A method for constructing a vegetable oil skin feel evaluation model and application

Abstract

The application discloses a kind of plant oil skin feeling evaluation model construction method and application, including the following steps: obtaining the sensory evaluation data of multiple plant oils in the stages of application, absorption and residual about multidimensional skin feeling attributes;(2) carry out phased physical property test on each plant oil and extract physical property characteristics;The correlation screening and collinearity screening of the physical property characteristics of each stage and the sensory evaluation data of the corresponding stage are carried out, to obtain the key feature set of each stage;Based on the key feature set and sensory evaluation data of each stage, a phased submodel for fitting the skin feeling of each stage and a skin feeling attribute submodel are constructed;Determine the fusion weight by combinatorial optimization, the fusion weight, the phased submodel and the skin feeling attribute submodel constitute the plant oil skin feeling evaluation model, which is used to output comprehensive skin feeling value and multidimensional skin feeling image, which can provide technical support for product oil selection, formula design, batch consistency control and alternative oil screening in the field of cosmetics.

Description

Technical Field

[0001] This invention belongs to the field of cosmetic raw materials and sensory evaluation technology, specifically relating to a method for constructing and applying a plant oil skin feel evaluation model. Background Technology

[0002] Plant oils are important natural raw materials in cosmetics and personal care products. They are rich in active ingredients such as unsaturated fatty acids, vitamins, and phytosterols, and have multiple effects such as moisturizing, anti-oxidation, and repairing the skin barrier. They are widely used in various products such as lotions, creams, serums, and body lotions.

[0003] The skin feel of plant oils (such as fluidity, spreadability, absorption, greasiness, moisturizing effect, and thickness) is a key factor influencing consumer experience and product market competitiveness. However, existing methods for evaluating the skin feel of plant oils have many limitations, including: First, it is highly subjective: traditional skin feel evaluation mainly relies on empirical descriptions or sensory panel scoring. The evaluation results are greatly affected by factors such as the evaluator's personal experience, physiological state, and subjective preferences. The evaluation results of different evaluators and different laboratories are not comparable, making it difficult to form a standardized evaluation system, such as (Marque, Pensé-Lhéritier, & Bacle, 2022, Sensory methods for cosmetic evaluation).

[0004] Second, there is a lack of systematicity: existing evaluation methods focus on the skin feel at a single stage (such as the refreshing feeling when applying), ignoring the dynamic changes in the skin feel of plant oils throughout the entire process of application, absorption, and residue, making it difficult to fully reflect the actual user experience of consumers (Giuffrè et al., 2025, a time-dominated method for capturing the sensory and emotional dynamics of cosmetic creams during use).

[0005] Third, the relationship between physical properties and skin feel is not close: In existing studies, some scholars have tried to evaluate skin feel through physical property tests (such as viscosity and contact angle), but they often only use a single physical property index and have not established a systematic relationship between physical property index and multi-dimensional skin feel attributes. The evaluation accuracy is low and it is difficult to use it for skin feel prediction.

[0006] Fourth, the application scenarios are limited: traditional evaluation methods are difficult to quickly screen plant oils that meet the target skin feel requirements, and cannot provide refined skin feel references for formula optimization, which restricts the efficiency of product development.

[0007] To address the aforementioned issues, there is an urgent need to develop an objective, accurate, and repeatable plant oil skin feel evaluation model that quantifies the three-stage skin application process (application, absorption, and residue) and can output a comprehensive score and multi-dimensional skin feel profile. This model would provide a scientific basis for cosmetic raw material screening, formulation design, and quality control. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for constructing and evaluating a plant oil skin feel evaluation model. This method achieves objective, repeatable, and predictable quantitative characterization of plant oil skin feel, outputting a comprehensive skin feel value and a multi-dimensional skin feel profile. It provides technical support for product oil selection, formulation design, batch consistency control, and alternative oil screening in the cosmetics industry. It is applicable to scenarios such as plant oil raw material screening, formulation design, and product quality control in the cosmetics and personal care products fields.

[0009] To achieve the above-mentioned objectives, the embodiments provide a method for constructing a plant oil skin feel evaluation model, comprising the following steps: (1) Obtain sensory rating data on the multidimensional skin feel properties of various plant oils during the application, absorption, and residue stages; (2) Perform phased physical property tests on each type of plant oil and extract its physical characteristics; (3) Correlation and collinearity screening were performed on the physical properties of each stage and the sensory score data of the corresponding stage to obtain the key feature set of each stage; (4) Based on the key feature set and sensory score data of each stage, construct a staged sub-model and a skin feel attribute sub-model for fitting the skin feel of each stage; (5) The fusion weight is determined by combination optimization. The fusion weight, together with the phased sub-model and the skin feel attribute sub-model, forms a plant oil skin feel evaluation model, which is used to output the comprehensive skin feel value and multi-dimensional skin feel profile.

[0010] Preferably, the multidimensional skin feel properties include fluidity, spreadability, mobility, elasticity, absorption, skin adherence, oiliness, moisturizing effect, and thickness; Among them, fluidity refers to the speed at which a drop of test sample flows downward when the palm is held upright with fingers pointing upwards. Spreadability refers to the speed at which the test sample spreads and forms a thin layer the instant it is dropped. The feeling of resistance when gently pushing a drop of oil sample with a single finger; The elasticity refers to the supporting force of the oil sample on the finger when pressing 2 drops of oil sample; Absorbency refers to the rate at which one drop of oil is absorbed when massaged in circular motions. Skin-fit refers to the degree of tightness and uniformity of the sample's adhesion to the skin; The oily feeling refers to the oily film on the skin surface after swirling two drops of oil in circles on the entire back of the hand; Moisturizing degree refers to the level of hydration on the skin after massaging 2 drops of oil in circular motions with the palm of your hand until fully absorbed; Thickness refers to the thickness of the oil between the finger and the back of the hand after rubbing back and forth five times with two drops of oil.

[0011] Preferably, the physical properties extracted during the application stage include the consistency coefficient, flow index, apparent viscosity, flow time, drop radius, and initial contact angle of the vegetable oil. The physical properties extracted during the absorption phase include: contact angle decay rate, absorption rate constant, and half-absorption time; The physical properties extracted during the residual stage include: initial friction coefficient, stable friction coefficient, change in friction coefficient, and residual film thickness.

[0012] Preferably, correlation and collinearity screening are performed on the physical properties of each stage and the sensory score data of the corresponding stage, including: During the correlation screening, Pearson correlation analysis was used for physical property features whose data conformed to a normal distribution, and Spearman rank correlation analysis was used for physical property features whose data did not conform to a normal distribution. The correlation between each physical property feature and the sensory score of each skin feel attribute was analyzed, and candidate features that were significantly correlated with the skin feel attribute score were screened. During collinearity screening, the variance inflation factor method is used to test the collinearity of candidate features. When the variance inflation factor is greater than the threshold (preferably 10), it indicates that there is serious collinearity among the features. For candidate features with serious collinearity, principal component analysis is used to fuse the features, transforming multiple collinear features into a few uncorrelated principal components. These principal components, together with candidate features without serious collinearity, form the key feature set for each stage.

[0013] Preferably, a phased sub-model for fitting the skin feel at each stage is constructed based on the key feature set and sensory rating data of each stage, including: For each stage, after determining the weights based on the importance of each skin feel attribute in each stage, the sensory scores of each skin feel attribute are weighted and summed to obtain the skin feel score as the label. The model is then optimized by taking the key feature set of each stage as the model input and the skin feel score as the label as the model output, to obtain a staged sub-model that fits the skin feel of each stage.

[0014] Preferably, the importance weights of each skin feel attribute at each stage are calculated using the analytic hierarchy process (AHP).

[0015] Preferably, a sub-model for fitting each skin feel attribute corresponding to a skin feel attribute is constructed based on the key feature set and sensory score data of each stage, including: The training process uses the physical properties of vegetable oil as input and individual skin feel attribute scores as supervision labels. The physical properties are input in stages, with different skin feel attributes predicted based on the physical properties corresponding to their main occurrence stages, as shown below: For fluidity, spreadability, propulsion, and elasticity, the input comes from the physical properties of the coating stage; For absorptivity, the input comes from the physical properties of the absorption stage; To address the greasy feeling, the input comes from the physical properties of the residue stage; For thickness, the input consists of physical properties from the coating and residue stages; Regarding the feel and hydration, input is based on the physical properties of the absorption and residue stages.

[0016] Preferably, for each stage and each skin feel attribute, one or more combinations of linear regression, partial least squares regression, random forest, or gradient boosting decision tree are selected as candidate models, and the coefficient of determination (r) of each candidate model on the validation dataset is evaluated through cross-validation. 2 The absolute error and root mean square error are used to determine the primary model: Main criterion: Select the candidate model with the smallest mean RMSE as the primary model; Secondary criterion: If the mean difference of RMSE among multiple candidate models is less than a threshold, they are considered performance equivalent, and the model with the smallest standard deviation of RMSE within the equivalent set is prioritized as the primary selection model; if they still cannot be distinguished, then R... 2 The model with the higher priority is the primary selection model; Alternative criteria: If neither the primary nor secondary criteria can determine the model, the candidate model with lower complexity is selected as the primary model. The complexity is measured by the size of the model parameters and the degree of structural complexity.

[0017] Preferably, the fusion weights are determined through combinatorial optimization, including: First, an initial weight is given by a stage importance questionnaire that combines consumer surveys and expert scoring; second, the optimal fusion weight is obtained by grid search and fitting optimization on the validation dataset.

[0018] To achieve the above-mentioned objectives, the embodiments also provide an application of a plant oil skin feel evaluation model, wherein the plant oil skin feel evaluation model is a model constructed by the above method. The model is used to evaluate the skin feel of the plant oil to be tested. Specifically, the standardized physical property test results of the plant oil to be tested are input into the plant oil skin feel evaluation model. The total skin feel score of each stage output by the phased sub-model is weighted with the corresponding fusion weight and used as the comprehensive skin feel value. The skin feel attribute score output by the skin feel attribute sub-model corresponding to each attribute forms a multi-dimensional skin feel profile. It is also graded according to the overall skin feel value, including: when the overall skin feel value is [0,5), the corresponding grade is thick and moisturizing; when the overall skin feel value is [5,10), the corresponding grade is balanced; when the overall skin feel value is [10,15], the corresponding grade is refreshing and light.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) Phased modeling to improve the comprehensiveness and accuracy of evaluation: The skin application process is divided into application, absorption, and residue stages, fully considering the dynamic changes in skin feel at each stage, avoiding noise and uninterpretability of a single comprehensive score, and can more comprehensively and accurately reflect the actual user experience of consumers. Correlation analysis shows that the training set R of the model constructed in this invention... 2 ≥0.85, validation set R 2 With an accuracy of ≥ 0.80, MAE ≤ 0.3 points, and RMSE ≤ 0.4 points, the evaluation precision is significantly higher than that of existing single-stage evaluation methods.

[0020] (2) Standardize physical property parameters to ensure repeatability and comparability: adopt standardized physical property parameters (K, n, η) 10 ,k sp , μ stable , h res With parameters such as (etc.) as input, the test conditions and parameter calculation methods are clearly defined, which facilitates cross-batch and cross-laboratory reproduction and solves the problems of strong subjectivity and poor comparability in traditional sensory evaluation.

[0021] (3) Output multi-dimensional skin feel profiles to support refined formula optimization: In addition to the comprehensive score S, output eight-dimensional attribute profiles P, and also output nine-dimensional skin feel profiles (fluidity, spreadability, propulsion, elasticity, absorption, skin adherence, oiliness, moisturization, thickness), which can accurately locate the skin feel advantages and disadvantages of plant oils, provide refined references for the targeted optimization of formula strategies (such as refreshing, moisturizing, and balanced products), and improve product development efficiency.

[0022] (4) Multi-scenario application and strong practicality: The evaluation model and evaluation method of the present invention can be widely applied to product oil selection (quickly screening plant oils that meet the target skin feel), formula design (optimization of plant oil compounding), batch consistency control (monitoring the skin feel stability of different batches of plant oils) and alternative oil screening (screening alternative oils with similar skin feel when raw materials are in short supply). At the same time, it can be integrated into an evaluation device and system to realize automated evaluation and rapid screening, which has strong practicality and industrialization prospects. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1This is a flowchart illustrating the method for constructing the plant oil skin feel evaluation model provided in the embodiments; Figure 2 This is a schematic diagram of the viscosity-shear rate curve and power-law fitting during the application stage provided in the embodiment; Figure 3 This is a schematic diagram of the contact angle θ(t) decay curve during the absorption stage provided in the embodiment; Figure 4 This is a schematic diagram of the residual stage friction coefficient μ(t) versus time provided in the embodiment; Figure 5 k is provided in the embodiment sp Plot a scatter plot with the absorbability score on the horizontal axis and perform linear regression to show the positive correlation. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0026] This invention provides a method for constructing a plant oil skin feel evaluation model. It revolves around the entire skin application process of plant oil—three stages: application, absorption, and residue. The model uses staged physical property testing as input and staged sensory scores as annotations, ultimately outputting a comprehensive skin feel value and a multi-dimensional skin feel profile. This method is suitable for precise evaluation of plant oil skin feel, product oil selection, and formulation design. Figure 1 As shown, it includes the following steps: S1, obtain sensory rating data on the multidimensional skin feel properties of various plant oils during the application, absorption, and residue stages.

[0027] In the examples, at least 50 representative plant oils were selected, covering varieties with different sources, fatty acid compositions, and physicochemical properties, such as herbaceous, woody, and nut oils, as training samples.

[0028] Standardized skin application conditions were determined as follows: ambient temperature 25 ± 2℃, relative humidity 50 ± 5%, test area was a uniform 5×5cm skin area on the inner forearm, the amount used each time was precisely controlled to be 50 μL, and it was applied in a uniform rhythm of 10 back-and-forth motions for about 10 seconds.

[0029] The sensory evaluation panel consists of at least 10 evaluators trained on anchored samples. The evaluators’ scoring error for each attribute is controlled within ±0.3 points. The final score for each sample is the average of all evaluators’ scores, accurate to two decimal places.

[0030] Under standardized skin application conditions, a professional sensory evaluation team conducted a three-stage sensory evaluation. Specifically, this included: in the application stage (Stage A: 0-60 s, focusing on immediate skin feel experience), the absorption stage (Stage B: 1-10 min, focusing on dynamic changes in skin feel), and the residue stage (Stage C: 10-30 min, evaluating the final residual skin feel), nine core skin feel attributes were quantitatively scored from 0 to 5 points (0 points being extremely weak, and 5 points being extremely strong) as sensory score data. Among them, 0 points represent poor fluidity, poor spreadability, poor elasticity, poor absorption, poor adhesion, non-greasy, non-moisturizing, and non-heavy; 5 points represent good fluidity, good spreadability, good elasticity, good absorption, good adhesion, greasy, moisturizing, and heavy.

[0031] The nine skin feel attributes include: fluidity, spreadability, propulsion, elasticity, absorption, adherence, greasiness, moisturizing effect, and thickness. Fluidity refers to the speed at which one drop of the test sample flows downwards when the palm is held upright with fingers pointing upwards; spreadability refers to the speed at which the test sample spreads and forms a thin layer upon being dropped; propulsion refers to the resistance felt when pushing one drop of oil sample with a single finger; elasticity refers to the support force of the oil sample on the finger when pressing two drops of oil sample; absorption refers to the rate at which one drop of oil sample is absorbed after circular massage; adherence refers to the tightness and uniformity of the sample's adhesion to the skin; greasiness refers to the oil film felt on the skin surface after two drops of oil sample are circularly massaged across the entire back of the hand; moisturizing effect refers to the level of hydration the skin exhibits after two drops of oil sample are completely absorbed by circular massage; and thickness refers to the thickness of the oil between the fingers and the back of the hand after five back-and-forth massages with two drops of oil sample. Table 1 provides an example of the nine skin feel scores (0-5 points) for a representative oil sample (grapeseed oil) within a three-stage time window. Table 1

[0032] In this embodiment, to improve the consistency and accuracy of sensory evaluation, high-end and low-end anchor samples are selected from the candidate standard oil set for each skin feel attribute. Preferably, preliminary sensory scoring is conducted on more than 50 standard oils through preliminary experiments. The oil sample with the highest average score for each skin feel attribute is used as the high-end anchor sample for that attribute (e.g., Cetiol CC is selected as the high-end anchor sample for fluidity, and castor oil is selected as the low-end anchor sample), and the oil sample with the lowest average score is used as the low-end anchor sample for evaluator training and scale calibration. During evaluator training, evaluators are required to repeatedly evaluate the anchor samples until the scoring error for each attribute is controlled within ±0.3 points, ensuring the stability and consistency of evaluator scores. Table 2 presents the recommended high / low-end anchor oils for each skin feel attribute (push-off = easy to spread, skin feel = easy to adhere): Table 2

[0033] Table 2 shows the anchoring endpoints. The final anchoring sample is preferably determined by the maximum / minimum value of the pre-experimental score to accommodate batch differences and test conditions.

[0034] To ensure consistency in the overall rating direction, a reverse transformation is used for certain skin feel attributes (such as oiliness, heaviness, etc.) to unify the direction of all ratings. Specifically, the reverse transformation is performed using the following formula: x*=5-x, where x is the original rating (0-5 points), and x* is the rating after the reverse transformation, ensuring that higher ratings indicate a more refreshing skin feel.

[0035] Single attribute level classification: Based on the quantile of the scores of each skin feel attribute in the training set, the 9 skin feel attributes are divided into three levels: low, medium and high (e.g., fluidity: 0-1.5 points is low, 1.6-3.4 points is medium, and 3.5-5.0 points is high) for targeted optimization of formulation strategies (e.g., when developing a refreshing product, plant oils with high fluidity, low greasiness and low heaviness are preferred).

[0036] S2, staged physical property tests were conducted on each plant oil to extract its physical characteristics.

[0037] In this embodiment, the vegetable oils used as training samples underwent phased standardized physical property testing using standardized instruments and methods. To ensure the repeatability and comparability of the test data, the calculation methods for each key parameter were clearly defined, and all physical characteristics were extracted to ensure the objectivity and repeatability of the data. Stage A (Application Stage): Shear rate scanning (0-100s) was performed using a rotational rheometer at 32 ± 1℃ (simulated skin surface temperature). -1 Obtain the viscosity-shear rate curve. Power-law model is adopted Viscosity-shear rate curve By performing a fitting, we obtain the following: Figure 2 The fitted curve shown is used to obtain the consistency coefficient K and the flow index n. Figure 2 Representative data points are shown in Table 3: Table 3

[0038] Simultaneously, the shear rate is directly extracted from the rheometer test data. apparent viscosity η at the point 10 Using a contact angle meter, vegetable oil samples were dropped from a uniform height (e.g., 2 cm) onto a standardized glass plate (roughness Ra = 0.8 μm) and an inclined plate (tilt angle of 30°), and the flow time T of the vegetable oil samples was recorded. flow (Time from dripping to stopping flow), drip radius Rspread The spreading radius (after sample stabilization) and initial contact angle θ0 (the contact angle between the sample and the inclined plate surface at the moment of drop) comprehensively characterize the spreading and flow properties of the sample during the application stage.

[0039] Stage B (Absorption Stage): The contact angle of the vegetable oil sample on the artificial skin surface was continuously recorded as a function of time using a contact angle meter, resulting in the curve θ(t). The curve was then fitted using a nonlinear model (e.g., exponential decay model) to form θ(t) = θ∞ + (θ...). 0- θ ∞ )exp(-k sp t), goodness of fit R 2 ≥0.90, where θ0 is the initial contact angle at t=0, θ∞ is the stable contact angle (plateau value) after a long time, and the resulting contact angle decay rate k sp It reflects the combined rate of sample spreading and penetration. Figure 3 Table 4 shows the contact angle θ(t) decay curve during the absorption stage. Figure 3 Representative data points: Table 4

[0040] An in vitro stratum corneum model was constructed (stratum corneum sheets obtained by thermal or enzymatic separation of pig ear skin, with a thickness of 15–25 μm). A Franz diffusion cell was used to record the amount of sample permeation through the stratum corneum model, Q(t), and the remaining mass M(t) on the model surface over time t. The permeation amount Q(t) curve was fitted using a first-order kinetic model. goodness of fit The absorption rate constant k was obtained. abs (Characterizing the rate at which the sample is absorbed by the skin) and half-absorption time t 50 (t) 50 =ln2 / k abs ).

[0041] Stage C (Residual Stage): Using a skin friction instrument, the friction coefficient change curve μ(t) over time was recorded under standardized friction conditions (100 Hz, 10 reciprocations, speed 10 mm / s, distance 20 mm). The initial friction coefficient μ0 and the stable friction coefficient μ0 were extracted. stable and the change in friction coefficient Δμ=μ0-μ stable Among them, the initial friction coefficient μ0 is the average friction coefficient in the initial stage of friction (within 0-0.5s after the start of friction), reflecting the initial lubrication state of the residual film; the stable friction coefficient μ stableThe friction coefficient is the average friction coefficient during the stable phase of the friction process (i.e., within 5-10 seconds after the start of friction), which characterizes the long-term lubrication performance of the residual film; the change in friction coefficient Δμ reflects the stability of the residual film during the friction process. Figure 4 Table 5 shows the curve of friction coefficient μ(t) versus time during the residual stage. Figure 4 Representative data points: Table 5

[0042] The residual sample on the skin surface was imaged using a laser confocal microscope, and the residual film thickness h was calculated using image analysis software. res The residual amount of the sample was quantified. Among them, the residual film thickness h... res The average film thickness at 6 different measurement points was selected using a laser confocal microscope, with a spacing of ≥ 1 mm between the measurement points to avoid data deviation caused by excessive concentration of measurement points and to ensure data reliability.

[0043] Table 6 shows the physical properties of the training weights (12 kinds of plant oils) extracted at 32 ± 1℃.

[0044] Table 6

[0045] S3. Correlation and collinearity screening are performed on the physical properties of each stage and the sensory score data of the corresponding stage to obtain the key feature set of each stage.

[0046] In this embodiment, to ensure the validity and independence of the model input features, correlation and collinearity screening are performed on the physical property features of each stage and the corresponding sensory scores of each stage: During the correlation screening, Pearson correlation analysis was used for physical property features whose data conformed to a normal distribution, and Spearman rank correlation analysis was used for physical property features whose data did not conform to a normal distribution. The correlation between each physical property feature and the sensory score of each skin feel attribute was analyzed, and candidate features that were significantly correlated with the skin feel attribute score were screened.

[0047] Specifically, for datasets with a sample size greater than 100, the significance threshold can be adjusted to... p < 0.01, will p Physical property features with values ​​less than 0.01 are used as candidate features to further enhance the stringency of feature selection; for datasets with small sample sizes (50 ≤ sample size ≤ 100), a significance threshold is used. p < 0.05, to avoid missing important features.

[0048] For example, calculating the contact angle decay rate k spCorrelation between sensory ratings and absorbability Figure 5 For k sp Scatter plots were created with sensory scores for absorbency on the horizontal axis and linear regression was performed to demonstrate a positive correlation.

[0049] During collinearity screening, the variance inflation factor (VIF) method is used to test the collinearity of candidate features, with a VIF threshold of 10. When VIF > 10, it indicates severe collinearity among features. For candidate features with severe collinearity, principal component analysis (PCA) is used for feature fusion, transforming multiple collinear features into a few uncorrelated principal components. These principal components, along with candidate features without severe collinearity, form the key feature set X for each stage. A X B X C , where X A Key features of the application stage, X B Key features of the absorption phase, X C These are key characteristics of the residual stage.

[0050] S4, based on the key feature set and sensory score data of each stage, constructs a staged sub-model and a skin feel attribute sub-model to fit the skin feel of each stage.

[0051] In this embodiment, based on the key feature sets of each stage and the sensory rating data of the corresponding stage, a stage-specific sub-model is established to ensure that the model accurately fits the skin feel at each stage: For each stage, after determining the weighted weights based on the importance of each skin-feeling attribute at each stage, the sensory scores of each skin-feeling attribute are weighted and summed to obtain the skin-feeling score as a label. Specifically, the Analytic Hierarchy Process (AHP) is used to calculate the weighted weights of importance for each skin-feeling attribute at each stage. AHP is a decision analysis method based on expert knowledge and qualitative assessment, widely used in multi-criteria decision-making. AHP decomposes complex problems into multiple hierarchical structures and calculates relative weights by comparing the importance of each element. Specifically applied to skin-feeling assessment, AHP is used to weight various skin-feeling attributes (such as fluidity, absorbency, and heaviness) to determine their relative importance in skin-feeling scores at different stages.

[0052] Step 1: Construct a hierarchical structure First, the hierarchical structure of AHP is determined. For this invention, the skin feel evaluation at each stage will serve as the target layer (first layer), while the skin feel attributes at each stage (such as fluidity, absorbency, and heaviness) will serve as the criterion layer (second layer). The specific structure is as follows: Target layer: Overall skin feel score for each stage (e.g., overall score for the application stage). Criterion layer: Various skin feel attributes (such as fluidity, absorbency, and heaviness).

[0053] Step 2: Construct the judgment matrix In the AHP calculation process, the first step is to construct an importance judgment matrix between skin feel attributes at each stage. Using expert or empirical data, the importance of each skin feel attribute relative to other attributes is compared in pairs, typically using a scale of 1–9 for scoring (e.g., 1 indicates both are equally important, 3 indicates one attribute is slightly more important than the other, and 9 indicates one attribute is extremely important). For example, assuming fluidity is more important than absorbency, an expert might rate it 3 in the judgment matrix.

[0054] Step 3: Calculate the weights Based on the judgment matrix, AHP uses the eigenvalue method (or other matrix calculation methods) to calculate the weights of each skin feel attribute. By normalizing the eigenvector of each skin feel attribute, the relative weight of each attribute is obtained (e.g., fluidity might account for 50%, absorbency for 30%, and heaviness for 20%). These weights reflect the relative importance of each attribute in the skin feel evaluation.

[0055] Step 4: Consistency Check In Expert Hierarchy Process (AHP), ensuring the consistency of the judgment matrix is ​​crucial. The Consistency Ratio (CR) is calculated to determine whether the experts' judgments are consistent. If the CR value is greater than 0.1, it indicates inconsistency in the matrix, requiring re-evaluation. If the CR value is less than 0.1, the judgment matrix is ​​consistent, and weight calculation can continue.

[0056] Step 5: Weighted summation of skin feel scores The weights of each skin feel attribute calculated using AHP will be used to weighted sum the skin feel scores for each stage. Specifically, the skin feel scores for the application, absorption, and residue stages will be weighted based on the corresponding skin feel attribute scores and their respective weights.

[0057] in, It is a score of various skin feel attributes (e.g., fluidity, absorbency, heaviness, etc.). These correspond to the weights of the skin feel attributes. The total skin feel score for each stage is the weighted overall score.

[0058] Step 6: Calculation of final skin feel score The final overall skin feel value S is obtained by weighting and combining the scores of each stage (application, absorption, residue). This can be achieved by assigning different weights to the scores of each stage:

[0059] Where α, β, and γ are the weights of each stage, α + β + γ = 1.

[0060] For the application stage, key sensory attributes include: fluidity, spreadability (ease of application), spreadability, elasticity, and thickness (initial feel); for absorption, key sensory attributes include: absorbency, adherence to the skin, and moisturizing effect; for residue, key sensory attributes include: greasiness, adherence to the skin, moisturizing effect, and thickness. The selection and definition of these sensory attributes combined expert experience with the Analytic Hierarchy Process (AHP). In the initial stage, through discussions and analysis with skincare product development experts, key sensory attributes for the application, absorption, and residue stages were selected, covering the sensory experiences of the skin at different stages (e.g., fluidity, absorbency, heaviness, greasiness, etc.). The definitions of these attributes referenced existing sensory evaluation standards and common practices in the cosmetics industry.

[0061] The skin feel attributes at each stage (e.g., fluidity in the application stage) are numerically obtained through feature extraction and sensory scoring, and these values ​​are input features into each sub-model (application sub-model, absorption sub-model, residue sub-model), influencing the final stage-specific skin feel score. These stage-specific scores are then weighted and summed based on weights determined by AHP to obtain the total skin feel score for each stage. For example, in the application stage, scores for fluidity, spreadability, and other properties are weighted according to their respective weights to obtain the overall score for the application stage. The scores for each stage (application, absorption, residue) are further weighted and combined to obtain the final comprehensive skin feel score S, which characterizes the overall skin feel quality of the product.

[0062] The model takes the key feature set of each stage as input and the corresponding skin feel score as output, and obtains a sub-model for each stage through training. Specifically, the model structure and loss function are as follows: Smearing Sub-model S A =f A (X A ), with the key feature set X of the smearing stage A As input, output a weighted sum of the scores for the main skin feel attributes during the application phase (such as fluidity, spreadability, and pushability). The application sub-model is constructed using one or more algorithms: linear regression, partial least squares regression (PLSR), random forest (RF), or gradient boosting decision tree (GBDT).

[0063] Absorber Model S B =f B (X B ), with the key feature set X of the absorption stage BAs input, the output is a weighted sum of the main skin-feel attribute scores (such as absorbency, skin-feel, and moisturizing properties) for the absorption phase. The same candidate algorithm as the application sub-model is used for training.

[0064] Residual Submodel S C =f C (X C The key feature set X of the residual stage C As input, the output is a weighted sum of the main skin feel attributes scores (such as heaviness, oiliness, absorbency, etc.) during the residual stage. The algorithm is trained using the same combination described above.

[0065] Training process and loss function: Loss function: For each stage of the sub-model, the mean squared error (MSE) is used as the loss function to optimize the model parameters to minimize the error between the predicted score and the actual score.

[0066] Model selection: For each sub-model, one or more of the following algorithms are selected as candidate models: linear regression, partial least squares regression (PLSR), random forest (RF), or gradient boosting decision tree (GBDT). These are evaluated on the validation dataset using 5-fold cross-validation, and the following evaluation metrics are calculated: coefficient of determination. Mean absolute error (MAE) and root mean square error (RMSE); and the primary model is determined by the following methods: Main criterion: Select the candidate model with the smallest mean RMSE as the primary model; Secondary criterion: If the mean difference in RMSE among multiple candidate models is less than a threshold (e.g., 30%), they are considered performance-equivalent, and the model with the smallest RMSE standard deviation within the equivalent set is prioritized as the primary selection model; if they still cannot be distinguished, then... The model with the higher priority is the primary selection model; Alternative criteria: If neither the primary nor secondary criteria can determine the model, the candidate model with lower complexity is selected as the primary model. The complexity is measured by the size of the model parameters and the degree of structural complexity (such as the number of variables).

[0067] Alternative criteria: If neither the primary nor secondary criteria can determine the model, the model with lower complexity is selected as the final model. The complexity is measured by the size of the model parameters and the degree of structural complexity (such as the number of variables).

[0068] Through this series of model selection and evaluation, the optimal sub-model for each stage is finally determined, and the optimal sub-model for each stage is used to output the total score S for each stage. A S B S C .

[0069] To obtain a nine-dimensional skin feel profile, this invention, in addition to the overall stage scoring sub-model, further sets up skin feel attribute sub-models for each attribute to directly output the predicted scores of each skin feel attribute, thus forming the nine-dimensional skin feel profile P. The nine-dimensional skin feel profile P is a set of predicted scores (range, for example, 0-5 points) for the oil sample under test on nine skin feel attributes, including the following nine dimensions: fluidity, spreadability (easiness to spread), spreadability, elasticity, thickness, absorption, adherence to the skin, oiliness, and moisturizing effect. Therefore, the nine-dimensional skin feel profile can be expressed as: P = [p 流动 , p 推动 , p 铺展 , p 弹感 , p 厚度 , p 吸收 , p 贴肤 ,p 油腻 , p 滋润 ], where each dimension p is the predicted score of the corresponding skin feel attribute output by the skin feel attribute sub-model.

[0070] The skin feel attribute sub-model takes the physical properties of the oil sample as input and uses the individual skin feel attribute scores given by the raters in the training set as supervision labels for training. Specifically, each oil sample in the training set corresponds to: (1) physical properties: including the set of physical properties X of the application stage. A The set of physical properties X during the absorption stage B The set of physical properties of the residual stage X C (2) True rating of nine skin feel attributes: Y = [y 流动 , y 推动 , y 铺展 , y 弹感 , y 厚度 , y 吸收 , y 贴肤 , y 油腻 , y 滋润 ].

[0071] Similarly, the skin feel attribute sub-model can be implemented using regression models such as linear regression, partial least squares regression, random forest, or gradient boosting decision tree. Its training objective is to minimize the error between the predicted value and the true score (such as minimizing the mean square error), and the optimal model is selected through cross-validation grouped by oil sample category.

[0072] To ensure consistency between the nine-dimensional image and the application-absorption-residue testing process, this invention preferably employs a phased input method: different skin texture attributes are predicted using the physical properties corresponding to their main occurrence stages as input. Examples are shown in Table 7: Table 7. Main Input Feature Sources for Nine-Dimensional Skin Feel Attributes

[0073] When evaluating an oil sample to be tested, the three-stage physical property characteristics are first obtained according to the method of this invention to obtain X. A X B X C Then: X A Input the skin feel attribute sub-model related to the application stage to obtain predicted scores for fluidity, pushability, spreadability, elasticity, and initial thickness; X B Input the skin feel attribute sub-model related to the absorption stage to obtain predicted scores for absorbency, skin feel, and moisturization. X C Input the skin feel attribute sub-model related to the residue stage to obtain predicted scores for oiliness, skin feel, moisturization and residue thickness; For attributes where predicted values ​​can be obtained at multiple stages (such as skin feel, moisturization, and thickness), the final attribute value is determined according to preset rules. Rule: Predicted values ​​from the residual stage are preferred as the final values, as this better reflects the difference between the later stages of use and the final skin feel.

[0074] The final predicted scores of the nine attributes are combined in a fixed order to form a nine-dimensional skin-feel profile P.

[0075] S5 determines the fusion weight through combination optimization. This fusion weight, together with the phased sub-model and the skin feel attribute sub-model, forms the plant oil skin feel evaluation model, which is used to output the comprehensive skin feel value and multi-dimensional skin feel profile.

[0076] In this embodiment, to comprehensively reflect the skin feel experience throughout the entire process of applying plant oil to the skin, the output results of the three stage sub-models are fused according to weights to obtain a comprehensive skin feel value S. The fusion formula is S=α·S A +β·S B +γ·S C , where α, β, and γ are the fusion weights of the application stage, absorption stage, and residue stage, respectively, and satisfy α+β+γ=1.

[0077] Among them, the fusion weights are determined through combinatorial optimization, including: First, an initial weight was given by a questionnaire on the importance of each stage, which combined consumer research and expert scoring. Specifically, 15-20 cosmetic R&D experts, sensory evaluation experts and consumer representatives were invited to score the importance of the skin feel in the three stages of application, absorption and residue (1-10 points). The average score of each stage was calculated and normalized to obtain the initial weights α0, β0, γ0. Secondly, the optimal fusion weights are obtained through grid search and fitting optimization on the validation dataset. Specifically, the initial weights are used as the initial range for the grid search, with α, β, and γ ranging from 0.1 to 0.5. The R-squared of the plant oil skin feel evaluation model under different weight combinations is calculated using the validation set data. 2 After considering MAE and RMSE, the overall performance (R) is selected as the optimal choice. 2 The final fusion weight is the combination of the largest (and smallest MAE and RMSE) values ​​to ensure that the overall skin feel value can reflect the actual skin feel experience of consumers to the greatest extent.

[0078] Table 8 shows the output of the three-stage sub-model and the fused overall skin feel value S. Example of fusion weights: α=0.35, β=0.35, γ=0.30; the overall S is mapped to 0-15 for grading.

[0079] Table 8

[0080] In the embodiment, in addition to outputting the total score S for the application stage, A Total score for the absorption phase (S) B Residual stage total score S C In addition, it also outputs a nine-dimensional skin feel profile P to guide alternative oil screening and targeted formulation optimization.

[0081] The embodiment also provides an application of the plant oil skin feel evaluation model constructed by the above method. The model is used to evaluate the skin feel of the plant oil to be tested. Specifically, the standardized physical property test results of the plant oil to be tested are input into the plant oil skin feel evaluation model. The total skin feel score of each stage output by the stage sub-model is weighted with the corresponding fusion weight and used as the comprehensive skin feel value. The skin feel attribute score output by the skin feel attribute sub-model corresponding to each attribute forms a multi-dimensional skin feel profile.

[0082] Specifically, following steps S2 and S3 above, the physical properties of the vegetable oil to be tested are tested in stages, and the physical property characteristics of the tested samples are screened to obtain the key feature set X of the sample to be tested. A X B X C Then, the key feature set is input into the trained smearing sub-model, absorption sub-model, and residue sub-model to obtain the output S of each stage sub-model. A S B S C The overall skin feel value S is calculated based on the fusion weights α, β, and γ, and the overall skin feel value S is classified into levels based on a preset threshold or quantile rule.

[0083] Overall Skin Feel Value S (0-15 points) Grading: [0,5) are classified as heavy and moisturizing (significant residue, high moisturizing effect, suitable for winter or dry skin formulas); [5,10) are classified as balanced (neutral skin feel, wide applicability, suitable for all-season formulas); [10,15] are classified as refreshing and lightweight (fast absorption, low residue, suitable for summer skincare formulas); Table 9 presents the grading rules for the overall skin feel value S: Table 9

[0084] Table 10 shows the input characteristics, output comprehensive skin feel value S, and grade of the six types of oils to be tested. Grade: S∈[0,5) is III (heavy / oily), S∈[5,10) is II (balanced), and S∈[10,15] is I (refreshing / light).

[0085] Table 10

[0086] The above-mentioned plant oil skin feel evaluation model can be widely used in product oil selection (quickly screening plant oils that meet the target skin feel requirements), formulation design (blending plant oils according to different skin feel grade requirements), batch consistency control (monitoring the skin feel stability of different production batches of the same plant oil), and alternative oil screening (screening alternative plant oils with similar skin feel when raw materials are in short supply).

[0087] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a plant oil skin feel evaluation model, characterized in that, Includes the following steps: (1) Obtain sensory rating data on the multidimensional skin feel properties of various plant oils during the application, absorption, and residue stages; (2) Perform phased physical property tests on each type of plant oil and extract its physical characteristics; (3) Correlation and collinearity screening were performed on the physical properties of each stage and the sensory score data of the corresponding stage to obtain the key feature set of each stage; (4) Based on the key feature set and sensory score data of each stage, construct a staged sub-model and a skin feel attribute sub-model for fitting the skin feel of each stage; (5) The fusion weight is determined by combination optimization. The fusion weight, together with the phased sub-model and the skin feel attribute sub-model, forms a plant oil skin feel evaluation model, which is used to output the comprehensive skin feel value and multi-dimensional skin feel profile.

2. The method for constructing the plant oil skin feel evaluation model according to claim 1, characterized in that, The multidimensional skin feel properties include fluidity, spreadability, mobility, elasticity, absorption, adherence to the skin, oiliness, moisturizing effect, and thickness; Among them, fluidity refers to the speed at which a drop of test sample flows downward when the palm is held upright with fingers pointing upwards. Spreadability refers to the speed at which the test sample spreads and forms a thin layer the instant it is dropped. The feeling of resistance when gently pushing a drop of oil sample with a single finger; The elasticity refers to the supporting force of the oil sample on the finger when pressing 2 drops of oil sample; Absorbency refers to the rate at which one drop of oil is absorbed when massaged in circular motions. Skin-fit refers to the degree of tightness and uniformity of the sample's adhesion to the skin; The oily feeling refers to the oily film on the skin surface after swirling two drops of oil in circles on the entire back of the hand; Moisturizing degree refers to the level of hydration on the skin after massaging 2 drops of oil in circular motions with the palm of your hand until fully absorbed; Thickness refers to the thickness of the oil between the finger and the back of the hand after rubbing back and forth five times with two drops of oil.

3. The method for constructing the plant oil skin feel evaluation model according to claim 1, characterized in that, The physical properties extracted during the application stage include the consistency coefficient, flow index, apparent viscosity, flow time, drop radius, and initial contact angle of the vegetable oil. The physical properties extracted during the absorption phase include: contact angle decay rate, absorption rate constant, and half-absorption time; The physical properties extracted during the residual stage include: initial friction coefficient, stable friction coefficient, change in friction coefficient, and residual film thickness.

4. The method for constructing the plant oil skin feel evaluation model according to claim 1, characterized in that, Correlation and collinearity screening were performed on the physical properties of each stage and the corresponding sensory score data, including: During the correlation screening, Pearson correlation analysis was used for physical property features whose data conformed to a normal distribution, and Spearman rank correlation analysis was used for physical property features whose data did not conform to a normal distribution. The correlation between each physical property feature and the sensory score of each skin feel attribute was analyzed, and candidate features that were significantly correlated with the skin feel attribute score were screened. During collinearity screening, the variance inflation factor method is used to test the collinearity of candidate features. When the variance inflation factor is greater than the threshold, it indicates that there is serious collinearity among the features. For candidate features with serious collinearity, principal component analysis is used to fuse the features, transforming multiple collinear features into a few uncorrelated principal components. These principal components, together with the candidate features without serious collinearity, form the key feature set for each stage.

5. The method for constructing the plant oil skin feel evaluation model according to claim 1, characterized in that, Based on the key feature sets and sensory rating data of each stage, a phased sub-model is constructed to fit the skin feel at each stage, including: For each stage, after determining the weights based on the importance of each skin feel attribute in each stage, the sensory scores of the main skin feel attributes corresponding to each stage are weighted and summed to obtain the skin feel score as the label. The model is then optimized by taking the key feature set of each stage as the model input and the skin feel score as the label as the model output, to obtain a staged sub-model that fits the skin feel of each stage.

6. The method for constructing the plant oil skin feel evaluation model according to claim 5, characterized in that, The importance weights of each skin feel attribute at each stage are calculated using the analytic hierarchy process (AHP).

7. The method for constructing the plant oil skin feel evaluation model according to claim 1, characterized in that, Based on the key feature sets and sensory score data of each stage, a sub-model for fitting the skin feel attribute corresponding to each skin feel attribute is constructed, including: The training process uses the physical properties of vegetable oil as input and individual skin feel attribute scores as supervision labels. The physical properties are input in stages, with different skin feel attributes predicted based on the physical properties corresponding to their main occurrence stages, as shown below: For fluidity, spreadability, propulsion, and elasticity, the input comes from the physical properties of the coating stage; For absorptivity, the input comes from the physical properties of the absorption stage; To address the greasy feeling, the input comes from the physical properties of the residue stage; For thickness, the input consists of physical properties from the coating and residue stages; Regarding the feel and hydration, input is based on the physical properties of the absorption and residue stages.

8. The method for constructing the plant oil skin feel evaluation model according to claim 1, characterized in that, For each stage and each skin feel attribute, one or more combinations of linear regression, partial least squares regression, random forest, or gradient boosting decision tree are selected as candidate models. The coefficient of determination (r) of each candidate model on the validation dataset is evaluated through cross-validation. 2 The absolute error and root mean square error are used to determine the primary model: Main criterion: Select the candidate model with the smallest mean RMSE as the primary model; Secondary criterion: If the mean difference of RMSE among multiple candidate models is less than a threshold, they are considered performance equivalent, and the model with the smallest standard deviation of RMSE within the equivalent set is prioritized as the primary selection model; if they still cannot be distinguished, then R... 2 The model with the higher priority is the primary selection model; Alternative criteria: If neither the primary nor secondary criteria can determine the model, the candidate model with lower complexity is selected as the primary model. The complexity is measured by the size of the model parameters and the degree of structural complexity.

9. The method for constructing the plant oil skin feel evaluation model according to claim 1, characterized in that, The fusion weights are determined through combinatorial optimization, including: First, an initial weight is given by a stage importance questionnaire that combines consumer surveys and expert scoring; second, the optimal fusion weight is obtained by grid search and fitting optimization on the validation dataset.

10. An application of a plant oil skin feel evaluation model, characterized in that, The plant oil skin feel evaluation model is a model constructed by the method described in claims 1-9. The model is used to evaluate the skin feel of the plant oil to be tested. Specifically, the standardized physical property test results of the plant oil to be tested are input into the plant oil skin feel evaluation model. The total skin feel score of each stage output by the phased sub-model is weighted by the corresponding fusion weight and used as the comprehensive skin feel value. The skin feel attribute score output by the skin feel attribute sub-model corresponding to each attribute forms a multi-dimensional skin feel profile. It is also graded according to the overall skin feel value, including: when the overall skin feel value is [0,5), the corresponding grade is thick and moisturizing. When the overall skin feel score is [5, 10), the corresponding level is balanced; when the overall skin feel score is [10, 15], the corresponding level is refreshing and lightweight.

Images