triadic competitions exist and successfully integrate it into metaheuristic optimization to support the search for optimal design configurations.
The rest of the paper is organized as follows. In Sec. 2, we introduce the proposed design framework. Then, Sec. 3 demonstrates the design framework using the vacuum cleaner case study. In Sec. 4, we discuss the generalizability and limitations of the proposed framework. Finally, the paper is concluded in Sec. 5 with closing thoughts on our future work.
2 Methodology
In Fig. 2, we compare a typical system design process with our proposed network-based approach. The traditional method begins by analyzing the system requirements from which the design goal is set and the design variables and constraints are identified [46,47]. The design goal guides the formulation of the design objective function and the design variables and constraints define the design space. Subsequently, optimal/suboptimal design solutions are found by exploring and exploiting the design space by evaluating design candidates with the objective function. Finally, the design solution is validated against system-level requirements. Building upon the traditional framework, the proposed network-based design framework consists of six major steps, each of which is elaborated below.
2.1 Step 1: Network Modeling.
The primary objective of the first step is to create a network representation, labeled as , in which corresponds to the network’s adjacency matrix and represents the vector of system design attributes. changes with . Let us take the customer-product market system as an example, in which the co-consideration relations among products can be represented by a unidimensional network shown in Fig. 3. This network is built using data from customers’ considerations of vacuum cleaners and following the approach described in Ref. [45]. In such a network, each node represents a unique product model that the customers considered. The dashed links denote that two products from different brands (e.g., Dyson versus iRobot) are co-considered by at least one customer. In contrast, the solid links denote co-consideration within the same brand. We assume that the design attribute vector in this example only includes the suction power and weight of each product. Updating the design attribute vector for any product will influence the co-consideration and therefore the network structure .
2.2 Step 2: Representing the Design Goal Using Network Motifs.
The objective of Step 2 is to represent the design goal using local networks. This involves transforming the original design objective function into a function of local networks, denoted as , where indicates the derived local network-based design variable (either a scalar or a vector). This step is essential to incorporate the significant dependencies between individual entities (represented by local networks) into the design process. To achieve the objective, we first identify significant local networks based on network motif theory [48].
Having identified the significant motifs, we proceed to convert the original design goal into the new representation in the form of local networks , which capture interdependencies among individual entities (e.g., product models). To illustrate this, let us use an example of the vacuum cleaner market system. Suppose that we take on the role of Dyson, with the original design goal being to participate in the dominant product competition as much as possible. If a significant network motif (shown in the top-left corner of Fig. 3), representing the inter-brand triadic closure competition, is identified as the significant competition pattern in this vacuum cleaner market, then we transform our goal into maximizing the number of inter-brand triadic closure competitions a product is involving in. Next, assume that the number is negatively correlated with the average suction power difference between products within the triadic closure. For a target product (e.g., product 2), the local network-based design variable becomes a scalar and (the values of and are given). The transformed design objective function represented by local networks is thus . The negative sign indicates the assumed negative correlation between and . is the total number of inter-brand triadic closure competitions in which product 2 is involved.
2.3 Step 3: Optimization Problem Formulation.
In Step 3, an optimization problem is formulated based on the local network-based design objective function obtained in Step 2. Figure 4 illustrates an example of the optimization problem associated with the transformed design objective . The objective function is defined to maximize the number of participation of all Dyson products in inter-brand triadic competitions, , by adjusting the design attributes, and , for each product. stands for the product IDs of all Dyson products in Fig. 3. However, solving this network-transformed optimization problem is a challenge. This is because to obtain the number of triadic competitions in which the product participates, we must know the network topology. But it changes every time when we change the design attributes (e.g., suction power stored as a node feature), and there is a lack of analytical expression between the design attributes and the network structure. Therefore, solving such an optimization problem that contains design variables in network representations (that could be non-linear) necessitates the employment of a surrogate model to predict new network structures when a node feature changes.
2.4 Step 4: Predictive Model Training and Evaluation.
Category | Examples | Interpretation |
---|---|---|
Nodal attributes effects | Nodecov | Main effect of a covariate. For example, nodecov.suction understands how suction power influences a vacuum cleaner being co-considered with other vacuum cleaners. |
Relational attributes effects | Absdiff | Absolute difference between two connected nodes’ attributes. For example, absdiff.weight looks into whether a large or small weight difference between two vacuum cleaners motivates them to be co-considered. |
Network structural effects | Edges | Equal to the number of links in the network, equivalent to the intercept term in the regression model. In the context of the vacuum cleaner co-consideration network, it estimates the likelihood that two vacuum cleaners will be co-considered randomly. |
GWESP | Geometrically weighted edgewise shared partner. In Ref. [53], it is also called -triangle, which is defined to be a set of distinct triangles that share a common edge. The GWESP term models the tendency for edges that close triangles to be more probable than edges that do not close triangles. In the context of the vacuum cleaner co-consideration network, it investigates whether two vacuum cleaners co-considered with the same set of vacuum cleaners are more likely to be co-considered or not. |
Category | Examples | Interpretation |
---|---|---|
Nodal attributes effects | Nodecov | Main effect of a covariate. For example, nodecov.suction understands how suction power influences a vacuum cleaner being co-considered with other vacuum cleaners. |
Relational attributes effects | Absdiff | Absolute difference between two connected nodes’ attributes. For example, absdiff.weight looks into whether a large or small weight difference between two vacuum cleaners motivates them to be co-considered. |
Network structural effects | Edges | Equal to the number of links in the network, equivalent to the intercept term in the regression model. In the context of the vacuum cleaner co-consideration network, it estimates the likelihood that two vacuum cleaners will be co-considered randomly. |
GWESP | Geometrically weighted edgewise shared partner. In Ref. [53], it is also called -triangle, which is defined to be a set of distinct triangles that share a common edge. The GWESP term models the tendency for edges that close triangles to be more probable than edges that do not close triangles. In the context of the vacuum cleaner co-consideration network, it investigates whether two vacuum cleaners co-considered with the same set of vacuum cleaners are more likely to be co-considered or not. |
ERGM training involves estimating the model parameters by feeding the observed network data obtained from Step 1. Once the estimated parameters are obtained, the predictive performance of the estimated model can be evaluated in four steps. The first step is to use the estimated model to simulate number of networks (, ,, ). Then, in the second step, each simulated network () is compared against the observed network (that is, the ground truth) in Step 1 to classify all possible links of the network in a confusion matrix [55]. According to the confusion matrix, we can then calculate common metrics, such as Recall, Precision, and F1-Score, to evaluate the predictive performance of the model [55]. Finally, the mean values of Recalls, Precisions, and F1-Scores of those simulated networks are used to represent the performance of a trained ERGM.
2.5 Steps 5 and 6: Optimal Problem Solving and Solution Validation.
In Step 5, the model obtained from Step 4 will be used as a surrogate model to predict new network structures in order to re-evaluate the objective value after modifying the associated node features (i.e., design attributes). Consequently, the computational search process for the optimal solution is performed through metaheuristic approaches, such as the genetic algorithm [56] or particle swarm optimization [57]. After finding the optimal design solution in Step 6, it can be validated by implementing the new designs and observing if the desired system performance can be achieved or not. For example, in the vacuum cleaner example, we will count if the new design attributes of a particular product would help increase its participation in the desired competition relations on the market. Validation is often challenging as it requires real-world implementation and testing. In this paper, we focus primarily on the verification of the optimization results computationally by recalculating the design objective with the optimal design variables and checking if the objective value indeed increases or not.
3 Case Study
In this section, the US household vacuum cleaner market system is used as a case study to demonstrate the proposed network-based system design framework.
3.1 Data Source and Network Modeling.
Data Source: The dataset used in this study is obtained from our previous survey study [45] launched in 2021 on Cint, a provider of digital survey solutions. The publicly available dataset contains 1002 responses from vacuum cleaner buyers, covering 624 unique models of household vacuum cleaners [58]. The dataset covers a diverse array of attributes, spanning from customer demographics and social network information to technical specifications of vacuum cleaners. The participants were tasked with detailing both the options that they initially considered and the ones they ultimately purchased.
Co-consideration Network Model: In this study, we are interested in competition analysis in the vacuum cleaner market system. We, therefore, construct the co-consideration network following our previous study [59]. In this unidimensional network, the nodes are unique vacuum cleaners from the top ten dominant brands and are considered by customers. Similar to the example given in Fig. 3, the undirected links represent that two vacuum cleaners are co-considered by at least one customer. The visualization of the co-consideration network is shown in Fig. 5. This network contains 386 unique vacuum cleaner models and 1259 co-consideration links. Product 369, Dyson Ball Multi-floor 2, has the largest degree, indicating that it is co-considered most frequently.
3.2 Deriving the Local Network-Based Design Goal and Formulating the Optimization Design Problem.
Definition of Derived Local Network-Based Design Variable: As described in Sec. 2.2, the first step of deriving the local network-based design goal is to identify significant local network structures [59]. As shown in Table 2, three significant network motifs of the co-consideration network are identified by the motif mining tool, FANMOD [60], each of which represents distinct competition relationships between brands (inter-brand) and within a brand (intra-brand). They are named for their edge types and topological characteristics. A real-world example for each significant motif is also given in Fig. 5. Among the three motifs, the inter-brand triadic closure competition with the highest -score is found to be the most significant competition structure in the co-consideration network. Based on the positional characteristics in these significant motifs, we can define four unique node roles: , , , and . For example, in the inter-brand triadic closure, all three node positions share the same type of node role, , because each node is co-considered with two products from two other brands in a closed triangle competition. Accordingly, we define the derived local network-based design variable of each product as , where (for ) is the number of times a product is involved in the node role . This defined network-based design variable can be easily extended. For example, if additional node roles such as and are discovered, we can extend the derived variable by concatenating and , resulting in . Similarly, the vector can be shortened by omitting less important node roles.
Optimization Problem Formulation: Now, let us pick one particular product model, e.g., Dyson Ball Multi-floor 2 (Product 369)—the one with the most co-consideration connections in the observed network—to continue the demonstration due to its increasing popularity in the US market. Assuming that Dyson is interested in maximizing a product’s market share, we use the number of purchases as an indicator of that product’s market share. Next, we formulate the network-based design objective function by estimating the relationships between the number of times product purchases and the local network-based design variable derived . As shown in Table 2, given that inter-brand triadic closure shows the highest -score, we simplify the derived design variable by focusing only on the node role in our first test case, resulting in . In this study, since the data format of the number of times product purchases is a count, following the method introduced in Ref. [61], negative binomial regression is selected to estimate the relationship between and . To ensure the reliability of the estimate model, four models corresponding to the combination of polynomial terms of the independent variable up to cubic are tested. Both the mean absolute error (MAE) [62] and Akaike’s information criterion (AIC) [63] are applied to measure the goodness of fit of the model while considering a balance between the goodness of fit and model complexity. The model with the lowest AIC and MAE is finally selected, which is provided in Table 3.
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 0.316 | 0.075 |
0.117 | 0.017 | |
−0.002 | 0.0005 |
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 0.316 | 0.075 |
0.117 | 0.017 | |
−0.002 | 0.0005 |
0.000 level of significance.
In this equation, represents the number of all potential inter-brand triadic closure competitions in which Product 369 is involved. These competitions can be enumerated when we know the total number of products on the market. Suppose that there are products on the market, and we denote the product set as where each element is a product model named by its ID (e.g., product 369’s ID name is ). In that case, we can pre-save the inter-brand triadic closure competitions in which is involved as a set denoted as where and represent the and products in with which Product 369 competes. For instance, in the example shown in Fig. 3, denotes , so . Among these products, , and are the only products that can form the inter-brand triadic closure competition with product 2. As a result, for product 2, and .
3.3 ERGM-Based Network Prediction.
As aforementioned, this study focuses on two design attributes and one constraint: suction power (), weight (), and price (). Therefore, they are taken into account by incorporating their associated Nodecov and Absdiff terms [50], as introduced in Table 1. To ensure model convergence, a trial-and-error process is performed. We evaluated 27 models with varying combinations of Nodecov and Absdiff, ultimately pinpointing a converged ERGM with all terms achieving a level of significance (p-value) close to 0. As presented in Table 4, in addition to three nodal effect terms, two network effect terms given in Table 1, Edges and the GWESP, are also included. To facilitate the convergence of the model and improve its performance, max–min normalization is applied to preprocess the attribute data [65]. The estimated results, i.e., the estimated model parameters in Eq. (3), are shown in Table 4. For example, the negative sign of Absdiff.price shows that two vacuum cleaners with less difference in their prices are more likely to be co-considered. In contrast, the positive sign of GWESP means that two vacuum cleaners that share the same set of co-consideration products are more likely to be co-considered with each other. It implies that customer’s consideration decisions involve a form of multiway grouping and comparison [39].
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Edges/intercept | −6.717 | 0.105 |
Absdiff. price | −0.562 | 0.148 |
Absdiff. weight | −1.287 | 0.194 |
Nodecov. suction | 0.187 | 0.035 |
GWESP | 2.671 | 0.092 |
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Edges/intercept | −6.717 | 0.105 |
Absdiff. price | −0.562 | 0.148 |
Absdiff. weight | −1.287 | 0.194 |
Nodecov. suction | 0.187 | 0.035 |
GWESP | 2.671 | 0.092 |
0.000 level of significance.
Once a trained model is obtained, we follow the process introduced in Sec. 2.4 to simulate 100 networks and validate its predictive power. By comparing the simulated networks against the ground truth, we calculate the means of the Precisions, Recall, and F1-Scores of all 100 simulated networks, and the results are 0.158, 0.192, and 0.173, respectively. The Precision 0.158 indicates that 15.8% predicted co-consideration links are correctly predicted on average. The Recall 0.192 means that 19.2% truly existing links are correctly predicted. Lastly, F1-Score, the harmonic mean of Precision and Recall, evaluates a balanced predictive accuracy of the model. It should be noted that in this study we did not spend excessive resources to find the best ERGM model and only included four independent variables in the model, with the objective of this case study being to demonstrate the proposed design framework. Therefore, we focused on more model convergence and stopped testing additional models (which required more data collection efforts) for better prediction. A further discussion of this is presented in detail in Sec. 4.
3.4 Optimal Design Solutions.
After having a predictive model, the next step in solving the optimization problem involves calculating the objective value each time the design variables (i.e., the values of suction power and weight of Product 369) are explored. As illustrated in Algorithm 1, we first employ the trained ERGM model with the estimated model parameters provided in Table 4 to simulate 100 networks . Within this set of networks, we examine each triadic closure contained in the set . Our objective is to count its occurrence across the 100 networks and compute its occurring ratio, representing its probability of existence, denoted as . Given that typically exhibits a skewed distribution with most values low, we set the median of these probabilities as the threshold to determine the existence of each triadic closure. This choice ensures a more accurate measure of central tendency and provides robustness to outliers [66]. Accordingly, the number of existing inter-brand triadic closures, equivalent to , is obtained and incorporated into Eq. (4) to compute the objective value. Next, as described in Algorithm 2, the evaluation of the objective function is built into the GA [56] which helps to find the optimal level of suction power and weight with the constraint on price to maximize the objective value. The initialization of the GA algorithm is detailed in Algorithm 2, where “popSize” denotes the population size in each round of search, “maxiter” indicates the defined maximum number of generations to run before the GA search stops, and “run” means the maximum number of consecutive generations for which the best objective value (fitness) has no improvement, leading to the termination of the GA search [56].
Objective value calculation
1: Given, , , , ,
2: Initiate
3: Simulate networks with the given , , , and estimated ERGM parameters
4: for to do
5:
6: for to do
7: if exists in then
8:
9: end if
10: end for
11:
12: end for
13:
14: for to do
15:
16: ifthen
17:
18: end if
19: end for
20: Return
Optimization process
1: Constraint
2: Variable
3:
4: GA (
5: fitness,
6: ,
7: ,
8: , , )
9: Summary (GA)
3.5 Comparison Between the Traditional and Proposed Design Methods.
In this section, we compare and evaluate the design outcomes between the traditional and proposed design methods. The key difference between these methods is that the traditional approach optimizes product design by treating each product independently, relying solely on the relationship between the design objective (e.g., maximizing market share) and product attributes. In contrast, the proposed method also considers local dependencies between products, such as competition, during the optimization process. Two design cases were analyzed using both methodologies. Case 1 focuses on designing the suction power of Product 369 to enhance its market competitiveness, considering the constraints on weight and price. In case 2, we optimize both the suction power and the weight of Product 369 with the objective of increasing its likelihood of being purchased, while keeping its price unchanged.
Results of the Traditional Design Method: Regarding the traditional method, we adhere to the procedure conducted in Sec. 3.2 to directly estimate the relationship between the number of times product purchases and the original design vector using a negative binomial regression model, without local network representation of competition relations between products. Specifically, for case 1, , and for case 2, . Furthermore, we incorporate the price () into the model due to its role as the constrained variable. The estimated results for both cases are presented separately in Tables 5 and 6.
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 1.110 | 0.310 |
−0.432 | 0.231 | |
0.087 | 0.039 | |
0.0003 | 0.0003 |
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 1.110 | 0.310 |
−0.432 | 0.231 | |
0.087 | 0.039 | |
0.0003 | 0.0003 |
0.000 level of significance.
0.01 level of significance.
0.05 level of significance.
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 0.764 | 0.295 |
−0.317 | 0.135 | |
0.0004 | 0.0003 | |
0.038 | 0.035 | |
−0.004 | 0.002 | |
0.028 | 0.011 |
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 0.764 | 0.295 |
−0.317 | 0.135 | |
0.0004 | 0.0003 | |
0.038 | 0.035 | |
−0.004 | 0.002 | |
0.028 | 0.011 |
0.001 level of significance.
0.01 level of significance.
Next, in case 1, we take the original price and weight of Product 369 (, ) into the estimated regression model in Table 5, aiming to search the maximum value within the specified suction power design range [1, 5].4 In case 2, we maintain the original price constraint while relaxing the weight constraint. Our goal is to identify the maximum value within the design space defined by the suction power range [1, 5] and the weight range [3.34 LB, 29.3 LB]. According to Fig. 7(a), the highest occurs at . In Fig. 7(b), the highest is achieved when and .
Results of the Proposed Design Method: Following the methodology outlined in the preceding subsections, we execute the algorithm detailed in Sec. 3.4 to explore optimal design solutions. Figure 8 shows the converged search process for optimal values. In case 1, the search terminates in the generation as there is no improvement in the best objective value for 15 consecutive generations. Throughout these generations, we identify ten optimal suction power values at [2.676, 2.680, 2.684, 2.688, 2.692, 2.696, 2.704, 2.708, 2.712, 2.716] (decrease by 1.284–1.324 from its original design), which corresponds to the best objective value of . In case 2, convergence is achieved by the generation as well. We identify three sets of optimal solutions characterized by weight and suction power values : [4.291, 22.317 LB], [4.445, 28.457 LB], and [1.985, 20.078 LB]. These solutions align with the best objective value of 7.729. Lastly, according to the mean value curves of the iterative search processes for both cases, we can observe that the search process in case 2 is more fluctuant. This fluctuation could be attributed to the fact that the design space of case 2, corresponding to two product attributes, is 2D and therefore much larger and more complex than the 1D design space of case 1.
Results Comparison: A comparison of the final results between the traditional and proposed design methods is summarized in Table 8 (the columns corresponding to the traditional design method and proposed design method for ). According to the table, the proposed design method achieves objective values approximately twice as high as those of the traditional design method in both cases. For instance, in case 2, the proposed design method, considering local inter-brand triadic closure competition relationships, achieves an objective value of 7.729. If translating to the number of times purchased, this is approximately eight times, about two times higher than the traditional design method (4.033). Moreover, the objective values obtained using the proposed method for both cases significantly exceed those derived from applying the original design to the proposed algorithm. This further demonstrates the efficiency of the proposed design method. Lastly, the optimal design of case 2 using the proposed design method provides three design options, which offer varying trade-offs between suction power and weight, allowing for tailored solutions catering to different customer preferences while maximizing the product’s market appeal.
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 0.227 | 0.077 |
0.097 | 0.017 | |
−0.002 | 0.0005 | |
0.204 | 0.056 | |
−0.014 | 0.006 |
Independent variables | Est. Coeff. | Std. error |
---|---|---|
Intercept | 0.227 | 0.077 |
0.097 | 0.017 | |
−0.002 | 0.0005 | |
0.204 | 0.056 | |
−0.014 | 0.006 |
0.000 level of significance.
0.001 level of significance.
0.01 level of significance.
Traditional design method | Proposed design method | ||||||||
---|---|---|---|---|---|---|---|---|---|
Case 1: | |||||||||
Original design | 4 | 2.364 | Original design | 4 | 0.854a | Original design | 4 | 2.146a | |
Optimal design | 5 | 3.353 | Optimal design | [2.676, 2.680, 2.684, 2.688, 2.692, 2.696, 2.704, 2.708, 2.720, 2.724, 2.728, 2.732, 2.736, 2.740, 2.744, 2.748, 2.752] | 6.493 | Optimal design | [2.664, 2.668, 2.672, 2.676, 2.684, 2.692, 2.696, 2.704, 2.716, 2.712, 2.716] | 11.555 | |
Case 2: | |||||||||
Original design | 4 | 15.6 LB | 2.802 | Original design | [4, 15.6 LB] | 0.854b | Original design | [4, 15.6 LB] | 2.146b |
Optimal design | 5 | 23.78 LB | 4.033 | Optimal design | [4.291, 22.317 LB] [4.445, 28.457 LB] [1.985, 20.078 LB] | 7.729 | Optimal design | [2.904, 18.723 LB] [3.216, 18.760 LB] [3.525, 18.797 LB] [3.164, 18.723 LB] [3.236, 18.760 LB] [3.280, 18.797 LB] [3.200, 18.723 LB] [3.296, 18.760 LB] [3.316, 18.797 LB] [3.200, 18.723 LB] [3.296, 18.760 LB] [3.316, 18.797 LB] [3.156, 18.760 LB] [3.384, 18.760 LB] [3.408, 18.797 LB] [3.176, 18.760 LB] [2.976, 18.797 LB] [3.164, 19.093 LB] [3.188, 18.760 LB] [3.228, 18.797 LB] [3.372, 19.204 LB] | 12.525 |
Traditional design method | Proposed design method | ||||||||
---|---|---|---|---|---|---|---|---|---|
Case 1: | |||||||||
Original design | 4 | 2.364 | Original design | 4 | 0.854a | Original design | 4 | 2.146a | |
Optimal design | 5 | 3.353 | Optimal design | [2.676, 2.680, 2.684, 2.688, 2.692, 2.696, 2.704, 2.708, 2.720, 2.724, 2.728, 2.732, 2.736, 2.740, 2.744, 2.748, 2.752] | 6.493 | Optimal design | [2.664, 2.668, 2.672, 2.676, 2.684, 2.692, 2.696, 2.704, 2.716, 2.712, 2.716] | 11.555 | |
Case 2: | |||||||||
Original design | 4 | 15.6 LB | 2.802 | Original design | [4, 15.6 LB] | 0.854b | Original design | [4, 15.6 LB] | 2.146b |
Optimal design | 5 | 23.78 LB | 4.033 | Optimal design | [4.291, 22.317 LB] [4.445, 28.457 LB] [1.985, 20.078 LB] | 7.729 | Optimal design | [2.904, 18.723 LB] [3.216, 18.760 LB] [3.525, 18.797 LB] [3.164, 18.723 LB] [3.236, 18.760 LB] [3.280, 18.797 LB] [3.200, 18.723 LB] [3.296, 18.760 LB] [3.316, 18.797 LB] [3.200, 18.723 LB] [3.296, 18.760 LB] [3.316, 18.797 LB] [3.156, 18.760 LB] [3.384, 18.760 LB] [3.408, 18.797 LB] [3.176, 18.760 LB] [2.976, 18.797 LB] [3.164, 19.093 LB] [3.188, 18.760 LB] [3.228, 18.797 LB] [3.372, 19.204 LB] | 12.525 |
The value corresponding to the original design is calculated by inputting the original design values of Product 369 into the estimated negative binomial regression model for the traditional design method or into the developed algorithm for the proposed design method.
Corresponding to a specific , cases 1 and 2 share the same algorithm, resulting in the same value of for the original design of both cases.
3.6 Extensibility of the Proposed Design Method.
In this section, we demonstrate the extensibility of the proposed design method by considering the second most important node role, , as shown in Table 2. Consequently, the derived local network-based design variable changes from to . Accordingly, the optimal design problem is reformulated by estimating the relationship between and the new .
Results of the Extensibility Test: Figure 9 shows the converged search process for optimal values. In case 1, the search terminates in the generation as there is no improvement in the best objective value for 15 consecutive generations. Throughout these generations, we identify 18 optimal suction power values, as listed in Table 8 corresponding to column , which decrease by 1.248–1.336 from the original design and correspond to the best objective value of 11.555. In case 2, convergence is achieved by the generation. We identify 21 sets of optimal solutions characterized by weight and suction power values given in Table 8. These solutions align with the best objective value of 12.525. Comparing these results with the traditional design method and the proposed method before extension (i.e., not including node role into account), the extended method achieves the highest objective values for both cases. For example, in case 2, highlighted in bold in Table 8, the number of times product purchases reaches around 13 when considering both inter-brand and intra-brand competitions. This is approximately three times higher than the traditional design method and five units more than the design that only considers inter-brand competition. This highlights the importance of a comprehensive understanding of the market competition environment for effective product design.
4 Discussion
In this section, we first discuss the generalizability of the proposed network-based design framework, and then discuss the limitations of the current work and suggest future directions for improvement.
4.1 Generalizability of the Proposed Method.
The generalizability of the proposed method is represented in two aspects: (1) generalizability in handling complexity. The proposed method is flexible and can handle different levels of complexity in optimization design problems. As illustrated in Sec. 3.5, it can optimize a varying number of product attributes. Additionally, Sec. 3.6 demonstrates that the derived local network-based design variable can be adjusted to include different local network structures. This adaptability allows the model to manage a wide range of complexities, making it applicable to various scenarios. (2) Generalizability across cases. Beyond vacuum cleaner product design, the proposed method can be directly applied to other product designs, such as vehicles and cellphones, by incorporating market competition information into the product design process. Additionally, the method can be generalized to guide the design of other networked systems, such as transportation systems and power grids. For example, in a shared mobility system, each docked bike station can be defined as a node in a trip network, with directed links representing trips between stations. Using a network motif mining tool, significant travel patterns can be identified to formulate the derived local network-based design variable for each station. This variable, determined by original station design parameters like dock numbers, can incorporate significant user travel patterns into station capacity design, improving system performance, such as user satisfaction scores.
4.2 Limitation.
The first limitation of the current work is the inadequate predictive accuracy of the ERGM obtained in Step 4. One key reason for this could be attributed to the data insufficiency. As stated in our previous study [59], the data for US household vacuum cleaners, including 945 customer responses to 612 unique vacuum cleaner models, are quite heterogeneous, i.e., most customers’ preferred vacuum cleaners are very different from others. This makes the co-consideration network have insufficient links to train an effective ERGM for prediction. Inspired by the existing study [42], where a GNN model was trained using a dataset aggregated from more than 40,000 vehicle survey responses to predict the co-consideration network for the vehicle market system with an -score of 0.65, we propose two potential solutions. First, we could collect more data. With more customer responses, we believe that the accuracy of the model will be improved. Second, we could use advanced deep learning models to replace ERGM since Step 4 of the proposed methodology only requires a network predictive model, and we plan to test more advanced deep learning models such as GNN as the surrogate model that is expected to further improve prediction accuracy.
Another limitation of this study lies in the slow computational efficiency of Algorithm 1. The computer used in the experiments is equipped with an 11th-gen Intel Core CPU (i9-11900 2.50 GHz, 8 cor es, 16 logical processors) and 32GB of RAM. Since ERGM is not compatible with graphics processing unit (GPU) calculation, we employ a parallel computing strategy utilizing 14 logical processors of the CPU. The computational time for each round of Algorithm 1 is approximately 4.3 min for only considering inter-brand competition and 4.8 min for considering both inter-brand and intra-brand competition. Consequently, each generation involving 30 populations of the genetic algorithm requires a total of 2.15 and 2.4 h, respectively. Therefore, solving the proposed optimal design problem and its extended version, which encompasses 15 generations, requires 32.25 h and 36 h to complete the calculations. Moreover, the inefficiency of Algorithm 1 will also hinder the applicability of the proposed method to systems with large network sizes. To address this computational challenge, there are two potential directions to explore. One approach involves extending the current ERGM package [68] to make it compatible with GPU computing. Another direction is to utilize the aforementioned GNN model, which is GPU-compatible, as a replacement for ERGM in calculating the objective value.
5 Conclusion
In this study, we introduce a network-based system design framework, consisting of six key steps. The first step involves generating a network representation for the complex systems. In the second step, we perform significant local network mining and articulate the local network-based design goal, which is an essential step in integrating interdependencies between individual entities into the design process. Subsequently, in Step 3, we formulate an optimization problem based on the proposed local network-based design goal. Moving on to the fourth step, we develop a network predictive model as a surrogate to evaluate the system objective to prepare for solving the optimization problem. In Step 5, we integrate the predictive model into optimization algorithms, such as the genetic algorithm, to address the optimization problem outlined in Step 3. Here, the predictive model plays a key role in updating the objective value, while the genetic algorithm is employed to search for the optimal objective value. Finally, the obtained optimal design solution is utilized to recompute the objective value for validation.
To demonstrate the applicability of our approach in real-world scenarios, we present a case study on the US household vacuum cleaner market. The objective is to optimize a specific product model’s design attributes to increase its sales. Following the proposed method, we first model the vacuum cleaner market competition as a unidimensional co-consideration network. Next, we employ network motif theory to identify three significant local competition patterns and define the derived local network-based design variable based on the unique node positions in those identified competition motifs. This derived variable is a function of the original product design variables, including suction power, weight, and price (as a constrained variable). With the goal of maximizing the number of times a vacuum cleaner is purchased, we formulate a local network-based design objective function by estimating the relationships between purchase times and the derived network-based design variable using negative binomial regression. We then frame an optimization problem based on this objective function and solve it using a typical genetic algorithm procedure. In this process, the ERGM-based predictive model works as a surrogate model to evaluate the system objective whenever a new value of the design attributes (i.e., weight and suction power) is explored.
We demonstrate the efficiency of our proposed design method by comparing it with the traditional design method. The results show that the optimal values of suction power and weight found by the proposed method can significantly enhance the number of times product is purchased, achieving about twice the increase compared to the traditional design method. Additionally, we demonstrate the extensibility of the proposed method by modifying the derived design variable to include more competition relations. The highest objective value illustrates the success of this extension and highlights the importance of comprehensively understanding the market competition environment for optimal product design.
To address the aforementioned limitations, an immediate plan is to develop a new graph neural network-based predictive model. This model will be seamlessly integrated into the optimization algorithm, helping to predict the network structure and compute the objective value.
Footnotes
For a rigorous comparison, each node in the random networks has the same number of degrees as the corresponding real-world network. Moreover, the random networks used to calculate the significance of size- local networks are generated to keep the same number of occurrences of all size- local networks as in the real-world network [48].
An ERGM failing to converge indicates that the parameter estimates are not settling down to stable values, and the iterative estimation process is not reaching a consistent solution [52]. Typical reasons for the convergence issue of ERGM include model degeneracy (an ill-fitting model in ERGM fails to adequately represent the observed network) [53] and inappropriate model specifications [52].
In the dataset, vacuum cleaners of different brands or categories have different units for suction power. Two commonly used units are horsepower and airflow [67]. To solve the problem, we first unify the suction powers of the same unit in the range without units. For example, if the original airflow interval is [21.2 CFM, 160 CFM] (CFM: cubic feet per minute), we evenly divide it into five subintervals. Products with airflow in the first subinterval [21.2 CFM, 48.96 CFM] are assigned a level value of 1, and the same operation applies to levels 2, 3, 4, and 5. In cases with multiple suction power units for the same product, the final level value is the average of the level values converted from multiple units.
Acknowledgment
The authors acknowledge collaborators Neelam Modi, Jonathan Haris Januar, Michael T. Cardone, and Gracia Cosenza for their assistance in data collection, data processing, and the inputs provided during research meetings. We also greatly acknowledge the funding support from NSF CMMI #2005661 and #2203080.
Conflict of Interest
There are no conflicts of interest.
Data Availability Statement
The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.
Appendix: Details of Optimization Problem Formulation and Solving for Extension Case
Figure 10 presents the formulated optimization problem for the extension case, where . The primary difference from the pre-update case is in the objective function, which now reflects the estimated relationship between the number of times product is purchased and the updated .
Algorithm 3 calculates the objective value for the extension case. The key differences from Algorithm 1 include: (1) in row 1, Algorithm 1, which contains all potential inter-brand triadic closures involving . In contrast, Algorithm 3 provides for all potential inter-brand triadic closures and for all potential intra-brand triadic closures involving . (2) From row 13 to row 21 in Algorithm 3, the existence probability of each intra-brand triadic closure is calculated. Correspondingly, rows 29–34 count the number of existing intra-brand triadic closures. (3) Row 35 in Algorithm 3 returns the objective value calculated by the updated objective function.
Objective value calculation
1: Given, , , , , ,
2: Initiate
3: Simulate networks with the given , , , and estimated ERGM parameters
4: for to then
5:
6: for to do
7: if exists in then
8:
9: end if
10: end for
11:
12: end for
13: for to
14:
15: for to
16: if exists in then
17:
18: end if
19: end for
20:
21: end for
22: Median
23: for to do
24:
25: if
26:
27: end if
28: end for
29: for to
30:
31: ifthen
32:
33: end if
34: end for
35: Return