A personalized fragrance recommendation method and system based on personality traits and skin electricity signals

Abstract

The application belongs to the technical field of product personalized recommendation, and discloses a personalized fragrance recommendation method and system based on personality traits and skin electricity signals, which comprises the following steps: obtaining personality trait data and demographic information, collecting EDA signals and PAD scores through a three-stage experiment, extracting features from the EDA signals and screening at least two key features, analyzing the influence and interaction of personality trait data, demographic information and fragrance conditions on physiological and emotional responses, clustering participants, refining fragrance recommendation rules based on the PAD scores of the cluster participants in each personality feature cluster obtained through clustering, training an XGBoost classifier, predicting the cluster to which a new user belongs according to the EPQ score, gender and age of the new user, and then matching the fragrance recommendation rules to realize personalized recommendation. The method not only captures objective physiological responses but also records subjective emotional experiences, and realizes comprehensive evaluation while solving the problem of individual differences.

Description

Technical Field

[0001] This application relates to the field of product personalization recommendation technology, and in particular to a personalized fragrance recommendation method and system based on personality traits and skin conductance signals. Background Technology

[0002] Aromatherapy has gained significant attention in clinical practice and wellness due to its stress-relieving and mood-enhancing effects. Scents such as lavender, peppermint, and lemon are particularly effective in regulating mood. However, most existing recommendation systems use standardized fragrance schemes, failing to adequately consider individual differences in how people respond to olfactory stimuli.

[0003] The potential of aromatherapy in relieving stress and regulating mood is gaining increasing attention, with its mechanism stemming from complex physiological and psychological interactions. Electrodermal response (EDA), a key indicator for measuring these interactions, detects the autonomic nervous system's response to stimuli such as fragrances. Compared to expensive methods like electroencephalography (EEG) or functional magnetic resonance imaging (fMRI), EDA equipment is inexpensive and easy to operate, making it widely applicable in various research scenarios. Studies have shown that different fragrances stimulate sympathetic nerve activity, thereby triggering significant changes in EDA.

[0004] Besides physiological indicators, emotional responses are typically quantified using tools such as the PAD scale (Pleasure-Arousal-Dominance). Compared to complex tools, the PAD scale accurately captures emotional fluctuations induced by environmental stimuli through three key dimensions, reflecting emotion regulation mechanisms in a simple and intuitive way. Studies have confirmed that aromas such as lemon, lavender, and peppermint have significant regulatory effects on emotional arousal and pleasure. Although these physiological and psychological mechanisms are recognized by the academic community, existing research often lacks a systematic exploration of individual differences and has failed to fully explore its application value in personalized interventions.

[0005] The effects of aromatherapy vary from person to person, with individual differences significantly influencing responses to scents. As a key influencing factor on emotions and behavior, personality traits play a decisive role in this process. The Eysenck Personality Questionnaire (EPQ) quantifies individuals using three core dimensions: extraversion (E), neuroticism (N), and psychoticism (P). Compared to the "Big Five" personality model, this assessment method is concise and precise, effectively capturing trait differences closely related to emotions and stress responses, and offering greater ease of operation and interpretation in physiological and psychological experiments. For example, extraversion is associated with physiological arousal levels, neuroticism with negative emotions and stress sensitivity, and psychoticism with impulsivity and risk preference. Furthermore, demographic factors such as gender and age also influence physiological and emotional responses to stress, and therefore must be considered in conjunction with personality traits.

[0006] In aromatherapy applications, ignoring the multidimensional characteristics of individuals makes it difficult to accurately match suitable fragrance solutions to different groups. This cognitive bias leads to significant flaws in existing personalized aromatherapy recommendation systems: many systems, while touting personalized services, only focus on users' subjective preferences or limited demographic information, failing to deeply analyze the differences in physiological responses and emotional experiences among different personality types, genders, and age groups. This raises two major problems—insufficient accuracy: ignoring differences in traits such as neuroticism and extraversion can cause significant fluctuations in the stress-reduction effects of recommended solutions; and weak adaptability: due to the lack of dynamic feedback mechanisms to integrate individual physiological and emotional data, most systems cannot quickly adapt to new users or new scenarios. Therefore, there is a critical gap in systematically incorporating comprehensive individual differences, including EPQ characteristics, into fragrance recommendation models.

[0007] To overcome the limitations of traditional "one-size-fits-all" aromatherapy and the superficial nature of personalized services in existing systems, advanced statistical modeling and machine learning (ML) methods offer promising directions for development. While personalized recommendation systems have achieved significant results, the complex application of aromatherapy (especially the integration of in-depth individual characteristic analysis) in fields such as healthcare and psychological counseling is still uncommon. For example, mixed-effects models have significant advantages in handling repeated measures data and individual differences, distinguishing between fixed effects (such as fragrance conditions and personality traits) and random effects (subject differences), thus accurately assessing the impact of each factor. Generalized linear mixture models (GLMMs) further enhance the analytical flexibility of discrete emotional data. Structural equation modeling (SEM) can reveal the interaction processes between personality traits, fragrance stimulation, and physiological-emotional responses from a causal perspective.

[0008] To address the specific needs of personalized recommendations, combining cluster analysis (used to identify fragrance sensitivity patterns in a multidimensional feature space) with machine learning techniques such as XGBoost yields significant results. XGBoost possesses powerful predictive capabilities for high-dimensional nonlinear data and can extract decision-making logic through interpretable methods such as feature importance analysis or SHAP values. However, despite the immense potential of these methods, a key deficiency remains in this field—the lack of a fragrance recommendation system that comprehensively integrates EPQ-based personality traits with physiological data such as electrodermal conductance mapping (EDA) and subjective affective response (PAD) within a unified and interpretable machine learning framework. Although various machine learning techniques are being actively explored, the market still needs a system that can accurately predict and explain the basis of recommendations—that is, explain why a particular fragrance is recommended to a specific personality trait and the underlying scientific basis. Therefore, bridging this gap by combining multidimensional personal information (EPQ, gender, age) with objective EDA and PAD indicators to establish an interpretable and scalable personalized fragrance recommendation method is a pressing technical problem that needs to be solved. Summary of the Invention

[0009] To address the aforementioned issues, this application provides a personalized fragrance recommendation method and system based on personality traits and skin conductance signals. This method combines electrical conductance analysis (EDA) with the PAD scale to capture both objective physiological responses and subjective emotional experiences, thereby achieving a comprehensive assessment while addressing individual differences.

[0010] According to the first aspect of this application, a personalized fragrance recommendation method based on personality traits and skin conductance signals is provided, the method comprising: Personality trait data and demographic information were acquired, and EDA signals and PAD scores were collected through a three-stage experiment; wherein, the demographic information included gender and age; The EDA signal is subjected to noise filtering, baseline correction, and median smoothing outlier detection to obtain a preprocessed EDA signal. Time-domain features, frequency-domain features, and nonlinear features are extracted from the preprocessed EDA signal. Multiple preliminary candidate features are obtained by combining PAD scoring. At least two key features are selected from the multiple preliminary candidate features. Using at least two key features, personality trait data, demographic information, and fragrance conditions as inputs, a mixed-effects model, hierarchical regression analysis, and multiple logistic regression are employed to analyze the influence and interaction of personality trait data, demographic information, and fragrance conditions on physiological and emotional responses. The participants were clustered using the K-prototype clustering algorithm to obtain... n Individual personality feature clusters; extract fragrance recommendation rules based on the PAD scores of cluster participants in each personality feature cluster; train an XGBoost classifier to predict the cluster to which a new user belongs based on their EPQ score, gender, and age, and then match the fragrance recommendation rules to achieve personalized recommendations.

[0011] Furthermore, the three-stage experiment includes a resting baseline stage, a stress-induced stage, and a fragrance exposure stage; wherein: The resting baseline phase includes: having participants wear EDA sensors and sit quietly for a set time, recording baseline EDA data and completing the initial PAD scale; The stress-inducing phase includes: having participants complete a timed mental arithmetic task, continuously recording EDA data, and completing the PAD scale before and after the task; The fragrance exposure phase includes: randomly assigning participants to four fragrance conditions, delivering fragrances at a rate of 0.3 μL / min using an ultrasonic diffuser, continuously monitoring EDA signals during the exposure period of several minutes, and completing the PAD scale after exposure; wherein, the four fragrance conditions are lemon, lavender, mint and unscented.

[0012] Further, at least two key features are selected from the plurality of preliminary candidate features, including: Spearman correlation analysis with a threshold of 0.8 was used to remove highly correlated redundant features, and an adaptive thresholding method based on mutual information index was used to select at least two preliminary candidate features that are most strongly associated with the emotion category as key features.

[0013] Furthermore, the structure of the mixed-effects model is: EDA function ~ E+N+P+fragrance code + (1|segment) + (1|file), where E is the extraversion score, N is the neuroticism score, P is the psychoticism score, the fragrance code is the code value corresponding to the fragrance condition, (1|segment) is the random effect item of the experimental segment, and (1|file) is the random effect item of the individual participant; Furthermore, the hierarchical regression analysis method includes: establishing a basic model of key EDA elements, then gradually adding personality trait data, fragrance effects and their interaction terms, and analyzing the comprehensive effect and moderating role of each variable on the emotional dimension through changes in F-values ​​and R-squared values; wherein, the emotional dimension includes pleasure, excitement and dominance.

[0014] Furthermore, the aforementioned multinomial logistic regression analysis uses emotion category as the dependent variable and personality trait data and fragrance conditions as independent variables to establish a regression model. The regression model expression is: Emotion Code ~ E + N + P + Fragrance Code. Before establishing the regression model, the correlation between emotion category, fragrance conditions and personality traits is explored through a chi-square independence test, and the Benjamin-Hockberg method is used to control the false discovery rate of multiple comparisons.

[0015] Furthermore, the K-prototype clustering algorithm was used to cluster the participants, resulting in... n Individual personality feature clusters; fragrance recommendation rules are extracted based on the PAD scores of participants in each personality feature cluster; an XGBoost classifier is trained to predict the cluster to which a new user belongs based on their EPQ score, gender, and age, and then the fragrance recommendation rules are matched to achieve personalized recommendations, including: Based on the aforementioned personality trait data, demographic information, and the influence and interaction of fragrance conditions on physiological and emotional responses, the K-prototype clustering algorithm was used to cluster participants. The optimal number of clusters was determined by analyzing the stability of the cost function under multiple k-values. n ,get n The system identifies clusters of personality traits with varying fragrance sensitivities. For each cluster, based on the PAD scores of participants within that cluster, it quantifies the impact of each fragrance on various emotional dimensions and calculates fragrance scores, thereby extracting personalized fragrance recommendation rules for each cluster. Using EPQ scores, gender, and age as input features, the system employs clustering-based methods to refine these rules. nUsing clusters as output categories, an XGBoost classifier is trained. The training process includes stratified sampling, cross-validation hyperparameter tuning, and independent test set evaluation. Simultaneously, feature importance analysis is used to quantify the relative contribution of input variables to cluster affiliation prediction, thereby enabling the prediction of the cluster to which a new user belongs.

[0016] Furthermore, the method also includes: A crossover experiment within the participants was conducted to compare personalized fragrance recommendations with random fragrance allocation, using the stress recovery index. SRI The effectiveness was evaluated based on EDA recovery time, PAD improvement score, and subjective satisfaction.

[0017] Furthermore, the stress recovery index SRI The calculation formula is: ; In the formula, This represents the peak level of skin conductance after stress induction. This represents the recovery value of skin electrical conductivity after fragrance exposure. This represents the baseline value of skin conductance during the resting baseline phase. The EDA recovery time is the time it takes for the EDA signal to recover from peak stress to baseline; the PAD improvement score is a weighted composite score of pleasure, excitement, and dominance scores from before to after the experiment; the subjective satisfaction is assessed using a 1-5 Likert scale, and the Wilcoxon signed-rank test is used to compare the differences between personalized and control conditions to verify the effectiveness of personalized recommendations.

[0018] According to the second technical solution of this application, a personalized fragrance recommendation system based on personality traits and skin conductance signals is provided to implement the method described above. The system includes: The data acquisition module is configured to acquire personality trait data and demographic information, and to collect EDA signals and PAD scores through a three-stage experiment; wherein, the demographic information includes gender and age; The feature selection module is configured to perform noise filtering, baseline correction, and median smoothing outlier detection on the EDA signal to obtain a preprocessed EDA signal, and extract time-domain features, frequency-domain features, and nonlinear features from the preprocessed EDA signal, combine them with PAD scoring to obtain multiple preliminary candidate features, and select at least two key features from the multiple preliminary candidate features. The statistical modeling module is configured to take at least two key features, personality trait data, demographic information and fragrance conditions as inputs, and use mixed-effects models, hierarchical regression analysis and multinomial logistic regression to analyze the influence and interaction of personality trait data, demographic information and fragrance conditions on physiological and emotional responses. The recommended execution module is configured to cluster participants using the K-prototype clustering algorithm to obtain... n Individual personality feature clusters are identified; fragrance recommendation rules are extracted based on the PAD scores of participants in each personality feature cluster; an XGBoost classifier is trained to predict the cluster to which a new user belongs based on their EPQ score, gender, and age, and then the fragrance recommendation rules are matched to achieve personalized recommendations. The personalized fragrance recommendation methods and systems based on personality traits and skin conductance signals according to the various solutions of this application have at least the following technical effects: This application integrates personality traits, demographic characteristics, and physiological-emotional responses (EDA, PAD) through statistical modeling and machine learning to achieve personalized fragrance recommendations. Experimental results show that fragrance generally affects physiological arousal, while emotional responses are mainly regulated by personality traits—especially neuroticism. Validation experiments confirm that the personalized fragrance recommendation system can accurately match user needs. Compared with general methods, fragrance recommendations significantly improve physiological stress recovery, emotional state, and user satisfaction. This application deepens the theoretical understanding of the interaction between aroma and personality and proposes a reproducible methodological framework to support aromatherapy interventions with emotional perception functions in practical applications.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a personalized fragrance recommendation method based on personality traits and skin conductance signals, provided for embodiments of this application; Figure 2 A schematic diagram of a three-phase experimental procedure including baseline, stress-induced and aroma-based recovery, as well as participant profiles and multimodal physiological monitoring, provided for embodiments of this application; Figure 3 A framework diagram of a personalized fragrance recommendation system that maps individual information to fragrance sensitivity rules, provided in an embodiment of this application; Figure 4 A graph showing the variation of R-squared values ​​for different emotion dimensions in a hierarchical regression analysis provided in this application embodiment. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solution of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Studies have shown that personality traits (especially neuroticism and extraversion) influence emotional sensitivity and physiological responses. Electrodermal response (EDA), a key indicator of the excitability of the autonomic nervous system, is frequently used to monitor these physiological changes. In the field of emotion assessment, the Pleasure-Arousal-Dominance (PAD) model describes the subjective emotional state evoked by environmental stimuli (such as fragrance) through three core dimensions; this model is widely used in perception and experience research. Despite these research findings, current perfume recommendation methods rarely combine multidimensional personal characteristics such as personality, gender, and age with physiological data and emotional self-assessment. This situation limits the personalized effectiveness and practical application value of perfume interventions.

[0023] To address the aforementioned problems, this application provides a personalized fragrance recommendation method based on temperature measurement data and deep learning, utilizing personality traits and skin conductance signals. The inventive concept for this recommendation method stems from in-depth research into the following four questions: RQ 1: How do personality traits (such as extraversion, neuroticism, psychoticism), gender, and age affect physiological and emotional responses to fragrance stimuli under stress?

[0024] RQ 2: Do different individuals have different effects on regulating emotions, and how are these effects regulated by personality traits?

[0025] RQ 3: Can machine learning frameworks extract interpretable scent-sensitive rules to enable personalized recommendations?

[0026] RQ 4: In terms of stress recovery and mood improvement, does the proposed personalized fragrance recommendation system outperform non-personalized recommendation systems?

[0027] By addressing the aforementioned issues, this method contributes to both theoretical and practical understanding. Theoretically, it elucidates how individual traits interact with aromatherapy stimuli, thereby influencing emotional and physiological responses. Practically, it provides a scalable, interpretable, and automated personalized aromatherapy solution applicable to the fields of health, psychology, and smart environments.

[0028] Please see Figure 1 This is a flowchart illustrating a personalized fragrance recommendation method based on personality traits and skin conductance signals, provided in an embodiment of this application. The method includes steps S10-S40.

[0029] S10: Obtain personality trait data and demographic information, and collect EDA signals and PAD scores through a three-stage experiment; wherein, the demographic information includes gender and age.

[0030] In this study, 168 participants (84 men and 84 women) aged 18 to 35 years were recruited. All participants completed the Eysenck Personality Questionnaire (EPQ) and provided basic demographic information before participating in the study. Inclusion criteria for the formal experiment required participants to be free from olfactory dysfunction, chronic diseases affecting the autonomic nervous system, or recent use of substances that may affect physiological responses (such as alcohol or caffeine). Exclusion criteria included allergies to the experimental fragrance and the use of medications that may interfere with autonomic responses or mood regulation.

[0031] In some embodiments, such as Figure 2 As shown, all experiments were conducted under controlled laboratory conditions (temperature 22 ± 1.C, noise <40dB, consistent illumination) following a standardized three-stage procedure: During the resting baseline phase, participants sat quietly for 5 minutes while wearing EDA sensors. Baseline physiological data were recorded during this phase. Participants completed an initial PAD mood scale to establish their baseline mood state.

[0032] During the stress induction phase, participants used E-prime software to complete a mental arithmetic task (calculating randomly generated two-digit numbers) for 5-7 minutes with a time limit. EDA (Emotional Analysis) was continuously recorded, and the PAD (Patient Assessment Scale) was administered before and after the task to quantify stress-induced changes.

[0033] During the fragrance exposure phase, participants were randomly assigned to experience four olfactory conditions (lemon, lavender, peppermint, or an unscented control group), delivered via an ultrasonic diffuser (0.3 μL / min). Electrical conductivity changes (EDA) were continuously monitored during the 5-minute exposure, followed by completion of the PAD scale to assess mood regulation effects. To prevent fragrance residual effects and response habituation, each participant was spaced at least three days apart between experimental phases, and the order of fragrance conditions was balanced among participants.

[0034] Multimodal data acquisition employed complementary physiological and psychometric methods: physiological data, electroskin activity (EDA), was recorded at a frequency of 500 Hz throughout all experimental phases. After noise filtering and baseline correction, three types of features were extracted: (1) time-domain features (e.g., response amplitude, latency, recovery time); (2) frequency-domain features obtained through Fourier transform (e.g., spectral center, peak frequency); and (3) nonlinear features capturing the dynamic complexity of EDA patterns (e.g., wavelet entropy, Lempel-Ziv complexity).

[0035] The psychological data-driven assessment of emotional responses employed the validated PAD (Pleasure, Excitement, Domination) emotional scale, which provides three orthogonal dimensions to quantify subjective emotional experiences. This self-report measurement method complements objective physiological data, enabling this embodiment to analyze conscious and unconscious responses to fragrance stimuli.

[0036] S20: Perform noise filtering, baseline correction, and median smoothing outlier detection on the EDA signal to obtain a preprocessed EDA signal, and extract time-domain features, frequency-domain features, and nonlinear features from the preprocessed EDA signal. Combine the PAD score to obtain multiple preliminary candidate features, and select at least two key features from the multiple preliminary candidate features.

[0037] This embodiment performs basic preprocessing on the raw electrodermal activity (EDA) signal, including noise filtering, baseline correction, and outlier detection using median smoothing. From the filtered signal, this embodiment extracts three types of features: (1) time-domain features (amplitude, latency, recovery time); (2) frequency-domain features obtained through Fourier transform; and (3) nonlinear features (entropy, complexity). Combined with the PAD mood scale score, approximately 135 preliminary candidate features are finally obtained, summarized in Table 1.

[0038] Table 1 Summary of Feature Selection Results

[0039] To overcome the challenges posed by high dimensionality while maintaining interpretability, this embodiment implements a three-step feature selection method: 1) Using Spearman correlation analysis with a threshold of 0.8, this embodiment identifies and removes redundant features that show high correlation.

[0040] 2) Mutual Information Screening This embodiment uses the mutual information (MI) index and an adaptive threshold method to identify the features most strongly associated with the emotion category.

[0041] 3) In the set after MI filtering, this embodiment eliminates the remaining redundancy by retaining only the features with the highest information content, thereby obtaining the final set of 14 features: 7 GSR, 3 SCR and 4 SCL features.

[0042] The above methods effectively balance feature richness and model simplification while addressing the inherent nonlinearity of physiological data. The specific selection results are summarized in Table 1.

[0043] S30: Using the aforementioned at least two key features, personality trait data, demographic information, and fragrance conditions as inputs, employ a mixed-effects model, hierarchical regression analysis, and multiple logistic regression to analyze the influence and interaction of personality trait data, demographic information, and fragrance conditions on physiological and emotional responses.

[0044] This embodiment uses a statistical model to systematically study the relationship between personality traits, demographic factors, fragrance exposure, and their effects on physiological and emotional responses. The analysis in this embodiment focuses on clarifying the main effects, interactions, and moderating effects, without making assumptions about the causal direction.

[0045] To account for individual differences and experimental phases in the repeated measures design, and to analyze electrodermal activity data, this embodiment implemented: In the mixed-effects model, perfume type and personality traits (E, N, P) are included as fixed effects, while subjects and experimental segments are included as random effects. The model structure is as follows: EDA Function ~ E+N+P+fragrance Encoding + (1|segment) + (1|file) (1) All models were estimated using the restricted maximum likelihood method (REML) and optimized using the BFGS algorithm.

[0046] Hierarchical regression analysis is a method that examines the combined effects and moderating role of environmental design elements (EDA), fragrance, and personality traits on emotional dimensions (pleasure, excitement, and dominance) by progressively adding variables. First, a basic model of key EDA elements is established. Then, personality traits, fragrance effects, and their interaction terms are systematically added. After each variable is added, the changes in F-values ​​and R-squared values ​​are used to compare different models.

[0047] To specifically assess how fragrance and personality traits influence categorized emotional responses: a multinomial logistic regression analysis was conducted, using emotion category as the dependent variable and personality traits (E, N, P) and fragrance type as independent variables. Emotional encoding ~ E + N + P + Perfume encoding (2) Before regression modeling, this embodiment conducted a preliminary analysis, using the chi-square independence test to explore the association between emotion categories, fragrance conditions, and personality traits. The Benjamin-Hockberg method was used to control for the false discovery rate of multiple comparisons.

[0048] S40: The participants are clustered using the K-prototype clustering algorithm to obtain... n Individual personality feature clusters; extract fragrance recommendation rules based on the PAD scores of cluster participants in each personality feature cluster; train an XGBoost classifier to predict the cluster to which a new user belongs based on their EPQ score, gender, and age, and then match the fragrance recommendation rules to achieve personalized recommendations.

[0049] This embodiment develops a three-step machine learning process to transform personal information into personalized perfume recommendations: 1) K-prototype clustering This embodiment employs the K-prototype clustering method, which identifies user groups with different personality traits by processing mixed data types such as continuous personality traits (E, N, P scores), age, and categorical demographic variables (gender). By analyzing the stability of the cost function under multiple k values, the optimal number of clusters was determined, thereby obtaining the fragrance sensitivity groups based on personality. Detailed characteristics are shown in Table 5.

[0050] 2) Rule Extraction For each identified user group, recommendation patterns are extracted by analyzing their emotional responses (PAD scores) to different fragrances. Specifically, this embodiment quantifies the impact of each fragrance on the emotional dimensions of each personality type, calculates a "fragrance score," and establishes a correspondence between personality traits and optimal fragrances. The final customized fragrance recommendation schemes for each group are summarized in Table 6.

[0051] 3) XGBoost Classification This embodiment trains an XGBoost classifier that automatically categorizes users into appropriate clusters based on their personality traits and demographic data. The model is validated using standard procedures (stratified sampling, cross-validation, and independent test set evaluation) to ensure reliable predictions, thus enabling rapid personalized recommendations without requiring physiological testing on new users.

[0052] In some embodiments, after step S40, the method further includes S50: comparing personalized fragrance recommendations with random fragrance assignments through an within-subjects crossover experiment, using the stress recovery index. SRI The effectiveness was evaluated based on EDA recovery time, PAD improvement score, and subjective satisfaction.

[0053] To evaluate the practical effectiveness of the personalized fragrance recommendation system, this embodiment conducted a controlled validation experiment with 30 healthy adult subjects (15 men and 15 women, aged 18-35 years). A strict within-subjects crossover design was used, with each subject experiencing both the personalized fragrance group and a randomized fragrance control group, and the experimental order was balanced. To minimize residual effects, a washout period of at least three days was set between each group.

[0054] This experiment employed a double-blind design (participants could not distinguish whether the fragrance was personalized or randomly assigned), while the experimenters were also in a single-blind state (unaware of the fragrance allocation logic). The experimental procedure strictly followed the standardized three-stage protocol of the previous study in this embodiment: (1) resting baseline stage (5 minutes); (2) stress induction stage (mental arithmetic task, 5 minutes); (3) fragrance exposure / recovery stage (fragrance contact for 5 minutes, recovery period for 3 minutes).

[0055] Two scenarios were tested: In the personalized recommendation condition, participants received a fragrance selected by the system based on their EPQ score, gender, and age, which predicted their cluster membership and applied a predefined best fragrance rule. In the control condition, participants were randomly assigned to one of four conditions: lemon, lavender, peppermint, or an unscented control.

[0056] Verification Experiment Analysis: This embodiment uses statistical tests and performance metrics to evaluate the effectiveness of the personalized system. For statistical testing, since the data structure does not conform to a normal distribution, the Wilcoxon signed-rank test is used to compare the differences between personalized and control conditions, and the statistical significance is set at p<0.05.

[0057] Stress Recovery Index (SRI): Calculated based on skin conductance level (SCL) data, used to quantify physiological stress recovery. ; In the formula, This represents the peak level of skin conductance after stress induction. This represents the recovery value of skin electrical conductivity after fragrance exposure. This represents the baseline value of skin conductance during the resting baseline phase.

[0058] Among them, the higher SCL peak - SCL baseline The value indicates a better recovery.

[0059] EDA recovery time: The time (in seconds) for the EDA signal to recover from peak stress to baseline. The shorter the time, the faster the recovery speed of physiological stress.

[0060] PAD Improvement Score: A weighted composite score derived from pleasure, excitement, and dominance scores from before to after the experiment.

[0061] Subjective satisfaction: Participants' overall satisfaction rating of the perfume (1-5 Likert scale).

[0062] This embodiment employs a systematic four-stage approach to develop and validate a personalized perfume recommendation system. Figure 3 ): Phase 1: Data collection. 168 participants (gender-balanced) were recruited for a three-phase controlled laboratory experiment: baseline resting, stress-induced, and scent-exposed. Measurements included electrical skin response (EDA), mood rating (PAD scale), personality traits (EPQ), and demographic information.

[0063] Phase Two: Statistical Modeling. This embodiment investigated the effects of personality traits (E, N, P), demographic factors, fragrance conditions, and their influence on physiological and emotional responses using mixed-effects models and hierarchical regression analysis. Multiple logistic regression further revealed interaction patterns.

[0064] Phase 3: Machine Learning Process This embodiment uses K-prototype clustering to identify user groups with different personalities, extracts fragrance sensitivity rules for each cluster, and develops an XGBoost classifier to predict the cluster to which a new user belongs based on their personal characteristics.

[0065] Phase 4: System validation. A crossover experiment was conducted within Group A, where 30 new participants compared personalized fragrance recommendations with random assignments. The results were evaluated using multiple indicators, including stress recovery index, EDA recovery time, mood improvement, and subjective satisfaction.

[0066] The personalized perfume recommendation system described above is used to implement the methods described in the above embodiments. Based on this personalized perfume recommendation system, the feasibility and advancement of the method proposed in this application will be explained in detail below with specific experimental data. First, the core influence of personality traits and demographic characteristics on physiological responses and emotional fluctuations will be explained. Next, the interaction mechanism between personality traits and fragrance stimulation at the emotional level will be analyzed in depth, and the correlation between electrical skin activity (EDA) and emotional state will be revealed. Finally, the fragrance recommendation algorithm based on machine learning will be explained in detail, and the results of the system verification experiment will be presented.

[0067] Mixed-effects model analysis showed that fragrance stimulation exhibited significant main effects on most of the selected 14 electrical skin activity (EDA) characteristics. As shown in Table 2, different fragrance types (lemon, lavender, peppermint, and unscented) were significantly associated with changes in characteristics based on skin conductance response (GSR), skin conductance response (SCR), and skin conductance level (SCL) (all p-values ​​were p < 0.05 after Benjamini-Hochberg multiple comparison correction based on the original analysis results). Regardless of individual personality, fragrance evoked relatively stable fluctuations in EDA at the physiological level. In contrast, personality traits (extraversion (E), neuroticism (N), and psychoticism (P)) did not show significant main effects on EDA indicators in the initial mixed-effects model, nor did their interactions with fragrance type show significant differences. Different personality groups did not show significant differences in EDA responses. This indicates that EDA responses are mainly driven by the common stimulus of fragrance and have low sensitivity to personality dimensions. Since the EDA indicator is primarily driven by the common stimulus of perfume, it has low sensitivity to different personality dimensions. Therefore, when exploring personalized emotional responses, EDA may not be effective in distinguishing different personality groups. In other words, if subsequent clustering or classification analysis focuses on the correlation between individual emotional states and personality traits, it is recommended to exclude the EDA indicator to avoid wasting model degrees of freedom due to the weak interaction between physiological indicators and personality traits.

[0068] Table 2: Fit statistics for all 14 selected features under four fragrance conditions

[0069]

[0070] Emotional response (PAD).

[0071] Generalized linear mixture model (GLLM) analysis revealed significant main effects of personality traits and scent type in predicting emotion categories. As shown in Table 3, neuroticism (N) and scent type exhibited particularly significant main effects on multiple emotion categories (especially categories 0, 4, 5, 7, and 10; all p-values ​​were 0.05). Extraversion (E) also showed a significant main effect on one emotion category (category 2). Psychopathy (P) did not produce a significant main effect on emotion categories in this model. These findings are derived from the results of multiple logistic regression analysis (see the original Table 4 for specific p-values). Notably, although extraversion may still have a significant impact on some emotions, the overall data suggest that its explanatory power for emotion distribution is not as prominent as that of the neuroticism dimension.

[0072] Table 3: Significant Effects of Polynomial Logistic Regression

[0073] Table 4: Chi-square test results of emotions and scents categorized by personality dimensions

[0074] Scent Modulation of Mood: Chi-square independence tests showed a significant association between neuroticism (N), scent conditions, and mood category (χ²(130) = 186.46, p = 0.0009, see Table 4). This indicates that individuals with different levels of neuroticism experience significantly different emotional experiences when exposed to scent stimuli. Extraversion or psychoticism did not show a similar strong association. These findings highlight neuroticism as a key personality dimension, playing a decisive role in influencing scent-induced emotional responses. GLLM results show that both personality (especially N) and scent have a significant impact on mood category, supporting their interaction.

[0075] EDA - Emotional Association.

[0076] Hierarchical regression analysis revealed a fundamental correlation between specific EDA features (such as wavelet entropy and SCR latency standard deviation) and the emotion dimension. By incorporating scent coding and personality traits, the model's explanatory power for emotion was significantly improved, as evidenced by the change in the R² value (see [link to analysis]). Figure 4 It is noteworthy that an interaction effect exists: when a fragrance associated with a high neuroticism level (N) is combined with specific conditions, changes in certain EDA indicators (such as transient phase standard deviation and the 75th percentile of spectral amplitude) may enhance or neutralize the intensity of specific emotions. For example, certain EDA traits increase the probability of negative emotions, while certain relaxing fragrances (such as lavender) have a "protective" effect on individuals with high N traits, significantly reducing the incidence of negative emotions. This suggests that fragrance and personality traits (especially neuroticism) play an important regulatory role in the relationship between physiological EDA responses and subjective emotional experiences.

[0077] Using the K-means clustering algorithm, combined with E, N, P scores and gender characteristics, this embodiment identified six unique personality type groups (where k = 6 was chosen based on the consideration of cost function stabilization). These groups exhibit significant differences in average E, N, and P scores, while P values ​​and gender distribution provide different group characteristics (see Table 5). For example, group 0 is characterized by high average neuroticism and is predominantly female (described as a high-anxiety female group), while group 5 is characterized by high average extraversion and is predominantly male (described as a confident extraverted male group).

[0078] Table 5: Summary of Clustering Characteristics (Personality Trait Scores). The groups derived from K-prototype clustering analysis based on EPQ scale scores and gender correspond to different personality traits: Group 0 is a high-anxiety female group; Group 1 is an independent female group; Group 2 is an extroverted and social male group; Group 3 is an introverted and sensitive male group; Group 4 is a stable and reserved female group; and Group 5 is a confident and extroverted male group.

[0079]

[0080] Analysis of the emotional indicators (P, A, D) of various user groups revealed a high degree of consistency in their emotional responses to fragrance stimulation. This embodiment calculates the cumulative scores of P, A, and D indicators exceeding set thresholds for each fragrance, creating a customized scoring system for each user group. This personalized scoring mechanism enables this embodiment to formulate precise fragrance recommendation rules based on the emotional needs of different groups (see Table 6). For example, lavender is identified as a tranquilizer for Group 0 (high-anxiety women), particularly suitable for those with dominant traits; while Group 5 (confident, extroverted men) showed the highest pleasure rating for unscented fragrances, and peppermint became the preferred awakening scent.

[0081] Table 6: Scent Recommendations Classified by PAD Dimension

[0082] An XGBoost classification model was trained to predict the group to which a new individual belongs based on their personality trait scores (E, N, P), age, and other information. The model was validated using features such as gender. Validation was performed using standard machine learning procedures, including stratified sampling, hyperparameter tuning based on cross-validation, and evaluation on an independent test set to ensure the model's robustness and generalization ability. The trained XGBoost model served as the core predictor for a personalized fragrance recommendation system, classifying new users into one of six personality trait clusters. Feature importance analysis of the XGBoost model quantified the relative contribution of input variables (such as neuroticism score, age, and gender) to cluster affiliation prediction, significantly improving the system's interpretability.

[0083] This embodiment specifically conducted a verification experiment to evaluate the actual effect of the personalized perfume recommendation system, and the results are summarized in Table 7.

[0084] Table 7: Comparison of stress recovery and mood improvement indicators between personalized recommendation and control conditions

[0085] Compared to the control group (randomly assigned fragrance or no fragrance), the personalized fragrance recommendation program significantly improved the recovery from physiological stress. The mean Stress Recovery Index (SRI) of the personalized group reached 0.78 ± 0.11, which was significantly improved by approximately 28% compared to the control group (0.61 ± 0.14). Simultaneously, the EDA recovery time (electrodermal response time) of the personalized group was shortened to 84.6 ± 22.3 seconds compared to the control group (mean 117.2 ± 29.5 seconds), a reduction of approximately 32.6 seconds on average. This indicates that personalized fragrance can more effectively accelerate the recovery process from physiological stress.

[0086] Participants experienced significant improvements in their emotional state through personalized fragrance recommendations. The personalized PAD improvement score was significantly higher in the personalized group than in the control group (2.35 ± 0.42, 67 ± 0.53), indicating an improvement of approximately 41%. This strongly suggests that personalized fragrances are superior in enhancing pleasure, regulating excitement, and strengthening dominance. Subjects also reported significantly higher subjective satisfaction with personalized fragrances (4.2 ± 0.6 on a 1-5 scale), far exceeding the control group's 3.1 ± 0.9, further confirming the superiority of the personalized approach.

[0087] The method proposed in this application aims to break through the "one-size-fits-all" approach in aromatherapy by developing and validating a personalized fragrance recommendation system. Through systematic research on the interaction between multidimensional individual information (especially personality traits, gender, and age) and fragrance stimuli in stress management and mood regulation, this embodiment has obtained several key findings, as follows.

[0088] The research presented in this application yielded several noteworthy findings that contribute to a deeper understanding of how fragrances can be used to promote health, highlighting the crucial role of individual differences.

[0089] This study is the first to confirm, through measurements of electrical skin response (EDA), that fragrance stimulation has an effect on life. Aroma arousal has a broad and generally consistent impact. This finding is consistent with existing literature on the direct effects of olfactory stimulation on the autonomic nervous system. However, it is noteworthy that this study found that personality traits did not significantly modulate these direct skin conductance responses. This suggests that while the level of physiological arousal from fragrance varies among individuals, subsequent emotional interpretation and regulation are more likely dominated by higher-order cognitive and emotional processes—in which personality traits play a decisive role.

[0090] One of the core findings of this application is that the neurotic trait (N) plays a crucial role in moderating people's emotional responses to perfume. This embodiment reveals a significant interaction between neurotic scores and perfume conditions on categorized emotional outcomes. This finding is highly consistent with Eysenck's theory.

[0091] Personality theory suggests that individuals with high neuroticism exhibit stronger emotional responses and are more susceptible to negative emotions. This study further reveals that neuroticism significantly influences an individual's response to therapeutic olfactory interventions. This means that individuals with high neuroticism levels are not only more susceptible to stress, but also experience individual differences in the emotional stabilizing effects of using specific fragrances—especially calming and soothing scents. This finding opens up important avenues for personalized therapy research, as traditional aromatherapy studies often focus on the overall population's primary response to fragrances, failing to fully explore the possibilities of individual differences.

[0092] Furthermore, the hierarchical regression analysis in this application shows that when the combined influence of scent type and personality traits (especially neuroticism) is considered, the correlation between physiological arousal level (conductivity analysis characteristics) and subjective emotional experience is significantly enhanced and becomes clearer. This indicates that the connection between bodily perception and emotional experience is not directly fixed, but rather a result of the dynamic shaping by both the external sensory environment (scent type) and internal persistent traits (personality). Ignoring these moderating factors, as in simplified models, may underestimate the complexity of individual emotional responses and obscure the true physiological mechanisms of emotional states. This finding challenges research designs that simply assess physiological or emotional responses, highlighting the value of integrative models—just as cognitive appraisal theory of emotion emphasizes the role of interpretation in emotional experience, this situation-individual interaction model is equally applicable to multidimensional analysis.

[0093] Finally, and most importantly from an application perspective, the machine learning process integrating K-prototype clustering and XGBoost classification has proven remarkably effective in transforming individual data into personalized fragrance recommendations, bringing tangible benefits to users. Specific validation experiments show that, compared to a randomly assigned fragrance or unscented control group, participants using personalized fragrance recommendations exhibited significantly improved physiological stress recovery abilities (stress recovery index increased by approximately 28%, and skin conductance recovery time shortened by an average of 32.6 seconds), more significant improvements in self-reported emotional state (PAD score), and higher satisfaction. While personalized recommendation systems are widely used in healthcare and e-commerce, their application in aromatherapy, particularly utilizing personality and physiological data, is an emerging field. The results of this study provide strong empirical evidence that this personalized approach is superior to traditional, non-personalized aromatherapy, validating the potential of data-driven technologies in creating more effective interventions.

[0094] This method is of great significance for future scientific research and the practical application of aromatherapy.

[0095] In terms of research methodology, this approach advocates a multimodal data collection strategy—integrating physiological indicators, subjective emotion reports, and stable personality traits, combined with advanced analytical techniques such as machine learning and mixed-effects models. This comprehensive approach can more accurately capture the complex dynamic interaction processes that humans undergo when responding to sensory stimuli. The research framework of this embodiment can provide a paradigmatic reference for subsequent research on individual differences in olfactory and emotion regulation. Theoretically, this approach deepens the understanding of how personality traits (especially neuroticism) modulate the impact of sensory interventions on well-being, pushing the aromatherapy theoretical system beyond generalizations and more explicitly incorporating individual difference variables. The system's scalability also paves the way for expanding research dimensions, such as exploring the mechanisms by which broader influencing factors like genetics, cultural background, or cognitive characteristics affect fragrance sensitivity.

[0096] Practical Applications: The personalized recommendation system developed in this embodiment has been validated and has direct commercialization potential. Clinicians in the fields of psychology and complementary medicine can utilize similar data-driven tools to tailor more precise and effective aromatherapy plans for patients experiencing stress, anxiety, or mood disorders. When used in conjunction with traditional therapies, this is expected to enhance treatment outcomes. In the broader health field, these research findings can provide a basis for the development of personalized fragrance products, intelligent home environment aroma diffusion systems, and even wearable devices that can provide customized olfactory experiences—devices that can help users proactively regulate their emotions. The significant improvements in objective stress indicators and subjective well-being observed in the validation study provide strong evidence for integrating personalized solutions into health services.

[0097] The module units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0098] The above embodiments are only used to illustrate this application and are not intended to limit this application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of this application. Therefore, all equivalent technical solutions also fall within the scope of this application, and the patent protection scope of this application should be defined by the claims.

Claims

1. A personalized fragrance recommendation method based on personality traits and skin conductance signals, characterized in that, The method includes: Personality trait data and demographic information were acquired, and EDA signals and PAD scores were collected through a three-stage experiment; wherein, the demographic information included gender and age; The EDA signal is subjected to noise filtering, baseline correction, and median smoothing outlier detection to obtain a preprocessed EDA signal. Time-domain features, frequency-domain features, and nonlinear features are extracted from the preprocessed EDA signal. Multiple preliminary candidate features are obtained by combining PAD scoring. At least two key features are selected from the multiple preliminary candidate features. Using at least two key features, personality trait data, demographic information, and fragrance conditions as inputs, a mixed-effects model, hierarchical regression analysis, and multiple logistic regression are employed to analyze the influence and interaction of personality trait data, demographic information, and fragrance conditions on physiological and emotional responses. The participants were clustered using the K-prototype clustering algorithm to obtain... n Individual personality feature clusters; fragrance recommendation rules are extracted based on the PAD scores of participants in each personality feature cluster; an XGBoost classifier is trained to predict the cluster to which a new user belongs based on their EPQ score, gender, and age, and then the fragrance recommendation rules are matched to achieve personalized recommendations, including: Based on the aforementioned personality trait data, demographic information, and the influence and interaction of fragrance conditions on physiological and emotional responses, the K-prototype clustering algorithm was used to cluster participants. The optimal number of clusters was determined by analyzing the stability of the cost function under multiple k-values. n ,get n The system identifies clusters of personality traits with varying fragrance sensitivities. For each cluster, based on the PAD scores of participants within that cluster, it quantifies the impact of each fragrance on various emotional dimensions and calculates fragrance scores, thereby extracting personalized fragrance recommendation rules for each cluster. Using EPQ scores, gender, and age as input features, the system employs clustering-based methods to refine these rules. n Using clusters as output categories, an XGBoost classifier is trained. The training process includes stratified sampling, cross-validation hyperparameter tuning, and independent test set evaluation. Simultaneously, feature importance analysis is used to quantify the relative contribution of input variables to cluster affiliation prediction, thereby enabling the prediction of the cluster to which a new user belongs.

2. The method according to claim 1, characterized in that, The three-stage experiment included a resting baseline stage, a stress-induced stage, and a fragrance exposure stage; wherein: The resting baseline phase includes: having participants wear EDA sensors and sit quietly for a set time, recording baseline EDA data and completing the initial PAD scale; The stress-inducing phase includes: having participants complete a timed mental arithmetic task, continuously recording EDA data, and completing the PAD scale before and after the task; The fragrance exposure phase includes: randomly assigning participants to four fragrance conditions, delivering fragrances at a rate of 0.3 μL / min using an ultrasonic diffuser, continuously monitoring EDA signals during the exposure period of several minutes, and completing the PAD scale after exposure; wherein, the four fragrance conditions are lemon, lavender, mint and unscented.

3. The method according to claim 1, characterized in that, From the plurality of preliminary candidate features, at least two key features are selected, including: Spearman correlation analysis with a threshold of 0.8 was used to remove highly correlated redundant features, and an adaptive thresholding method based on mutual information index was used to select at least two preliminary candidate features that are most strongly associated with the emotion category as key features.

4. The method according to claim 1, characterized in that, The structure of the mixed-effects model is: EDA function ~ E+N+P+fragrance code + (1|segment) + (1|file), where E is the extraversion score, N is the neuroticism score, P is the psychoticism score, the fragrance code is the code value corresponding to the fragrance condition, (1|segment) is the random effect item of the experimental segment, and (1|file) is the random effect item of the individual participant.

5. The method according to claim 1, characterized in that, The hierarchical regression analysis includes: establishing a basic model of key EDA elements, then gradually adding personality trait data, fragrance effects and their interaction terms, and analyzing the comprehensive effect and moderating role of each variable on the emotional dimension through changes in F-values ​​and R-squared values; wherein, the emotional dimension includes pleasure, excitement and dominance.

6. The method according to claim 1, characterized in that, The aforementioned multiple logistic regression analysis uses emotion category as the dependent variable and personality trait data and fragrance conditions as independent variables to establish a regression model. The regression model expression is: Emotion Code ~ E + N + P + Fragrance Code. Before establishing the regression model, the correlation between emotion category, fragrance conditions and personality traits is explored through chi-square independence test, and the Benjamin-Hockberg method is used to control the false discovery rate of multiple comparisons.

7. The method according to claim 1, characterized in that, The method further includes: A crossover experiment within the participants was conducted to compare personalized fragrance recommendations with random fragrance allocation, using the stress recovery index. SRI The effectiveness was evaluated based on EDA recovery time, PAD improvement score, and subjective satisfaction.

8. The method according to claim 7, characterized in that, The stress recovery index SRI The calculation formula is: In the formula, This represents the peak level of skin conductance after stress induction. This represents the recovery value of skin electrical conductivity after fragrance exposure. This represents the baseline value of skin conductance during the resting baseline phase. The EDA recovery time is the time it takes for the EDA signal to recover from peak stress to baseline; the PAD improvement score is a weighted composite score of pleasure, excitement, and dominance scores from before to after the experiment; the subjective satisfaction is assessed using a 1-5 Likert scale, and the Wilcoxon signed-rank test is used to compare the differences between personalized and control conditions to verify the effectiveness of personalized recommendations.

9. A personalized fragrance recommendation system based on personality traits and skin conductance signals, used to implement the method as described in any one of claims 1 to 8, characterized in that, The system includes: The data acquisition module is configured to acquire personality trait data and demographic information, and to collect EDA signals and PAD scores through a three-stage experiment; wherein, the demographic information includes gender and age; The feature selection module is configured to perform noise filtering, baseline correction, and median smoothing outlier detection on the EDA signal to obtain a preprocessed EDA signal, and extract time-domain features, frequency-domain features, and nonlinear features from the preprocessed EDA signal, combine them with PAD scoring to obtain multiple preliminary candidate features, and select at least two key features from the multiple preliminary candidate features. The statistical modeling module is configured to take at least two key features, personality trait data, demographic information and fragrance conditions as inputs, and use mixed-effects models, hierarchical regression analysis and multinomial logistic regression to analyze the influence and interaction of personality trait data, demographic information and fragrance conditions on physiological and emotional responses. The recommended execution module is configured to cluster participants using the K-prototype clustering algorithm to obtain... n Individual personality feature clusters; extract fragrance recommendation rules based on the PAD scores of cluster participants in each personality feature cluster; train an XGBoost classifier to predict the cluster to which a new user belongs based on their EPQ score, gender, and age, and then match the fragrance recommendation rules to achieve personalized recommendations.

Images