The acquisition method of the user’s Kansei needs based on double matrix recommendation algorithm

Abstract

In the development of product design, one of the elements of market competition for products is to meet the Kansei needs of users. Compared to features, users pay more attention to whether products can match their emotions, which is Kansei needs. The product developers are eager to get the Kansei needs of users more accurately and conveniently. This paper takes the computer cloud platform as the carrier and based on the collaborative filtering algorithm. We used personalized double matrix recommendation algorithm as the core, and the adjectives dimensionality reduction method to filter the image tags to simplify the users’ rating process and improve the recommendation efficiency. Finally, we construct a Kansei needs acquisition model to quickly and easily obtain the Kansei needs of users. We verify the model using the air purifier as a subject. The results of the case show that the model can find out the user’s Kansei needs more quickly. When the data is more, the prediction will be more accurate and timely.

Keywords

Kansei needs image tags double matrix recommendation algorithm adjectives dimensionality reduction cloud platform

1 Introduction

1.1 Kansei engineering (KE)

The Design is primarily based on user’s needs. With the improvement of living standards, the user’s needs conclude not only the good function of the products, but also the product modeling, color, related human scale, and style. These characteristics are expressed as a perceptual response to emotional needs. Emotion becomes the new direction of research product design. By using products directly, people can realize that products have been able to meet their basic needs, but their emotion from products is the key driver for their purchase [1]. The word ‘Kansei’ in Japanese can express the people feelings and psychological activities about a certain thing, including the process of judging uncertain information from intuitional preferences [2]. Users demand that the products meet their own aesthetic and intuitive judgments, which can be understood as the user’s Kansei needs. Users project their emotions on products, and the products can satisfy their subjective Kansei needs.

Kansei design translates the user’s emotions into product attributes. These attributes can evoke the subjective perception of the users. The selection of Kansei variables related to product design elements is the basis for the development of emotional product design [3], which quantifies the degree for products to meet the user’s Kansei needs. With the development of technologies such as emotion measurement and cognitive computing in the context of inter-disciplinary, people’s emotional awareness can be extracted, and the user’s Kansei needs for the products are gradually clear and expressed as a kind of conscious product, that is, the Kansei image of the product. Designers can design according to the Kansei images, convey the Kansei information of the product, and then put the degree of emotion fit as the standard for optimization [4, 5]. Kansei image is generally expressed by the Kansei adjectives. Determining the target imagery vocabulary of product design is the basis for excavating the perceptual positioning of products. At present, people interact more with the network, generating a large amount of data, which replaces the way of artificially extracting Kansei images. For example, from the online product descriptions and comments, the emotional characteristics and preferences of users can be effectively extracted and summarized.

The Kansei engineering technique transforms the theory of KE into a tool for assisting design and conveys the user’s emotions to the designers, which are divided into three types [6]. The first is a category classification for getting the design details. The second is the Kansei engineering computer system based on expert systems, neural networks, genetic algorithms, etc. It has a variety of databases concluding the Kansei database, image database, knowledge-base. The third is the Kansei engineering modeling that builds the rule base, using mathematical models.

Computer-aided Kansei engineering system (KES) includes an expert system and databases. Yukihiro Matsubara et al. proposes a hybrid Kansei engineering system that includes a forward KES oriented consumers and a backward KES that supports designers [7]. In the forward KES, the user enters the Kansei words and the system finds the candidates using the database, and then the designer designs the sketch according to them. In the backward KES, the designer enters the sketch into the system, and the system recognizes the pattern of the sketch and gives an evaluation. Based on this Kansei mathematical model, many new models have been derived, such as virtual Kansei engineering(VIKE), which can display the user-satisfied candidates in the virtual space [8]. The groupware design system is a collaborative Kansei engineering system. In addition to the Kansei database, it also includes a smart system that supports designers for collaborative design. Designers can communicate with each other and follow the recommendations proposed by the system, as shown in Fig. 1.

Fig. 1

The groupware design system.

These Kansei engineering systems contain ample databases. The most important thing is to investigate users, collect their Kansei, and get a lot of data to construct the database at the beginning of product development. Since human emotions are extremely subjective, Ota et al. believe that the adjectives related to emotions can express human emotions about objects independently [9]. However, subjective emotions are still difficult to express scientifically. Because it is difficult for users to clearly express their Kansei needs, scholars began to pay attention to finding the laws of human emotion and quantify Kansei needs [10].

To quantify the Kansei, the Semantic differential, the Cluster analysis, and the Genetic algorithm are usually adopted. Huang, Y.X. et al. [11 –13] introduced how to use the Semantic differential to obtain the human perception of the image of the product. And he also showed how to achieve the dimension reduction of the Kansei needs through using the Cluster method. However, for different products, the Semantic differential needs to be repeatedly implemented. And the amount of data analyzed in this method is generally small, which has a negative impact on accuracy. Nevertheless, the Semantic differential method is still considered to be an effective traditional method, which solved the problem of user’s inconsistent understanding of a vocabulary in the past. Lee H.C. et al. [14] combined the shape grammar with a genetic algorithm to process product morphological features through interactive operations, which helping in the expression of perceptual design intent. Zhai L.Y. et al. [15] proposed a systematic approach to Kansei engineering based on the dominance-based rough set theory, that is reasonably accurate in establishing the knowledge between product design elements and human Kansei. It can deal with uncertain information in Kansei engineering. Jyh-Rong Chou [16] presented a Kansei evaluation approach based on the technique of computing with words to quantify and prioritize human Kansei preferences. The user’s Kansei needs is a dynamic demand, which will change with the changes of the times, social trends, living environment, and personal concepts. Therefore, it is necessary to quantify and constantly update the user’s Kansei needs for the product designers [17], which can make the products more in line with the user’s “feeling in the heart”.

1.2 The personalized recommendation algorithm

The emergence of the Internet cloud platform, which aggregates various network resources and constitutes a service chain, has a large amount of user information in real time. It becomes the best choice for getting the latest user’s needs. The personalized recommendation algorithm is generated to deal with a large amount of Internet information, which can help users to filter massive amounts of network information. The personalized recommendation algorithm uses the group effect of users to gradually make users to groups and try to recommend a set of recommendations to a group of users. This method can provide users with accurate suggestions to increase user stickiness. The personalized recommendation algorithm have been widely used in e-commerce, film and food webpages, and so on. This kind of algorithm puts the user’s search object, access object, and evaluation as a database for relevance search and personalized recommendation. However, there are few webpages that use the users’ Kansei needs as a database for product recommendation. The Collaborative filtering algorithm (CF) is the most widely used recommendation algorithm [18]. The basic idea of this algorithm is to find specific neighboring user groups, and then other products with high scores in these user groups are found. After that the algorithm can predict user’s scorings. Finally, the user can be recommended to look at the high score item predicted by the algorithm. This algorithm is based on the users rating database to output recommendation results.

When the platform is just starting to operate, the amount of data is not large and the accuracy of the Collaborative filtering algorithm will decrease. Moreover, the new information generated by users cannot be effectively used in the platform immediately [19]. For these problems of the Collaborative filtering algorithms, many scholars conducted related research. Nikolaos Polatidis et al. [20] proposed a dynamic multi-level collaborative filtering method in order to solve the problem of reduced recommendation accuracy. This method is based on positive and negative adjustments of the user’s similarity values and may use either a static or a dynamic multi-level approach. Liu Yue et al. [21] proposed a novel method named Item Life Cycle based Collaborative filtering (ItemLC-CF), taking both user’s preference and item’s popularity into consideration. Toledo R.Y. et al. [22] detected noisy ratings by characterizing items and users by their profiles, and then fix these noisy ratings to increase the accuracy of such recommender systems. The noisy means exposing recommender systems data to inconsistencies. SM Yang et al. [23] developed a rule-based inference model for human sensibility engineering system. This method improves the efficiency of recommendation. These methods have effectively improved the recommendation quality of the Collaborative filtering algorithm.

As the Collaborative filtering algorithm continue to be improved, the improved algorithms can be divided into three categories: memory-based approaches for CF, model-based approaches for CF, and hybrid CF [24]. The memory-based approaches determine the similarity of two users by comparing the scores of two users on a set of objects to make mutual recommendation between users. The K-Nearest neighbor algorithm(KNN) is the main methods. The memory-based methods effectively solved the problem of recommendation efficiency, but they have data sparsity and do not conclude scalable algorithm [25]. The model-based approaches rely on a statistical modelling technique and can solve the problems [25]. The Smoothing-based model is based on the clusters generated from the training data, but the form of the data has constraints [26]. The Bayesian probabilistic model is for factoring the rating matrix [27]. The Similar function-based model rely on the database without additional information, and is suitable for non-negative matrix factorization. These methods improve the quality of recommendations and are helpful for classical matrix factorization.

The most popular model applied to Collaborative filtering is Matrix Factorization (MF) algorithms. It decomposes the rating matrix into 2 or more user-items matrix. The advantages of MF are as follows: (1) The scholars generally believe that the MF has higher accuracy compared to the memory-based approaches. (2) As the number of users increases, the results are more accurate. (3) Due to the small number of users in the early stage, the recommender requires a learning process, and the recommendation system is highly efficient after the learning phase ends.

Ortega F. et al. [24] performed group recommendations using Matrix Factorization based Collaborative filtering. He proposed a new recommendation algorithm that decomposes the user scoring matrix into two non-negative matrices and uses matrix reconstruction to predict the user’s other score. Matuszyk P. et al. [28] presented unsupervised forgetting techniques that make recommender systems adapt to changes of user’s preferences over time. They use an incremental matrix factorization algorithm and extended this algorithm by the forgetting techniques. The accuracy of the matrix decomposition algorithm is significantly improved [27]. The advantage of the double matrix recommendation algorithm is that it can decompose the scoring matrix as quickly as possible without updating the decomposition process. So we use a modified double matrix recommendation algorithm based on the Collaborative filtering in this study. In this paper, the objects to be scored are Kansei image and Kansei products, which the double matrix factorization is most suitable for this study. Moreover, the cloud platform has a large number of users and a large amount of information, which can help us improve the accuracy of recommendation. Meanwhile, if the recommendation is more accurate, the probability of the click and evaluation will increase, and the recommended accuracy rate will also increase.

As users browse the web, it is important to quickly generate results from a limited number of visits and reviews. In order to speed up the calculation process and reduce the complexity of the decomposition process, we reduce the dimension of the Kansei image to increase the efficiency of the double matrix recommendation algorithm.

This paper constructs a Kansei needs acquisition model based on the Internet cloud platform, uses the personalized double matrix recommendation algorithm as the core, improves the efficiency of the algorithm by reducing the adjectives dimension, which can acquire the Kansei needs of users accurately and quickly, and in real time.

2 Kansei needs acquisition method

2.1 Kansei needs acquisition procedure

The acquisition model of Kansei needs proposed in this paper is based on the Internet cloud platform. The cloud platform is different from other computer-aided Kansei engineering systems. It integrates users, design, manufacturing, marketing, and other resources, and has the advantages of aggregating resources, sharing opened, and collaborative communication. It has the function of input needs and output candidates, which is already available in the KE system. It also has the ability for users to browse, rate, collect, and purchase products. At the same time, the platform automatically collects the real-time preferences of every user to continuously update the database. It’s not just a tool for assisting designers to create, but also a platform that is useful to users and can attract users to browse. It forms a complete service chain.

The cloud platform needs to have the ability to allow users to view products and rate them. The acquisition model of Kansei needs in this paper is based on the user’s rating of the products. The acquisition procedure is shown in Fig. 2. First, we screen out adjectives with less relevant based on the Adjectives dimension reduction method in Section 2.2. These adjectives become the image tags of the products, and each product initially has several image tags. Then, the scoring given by the users for the product on the platform is put as a weight, and the “like”, “neutral” or “dislike” evaluation of the image tag of the products is put as matrix content to generate a user-image scoring matrix and a product-image scoring matrix. That forms the basis of the double matrix recommendation algorithm. Through the calculation of the algorithm in Section 2.3, we can finally get the Kansei needs of users for the product.

Fig. 2

Perceptual needs acquisition procedure.

2.2 Adjectives dimension reduction

The image tags are used to describe the Kansei characteristics of a product. We use adjectives as the image tags, and these words are added by the back-end. There are many adjectives describing Kansei images, and there are certain links between some adjectives, like ‘elegant’ and ‘pure’, which will make users feel confused about the choice of images. In order to eliminate the repeated adjectives and accurately confirm user preferences, we need to remove the association between adjectives and find relatively independent words as image tags. For the recommendation system, the first step is to reduce the dimensionality of the words and filter out independent kansei images [29].

After the secondary adjectives of the Kansei words database were initially selected, we designed the adjectives-color attribute questionnaire referring to the attribute-based semantic similarity measuring method. The attribute-based semantic similarity measurement method is based on the lexical emotional attributes to classify vocabulary. The data were counted and input into the Statistical Product and Service Solutions(SPSS) for correlation analysis to obtain independent primary adjectives.

Every object has certain characteristics, the way people distinguishing things is to judge by identifying their characteristics. These characteristics can be described as “attributes”. The more similar the attributes are between the kansei words, the more consistent their meaning is [30]. The attributes and the scores corresponding to attributes are used to describe the characteristics of Kansei words.

For each Kansei adjective, we set two different attributes for it, and each attribute corresponds to a score that describes it. A high score indicates that the word has this attribute very much. The characteristics of each adjective are represented by several < attribute, value>, reflecting the description of the Kansei word. Taking the word “natural” as an example, using colors for its attributes. The result is < green, yellow, 0.8, 0.2>, that is, the degree score of green corresponding to “natural”’ is 0.8, and the score of yellow is 0.2. Using the coldness and warmth for its attribute, the result is < cold, warm, 0.6, 0.4>, that is, the score of the coldness corresponding to “natural” is 0.6, and the score of the warmth is 0.4. We compare the degree of similarity between Kansei adjectives by comparing the corresponding values of their same attributes. Since there is no correlation between the different attributes, repeated evaluations of the words are avoided. The steps of the adjectives dimensionality reduction method are shown in Fig. 3.

Fig. 3

Adjectives dimensionality reduction method.

Based on the images proposed by Kobayashi [31], We invited five designers with more than ten years of design experience to help us screene out some adjectives that have a significant difference. This step is to use the designers’ Kansei judgment to reduce the workload. As shown in Table 1.

Table 1

Secondary adjectives

Young	Cheerful	Elegant	Romantic
Simple	Classical	Bright	Cute
Energetic	Active	Heavy	Delicate
Intense	Dynamic	Flexible	Firm
Happily	Funny	Sunny	Open
Lively	Free	Gorgeous	Full
Soft	Relaxing	Rough	Hard
Rural	Substantial	Pure	Fashionable
Soft	Archaistic	Mature	Natural
Quiet	Homely	Nostalgic	Solemn
Charming	Rigorous

We chose the primary vocabularies with higher independence based on the secondary adjectives.

Firstly, we designed the vocabulary-color attribute questionnaire referring to the attribute-based semantic similarity measure method. We published the questionnaire on the cloud platform and invited the industrial designers who have registered and have at least one year of design experience to complete the questionnaire. We have collected 50 valid questionnaires. For one word, they need to select three pairs of colors, and only one of each pair of colors should be selected. Each option is 2 points, that means if the designer selects“green” and two points for “green”. There are a total of 6 points for a word containing three choices. Finally, we summarize all the questionnaires and calculate the score of each color attribute of each adjective. Part of the questionnaires is shown in Table 2.

Table 2

Part of the questionnaires

Young	□green-red□
	□blue-yellow□
	□black-white□
Cheerfull	□green-red□
	□blue-yellow□
	□black-white□
Elegant	□green-red□
	□blue-yellow□
	□black-white□

Secondly, we find out the maximum and minimum values of each color and use the normalized Equation (1) to standardize the data for each color [32]. Then we use SPSS for correlation analysis to obtain the correlation between different perceptual words. According to the principle of the hypothesis test, we initially assume that any two words are completely uncorrelated, and test whether there is a clear relationship between the two words. When the probability that the hypothesis is rejected is less than 0.01, we believe that there is a significant relationship between the two words. The value of significance level is 0.01, that is, when the two are completely uncorrelated, the probability of wrongly rejecting the null hypothesis is 0.01. $M^{'} = \frac{M - M_{min}}{M_{max} - M_{min}}$ (1)

Where M′ is the standardized value. M is the score of a certain color of this word. M_min is the minimum value of the color and M_max is the maximum value of the color.

Finally, we confirmed the words that have no correlation to each other by comparing the significance. Take the data of the four words “full’, “substantial”, “natural”, and “rigorous” as an example. The correlation between them is shown in Table 3.

Table 3

Correlation of words

	Full	Substantial	Natural	Rigorous
Full	/	0.302	0.143	0.182
Substantial	0.302	/	0.323	0.003
Natural	0.143	0.323	/	0.159
Rigorous	0.182	0.003	0.159	/

As can be seen from the table, the correlation between “full” and “rigorous” is 0.003 which less than 0.01, indicating that the two words have a clear correlation. So only one word can be chosen as a primary adjective. There are 15 words that we filtered finally. As shown in Table 4.

Table 4

Primary adjectives

Cute	Fashionable	Relaxing	rough
Natural	Classical	Elegant	Soft
Rural	Bright	Simple	Mature
Active	Intense	Open

The 15 adjectives obtained by the dimension reduction will be applied in the cloud platform as the Kansei image tags of the product, and for users to evaluate and become the basis of the double matrix algorithm.

2.3 The double matrix algorithm

The 15 image tags selected in Section 2.2 are used as adjectives to describe every product. During visiting the website, the users evaluate the image tags of products according to his or her preference and scores the product to calculate the image tag weight. The user’s evaluation of the image tags in the platform includes: “like”,“neutral”, and “dislike”. For example, if a user likes product A and thinks that the image tags 1 and 2 make him like it, he can give this product a high score and give a “like” to the image tags 1 and 2. Then we build a user-image tag scoring matrix and product-image tag scoring matrix. Next, we use the double matrix algorithm to predict the vacancy in the matrix. Finally, we can restore the complete user-product scoring matrix to get the score of Kansei needs. The specific steps are as follows.

Step 1: The user-product scoring matrix is constructed based on the user’s score of the product in the platform. Assuming that the number of existing users is a, the user set is represented as U ={ user₁, user₂, user₃, ·· · , user_a }. The number of products that can be scored is b, then the product set is represented as I ={ item₁, item₂, item₃, ·· · , item_a }. The user-product scoring matrix is constructed as shown in Table 5. The total score is 5 points. The higher the score is, the more the users like a product. The vacancy in the matrix indicates that the user does not browse these products. Some of them may become the object of recommendation by the platform in the future. We will fill it up in the next steps.

Table 5
User-product scoring matrix

Item₁ Item₂ ...... Item_a

User₁ / 1 ...... 1

User₂ 3 / ...... 5

...... ...... ...... ...... ......

User_b 4 2 ...... 4

	Item₁	Item₂	......	Item_a
User₁	/	1	......	1
User₂	3	/	......	5
......	......	......	......	......
User_b	4	2	......	4

Step 2: Through the user’s evaluation of the image tags, including “like”, “neutral”, or “dislike”, the corresponding effective image tag of the product in the platform is determined. For product b, using $U^{b} = {u_{1}^{b}, u_{2}^{b}, . . ., u_{n}^{b}}$ to indicate the number of the user’s “like” for the image tag 1, tag 2, tag 3 ... , and using $D^{b} = {d_{1}^{b}, d_{2}^{b}, . . ., d_{n}^{b}}$ to indicate the number of the user’s ‘dislike’ for the image tag 1, tag 2, tag 3 ... . When the number of “dislike” of the image tags in product b is larger than the number of “like”, it means that most users do not agree with this tag and this tag belongs to the invalid image tag. We can calculated the quantity of user’s support for image tags based on Equation (2), where P is the quantity of user’s support. The P of the invalid image tag may be less than or equal to 0, and both are fixed to 0. $P^{b} = U^{b} - D^{b} = {u_{1}^{b} - d_{1}^{b}, u_{2}^{b} - d_{2}^{b}, . . ., u_{n}^{b} - d_{n}^{b}}$ (2)

Step 3: We can calculated the total number N^b of the user’s support for image tags in product b according to Equation (3). A product-image tag matrix is constructed based on the user’s support for image tags, as shown in Table 6. Invalid image tags are filled with 0.

Table 6

Product-image tag matrix

	Tag₁	Tag₂	......	Tag_c
Item₁	$(u_{1}^{1} - d_{1}^{1}) / N^{1}$	$(u_{2}^{1} - d_{2}^{1}) / N^{1}$	......	$(u_{c}^{1} - d_{c}^{1}) / N^{1}$
Item₂	$(u_{1}^{2} - d_{1}^{2}) / N^{2}$	0	......	$(u_{c}^{1} - d_{c}^{1}) / N^{2}$
......	.._....	......	......	......
Item_a	$(u_{1}^{b} - d_{1}^{b}) / N^{b}$	$(u_{2}^{b} - d_{2}^{b}) / N^{b}$	......	0

$N^{b} = (u_{1}^{b} - d_{1}^{b}) + (u_{2}^{b} - d_{2}^{b}) + . . . + (u_{n}^{b} - d_{n}^{b})$ (3)

Step 4: According to Equation (4) and Equation (5), we can get the score of User_x on Tag_x, respectively. And then a user-image tag scoring matrix is constructed based on the scores, as shown in Table 7.

Table 7

User-image tag matrix

	Tag₁	Tag₂	......	Tag_c
User₁	1	/	......	1
User₂	3	5	......	/
......	......	......	......	......
User_b	4	/	......	2

$N_{c}^{a} = \sum_{b = squo}^{i} λ_{1} λ_{2}$ (4) $S_{a, c} = \sum_{b = 1}^{i} \frac{λ_{1} λ_{2}}{N_{c}^{a}}$ (5)

Where S_a,c represents the rating on Tag_c from User_a, i is the number of products. R_b is the rating from User_a on product b, $N_{c}^{a}$ is the total support degree of User_a for product b. λ₁ = 1 and λ₂ are condition variables. When the image tags in the product contain Tag_c, λ₁ = 1, otherwise λ₁ = 0. When the user gives the tag “like”, λ₂ = 2, “dislike” for λ₂ = 0, “neutral” for λ₂ = 1. The score of “like” is set to 2 in order to indicate that ‘like’ has a higher degree of approval.

Step 5: There are still vacancies in the user-image tag scoring matrix obtained in step4. We use the method of the score prediction to construct a matrix with high similarity to the original matrix, and predict the scores and fill the scores into the vacancy. The calculation method of the score prediction is as shown in Equation (6) and Equation (7). $N_{j} = \sum_{W \in C_{k (j)}} λ_{W} N_{j, W}$ (6) $S_{a, j} = \sum_{W \in C_{k (j)}} \frac{N_{j, W}}{N_{j}} λ_{W} S_{a, W}$ (7)

Where S_a,j is the score of the vacant part to be filled, N_j is the number of image tags that have scored in a product. C_k(x) is a set containing x image tags, j and w are image tag items. N_j,w. S_a,w is the quantity of Tag_W and Tag_j in a product. is the score of User_a for Tag_W in the user-image tag scoring matrix obtained in step 4. λ_W is a condition variable that indicates whether there is a score. When the score is known, λ_W = 1, otherwise λ_W = 0.

Step 6: When the user-image tag scoring matrix is filled, we can obtain the user-product scoring matrix which is fully filled according to Equation (8). The filled user-image tag scoring matrix and user-product scoring matrix are all non-negative matrices. $R_{a, b} = S_{a, c} \times S_{b, c}^{T}$ (8)

Where R_a,b is filled user-product scoring matrix, S_a,c is the user-image tag scoring matrix, $S_{b, c}^{T}$ is the anti-matrix of product-image tag scoring matrix.

Step 7: After obtaining the user-product scoring matrix, we can calculate the score of the Kansei needs of users for the product according to Equation (9). $P = R_{a, b} \times S_{b, c}^{T}$ (9)

Where P is the image tag scoring matrix of products, R_a,b is the user-product scoring matrix, $S_{b, c}^{T}$ is the anti-matrix of product-image tag scoring matrix.

According to the scores of Kanseineeds, the platform can get the user’s affection degree for a kind of style, and recommend other products in the same style. With the increase of users and the evaluation, the Kansei needs of users become more and more accurate, and the accuracy of recommendation is higher, which will attract more users to log in. At the same time, for a specific product, the image tags that have a great deal of preference will also be recorded in the platform. And the Kansei needs data of users will be provided to the product designers and manufacturers.

3 Application

We verify the proposed method in the cloud platform and compare the results with the Semantic differential.

The air purifier is selected as the application object. We use the Kansei needs acquisition method in this paper and the Semantic differential method to obtain the Kansei imageries of the air purifiers, and then we contrast the two results.3.1. Subsection.

3.1 The cloud platform search

The “ShengHong” cloud platform provided us with a platform to obtain the Kansei needs of users. The consumers use this platform to search for demanded products, collect and purchase products they like, and evaluate them. They can also search for a vocabulary to find products.

We posted this “Search for air purifiers” message on the platform. When the consumers browsed the webpages, they rated the air purifiers and evaluated the imagery tags. We reserved a month to enrich the amount of user’s data. The steps of user search are as shown in Fig. 4.

Fig. 4

The step of user search.

We summarized all the Kansei needs reports generated by the platform finally. Among all the Kansei imagery tags, “elegant”, “relaxing”, and “fashionable” are the three most popular tags for users. The “rough”, “bright”, and “intense” are the three tags that users are least interested in. The Kansei needs report is shown in Fig. 5.

Fig. 5

The Kansei report.

3.2 The Kansei images based on the semantic differential

The Semantic differential method as a traditional method to recognize Kansei images is widely used in Kansei engineering and considered to be reliable. It proposed a basic emotional systems to identify kansei variance and mapping functions in determining transforned values. Although the Semantic differential can process a smaller amount of data than other algorithms, it is very effective in identifying the emotional words of users. We compare the results of the Kansei needs acquisition method in this paper with the results of the Semantic differential, to verify whether the results are abnormal. Firstly, we selected the 15 words after dimension reduction above and added the different words to form the phrases. Secondly, eight air purifiers were selected as the sample library for the Semantic differential method. Thirdly, we designed a rating form to investigate the degree of bias in each of the 15 pairs of imagery words, as shown in Table 8, and publish it on the Internet cloud platform and wait for users evaluation and collect the rating data. The respondents selected the scores of the imageries according to their style. 7 points indicate the user extremely prefer to the left adjective, which is the imagery tag in the platform. 1 point indicates the user extremely prefer to the right.

Table 8
The degree of bias in the opposing image words

7 6 5 4 3 2 1

Cute □ □ □ □ □ □ □ Naughty

Fashionable □ □ □ □ □ □ □ Old-fashioned

Open □ □ □ □ □ □ □ Traditional

Soft □ □ □ □ □ □ □ Rigid

Elegant □ □ □ □ □ □ □ Vulgar

Mature □ □ □ □ □ □ □ Childish

Bright □ □ □ □ □ □ □ Plain

Rural □ □ □ □ □ □ □ Urban

7 6 5 4 3 2 1

Relaxing □ □ □ □ □ □ □ Hurry

Natural □ □ □ □ □ □ □ Artificial

Classical □ □ □ □ □ □ □ Modern

Active □ □ □ □ □ □ □ Dreary

Simple □ □ □ □ □ □ □ Gorgeous

Rough □ □ □ □ □ □ □ Introverted

Intense □ □ □ □ □ □ □ Peaceful

Overall product scoring(1– 5):

	7	6	5	4	3	2	1
Cute	□	□	□	□	□	□	□	Naughty
Fashionable	□	□	□	□	□	□	□	Old-fashioned
Open	□	□	□	□	□	□	□	Traditional
Soft	□	□	□	□	□	□	□	Rigid
Elegant	□	□	□	□	□	□	□	Vulgar
Mature	□	□	□	□	□	□	□	Childish
Bright	□	□	□	□	□	□	□	Plain
Rural	□	□	□	□	□	□	□	Urban
7	6	5	4	3	2	1
Relaxing	□	□	□	□	□	□	□	Hurry
Natural	□	□	□	□	□	□	□	Artificial
Classical	□	□	□	□	□	□	□	Modern
Active	□	□	□	□	□	□	□	Dreary
Simple	□	□	□	□	□	□	□	Gorgeous
Rough	□	□	□	□	□	□	□	Introverted
Intense	□	□	□	□	□	□	□	Peaceful
Overall product scoring(1– 5):

We recalled all the rating forms and calculated the average score of each group of image words using computer. The higher the score, the more the users prefer the imagery tags in the platform. The lower the score, the less the users agree with the adjectives in the platform. The last item in the evaluation is the user’s scoring of the product. The higher the score, the more the designer likes the product. The scoring will be used as the weight of the image words. Calculating the score of the image words according to Equation (10). $S^{'} = \frac{\sum_{x \in A} R \times S}{R_{a} \times x}$ (10)

Where S′ is the final score of a certain image word, A is the questionnaire set, x is the number of samples, and x = 8. R represents the user’s scoring of the product, S is the scoring of this image word, is the average by weight of the scores of a certain word.

In the Kansei images of the air purifier obtained through the semantic differential method, the Kansei images that the users prefer are “elegant”, “fashionable” and “open”, which become the user’s Kansei needs for the air purifier. The most unacceptable images are “bright”, “rough”, and “intense”. The results obtained by the Kansei needs acquisition method based on the cloud platform are basically the same as those obtained by the semantic differential method.

3.3 Discussion

Firstly, the results obtained by the method proposed in this paper are basically the same with the Semantic differential when the amount of data is not large in the initial stage of the platform. This indicates that the results obtained by this method are feasible within the normal range. As user data is updated, it can have more database volume to obtain a result that is more responsive to user preferences. This can help designers understand the real needs of users.

Secondly, the model-based collaborative filtering algorithm is widely considered as a approach with highly accurate in recommendation algorithms. The greater the number of users on the Internet cloud platform, the higher the accuracy of the method. Although the system requires a period of learning during the initial operation of the platform, the system operation will speed up afterwards. The Collaborative filtering algorithm based on double matrix factorization is not only efficient, but also very suitable for this paper which users’ evaluation objects include intent tags and products. It can easily decompose the user-item matrix, and the algorithm does not need to update the decomposition process.

Thirdly, the user preferences are not constant. When the new users browse the platform, or the old users revisit the platform, they create new data. The existing data will be re-analyzed in the back-end. Therefore, the results we obtained in the platform are synchronous with the user’s Kansei needs at the time. These needs are constantly updated as the user’s viewpoints change, and can better express the current Kansei needs of users. As the amount of user data will gradually increase, in order to encourage users to evaluate image tags and products, and reducing computing time, and simplifying the scoring matrix, we reduced the dimensions of the image tags, which can increase the efficiency of the double matrix recommendation algorithm.

Fourthly, the platform can recommend for users other products according to the Kansei needs of users. These products may be other air purifiers of similar style, or may not be limited to air purifiers. They may be other products but have the same Kansei images that users like. The users may not have considered these products before, but the products are consistent with the style that users prefer. These products may give users a new aesthetic experience and be their next purchase target.

In this study, we reduced the dimension of the provided adjectives to get vocabulary with a big difference and to avoid users being confused about adjectives with similar meanings and difficult to distinguish. But users have different educational backgrounds, living environments, and personalities. They have different understandings of each imagery tag. The meaning of image tags also changes over time. Therefore, there may be such a situation, that is, the same imagery tag, different people think it corresponds to different styles of products. Although this problem relates to psychology and linguistics, beyond the scope of this article. It has a reminder for the research, that is, in the subsequent research, the meaning of the imagery tag should be standardized, to avoid making the user confused about the adjective given by the platform. We may be able to provide some samples corresponding to different tags for users so that they can visually know the meaning of the tag.

At the same time, we also need to classify users according to their occupation, age, and other imformation based on their registration content in the future work. Because the user’s living environment is different, their needs are different. By classifying users, we can get the results of different groups of perceptual needs, or you can set weights according to the group effect of users to estimate the overall perceptual needs.

The classic matrix rating has no weight effect. But due to the items in the matrix which have different importance to the recommendation results, there have been some related research that pointed out the importance of giving weight to the items in the matrix. In the future work, if the number of user ratings and the rating deviation can be integrated into the weight of the items, we can work out weighting matrix, which can effectively eliminate unimportant information and strengthen the role of key information in the recommendation process.

4 Conclusion

Acquiring Kansei needs of the users is the first stage in the development of a Kansei engineering system. This paper puts forward a method to quickly obtain the Kansei needs of users, using the big data in the computer cloud platform, and taking the personalized double matrix recommendation algorithm as the core. We use the air purifier as a case for application in the platform. We compare the results obtained by the method proposed in this paper with the Semantic differential. The results are not abnormal, it is feasible. The method proposed in this paper is based on the double matrix factorization algorithm. Due to the users’ rating databas, the results obtained are more accurate for the preference of the user. In order to increase the recommendation efficiency of the algorithm, we reduce the adjectives dimension to simplify the calculation of the rating matrix. In addition, we use the Internet to get the user’s data which is real-time updated. The method proposed in this paper improves the efficiency of obtaining the Kansei needs of users, and can acquire the Kansei needs of users accurately and quickly, and in real time. In the future work, how to standardize the Kansei image tags and how to weight items in the rating matrix are worthy problem to study.

Footnotes

Acknowledgment

We are grateful to ZhengJiang Shenghong Industrial Design Creative Co., Ltd. for providing the platform for the verification and helping us with the implementation of our approach.

References

Song

, Intelligent system for products personalization and design using genetic algorithm, Journal of Intelligent & Fuzzy Systems 37 (2019), 63–70.

Yang

C.C.

, A classification-based Kansei engineering system for modeling consumers’ affective responses and analyzing product form features, Expert Systems with Applications 38(2011), 11382–11393.

Monhanty

S.N.

, Parvin

J.R.

, Kumar

K.V.

, et al., Optimal rough fuzzy clustering for user profile ontology based web page recommendation analysis, Journal of Intelligent & Fuzzy Systems 37 (2019), 205–216.

Jianning

, Qinwei

, Jianghua

and Yun

, Research on evolutionary design of product multi-image styling, Computer Intergrated Manufacturing Systems 20 (2014), 2675–2682.

Hayat

, Ali

M.I.

, Alcabtud

J.C.R.

, Cao

B.Y.

and Tariq

K.U.

, Best concept selection in design process: An application of generalized intuitionistic fuzzy soft sets, Journal of Intelligent & Fuzzy Systems 35 (2018), 5707–5720.

Razza

and Paschoarelli

L.C.

, Affective Perception of Disposable Razors: A Kansei Engineering Approach, Procedia Manufacturing 3 (2015), 6228–6236.

Matsubara

and Nagamachi

, Hybrid kansei engineering system and design support, International Journal of Industrial Ergonomics 19 (1997), 81–92.

Nagamachi

, Kansei engineering of a powerful consumer-oriented technology for product development, Applied Ergonomics 33 (2002), 289–294.

Ota

and Aoyama

, Aesthetic Design Based on “Kansei Language”, Initiatives of Precision Engineering at the Beginning of a Millennium Jul, (2002), pp. 917–921.

10.

Q.X.

, Kansei knowledge extraction based on evolutionary genetic algorithm: an application to e-commerce web appearance design, Theoretical Issues in Ergonomics Science 16 (2015), 299–313.

11.

Huang

Y.X.

, Chen

C.H.

and Khoo

L.P.

, Products classification in emotional design using a basic-emotion based semantic differential method, International Journal of Industrial Ergonomics 42 (2012), 569–580.

12.

Lin

M.C.

, Wang

C.C.

and Chen

M.S.

, Using AHP and TOPSIS approaches in customer-driven product design process, Computers in Industry 59 (2008), 17–31.

13.

Kess

A.O.

and Rossitza

, Evaluating the affective aspect of intuitive interactions using image schemes, Journal of Computing and Information Science in Eingeering 18 (2018), 1–12.

14.

Lee

H.C.

, Herawan

and Noraziah

, Evolutionary grammars based design framework for product innovation, Procedia Technology 1 (2012), 132–136.

15.

Zhai

L.Y.

, Khoo

L.P.

and Zhong

Z.W.

, A dominance-based rough set approach to Kansei Engineering in product development, Expert Systems with Applications 36 (2007), 393–402.

16.

Chou

J.R.

, A Kansei evaluation approach based on the technique of computing with words, Advanced Engineering Informatics 30 (2016), 1–15.

17.

Yanagisawa

and Fukuda

, Interactive reduct evolutional computation for aesthetic design, Journal of Computing and Information Science in Engineering 5 (2005), 72–78.

18.

Baltrunas

, Makcinskas

and Ricci

, Group recommendations with rank aggregation and collaborative filtering, Proceedings of the 4th ACM Conference on Recommender Systems 10 (2010), 119–126.

19.

Mishra

, Kumar

and Bhasker

, A web recommendation system considering sequential information, Decision Support Systems 75 (2015), 1–10.

20.

Polatidis

and Georgiadis

C.K.

, A dynamic multi-level collaborative filtering method for improved recommendations, Computer Standars & Interfaces 51 (2017), 14–21.

21.

Liu

, Cai

, Ren

P.F.

and Gu

Z.Z.

, Item life cycle based collaborative filtering, Journal of Intelligent & Fuzzy Systems 36 (2019), 2743–2755.

22.

Toledo

R.Y.

, Mota

Y.C.

and Martínez

, Correcting noisy ratings in collaborative recommender systems, Knowledge-Based Systems 76 (2015), 96–108.

23.

Yang

S.M.

, Ahn

and Seo

K.K.

, Development of a rule-based inference model for human sensibility engineering system, Journal of Mechanical Science & Technology 19 (2005), 743–755.

24.

Ortega

, Hernando

, Bobadilla

and Kang

J.H.

, Recommending items to group of users using Matrix Factorization based Collaborative Filtering, Information Sciences 345 (2016), 313–324.

25.

Bobadilla

, Ortega

, Hernando

, et al., Improving collaborative filtering recommender system results and performance using genetic algorithm, Knowledge-Based Systems 24 (2011), 1310–1316.

26.

Xue

G.R.

, Lin

, Yang

, et al., Scalable collaborative filtering using cluster –based smoothing, Proceedings of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (2005), pp. 114–121.

27.

Hernando

and Ortega

, A non negative matrix factorization for collaborative filtering recommender systems based on a Bayesian probabilistic model, Knowledge-Based Systems 94 (2016), 188–302.

28.

Matuszyk

, Vinagre

, Spiliopoulou

, Jorge

A.M.

and Gama

, Forgetting techniques for stream-based matrix factorization in recommender systems, Knowledge and Information Systems 55 (2018), 275–304.

29.

Cynthia

G.C.

and Tanati

T.N.

, Variation in the strength of lexical encoding across dialects, Journal of Phonetics 58 (2016), 87–103.

30.

Wan

and Angryk

R.A.

, Measuring semantic similarity using Word Net- based Context Vectors, In Proceeding of systems on Man and Cybernetics, Montreal, Que., Canada OCT. (2007), 908–913.

31.

Kobayashi

, Color image coordinatesl, People’s Fine Arts Publishing House: Beijing, China, 2006.

32.

Golov

and Rönnbäck

, Big Data Normalization for Massively Parallel Processing Databases, Computer Standards & Interfaces 54 (2017), 86–93.