A Feasibility Discussion: Is ML Suitable for Predicting Sustainable Patterns in Consumer Product Preferences?

Abstract

In the era when product design must meet the needs of consumers, the products preferred by consumers are an important source of design creativity and design reference for product designers to design products. Therefore, how to effectively grasp the products that consumers prefer has become an important issue for product designers. In order to allow designers to have more convenient and accurate consumer preference product prediction tools, this study proposed machine learning (ML) to analyze and predict sustainable patterns in consumer product preferences and conducted a feasibility study on the use of ML for predicting sustainable patterns in consumer product preferences. A total of three experiments were carried out in this study: the KJ method to predict consumer product preference experiment, the AHP method to predict consumer product preference experiment, and ML to predict consumer product preference experiment. This study uses the three experiments to discuss and compare the prediction ability of ML and the current commonly used forecasting tools, namely the KJ method and AHP method. The research results show that no matter what kind of consumer product attribute preference is predicted, the accuracy rate of consumer product preference prediction by ML is much higher than that of the KJ method and AHP method. These research results show that no matter the product attribute dimension, ML has the ability to predict consumer preferences, and ML has a better ability to predict consumer preferences than traditional tools. Therefore, this study believes that ML can be used to analyze and predict sustainable patterns in consumer product preferences. Therefore, this study suggests that product designers can use ML technology to assist in the analysis and prediction of consumer product preferences, so as to improve the grasp of consumer preference products.

1. Introduction

Under the fierce global competition and rapidly changing economic system, in order to respond to changes in consumer demand, companies must seek new product design strategies to create differentiated products and create product value to meet consumer needs. Since today’s product design focus on meeting consumers’ preferences [1,2], therefore, designers usually have to rely on the products preferred by consumers as design references in order to develop products that satisfy consumers. Therefore, how to effectively grasp the products preferred by consumers is an important issue for product designers [3,4]. This means that a technical tool that can effectively and accurately predict sustainable patterns in consumer product preferences will be very helpful and important to product designers [5,6].

The uncertainty and complexity of sustainable patterns in consumer product preferences make it difficult to predict consumer product preference behavior patterns. The prediction of consumer product preference behavior patterns requires technology that can sort out rules from the unknown. Sustainable patterns in consumer product preferences refers to the continuous preference behavior of consumers for a commodity [7,8]. Sustainable patterns in consumer product preferences are a personalized preference that reflects consumers’ preference for different products and services [9,10]. It is an important factor affecting market demand [11]. Sustainable patterns in consumer product preferences is mainly determined by the impact of the local social environment [12], customs [13], and fashion changes [14,15,16,17,18] on the entire consumer group or a specific group [19,20]. Consumers will rank consumption goods or commodity combinations according to their own wishes, and this sorting reflects consumers’ individual needs, interests, and preferences [21,22]. Generally speaking, sustainable patterns in consumer product preferences can be divided into the following four basic types: Type 1—Consumers who do not know their preferences and are fluid and ambiguous. If consumers’ preferences are unstable and ambiguous, it is impossible to provide them with a satisfactory solution to satisfy their preferences. However, because they do not know their own preferences, they are easily influenced and persuaded by companies to believe that their customized supply is satisfactory and truly in line with their preferences [23]. Furthermore, if the customization supply is successful, these consumers will think that the customization satisfies their previous preferences. This will form the basis of their future preferences. Type 2—Consumers who are aware that they do not have stable and clear preferences. These consumers are likely to base their evaluations of offers on the attractiveness of the product’s appearance rather than whether the product actually meets their preferences. Furthermore, this type of consumer is most receptive to suggestions and assistance that help them discern their preferences. For example, a consumer who likes to drink wine, but knows he does not have the knowledge, may be more than happy to receive education and consumption advice about wine. Type 3—Consumers who are not aware that they have stable preferences. This type of consumer is probably the least numerous. Such consumers have stable consumption preferences, which guide their choices. However, they are not clearly aware of the driving force of preferences on their consumption choices [21]. For example, they may think that their choices are based on rational and objective judgments. In fact, their choice is mainly based on emotional factors or aesthetic factors. Consequently, these consumers may wrongly accept either customized offers or selection criteria that do not actually meet their preferences. Alternatively, they may choose to reject customized offers or selection criteria that really fit their preferences. Type 4—Consumers who know they have stable preferences. Such consumers not only have clear preferences, but also have a sufficient understanding of their own preferences. This enables them to correctly judge whether a customized offer really meets their preferences. Therefore, these consumers may be good potential customers for customization. They will generate more satisfaction with marketers’ efforts to understand their preferences [24]. However, because of their knowledge of their preferences, they may rely less on marketer recommendations [25]. These studies on consumer product preference behavior patterns show that sustainable patterns in consumer product preferences have multiple types, which are difficult to predict and grasp. This means that unless the prediction technology used to predict consumer product preference behavior patterns has the effect of learning sustainable patterns in consumer product preferences, it is a very difficult task to perform, as we only rely on personal subjective evaluation methods or tools to analyze and predict consumer preferences. At present, most studies on sustainable patterns in consumer product preferences mostly discuss the causes or influencing factors of sustainable patterns in consumer product preferences, but rarely discuss the effectiveness and accuracy of analysis and prediction tools, such as a predictive tool with a learning function for sustainable patterns in consumer product preferences. Therefore, a technical method of a learning function for consumer product preference behavior patterns will be explored in this paper.

This paper mainly proposes a machine learning (Abbreviation ML) technology for predicting sustainable patterns in consumer product preferences and discusses its feasibility. Since the product factors that affect consumer product preference usually include product shape, product color, and overall product feeling, in this study, we used ML to analyze and predict three types of sustainable patterns in consumer product preferences: overall product preference behavior pattern, product modeling preference behavior pattern, and product color preference behavior pattern. In addition to using ML analysis to predict consumer product preference behavior patterns, this paper also uses the KJ method and AHP method to analyze and predict three consumer product preference behavior patterns. This research will first understand and compare the correctness of ML, the KJ method, and the AHP method to analyze and predict the three types of consumer product preference behavior patterns. This analysis will help to understand the capability and feasibility of ML prediction for sustainable patterns in consumer product preferences.

This study hypothesizes that ML is better at predicting sustainable patterns in consumer product preferences than traditional prediction methods. This study will compare the correctness of ML in predicting sustainable patterns in consumer product preferences with the correctness of traditional forecasting methods in predicting sustainable patterns in consumer product preferences. This study expects that the accuracy rate of ML in predicting sustainable patterns in consumer product preferences will be higher than that of traditional prediction methods in predicting sustainable patterns in consumer product preferences.

Typically, ML is mainly used to design and analyze some algorithms that allow computers to “learn” automatically. An ML algorithm is a kind of algorithm that automatically analyzes and obtains laws from data and then uses the laws to predict unknown data [26,27]. The ML approach emphasizes the context of analysis, learning, and prediction from “reasoning” to “knowledge” and then to “learning”. Therefore, this paper believes that it is feasible to use ML to analyze and predict sustainable patterns in consumer product preferences. It will make the following contributions:

• Giving the prediction of the consumer product preference behavior model based on artificial intelligence architecture and paying attention to the details of different consumer product preference behaviors.
• Demonstrating the advantages and characteristics of ML that can be applied to predict sustainable patterns in consumer product preferences.
• Demonstrating ML’s ability to predict consumer product preference characteristics in multiple product attribute dimensions.
• Providing product designers with an efficient and convenient consumer product preference prediction technology. This allows product designers to more easily, quickly, and accurately predict consumer preference products.
• Providing a new and reliable analysis and prediction tool for consumer preference behavior that is not easy to grasp and accurately predict.

The chapters of this article are as follows. Section 2, the literature review, explains the theoretical background and previous research. Section 3, Methodology, describes the theory and details of the analysis and prediction work of the ML method, KJ method, and AHP method. Section 4, Results, describe the effect of the ML method, KJ method, and AHP method in predicting sustainable patterns in consumer product preferences. Section 5 discusses and explains the reasons for the differences among the prediction results of the ML method, KJ method, and AHP method. Finally, in Section 6, conclusions and future research directions are presented.

2. Theory Literature Review of Prediction Methods

The “Kawakita Jiro Method (abbreviation KJ Method)” is also known as the A-type graphical method and affinity graph method. It was invented by Japanese anthropologist Kawakita Jiro, and KJ is his English abbreviation. The so-called A-type diagram is to collect information on facts, opinions, and ideas about unknown or untested chaotic problems, and then to use the mutual affinity between the materials to make a classification and merger diagram, as well as to find a solution for the problem [28]. The operation steps of the KJ method are as follows: Step 1—decide on the subject of the question; Step 2—conduct brainstorming discussions; Step 3—make keywords; Step 4—organize; Step 5—name each group; Step 6—review after classification; Step 7—find out essential categories and sorting; Step 8—the planning process; Step 9—conception plan [29]. The KJ method was used to spread and converge ideas, and the team selected the “lively” color [30]. In addition, the KJ method was used to study the consumer’s acceptance behavior of blue and white porcelain-style casual clothing [31]. Finally, the KJ method was also used to identify customers’ needs for the replacement project of aluminum windows and doors [32].

The “Analytic Hierarchy Process method (abbreviation AHP method)” was developed by Thomas L. Saaty (a professor at the University of Pittsburgh) in 1971. It is mainly applied to decision-making problems under uncertain conditions and with multiple evaluation criteria [33,34,35]. The purpose of the development of the “AHP method” is to systematize complex problems, decompose them at different levels and, through quantitative calculations, find out the context and make a comprehensive evaluation. Once the hierarchy is established, experts will systematically evaluate the scale and assign weight values to the relative importance of each part. Afterwards, a pairwise comparison matrix is established, and the eigenvectors and eigenvalues are obtained. Finally, the feature vector is used to represent the priority of each part in each level [32,36]. These data, such as criteria (criteria), weight (weight), and analysis (analysis), can provide decision makers with sufficient decision-making selection conditions. This can provide the decision maker with the right decision and reduce the risk of a wrong decision [37]. The AHP method and fuzzy method are used to evaluate consumer satisfaction [37]. In addition, the AHP method is also used to explore consumers’ preferences for green tourism [26], as a social performance measurement of mutual funds [27], and to evaluate employee performance [38].

The ML approach is a branch of artificial intelligence. The history of artificial intelligence research has a natural and clear vein from focusing on “reasoning”, to focusing on “knowledge”, to focusing on “learning”. The ML method is a path to carry out artificial intelligence. It uses algorithms as a method to solve problems in artificial intelligence. The ML method has developed into a multi-field interdisciplinary subject in the past 30 years. The ML method involves many disciplines, such as probability theory, statistics, approximation theory, convex analysis, computational complexity theory, etc. [39,40,41]. The theory of ML is mainly to design some algorithms that allow computers to “learn” automatically. An ML algorithm is a kind of algorithm that automatically analyzes and obtains laws from data, and uses the laws to predict unknown data [42,43]. Because a large number of statistical theories are involved in the ML algorithm, ML is particularly closely related to inferential statistics, also known as statistical learning theory [44,45,46,47,48,49,50]. The ML method uses different types of learning models and appropriate algorithms according to the nature of the data and the desired results [51,52,53]. The learning patterns of ML can be divided into four learning modes: supervised, unsupervised, semi-supervised, and reinforced [54,55,56,57]. In each learning pattern, ML can apply one or more algorithms. Which algorithm ML applies in the learning pattern depends on the dataset used and the expected results. The ML method is mainly used to classify things [58], identify patterns [59], predict results [60], and make comprehensive judgments. For example, ML was used by Odey Alshboul [61] to predict green building cost. This study presents ML-based algorithms, including extreme gradient boosting (XGBOOST), deep neural network (DNN), and random forest (RF), to predict green building costs. The proposed models are designed to consider the influence of soft and hard cost-related attributes. In addition, ML has been used by Ardvin Kester S. [62] to predict the factors that affect STEM students’ intention to take chemistry-related courses in the future. In this study, using ML algorithms, such as a random forest classifier and an artificial neural network, a total of 40,782 datasets were analyzed. Results showed that attitude toward chemistry and perceived behavioral control represent the most influential factors, followed by autonomy and affective behavior. This demonstrated that students’ interest, application in real life, and the development of knowledge and skills are key indicators that would lead to a positive future intention for pursuing the course in higher education. Moreover, Arash Heidari’s [63] systematic literature review, “Applications of ML/DL in the management of smart cities and societies based on new trends in information technologies”, explains that AI (artificial intelligence), ML (machine learning) and DL (deep learning) methods are helpful for managing automation in smart cities. Shu-Rong Yan et al. [64] also discussed implementation of a product-recommender system in an IoT-based smart shopping using fuzzy logic and an Apriori algorithm. This article presents a new IoT-based smart product-recommender system based on an Apriori algorithm and fuzzy logic. The results revealed that the suggested technique had achieved acceptable results in terms of mean absolute error, root-mean-square error, precision, recall, diversity, novelty, and catalog coverage when compared to cutting-edge methods. Finally, the method helps increase recommender systems’ diversity in IoT-based smart shopping.

From the concept introduction and application results of the KJ method and AHP method, it can be seen that consumer product preferences cannot be determined via the method of using accumulation and experience to make predictions. Therefore, this study believes that the KJ method and AHP method are not suitable to analyze and predict sustainable patterns in consumer product preferences. Otherwise, since ML prediction has the functions of analysis, learning, and prediction, it is suitable for analyzing the situation that is difficult to grasp due to chaos. Therefore, this study believes that ML is more suitable for analyzing and predicting sustainable patterns in consumer product preferences.

3. Methodology

3.1. ML Predicts Sustainable Pattern in Consumer Product Preferences

3.1.1. Sample Collection and Processing

In this study, we collected a total of 60 photos of coffee machines of different brands and models for ML training, and at the same time we collected 40 photos of coffee machines of different brands and models to verify the effectiveness of ML training. In order to improve the effect of ML, this research continued to shave the blurred and poor-quality photos in 60 coffee machine photos for ML training purposes and 40 coffee machine photos for verification ML training purposes. Finally, there were only 40 photos (Figure 1) of the coffee machine used for ML training, and only 20 photos (Figure 2) of the coffee machine used to verify the effectiveness of ML training.

3.1.2. Analysis and Predicting Work

The process of using ML to predict consumers’ coffee machine preference in this study is as follows:

• Step 1. Input photos for ML training. First, open the 40 photos (Figure 1) and the 20 photos (Figure 2) in the ML system to prepare for ML analysis and training.
• Step 2. Mark photos. Ask a consumer who has experience in buying a coffee machine to use a score from 1 to 5 to rate each of the 40 photos (Figure 1) and the 20 photos (Figure 2) in the ML system. Score notation is based on the consumer’s personal preference for the coffee machine in the photo. The 1–5 points represent consumers’ preference scores for coffee machines (1 point: do not like it very much, 2 point: do not like it, 3 point: normal, 4 point: like it, 5 point: like it very much).
• Step 3. ML training: First set of ML parameters. (A) Set network model to deep learning; (B) set image information to image width 300 dpi, image height 300 dpi, image depth 3 dpi; (C) set learning parameters iteration to 100, batch quantity 1, learning rate 1 × 10⁻¹. After the parameters were set, this study began to perform ML training on the 40 photos in Figure 1 in the ML system. When the accuracy rate of ML reaches 82.8%, this research will store the training model obtained by ML in the ML system.
• Step 4. Training model verification. Input the 20 photos (Figure 2). After the pictures are inputted, this study will then input the training model stored in the ML system and use it to judge consumers’ coffee machine preferences for the 20 photos (Figure 2). The results of consumer preferences predicted by the ML system for 20 coffee machines are presented in

  Table 1. ML predicts the results of consumer coffee machine preferences.

  Photo No.	No.1	No.2	No.3	No.4	No.5	No.6	No.7	No.8	No.9	No.10	No.11	No.12	No.13	No.14	No.15	No.16	No.17	No.18	No.19	No.20
  Predict consumer overall product preference	3	4	5	5	4	3	2	4	2	3	4	2	4	5	3	2	4	3	3	1
  Predict consumer product shape preference	3	3	5	5	4	4	4	4	2	3	5	1	5	5	2	1	5	4	3	2
  Predict consumer product color preference	2	2	3	2	1	4	1	5	3	2	3	1	4	5	1	3	4	5	3	1

  .
• Step 5. Then, ML predicts consumers’ coffee machine shape preferences. Using the same steps of Step 2–Step 4, this study invites the same consumer to rate each of the 40 photos (Figure 1) and the 20 photos (Figure 2) with 1–5 points according to his personal styling preference. Then, we make the ML algorithm train, and the output is verified again. Table 1 shows the results of the ML system predicting consumers’ preferences for 20 coffee machine shapes.
• Step 6. ML predicts consumers’ coffee machine color preferences. Using the same steps of Step 2–Step 4, this study invites the same consumer to rate each of the 40 photos (Figure 1) and the 20 photos (Figure 2) with 1–5 points according to his personal color preference. Then, we make the ML algorithm train, and the output is verified again. Table 1 shows the results of the ML system predicting consumers’ preferences for 20 coffee machine colors.

3.2. KJ Method Predicts Sustainable Pattern in Consumer Product Preferences

3.2.1. Sample Collection and Processing

This study also uses the 20 photos of coffee machines in Figure 2 as samples for predicting consumer product preferences using the KJ method. This method uses the 20 photos of coffee machines in Figure 2 as the prediction samples of the KJ method in order to compare the consumer product preferences results predicted using the KJ method as well as the results predicted using ML.

3.2.2. Analysis and Predicting Work

The process of using the KJ method to predict consumers’ coffee machine preference method in this study is as follows:

• Step 1. Decide on the topic of prediction. This study first decides on the prediction topics as follows: using the KJ method to predict the overall preference of consumers’ coffee machines, to predict the preferences of consumers’ coffee machines shape, and to predict the preferences of consumers’ coffee machines color.
• Step 2. Watch samples. Product designers who have experience in designing coffee machines are invited to carefully look at the 20 photos of coffee machines in Figure 2. When the product designer watches, the product designer is required to pay special attention to the shape of the coffee machine, the color of the coffee machine, and the overall feeling of the coffee machine in each photo.
• Step 3. Make a preference scoring word card. Product designers are asked to write 1 point, 2 points, 3 points, 4 points, and 5 points on five 5 cm × 2 cm white cards. The scores on these five white cards represent the degree to which the designer predicts consumer product preference. Among them, 1 point represents “don’t like it very much”, 2 points represents “don’t like it”, 3 points represents “normal”, 4 points represents “like it”, and 5 points represents “like it very much”.
• Step 4. Make groups. Ask the product designer to spread out all the 20 coffee machine photos (Figure 2) and 5 preference scoring word cards, and then ask the product designer to predict the degree (score) of consumer’s overall product preference for each coffee machine photo. After the prediction work is completed, the product designer is asked to put the coffee machine photos with the same overall product preference scores in the same group. If there is a single photo of a coffee machine that cannot be classified, it will be regarded as a group as it is, and finally the product designer will be asked to classify it according to the similarity between the groups.
• Step 5. Match preference scoring word card for each group. After all the coffee machine photos are grouped, the product designer can be asked to match the preference scoring word cards with the coffee machine photo group according to the degree of deviation of the predicted consumer coffee machine. When the five preference scoring word cards are completely macheted, the KJ method prediction for the overall coffee machine preference of consumers is completed.
• Step 6. Check after classification. Check the classified groups to see if there are any duplications, errors, or omissions in the classifications, and then make corrections.
• Step 7. Record classification results. Record the number of photos in the classified group and the group’s consumer product preference prediction score (

  Table 2. KJ method predicts the results of consumer coffee machine preferences.

  Consumer Product Preference	1 Point	2 Points	3 Points	4 Points	5 Points
  Predict consumer overall product preference	(16)	(2, 12)	(6, 7, 10, 17, 20)	(1, 4, 8, 11, 14, 15, 19)	(2, 3, 5, 13, 18)
  Predict consumer product shape preference	(16,)	(7, 10, 12, 17, 20)	(3, 6, 19)	(4, 8, 11, 13)	(1, 2, 5, 9, 14, 15, 18)
  Predict consumer product color preference	(7, 12, 16)	(3, 4, 9, 10, 20)	(6, 19)	(8, 11, 13, 14, 17)	(1, 2, 5, 15, 18)

  ( ) indicates the photo number.

  ).
• Step 8. Predicting consumer coffee machine shape preferences. Invite product designers to complete the prediction of consumers’ coffee machine shape preference according to Step 4–Step 6 again and make records (Table 2).
• Step 9. Predicting consumer coffee machine color preferences. Invite product designers to complete the prediction of consumers’ coffee machine color preference according to Step 4–Step 6 again and make records (Table 2).

3.3. AHP Method Predicts Sustainable Pattern in Consumer Product Preferences

3.3.1. Sample Collection and Processing

In order to compare the results of the AHP method to predict consumer product preferences with the KJ method and ML methods to predict consumer product preferences, this study also uses the 20 coffee machine photos in Figure 2 as samples for the AHP method to predict consumer product preferences.

3.3.2. Analysis and Predicting Work

The process of using the AHP method to predict consumers’ coffee machine preference in this study is as follows:

• Step 1. Confirm the problem. In the research on the effectiveness of the AHP method in predicting consumers’ product preferences, the research topics are as follows: the effectiveness of the AHP method in predicting the overall preferences of consumers for coffee machines, the effectiveness of the AHP method in predicting consumers’ coffee machine shape preferences, and the effectiveness of the AHP method in predicting consumers’ coffee machine color preferences.
• Step 2. List the evaluation factors. In the study, we take the 20 coffee machine photos in Figure 2 as evaluation factors.
• Step 3. Establish the evaluation matrix. In this step, this study establishes three evaluation matrices, as follows: the consumer coffee machine overall preference prediction evaluation matrix, consumer coffee machine shape preference prediction evaluation matrix, and consumer coffee machine color preference prediction evaluation matrix.
• Step 4. Paired comparison evaluation. In this study, the implementation method of the paired comparison evaluation of consumers’ overall preference for coffee machines is as follows: two coffee machine photos are randomly selected in the prediction evaluation matrix of consumers’ overall coffee machine preferences, and without considering the influence of other factors, these two evaluation coffee machine photos are compared for predictive consumer product overall preference. This study uses 1 point, 2 points, 3 points, 4 points, and 5 points (1 point represents “don’t like it very much”, 2 points represents “don’t like it”, 3 points represents “normal”, 4 points represents “like it”, and 5 point represents “like it very much”) to present the result of the paired comparison evaluation. After a set of paired evaluation coffee machine photos completes the comparative evaluation, the subjects will conduct a comparison of another set of random two coffee machine photos to compare the overall predictive consumer product preference again. After completing the comparative evaluation of a pair of evaluation factors, the subjects will conduct another group of random selections of two coffee machines photos to carry out the overall preference prediction and comparison of consumer products again. When all the coffee machine photos in the matrix have completed the paired evaluation, the paired comparison evaluation of the overall consumer coffee machine preference is completed. Next, this study uses the same pairwise comparison evaluation method to let the subjects conduct pairwise comparison evaluation of consumer coffee machine shape preference and consumer coffee machine color preference.
• Step 5. Create a pairwise comparison matrix. The pairwise comparison matrix is built as follows: (A) Put the paired comparison evaluation of the consumer’s overall coffee machine preference in the consumer coffee machine overall preference prediction evaluation matrix (Figure 3); (B) put the paired comparison evaluation of the consumer coffee machine shape preference in the consumer coffee machine shape preference prediction evaluation matrix (Figure 4); (C) put the paired comparison evaluation of the consumer coffee machine color preference in the consumer coffee machine color preference prediction evaluation matrix (Figure 5).

• Step 6. Calculate the pairwise comparison matrix eigenvalues and eigenvectors. This step is mainly to calculate and obtain the eigenvalues and eigenvectors of the pairwise comparison matrix with the paired comparison evaluation of the consumer overall coffee machine preference, the pairwise comparison matrix with the paired comparison evaluation of the consumer coffee machine shape preference, and the pairwise comparison matrix with the paired comparison evaluation of the consumer coffee machine color preference. In this study, Formula (1) is used to obtain the eigenvalues of each pairwise comparison matrix, and Formula (2) is used to obtain the eigenvectors of each pairwise comparison matrix.

$$w_i=\frac{{\left\{{{\displaystyle \prod }}_{j=1}^n{a}_{ij}\right\}}^{\frac{1}{n}}}{{{\displaystyle \sum }}_{i=1}^n\left\{{{\displaystyle \prod }}_{j=1}^n{a}_{ij}\right\}}$$

(1)

i, j = 1, 2, K ………n

$${\lambda }_{max}=\frac{1}{n}\left\{\frac{W_1}{W_1}+\frac{W_2}{W_2}+\dots +\frac{W_n}{W_n}\right\}$$

(2)

• Step 7. Calculate the weights of each level and factor. This study uses the geometric mean method to calculate the preference weight for each coffee machine. The weight of each product attribute of the 20 coffee machines in this study is shown in

  Table 3. AHP method predicts results of consumer coffee machine preferences.

  Photo No.	No 1	No 2	No 3	No 4	No 5	No 6	No 7	No 8	No 9	No 10	No 11	No 12	No 13	No 14	No 15	No 16	No 17	No 18	No 19	No 20
  Predict consumer overall product preference	0.19 (13)	0.19 (13)	0.20 (12)	0.20 (12)	0.20 (12)	0.23 (11)	0.26 (9)	0.24 (10)	0.48 (6)	0.51 (5)	0.37 (8)	0.68 (2)	0.44 (7)	0.44 (7)	0.57 (3)	0.56 (4)	0.10 (15)	0.69 (1)	0.10 (15)	0.15 (14)
  Predict consumer product shape preference	0.26 (16)	0.26 (16)	0.32 (13)	0.51 (9)	0.25 (17)	0.19 (9)	0.37 (11)	0.27 (15)	0.64 (5)	0.55 (6)	0.31 (14)	0.87 (3)	0.38 (10)	0.33 (12)	0.54 (7)	0.92 (2)	0.86 (4)	0.53 (8)	0.53 (8)	0.93 (1)
  Predict consumer product color preference	0.43 (12)	0.41 (13)	0.51 (6)	0.48 (9)	0.40 (14)	0.45 (11)	0.60 (3)	0.46 (10)	0.49 (8)	0.56 (5)	0.56 (5)	0.48 (9)	0.59 (4)	0.46 (10)	0.66 (2)	0.71 (1)	0.50 (6)	0.51 (7)	0.40 (14)	0.35 (15)

  ( ) indicates the order of preference.

  .

• Step 8. Calculate C.R. The C. R. refers to the consistency ratio. Its calculation formula is given in Formula (3). The C.I. refers to the consistency index. Its calculation formula is given in Formula (4). The R. I. refers to the random index, and its values are shown in

  Table 4. Random index (when the order of Matrix A and the number of evaluation scales are known, the C.I. value generated is called R.I.).

  Order	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
  R.I.	0	0	0.58	0.9	1.12	1.24	1.32	1.41	1.45	1.49	1.51	1.48	1.58	1.57	1.58

  . According to Formulas (3)–(5), the C.R. value of the coffee machine overall preference pairwise comparison matrix is 0.063 (

  Table 5. C.R. value of three consumer coffee machine shape preference pairwise comparison matrices.

  	The Coffee Machine Overall Preference Pairwise Comparison Matrix	The Consumer Coffee Machine Shape Preference Pairwise Comparison Matrix	The Coffee Machine Color Preference Pairwise Comparison Matrix
  C.R.value	0.063	0.077	0.028

  ), the C.R. value of the consumer coffee machine shape preference pairwise comparison matrix is 0.077 (Table 5), and the C.R. value of the coffee machine color preference pairwise comparison matrix is 0.028 (Table 5).

$$C.R.=\left(C.I.\right)/\left(R.I.\right)$$

(3)

If C.R. ≤ 0.1, the consistency of the matrix is satisfactory.

$$C.I.=\left(\lambda \_max-n\right)/\left(n-1\right)$$

(4)

When C.I. = 0, it means that the judgments before and after are completely consistent. When C.I. = 1, it means that the judgments before and after are inconsistent. Furthermore, C.I. ≤ 0.1, is allowable deviation.

4. Results

4.1. Differences in Designer’s Prediction Results Caused by Different Product Attributes Used to Evaluate Cousumers’ Product Preferences

Table 6. Ranking results of the KJ method in predicting sustainable patterns in consumer product preferences.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20
Consumer product preferences	4	5	5	4	5	3	3	4	2	3	4	2	5	4	4	1	3	5	4	3
Consumer product preferences	5	5	2	2	5	2	1	4	2	2	4	1	4	4	5	1	4	5	3	2
Consumer product preferences	5	5	3	4	5	3	2	4	5	2	4	2	4	5	5	1	2	5	3	2

shows the ranking results of the KJ method in predicting sustainable patterns in consumer product preferences.

Table 7. Ranking results of the method in predicting sustainable patterns in consumer product preferences.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20
Consumer product preferences	13	13	12	12	12	11	9	10	6	5	8	2	7	7	3	4	15	1	15	14
Consumer product shape preferences	16	16	13	9	17	9	11	15	5	6	14	3	10	12	7	2	4	8	8	1
Consumer product color preferences	12	13	6	9	14	11	3	10	8	5	5	9	4	10	2	1	6	7	14	15

shows the ranking results of the KJ method in predicting sustainable patterns in consumer product preferences.

Table 8. Ranking results of the ML method in predicting sustainable patterns in consumer product preferences.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20
Consumer product preferences	3	4	5	5	4	3	2	4	2	3	4	2	4	5	3	2	4	3	3	1
Consumer product shape preferences	3	3	5	5	4	4	4	4	2	3	5	1	5	5	2	1	5	4	3	2
Consumer product color preferences	2	2	3	2	1	4	1	5	3	2	3	1	4	5	1	3	4	5	3	1

shows the ranking results of ML methods in predicting sustainable patterns in consumer product preferences. From Table 6, Table 7 and Table 8, we can find that designers predict that consumers will have different preference evaluation results on different product attributes as a whole, and on the shape and colors, regardless of the method used to predict consumers’ product preferences. This means that designers predict that consumers use different perception channels to evaluate the product as a whole, as well as the shape and colors of the product.

4.2. Differences in Designer’s Prediction Results Caused by Different Evaluation Methods Used to Evaluate Consumer Product Preferences

Table 9. Comparison of ranking results of various forecasting methods in predicting sustainable patterns in consumer product preferences.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20
KJ Prediction	4	5	5	4	5	3	3	4	2	3	4	2	5	4	4	1	3	5	4	3
AHP Prediction	13	13	12	12	12	11	9	10	6	5	8	2	7	7	3	4	15	1	15	14
ML Prediction	3	4	5	5	4	3	2	4	2	3	4	2	4	5	3	2	4	3	3	1

shows the comparison of ranking results of various forecasting methods in predicting sustainable patterns in consumer product preferences.

Table 10. Comparison of ranking results of various forecasting methods in predicting sustainable patterns in consumer product shape preferences.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20
KJ Prediction	5	5	3	4	5	3	2	4	5	2	4	2	4	5	5	1	2	5	3	2
AHP Prediction	16	16	13	9	17	9	11	15	5	6	14	3	10	12	7	2	4	8	8	1
ML Prediction	3	3	5	5	4	4	4	4	2	3	5	1	5	5	2	1	5	4	3	2

shows the comparison of ranking results of various forecasting methods in predicting sustainable patterns in consumer product shape preferences.

Table 11. Comparison of ranking results of various forecasting methods in predicting sustainable patterns in consumer product color preferences.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20
KJ Prediction	5	5	2	2	5	2	1	4	2	2	4	1	4	4	5	1	4	5	3	2
AHP Prediction	12	13	6	9	14	11	3	10	8	5	5	9	4	10	2	1	6	7	14	15
ML Prediction	2	2	3	2	1	4	1	5	3	2	3	1	4	5	1	3	4	5	3	1

shows the comparison of ranking results of various forecasting methods in predicting sustainable patterns in consumer product color preferences. From Table 9, Table 10 and Table 11, it is not difficult for researchers to find that designers have quite different prediction results of consumer product preference when they use different prediction methods, such as the KJ method, AHP method, or ML method, regardless of if the product attributes are considered in terms of product attributes as a whole or the shape and colors. For example, when the designer uses the KJ method to predict the degree of consumers’ product preferences, the degree of consumers’ product preferences can only be divided into five different degrees. However, the degree of consumers’ product preferences was divided into more than 10 levels when designers used the AHP method to predict the consumers’ product preferences.

4.3. A High Gap between the Prediction Results of Traditional Prediction Methods and Actual Consumer Preferences

Table 12. Comparison of ranking results of actual consumer preferences and consumer product preferences that the KJ and AHP methods predicted for the whole product.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20	Accuracy Rate
Actual Preference	3	3	5	5	4	2	2	5	2	3	4	1	4	5	3	2	4	4	3	1
KJ Prediction	4	5	5 (O)	4	5	3	3	4	2 (O)	3 (O)	4 (O)	2	5	4	4	1	3	5	4	3	20%
AHP Predict	13	13	12	12	12	11	9	10	6	5	8	2	7	7	3 (O)	4	15	1	15	14	5%

(O): consistency between predictions and actual preferences.

shows the comparison of ranking results of actual consumer preferences and the consumer product preferences that the KJ and AHP methods predicted for the whole product.

Table 13. Comparison of ranking results of actual consumer preferences and consumer product preferences that the KJ and AHP methods predicted for product shape.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20	Accuracy Rate
Actual Preference	3	3	5	5	4	4	4	5	2	3	5	1	5	5	2	1	5	4	3	1
KJ Prediction	5	5	3	4	5	3	2	4	5	2	4	2	4	5 (O)	5	1 (O)	2	5	3 (O)	2	15%
AHP Prediction	16	16	13	9	17	9	11	15	5	6	14	3	10	12	7	2	4	8	8	1	0%

(O): consistency between predictions and actual preferences.

shows the comparison of ranking results of actual consumer preferences and the consumer product preferences that the KJ and AHP methods predicted for the product shape.

Table 14. Comparison of ranking results of actual consumer preferences and consumer product preferences that the KJ and AHP methods predicted for product color.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20	Accuracy Rate
Actual preference	2	2	4	2	1	3	1	5	3	2	4	1	4	5	1	3	5	5	3	1
KJ Prediction	5	5	2	2	5	2	1 (O)	4	2	2 (O)	4 (O)	1 (O)	4 (O)	4	5	1	4	5	3	2	20%
AHP Prediction	12	13	6	9	14	11	3	10	8	5	5	9	4	10	2	1	6	7	14	15	0%

(O): consistency between predictions and actual preferences.

shows the comparison of ranking results of actual consumer preferences and the consumer product preferences that the KJ and AHP methods predicted f product color. It is not difficult for researchers to find that no matter what the product attributes are, there is a large gap between the actual consumer preferences and the consumer preferences that designer predicted when he used a traditional prediction method. For example, for the consumer preference of overall product, the accuracy rates that the KJ method predicted, as well as those that the AHP method predicted, are 20% and 5%, respectively. For the consumer preference of product shape, the accuracy rates that the KJ method predicted and that the AHP method predicted are 15% and 0%, respectively. For the consumer preference of product color, the accuracy rates that the KJ method predicted and that the AHP method predicted are 20% and 0%, respectively. This means that sustainable patterns of consumer product preferences cannot be easily predicted using traditional methods.

4.4. A High Consistency between the Prediction Results of ML Predictions Methods and Actual Consumer Preferences

Table 15. Comparison of ranking results of actual consumer preferences and consumer product preferences that the ML method predicted for the whole product.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20	Accuracy Rate
Actual Preference	3	3	5	5	4	2	2	5	2	3	4	1	4	5	3	2	4	4	3	1
ML Prediction	3 (O)	4	5 (O)	5 (O)	4 (O)	3	2 (O)	4	2 (O)	3 (O)	4 (O)	2	4 (O)	5 (O)	3 (O)	2 (O)	4 (O)	3	3 (O)	1 (O)	75%

(O): consistency between predictions and actual preferences.

shows the comparison of ranking results of actual consumer preferences and consumer product preferences that the ML method predicted for the whole product.

Table 16. Comparison of ranking results of actual consumer preferences and consumer product preferences that the ML method predicted for product shape.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20	Accuracy Rate
Actual Preference	3	3	5	5	4	4	4	5	2	3	5	1	5	5	2	1	5	4	3	1
ML Prediction	3 (O)	3 (O)	5 (O)	5 (O)	4 (O)	4 (O)	4 (O)	4	2 (O)	3 (O)	5 (O)	1 (O)	5 (O)	5 (O)	2 (O)	1 (O)	5 (O)	4 (O)	3 (O)	2	90%

(O): consistency between predictions and actual preferences.

shows the comparison of ranking results of actual consumer preferences and consumer product preferences that the ML method predicted for product shape.

Table 17. Comparison of ranking results of actual consumer preferences and consumer product preferences that the ML method predicted for product color.

Photo No.	No. 1	No. 2	No. 3	No. 4	No. 5	No. 6	No. 7	No. 8	No. 9	No. 10	No. 11	No. 12	No. 13	No. 14	No. 15	No. 16	No. 17	No. 18	No. 19	No. 20	Accuracy Rate
Actual Preference	2	2	4	2	1	3	1	5	3	2	4	1	4	5	1	3	5	5	3	1
ML Prediction	2 (O)	2 (O)	3	2 (O)	1 (O)	4	1 (O)	5 (O)	3 (O)	2 (O)	3	1 (O)	4 (O)	5 (O)	1 (O)	3 (O)	4	5 (O)	3 (O)	1 (O)	80%

(O): consistency between predictions and actual preferences.

shows the comparison of ranking results of actual consumer preferences and consumer product preferences that the ML method predicted for product color. It is not difficult for researchers to find that no matter what the product attributes are, there is a high consistency between the actual consumer preferences and the consumer preferences that the designer predicted when he used the ML method. For example, for the consumer preference of the overall product, the accuracy rate that the ML method predicted was 75%. For the consumer preference of product shape, the accuracy rate that ML method predicted was 90%. For the consumer preference of product color, the accuracy rate that ML method predicted was 80%. This means that sustainable patterns of consumer product preferences can be easily predicted by using ML methods.

4.5. ML Has a Better Ability to Predict Consumer Product Preferences Than Traditional Tools

Figure 6 is the accuracy rate of three methods for analyzing and predicting consumer product preferences. Figure 6 shows that, no matter what the product attributes are, the prediction accuracy of the ML method for consumer product preferences is much higher than that of the KJ method and AHP method for consumer product preferences. For example, for the overall product preference, the ML method has 75% prediction accuracy for the consumer product preferences, while the KJ method and AHP method predict the method consumer product preferences correctly at a rate of 20% and 5%, respectively. Moreover, for the product shape preference, the ML method has 90% prediction accuracy for the consumer product preferences, while the KJ method and AHP method predict the method consumer product shape preferences correctly 15% and 0% of the time, respectively. Finally, for the product color preference, the ML method has an 80% prediction accuracy for the consumer product color preferences, while the KJ method and AHP method predict the consumer product color preferences correctly 20% and 0% of the time, respectively. This means that ML has a better ability to predict consumer product preferences than traditional tools.

5. Discussion

5.1. Implications of Differences in Designer’s Prediction Results Caused by Different Product Attributes

The differences in the designer’s prediction results caused by the different evaluation methods used to evaluate consumer product preferences means that designers predict that consumers use different perception channels to evaluate the product, regardless of the method used to predict sustainable patterns of consumer product preferences. This implies that a designer considers that consumers have the ability to segment and use product attributes to make preference judgments independently. This ability to independently segment the evaluation of product attributes can be supported by other research results, such as research on consumer preferences for sustainable product attributes and farm program features by Yue et al. (2020) [65].

5.2. Implications of Differences in Designer’s Prediction Results Caused by Different Traditional Forecasting Methods

If the designer uses the KJ method to predict the consumer product preference ranking prediction results compared with using the AHP method to predict the consumer product preference ranking prediction results, the researchers find that there is a large gap between the two prediction results. The reason why there is a large gap between the two prediction results is explained as follows. In this study, the AHP method predicts the preference of consumers’ coffee machine samples by comparing the preferences of consumers’ coffee machine samples one by one in pairwise pairs, unlike the KJ method that predicts consumers’ coffee machine samples by grouping coffee machine samples. According to this, the AHP method will obtain the consumer preference results of each coffee machine sample, but the KJ method cannot. Since there are 20 coffee machine samples in this study, the AHP method analysis obtains a total of 20 observations (the 20 observations are the predicted values of the coffee machine sample preferences of 20 consumers). This is the reason why the AHP method predicted values higher than 5. This means that the AHP method observes consumer preferences in an individual way, unlike the KJ method, which observes consumer preferences in a whole way. The accuracy rate of the AHP method’s prediction of consumer preference is very low (0%), and it can only be said that the evaluation pattern of consumer preference predicted by AHP method is different from the actual evaluation pattern of consumers’ preference. Consumers’ actual preference evaluation patterns tend to be carried out in an intuitive and classified way, rather than through one-by-one comparison.

5.3. The Meaning of the High Gap between the Prediction Results of Traditional Forecasting Methods and Consumers’ Actual Preferences

It is not difficult for the researchers to find that no matter which traditional forecasting methods a designer uses, there is a large gap between the prediction results of the forecasting method and the actual sustainable patterns in consumer product preferences, if the KJ method data and the AHP method data are compared with the consumer’s actual product preference ranking results. This means that a designer’s point of view is significantly different from a consumer’s point of view when evaluating a product. In fact, the result of this forecast gap is not difficult to understand. This can be explained by individual differences in cognitive styles. The results of many studies on cognitive styles show that everyone has a unique cognitive pattern, and even the cognitive pattern of the designer itself may be different from that of consumers [65,66,67]. This means that it is not easy for designers to accurately grasp the consumer’s point of view. Therefore, it is not surprising that there will be a significant gap between the prediction results of sustainable patterns in consumers’ product preferences predicted by designers based on subjective consciousness and the actual situation of sustainable patterns in consumer product preferences. Furthermore, this also means that it is not easy to be accurate, and that it easy for results to be distorted if designers themselves use forecasting tools to predict consumer preferences. Therefore, in order to accurately predict sustainable patterns in consumer product preferences, it is best to use the intuitive analysis method of simultaneous involvement of consumers, so that the prediction results are likely to be close to the truth of sustainable patterns in consumer product preferences.

6. Conclusions

The product preferred by consumers is an important design reference for product designers when designing products. Therefore, how to enable product designers to effectively predict the products preferred by consumers is very important for product designers. In order to allow designers to have more convenient and accurate consumer preference product prediction tools, this study conducted a feasibility study on the use of ML to analyze and predict sustainable patterns in consumer product preferences. This study first discusses the accuracy of ML for analyzing and predicting consumers’ overall product preferences, the accuracy of ML for analyzing and predicting consumer preferences of product shapes, and the accuracy of ML for analyzing and predicting consumer preferences of product colors. Then, this study also discussed the accuracy of the traditional forecasting tools—the KJ method and the AHP method—for analyzing and predicting the overall product preference of consumers, the accuracy of the KJ method and AHP method for analyzing and predicting consumer preference of product shapes, and the accuracy of the KJ method and AHP method for analyzing and predicting consumer preference of product colors. The results of the study show that in terms of predicting the overall preference of consumer products, the accuracy rate of the KJ method prediction is 20%, the accuracy rate of the AHP method is 5%, and the accuracy rate of the ML prediction is 75%. In terms of predicting consumers’ product shape preferences, the accuracy rate of the KJ method prediction is 15%, the accuracy rate of the AHP method is 0%, and the accuracy rate of the ML prediction is 90%. As for the prediction of consumer product color preference, the accuracy rate of KJ method prediction is 25%, the accuracy rate of AHP method is 0%, and the accuracy rate of ML prediction is 80%. These accuracy data show that no matter the product attribute dimension, ML has the ability to predict consumer preferences, and ML has a better ability to predict consumer preferences than traditional tools. Therefore, this study believes that ML can be used to analyze and predict sustainable patterns in consumer product preferences. Therefore, this study suggests that product designers can use ML technology to assist in the analysis and prediction of consumer product preferences, so as to improve their grasp of consumer preference for products.

In many predictive research fields, ML has shown excellent predictive ability, so ML has been receiving much attention and is being used for predictive research. In this study, by introducing ML to predict the effect of consumer product preference, this study proves that ML also has the ability to predict sustainable patterns in consumer product preferences. The higher accuracy of ML compared to current forecasting tools is due to the ability of ML to learn the behavioral characteristics of consumer preferences. Therefore, ML can be used in the future to predict consumer preferences for various types of products. On the other hand, besides the physical attributes of the product itself, the factors that affect consumers’ product preferences also include individual differences in consumers, such as value differences, cultural differences, gender differences, etc. However, the predictive ability of individual differences in consumer product preferences has not been tested and discussed in this study. Therefore, in future research, we will explore the feasibility of using ML to predict individual differences in consumer product preferences. This will make the dimension of ML’s ability to predict consumer product preference behavior more diverse and will provide product designers with more reference information on consumers’ product preferences. As such, the chances of product design success can be further improved.