Enhancing Urban Design through Kansei Engineering and Deep Learning: A Comprehensive Framework for Emotional Design

Abstract

This paper delves into the transformative impact of deep learning in the integration of geosocial network-derived urban imaginaries into the Kansei engineering framework. By harnessing the dynamic and rich datasets obtained from platforms such as Instagram and Twitter, the study constructs a comprehensive Kansei semantics database, capturing real-time, location-specific emotional responses tied to urban elements. Unlike traditional statistical methods, this approach allows for a nuanced understanding of the non-linear correlations between Kansei words and urban design properties.

The deep learning model’s architecture, featuring Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and attention mechanisms, outperforms statistical methods in deciphering intricate patterns within emotional responses. The model employs an embedding layer for continuous vector space representation of Kansei semantics and CNNs for extracting spatial hierarchies from urban design images. Incorporating Recurrent Neural Networks (RNNs) enables the capture of evolving sentiments over time, essential for a comprehensive understanding of emotional dynamics.

Attention mechanisms in the model focus on specific aspects of Kansei semantics or urban design images, enhancing precision in correlating emotional responses with design properties. The final regression layer synthesizes these representations, providing a sophisticated tool for generating emotionally resonant urban environments. A case analysis shows the model’s superiority in discerning complex, non-linear relationships, surpassing traditional statistical methods. This innovative approach advances the field by bridging the gap between subjective user experiences and objective design properties, contributing to the development of more user-centric, culturally sensitive urban designs.

Introduction

The current architectural design bases its project configuration on functional aspects typical of the type of architecture to be built, and on formal aspects typical of a design idealized by the designer architect. The differentiating concept of this type of architectural projection is emotional design, which sets the tone for design centered on the user and their subjective and collective conception, which will guarantee a structuring of the proposal that integrates the architectural design with the sensations experienced by the user when faced with certain elements or situations converging in the living space. The kansei methodology is one of the pioneering and most complete methodologies in the field of emotional design. As (Nagamachi, 1995) indicates, this engineering consists of capturing users’ expressions and translating them into correct design elements. The term Kansei is a Japanese word, whose meaning is related to the words “sensation”, “emotion” or “feeling”. It is used to denote the qualities that an object has to transmit pleasant emotions when using it. Therefore, Kansei is the impression or stimulus that a person may have when faced with a certain product, environment, or situation, when they use their senses. From this, it can be said that Kansei Engineering (KE) is a tool that allows us to capture the emotional needs of users and establish mathematical prediction models to relate the characteristics of products with these emotional needs. “In this regard, Kansei Design (KD) is a novel holistic approach to users’ perceptions at cognitive level which seems appropriate to applications in architectural design.” (Carateli & Misuri, 2020, p.1)

The conceptualization of the problem sets a clearly defined pattern, as a consequence of a superficial assessment of thoughts based on preceding ideologies, where the systematic approach to design development has as its genesis the experiences and situations of daily life, not from advanced visuals impressions. All of this configures a thought where the intrinsic senses of the human being provide a nuance of conceptual projection referred to the creation of architectural spaces. In this context (Fuentes, 2018) points out that the qualitative concepts of people mark a correlation between the experiences lived and perceived in the city, with conceptions of a formal nature, such as defining architectural forms as “heavy” or “light”, etc. And where “the criticism of key concepts that such as identity urban imaginaries, refer to the point of view of the city’s actors, as an indispensable factor to understand the “complex nature of design” (Fuentes, 2018, p . 1). At this point is where architects are called to establish a methodology as a basis for their design that synthesizes their own life experience as an urban-social actor, in order to interpret and understand the different points of view of others.

Applying Kansei Engineering to urban imaginaries could involve understanding how people emotionally respond to different aspects of urban spaces. For example, it could be used to understand how people feel about public transportation, parks, buildings, and other elements of the urban environment. This information could then be used to design urban spaces that better meet the needs and desires of their inhabitants. Although this methodology gives us a clear vision and abstraction of the user’s perceptions, these results are still theoretical and linguistic, and “cannot directly generate products, which is a fatal defect of product form design” (Quan et. Al, 2018). For instance, as (James, 2011) propose and explain the possibility of capturing patterns of individual perception, interaction, and sensation in a determinate space environment, accessing through data that are publicly accessible, surveys and other data sources.

This paper explores the way of how the relationship between urban imaginaries and Kansei engineering can be improved with the use of machine learning tecniques. Through a case study analysis and empirical research, investigating how Kansei engineering principles can enhance the emotional quality of urban spaces and contribute to the development of more user-centric and culturally sensitive city environments. Generating in this way a theoretical approach on how to establish an emotional methodology for the creation of urban spaces relating historical, emotional, cultural, subjective, and objective aspects.

Numerous studies have utilized Kansei evaluation, employing statistical analyses associated with the semantic differential (SD) method to quantify human perception and discern Kansei preferences. Common statistical methods applied in Kansei evaluation, such as correlation coefficient analysis, principal component analysis, factor analysis, and multiple regression analysis, assume a linear relationship between customer preferences and improvements or deteriorations in product attributes. (Chou, 2016)

Kansei engineering methodology

The fundamental concept involves describing the essence of a product from two distinct viewpoints, based on a previously chosen domain:

• Semantic Description: This pertains to portraying the product idea in terms of meaning and significance.
• Product Properties Description: This involves detailing the attributes and characteristics of the product.

Each of these descriptions creates a distinct vector space. In the subsequent synthesis phase, these spaces are amalgamated, indicating the relationship between the semantic impact and the elicited product properties. Only after these steps are completed can a validity test be conducted. Following this, the two vector spaces are updated, and the synthesis step is iterated. The model describing the association between the semantic aspect and the application space is developed when satisfactory results are obtained from this iterative process.

Domain selection

The domain within Kansei Engineering is related to the ideal concept of a certain product; includes the definition the type of market and the target audience, the market niche, as well as the specifications of the new product, environment, or situation. To establish the domain, it is necessary to collect samples of products that represent it, whether similar concepts or even unknown design solutions (Nagamachi, et al 2004). “The task in the first step is to define the research object and collect data, including Kansei words and product form images” (Quan, et al 2018).

Semantic space

The expression “Semantic space” was coined by Charles Osgood, indicating that products can be described within a vector space that is defined by semantic expressions. Kansei engineering in its model methodological uses what is known as “Semantic Differential”, which is basically an evaluation technique psychological theory on perception and emotional impacts on the human mind, created by Osgood himself, George Suci and Percy Tannenbaum in 1957. It is worth mentioning that this technique collects data on the subjective perceptions and emotions generated by a product, object or image, but not about its meaning conceptual. (Schütte, et al 2004).

His assumption was to make a separation between the object and an object-representing sign:

“The object, which is a pattern of stimulation which evokes reactions on the part of anorganism, and the sign, ‘which is any pattern of stimulation which is not the object but yet evokes reactions relevant to ‘object’-conditions under which these holds lying the problem for theory.”

This concept can be illustrated using a hammer as an example. The spoken term ‘hammer,’ for instance, does not equate to the physical object itself. The former consists of patterns of sound waves, while the latter involves a combination of visual, olfactory, and tactile experiences. The word ‘hammer’ triggers a specific kind of response that is somehow connected to the actual object ‘hammer.’ In essence, the spoken or read word ‘hammer’ serves as a sign for the object ‘hammer’ as Schütte explained.

Compilation of kansei words

As a first step the words or kansei of lower levels that define the established domain and semantically describe the object. “Kansei words” are generally adjectives, although they could also be verbs and can be collected from different sources (surveys and opinions of people related or not related to the domain in question, literature related, brainstorming, manuals, experts, among others). The number of adjectives varies depending on the product and the level or design in a conceptual or detailed manner, so the maximum and minimum number of the kansei words in this way the survey does not become tedious nor is the information too imprecise due to a small number of semantics. These kansei words are hierarchical, so the more specific kansei or lower level (“slow”, “fast”) are related and linked to those of the higher level (“kinetic”), which are connected to the property space in the diagnosis stage.

In order to facilitate the synthesis phase, the data is collected in a standard way, trying to select more words than necessary because the final result may have a certain restriction in its validity. Depending on the domain that was considered, the number of existing Kansei can vary generally between 50 and 600 words.

Kansei structure identification

Thoruhgt semantic differentiation technique we can identify these higher level kansei between different semantics or given adjectives. However, words or semantics can be recorded that are not related to or do not belong to the domain, so it is necessary to reduce the list according to different criteria and methods until obtaining the semantic space that best identifies the terminology with the greatest impact on the subjective perception of the user, so there are different methods.

One of them is through factor analysis, which is carried out in statistical analysis programs and software. Another method is the so-called “focus groups” or group concentration, which tries to bring together people who are experts in the subject to bring together the different words previously selected according to their affinity and determine a representative kansei word or higher level for each group. Once the number of kanseis has been reduced, a subjective judgment is made of each of the relationships of words formed. The result of this stage is a compiled list that includes the rankings of the chosen words in comparison to the artifacts used by each participant.

At this point, the Kansei of higher levels are identified through different methods that can be divided in two groups:

Manual method: it is generally developed by experts in Kansei Engineering, which tries to group the words according to the preferences and needs of the participating groups only, such as the affinity diagram.

Statistical method: allows quantifying the affinity of the different kansei words within the semantic space, collected through interviews or questionnaires. Some of the most used statistical methods are:

• Principal components analysis (Osgood and Suci, 1969)

• Factor analysis (Osgood and Suci, 1969)

• Group analysis (Hair et al., 1995)

Property space

To determine which properties chosen are most relevant to the user within a context evaluated, the importance of these properties must be established and make this a criterion for selection. Just as in the generation of semantic space, a procedure that has a model is structured.

Property Collection

At this point, the different possible properties that are related to the proposed domain are compiled from different bibliographic sources. Generally, existing products can be taken as inspiration for the identification of a wide variety of properties, which are similar to what is wanted as an identifying and characteristic point within a company or registered trademark. In addition, trying to identify new properties, integrating the previous sections, and bringing together creative thinking and new ideas as a method. Therefore, the creativity of designers is very important in this part, generating drawings, sketches or models of the entire product or a part of it.

Property selection

Here we proceed to select the most important properties, those that generate the greatest affective and emotional impact according to certain rules. Different tools can be used such as the mentioned “focus groups”, one-on-one interviews, the affinity diagram or also the “Pareto Diagrams” developed by Berman and Klefsjö (Nagamachi, 1995).

Typically, in Kansei engineering methodologies, morphological analysis is employed to break down a product into distinct elements (product properties) and further categorize these elements. In the context of product design, graphical definitions of product form features are often utilized, as they offer simplicity and ease of understanding for individuals dealing with intricate shapes and patterns (Quan, et al 2018).

Finally, the property space is obtained, which will be used in the synthesis phase to identify relationships between the Kansei words and the chosen properties. Likewise, each property chosen must be parameterized, which involves breaking down the design elements that make up each axis, therefore helping to define the factors that will be part of the design in each specific property.

Synthesis

In this step, semantic space and property space are related, so that for each kansei word you must find several properties that will affect this word, quantifying the size of the affective impact of these properties on each kansei. There are different quantitative and qualitative tools to identify the relationships of the two mentioned spaces.

Identifying Relationships

At the time a number of different qualitative and quantitative tools are available as Schütte state.

Manual methods: it is easy to use and requires fewer resources than the other methods, among which the following stand out: – Identification category (Nagamachi, 1995)

Statistical methods: they are used to process a large amount of data from the two spaces (semantic and properties), but in this case, the different tools used will have to be modified to meet the methodology of Kansei Engineering. Statistical tools include:

• Regression Analysis (Shütte, 2005)

• General Linear Model (Arnold, 2002)

• Type I Quantification Theory (Komazawa and Hayashi, 1976)

• Correlation coefficient by Spearman’s hierarchies (Spearman’s Rho)

Other methods: There are other tools that use classifications and valuation methods that are based on intelligent computer systems capable of classifying and defining similarities in data:

• Genetic algorithm (Nishino et al., 1999)

• Neural Networks (Ishihara, et al. 1996)

• Fuzzy Set Theory (Shimixzu and Jindo, 1995)

• Rough Set Theory (Mori, 2002; Nishino, Nagamachi and Ishihara, 2001)

The main aspect of the relationship is to precisely find the point of conjunction that determines a result of linear progression and statistical correlation based on similarity trends. Through this method it is possible to establish the existence or not of similarities between the variables – in this case the kansei and the properties – to make evident the joint variability and in this way to be able to typify what happens with the data. This method allows you to identify the Spearman correlation coefficient (Spearman’s Rho) which is a measure of linear association that does not use the real values of the variables, but the measurement by ranges by comparing them with each other. This coefficient allows measuring the degree of relationship or association between two or more random variables (Gonzales and Restrepo, 2007), determining whether two variables are correlated, “that is, whether values of one variable tend to be higher or lower for higher or lower values of the other variable” (Ortega, Pendás, Ortega, Abreu, & Cánovas, 2009, p.2).

In this step, we need to divide the products and construct the questionnaire by combining the samples and Kansei words with the 7-point SD scale. As an example, a research was conducted in Spain where the focus was to determinate the differences between architects and nonarquitects perception of urban design. The study involved 140 participants, including 70 architects and 70 non-architects, all affiliated with the Universidad Politécnica in Valencia. The sample size was chosen with the criterion of having 8 cases per adjective. Each participant, comprising professors, research staff, administration, and services staff, responded to three questionnaires, resulting in a total of 420 replies. The questionnaires contained 59 adjectives in Spanish to describe citizens’ emotional responses when evaluating city areas. Additionally, variables related to residential and investment choices were included. The field study utilized 74 images of different neighborhoods in Valencia, and the participants were asked to express their opinions in a spontaneous manner. The data processing involved discriminant analysis, ANOVA, and techniques such as differential semantics and regression analysis to identify and rank the semantic axes influencing residential and investment choices. The findings were visually represented through semantic profiles, allowing a comparative analysis of the perceptions of architects and non-architects regarding specific neighborhoods in the city. Statistical analyses were conducted using SPSS 16.0 (Llinares et. al, 2011).

Numerous studies have utilized Kansei evaluation, employing statistical analyses associated with the semantic differential (SD) method to quantify human perception and discern Kansei preferences. Common statistical methods applied in Kansei evaluation, such as correlation coefficient analysis, principal component analysis, factor analysis, and multiple regression analysis, assume a linear relationship between customer preferences and improvements or deteriorations in product attributes. (Chou, 2016).

Nevertheless, these preferences often exhibit a non-linear pattern attributed to uncertain, imprecise, or incomplete data arising from human error, recording discrepancies, or arbitrary guesses, leading to potentially unreliable outcomes. The distinctive non-linear behavior necessitates specialized analytical techniques to discern the diverse effects that variations in Kansei attributes might exert on customer preferences. To address the quantitative aspects of perceptual information, various non-linear inference techniques, such as neural networks, have been developed and utilized in modeling Kansei evaluation systems.

Deep learning model setup

Semantic space spanning

Within the context of Kansei engineering, a sophisticated methodology is employed to enhance emotional design in urban environments. The foundational structure of the deep learning model incorporates a database of Kansei semantics, which encapsulates the collective emotional responses and imaginaries associated with urban spaces. This database is intricately curated through an innovative process that includes leveraging APIs from geosocial networking platforms, employing advanced text mining techniques, and utilizing web scraping methods to extract rich and diverse datasets reflecting the nuanced emotional landscape of urban experiences. The incorporation of geosocial networking APIs allows for the collection of real-time, location-specific data, offering a dynamic and authentic representation of urban imaginaries. By tapping into platforms such as Instagram, Twitter, and others, the Kansei semantics database becomes enriched with user-generated content, including textual descriptions, images, and associated sentiments related to various urban elements and spaces. Urban design elements, visually represented in images, serve as pivotal attributes within this Kansei engineering framework. These elements contribute essential information to the database, offering a holistic understanding of the emotional dimensions embedded in various urban features. The Kansei semantics database, thus fortified with geosocially derived content, becomes a powerful repository of emotional insights connected to specific locations, creating a contextualized foundation for the subsequent deep learning process.

Properties space spanning

Urban design elements, visually represented in images, serve as pivotal attributes within this Kansei engineering framework. These elements contribute essential information to the database, offering a holistic understanding of the emotional dimensions embedded in various urban features. The Kansei semantics database, thus fortified with geosocially derived content, becomes a powerful repository of emotional insights connected to specific locations, creating a contextualized foundation for the subsequent deep learning process.

Relationship model building

To maximize the effectiveness of this wealth of information, the Kansei engineering structure adopts a nuanced strategy. First, it employs the semantic database to unravel the intricate emotional nuances associated with different urban elements, utilizing methods such as semantic clustering and sentiment analysis. This process ensures that the emotional responses are categorized and analyzed in a manner that reveals underlying patterns and relationships that signify emotional responses.

Subsequently, the structured knowledge gleaned from the Kansei semantics database is fed into a sophisticated deep learning model, which is specifically designed to understand and learn the complex mappings between emotional responses and visual representations of urban design elements. The model’s architecture comprises multiple layers capable of capturing hierarchical and abstract representations of the emotional aspects associated with urban design. This architecture is tailored to uncover latent patterns and associations within the Kansei semantics, enabling the model to generate a more refined and reliable emotional design for urban spaces, leveraging the power of neural networks to discern intricate and non-linear correlations between Kansei semantics and urban design properties, and contributing to the development of emotionally resonant urban environments.

The deep learning model’s architecture involves multiple layers capable of capturing hierarchical and abstract representations of the emotional aspects associated with urban design. Initially, the model utilizes an embedding layer to represent Kansei semantics in a continuous vector space, transforming textual and contextual information for subsequent effective analysis. Simultaneously, for urban design properties represented in images, Convolutional Neural Networks (CNNs) are incorporated to excel in capturing spatial hierarchies and patterns within visual data, extracting meaningful features from urban design elements. Subsequently, a merge layer integrates semantic representations from Kansei semantics with visual features extracted from urban design images, crucial for establishing cross-modal correlations and aligning textual descriptions with visual characteristics in emotional responses.

Considering the temporal sequences inherent in urban imaginaries collected through geosocial networking APIs, Recurrent Neural Networks (RNNs) are introduced to capture evolving sentiments and changing perceptions over time, providing a more comprehensive understanding of emotional dynamics to discern subtle correlations, attention mechanisms are integrated, enabling the model to focus on specific aspects of Kansei semantics or urban design images relevant to emotional responses. The model’s final layer is a regression layer that synthesizes learned representations and outputs predictions for emotional design responses based on urban design properties, capturing complex and non-linear relationships identified through preceding layers. Furthermore, transfer learning techniques may be applied to enhance generalization capabilities, allowing pre-trained models on large datasets related to urban imagery or emotional responses to be fine-tuned on the specific Kansei semantics dataset.

Conclusion

In conclusion, this paper explores the transformative integration of deep learning into Kansei engineering, specifically applied to urban design. Leveraging geosocial network-derived data, a comprehensive Kansei semantics database is constructed, capturing real-time, location-specific emotional responses tied to urban elements. The deep learning model, with its intricate architecture featuring CNNs, RNNs, and attention mechanisms, proves to outperform traditional statistical methods in deciphering complex patterns within emotional responses. The model’s ability to discern non-linear correlations between Kansei words and urban design properties surpasses conventional techniques.

The application of Kansei engineering to urban imaginaries provides a user-centric approach to architectural design, emphasizing emotional experiences in city spaces. The machine learning-driven enhancement of Kansei principles contributes to the development of culturally sensitive and emotionally resonant urban environments. Through empirical research and case study analysis, this paper establishes a theoretical framework for generating emotionally impactful urban spaces, addressing the limitations of traditional Kansei evaluation methods. This innovative approach facilitates a bridge between subjective user experiences and objective design properties, fostering a new era in urban design that prioritizes the diverse emotional needs of inhabitants.

References

Fuentes Farías, F. (2019). On the condition of urban and architectural design. Estoa, (15), 33– 42, J. L. G. (2018). Urban imaginaries and labor practices in the traders of the public route of the historical center of Querétaro. Andamios, 15(38), 91–112. https://doi.org/10.29092/uacm.v15i38.653

Nagamachi, M. (1995). Kansei engineering and comfort. Int. J. Ind. Ergon. Volume 15, Issue 1, January 1995, Pages 3-11. Department of Industrial and Systems Engineering, Hiroshima University, 1-4-1, Kagamiyama, Higashi-Hiroshima 724, Japan. https://doi.org/10.1016/0169-8141(94)00052-5

Caratelli, Paolo & Misuri, Alessandra. (2020). Kansei Design and Its Applications in Architecture and the Built Environment. 10.1007/978-3-030-44267-5_7.

Quan, H., Li, S., & Hu, J., (2018). Product Innovation Design Based on Deep Learning and Kansei Engineering. School of Mechanical Engineering, Guizhou University, Guiyang 550025, China. Department of Computer Science and Engineering, University of South Carolina, Columbia, SC 29208. https://doi.org/10.3390/app8122397

James, Matthew. (2011). The emergent urban imaginaries of geosocial media. GeoJournal. 78. 10.1007/s10708-011-9439-1.

Chou, J. (2016). A Kansei evaluation approach based on the technique of computing with words. Department of Creative Product Design, I-Shou University, Kaohsiung City 84001, Taiwan, ROC. https://doi.org/10.1016/j.aei.2015.11.001

Schütte, S., Eklund, J., Axelsson, J. & Nagamachi, M., (2004). Concepts, Methods and Tools inKansei Engineering. Quality and Human-Systems Engineering, Department of Mechanical Engineering,Linköping University, SE-58183 Linköping, Sweden. Collage of Human and Social Environment, Hiroshima International University,555-36 Gakuendai, Kurosecho, Hiroshima 724-0695, Japan.

Restrepo, L., & Gonzáles, J. (2007). From Pearson to Spearman. [De Pearson a Spearman]. Revista Colombiana de Ciencias Pecuarias, 20(2), 183–192. Retrieved from http://www.scielo.org.co/scielo.php?script=sci_arttext&pid=S0120-06902007000200010

Martínez Ortega, R. M., Tuya Pendás, L. C., Martínez Ortega, M., Pérez Abreu, A., & Cánovas, A. M.  (2009). EL COEFICIENTE DE CORRELACION DE LOS RANGOS DE SPEARMAN CARACTERIZACION. Revista Habanera de Ciencias Médicas, 8(2).

Carmen Llinares, Antoni Montañana, Elena Navarro, “Differences in Architects and Nonarchitects’ Perception of Urban Design: An Application of Kansei Engineering Techniques”, Urban Studies Research, vol. 2011, Article ID 736307, 13 pages, 2011. https://doi.org/10.1155/2011/736307