Individual care product process optimization method and system based on user behavior data

Abstract

The invention discloses an individual care product process optimization method and system based on user behavior data, and relates to the field of intelligent data processing, and the method comprises the steps: collecting a historical user behavior data set in a historical time range after a target product is put into the market through multiple channels, and carrying out the data preprocessing and product improvement correlation feature extraction, obtaining a historical product improvement associated feature set; determining an optimal process adjustment time period by combining a predicted product demand growth rate and a predicted product demand fluctuation coefficient within a preset time range; according to the user behavior data set in the optimal process adjustment time period, determining a product improvement index according to the product quality associated feature set; and obtaining an optimal process adjustment scheme by taking approximation to a product improvement index as an optimization target, and carrying out production process iterative optimization within a preset time range. According to the method and the device, the problems of poor user experience, low optimization strategy scene adaptability and inaccurate process optimization direction in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent data processing, and more specifically to a method and system for optimizing the process of personal care products based on user behavior data. Background Technology

[0002] In the research and development and production of personal care products, continuous optimization of product processes is key to maintaining market competitiveness and improving user satisfaction.

[0003] Traditional personal care product process optimization relies on fixed cycles or subjective experience adjustments, which cannot dynamically respond to market demand fluctuations and user behavior feedback. This leads to a disconnect between the optimization adjustment rhythm and market rhythm, resulting in resource waste and production disorder, or a mismatch with the window of change in user needs, and decision-making based on one-sided and lagging factors. Relying on subjective experience makes it difficult to obtain real user experience data in real time, resulting in poor user experience, weak adaptability of optimization strategies to different scenarios, and inaccurate direction of process optimization.

[0004] Therefore, an innovative technological solution is needed to transform massive amounts of user behavior data into a core driving force for precise, agile, and economical optimization of production processes, thereby achieving data-driven intelligent process optimization. Summary of the Invention

[0005] This application provides a method and system for optimizing the process of personal care products based on user behavior data, which addresses the problems of poor user experience, weak adaptability of optimization strategies to different scenarios, and inaccurate direction of process optimization in existing technologies.

[0006] In view of the above problems, this application provides a method for optimizing the process of personal care products based on user behavior data.

[0007] In a first aspect, this application provides a method for optimizing the manufacturing process of personal care products based on user behavior data, the method comprising: By collecting historical user behavior datasets from the target product within a historical time range since its market launch through multiple channels, and performing data preprocessing and product improvement-related feature extraction on the historical user behavior datasets, a historical product improvement-related feature set is obtained. The optimal process adjustment time period is determined based on the historical product improvement correlation feature set and the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range. By collecting user behavior datasets of the target product within the optimal process adjustment time period through multiple channels, and performing data preprocessing and product quality correlation feature extraction, product improvement indicators are determined based on the product quality correlation feature set. With the goal of approximating the product improvement indicators, the production process of the target product is optimized to obtain the optimal process adjustment scheme. Within the preset time range, the production process is iteratively optimized based on user behavior data collected in real time within the optimal process adjustment time period.

[0008] Secondly, the present invention provides a personal care product process optimization system based on user behavior data, comprising: The data acquisition module is used to collect historical user behavior datasets within a historical time range since the target product was launched on the market through multiple channels, perform data preprocessing and product improvement related feature extraction on the historical user behavior datasets, and obtain a historical product improvement related feature set. The time cycle prediction module is used to determine the optimal process adjustment time cycle based on the historical product improvement correlation feature set and the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range. The quality feature extraction module is used to collect user behavior datasets of the target product within the optimal process adjustment time period through multiple channels, perform data preprocessing and product quality correlation feature extraction, and determine product improvement indicators based on the product quality correlation feature set. The process optimization module is used to optimize the production process of the target product with the goal of approximating the product improvement indicators, obtain the optimal process adjustment scheme, and perform iterative optimization of the production process based on user behavior data collected in real time within the preset time range and the optimal process adjustment time period.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application achieves several key milestones. First, it acquires a historical set of product improvement-related features through multi-channel data collection, constructing a historical dataset covering all dimensions of user behavior, providing a high-quality data foundation for subsequent intelligent decision-making. Second, based on this historical set of product improvement-related features, it determines the optimal time cycle for process adjustment. By combining intrinsic characteristics reflecting historical product performance with external predictions reflecting future market trends, it calculates a dynamically changing adjustment cycle, achieving a precise match between the timing of process optimization and market dynamics. Third, through multi-channel data collection, it identifies product improvement indicators, transforming abstract user feedback into quantifiable product improvement metrics, ensuring that the direction of process optimization accurately improves user experience. Finally, it uses approximation of product improvement indicators as the optimization goal to iteratively optimize the production process, solving the problems of low efficiency and high cost in solution optimization. In a virtual environment, it utilizes intelligent algorithms to efficiently explore and evaluate the process parameter space, forming a data-driven, precise optimization scheme. Under resource constraints, this achieves optimal process optimization results, enhancing product competitiveness. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0011] Figure 1 This is a flowchart illustrating the personal care product process optimization method based on user behavior data presented in this application.

[0012] Figure 2 This is a schematic diagram of the structure of the personal care product process optimization system based on user behavior data in this application.

[0013] In the attached diagram, the components represented by each number are as follows: Data acquisition module 11; time cycle prediction module 12; quality feature extraction module 13; process scheme optimization module 14. Detailed Implementation

[0014] This application provides a process optimization method for personal care products based on user behavior data, which solves the problems of poor user experience, weak adaptability of optimization strategies to different scenarios, and inaccurate direction of process optimization in the prior art.

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0017] The present invention will now be described in detail with reference to the accompanying drawings.

[0018] Example 1, as Figure 1 As shown, this application provides a method for optimizing the process of personal care products based on user behavior data, the method comprising: S10: Collect historical user behavior datasets from the target product within a historical time range since its launch into the market through multiple channels, perform data preprocessing and product improvement related feature extraction on the historical user behavior datasets, and obtain a historical product improvement related feature set. In this embodiment, multi-channel collection refers to the act of collecting behavioral information generated by user interactions with the product through multiple paths; the historical user behavior dataset is the original set of all user behavior data accumulated across multiple channels from the time the target product was launched to the current time point; data preprocessing is a standardized operation of handling outliers, imputing missing values, and aligning multi-source data in the original dataset; product improvement related feature extraction is the process of selecting and quantifying features directly related to product process improvement from the preprocessed standard data to form a structured feature set; the historical product improvement related feature set is a set of feature data reflecting the relationship between user behavior and product process defects in historical stages, and is the core basis for subsequent process adjustment decisions.

[0019] Historical user behavior data was collected through official procedures, stores, and offline user interviews. Subsequently, features were extracted using process sensitivity indicators, multi-factor coupling influence indicators, and process window robustness indicators as constraints, ultimately forming a historical product improvement correlation feature set for this sunscreen.

[0020] Step S10 in the method provided in this application embodiment includes: Configure multiple data collection paths, wherein the data collection paths include at least the brand's official application, e-commerce transaction platform, social media content platform and user research interviews; Using the aforementioned multi-channel data collection path, historical user behavior datasets are collected within a historical time range since the target product was launched into the market. The target product is a sun protection personal care product, and the types of user behavior data include at least behavioral dimensions, scenario dimensions, efficacy experience dimensions, emotional tendency dimensions, and spatiotemporal correlation.

[0021] In this embodiment, a multi-channel data collection path is first configured. This path includes at least the brand's official application, e-commerce platforms, social media content platforms, and user research interviews. The brand's official application is a user-facing mobile application developed by the brand itself, serving as a private data channel for the brand to directly reach users. E-commerce platforms are third-party comprehensive or vertical e-commerce platforms that handle product sales, user reviews, returns, and other transaction-related activities, providing a public user feedback channel. Social media content platforms are user-generated content-centric social platforms that cover non-transactional user behavior, serving as a channel for uncovering potential user needs. User research interviews are qualitative / quantitative research channels that directly communicate with users through offline / online interviews, questionnaires, etc., allowing for in-depth analysis of the motivations behind user behavior.

[0022] Taking a certain brand of sunscreen as an example, a multi-channel data collection path was configured, including: through the brand's official application, the collection dimensions were determined to be user registration information, product purchase records, and usage check-in data, with a collection frequency of real-time collection, and the data was transmitted to the brand's data center through the application's backend interface; through e-commerce transaction platforms, the core collection channels were selected, and the collection dimensions were determined to be transaction order data, user reviews, and customer service chat records, with a collection frequency of daily incremental collection, and the data was obtained through the e-commerce platform's open interface and converted into a unified format; on social media content platforms, the collection dimensions were determined to be relevant notes / video content, interaction data, and topic participation, and the data was collected using network technology, with a full update once a week; through user research interviews, a semi-structured interview outline was designed, and the interviewees were determined to be core users of sunscreen products, with an interview frequency of once a quarter, using a combination of offline face-to-face interviews and online questionnaires.

[0023] At the same time, data from all channels is transmitted to the brand's private cloud database, and a data verification mechanism is designed to ensure that data fields from all channels match, thus completing the overall configuration of the multi-channel data collection path.

[0024] Secondly, by utilizing multi-channel data collection paths, historical user behavior datasets are collected within a historical time range since the target product was launched into the market. The target product is a sun protection personal care product, and the types of user behavior data include at least behavioral dimensions, scenario dimensions, efficacy experience dimensions, emotional tendency dimensions, and spatiotemporal correlation dimensions.

[0025] The historical timeframe spans from the date the target sunscreen product is officially launched to the date the process optimization decision is initiated, used to set the process adjustment time cycle; sunscreen personal care products refer to personal care products with sunscreen as their core function, distinguishing them from other personal care products to ensure the targeting of data collection; the user behavior data type is a structured classification of user behavior, including: behavior dimension: direct interaction between users and the product, including usage frequency, single application amount, and reapplication habits; scenario dimension: the specific environment and context in which users use the product, such as commuting, outdoor sports, and travel, and their corresponding UV intensity and temperature / humidity; efficacy experience dimension: users' perception and evaluation of the product's core efficacy, subjective evaluations of skin feel such as film-forming speed, refreshing feel, and water resistance, as well as objective feedback on effects such as redness or tanning after sun exposure; emotional tendency dimension: users' subjective emotions and attitudes towards the product, including keywords and emotional tendencies extracted from reviews and notes regarding oiliness, whitening, and irritation; spatiotemporal correlation dimension: the temporal and spatial attributes of user behavior, including the combination patterns of usage behavior with geographical location, season, and month.

[0026] By performing multi-channel data collection operations, the configured collection paths are transformed into actual user behavior data assets, providing comprehensive, specific, and analyzable data support for subsequent process optimization decisions. At the same time, by clarifying the boundaries of data types, it is ensured that the collected data focuses on core information related to the improvement of sunscreen product processes, avoiding invalid data redundancy.

[0027] Step S10 in the method provided in this application embodiment further includes: The historical user behavior dataset is processed for outlier handling, missing value imputation, and multi-source data alignment to generate a historical standard behavior dataset. Predefine product improvement-related indicators for the target product; Using the product improvement correlation index as a constraint, several sample user behavior datasets and several sample product improvement correlation feature label sets are collected, and a graph neural network with an attention mechanism is trained to construct an improvement correlation feature extractor. Using the improved correlation feature extractor, product improvement correlation features are extracted from the historical standard behavior dataset to obtain a historical product improvement correlation feature set.

[0028] In this embodiment, the historical user behavior dataset is first processed for outlier handling, missing value imputation, and multi-source data alignment to generate a historical standard behavior dataset. Outlier handling identifies and removes extreme data from the historical user behavior dataset that contradicts common sense, logical contradictions, or is incorrectly collected, aiming to eliminate data noise interference with subsequent analysis. Missing value imputation involves supplementing missing key fields in the dataset using appropriate methods, such as mean / median imputation, interpolation, user profile-based matching imputation, and model prediction imputation; the appropriate method must be selected based on the data type. Multi-source data alignment standardizes user behavior data from different channels in terms of data format, timestamps, field definitions, and identifier systems, with the core goal of achieving cross-channel data integrability and comparability. The historical standard behavior dataset, obtained after outlier handling, missing value imputation, and multi-source data alignment, is a structured, noise-free, and uniformly formatted user behavior dataset, serving as the foundational data carrier for subsequent feature extraction and model training.

[0029] Identify and remove extreme data operations from historical user behavior datasets that contradict common sense, logically inconsistencies, or are incorrectly collected, eliminating data noise that interferes with subsequent analysis. Then, for key fields missing in the dataset, employ appropriate methods to complete them, such as mean / median imputation, interpolation, user profile-based matching imputation, and model prediction imputation. The appropriate method must be selected based on the data type. Standardize user behavior data from different channels in terms of data format, timestamps, field definitions, and identifier systems to achieve cross-channel data integrability and comparability. Finally, use the data after outlier handling, missing value imputation, and multi-source data alignment as the user behavior dataset, serving as the foundation for subsequent feature model training.

[0030] Secondly, predefine the product improvement correlation indicators for the target product. Among them, the product improvement correlation indicators are quantitative evaluation indicators that are directly related to the direction of process optimization of sunscreen products, and are the core basis for screening product improvement correlation characteristics. Predefinition is an operation that clearly defines the product improvement correlation indicators based on the production process logic of sunscreen products, core user needs and industry standards, to ensure that the indicators are quantifiable and process-oriented.

[0031] For a specific sunscreen, predefine product improvement-related indicators, and clarify the data source, update frequency, and anomaly threshold for each indicator to complete the predefinement of product improvement-related indicators.

[0032] Next, using product improvement-related metrics as constraints, we collected several sample user behavior datasets and several sample product improvement-related feature label sets, trained a graph neural network with an attention mechanism, and constructed an improvement-related feature extractor.

[0033] The sample user behavior dataset is a subset of the historical standard behavior dataset used for model training. It needs to cover different user groups, usage scenarios, and feedback types to ensure the representativeness and diversity of the samples. The sample product improvement correlation feature label set is a set of labels obtained by labeling the sample user behavior dataset based on predefined product improvement correlation indicators. Each sample corresponds to a set of feature labels and serves as a supervision signal for model training. The graph neural network with attention mechanism is a deep learning model that introduces an attention mechanism on the basis of traditional graph neural networks. The graph neural network is used to model the complex correlations in user behavior data, and the attention mechanism is used to highlight key features that have a significant impact on product improvement correlation indicators, thereby improving the feature extraction accuracy of the model.

[0034] The specific steps for constructing an improved association feature extractor for a specific sunscreen are as follows: Sample collection: 100,000 sample user behavior datasets were extracted from the historical standard behavior dataset using stratified sampling principles, covering different skin types, usage scenarios, and feedback types; based on predefined product improvement correlation indicators, each sample was manually labeled to generate a corresponding sample product improvement correlation feature label set.

[0035] Model Construction: A graph neural network with an attention mechanism is employed, using users, scenarios, and feedback types as graph nodes, and user-scenario, scenario-feedback, and user-feedback relationships as graph edges. The attention mechanism module calculates the contribution weight of each node to product improvement metrics. The model output layer is set to the output dimension corresponding to predefined metrics. It typically consists of an input layer, multiple attention graph convolutional layers, and a task-specific output layer. Each layer first performs a linear transformation on the node features, then calculates the attention coefficients between the node and its neighbors, normalizes them to attention weights using a softmax function, and then uses these dynamic weights to weighted aggregate the neighbor features. Finally, a non-linear activation function is used to obtain the updated node representation. The model employs a multi-head attention mechanism to capture different relationship patterns, enhancing representational power by concatenating or averaging the outputs of multiple attention heads.

[0036] Model training: Input the sample user behavior dataset into the model, use the sample product improvement association feature label set as supervision, use mean squared error (MSE) as the loss function, set the learning rate to 0.001, and after 50 rounds of iterative training, the model converges (the loss value drops from the initial 2.5 to below 0.05), and finally obtains the improved association feature extractor.

[0037] Finally, using an improved correlation feature extractor, product improvement correlation features are extracted from the historical standard behavior dataset to obtain a historical product improvement correlation feature set.

[0038] A historical standard behavior dataset of a certain sunscreen is input into a trained improved correlation feature extractor. The model processes each data point row by row, outputting corresponding feature values ​​based on predefined product improvement correlation indicators. After feature extraction of all data, the output results are integrated according to dimensions such as user ID, usage time, and scenario to generate a structured historical product improvement correlation feature set. Each record corresponds to a product improvement correlation feature value, and the user group, usage scenario, and other attributes corresponding to the feature are also labeled to facilitate subsequent dimensional analysis.

[0039] In step S10 of the method provided in this application embodiment, the types of product improvement correlation indicators include at least process sensitivity indicators, multi-factor coupling influence indicators, and process window robustness indicators. Among them, the process sensitivity indicators include at least dosage-extensibility sensitivity coefficient and environment-film stability correlation degree. The multi-factor coupling influence indicators include at least superimposed use-pilling coupling index and high frequency recoating-tolerance attenuation gradient. The process window robustness indicators include at least homogeneous shear strength-skin feel consistency bandwidth and filling accuracy-dosage perception tolerance rate.

[0040] Product improvement correlation indicators are quantitative evaluation indicators that directly relate to the direction of process optimization for sunscreen products. They are the core basis for screening product improvement correlation characteristics and indicate the specific process correlation dimensions that need to be focused on and optimized.

[0041] In this embodiment of the application, behavior-process sensitivity index, multi-factor coupling influence index and process window robustness index are predefined. Among them, behavior-process sensitivity index measures the responsiveness of a specific user behavior pattern to the adjustment of a certain process parameter.

[0042] Behavioral-process sensitivity indicators include the dosage-spreadability sensitivity coefficient and the environment-film stability correlation. The dosage-spreadability sensitivity coefficient is the percentage reduction in dosage by the user per application, and the strength of the correlation between this percentage and adjustments to the ratio of silicon elastomer to volatile oil in the formulation. A higher value indicates a greater potential to guide users to use the correct dosage by adjusting the formulation's spreadability process. The dosage-spreadability sensitivity coefficient = change in spreadability score / change in dosage, with a threshold range of -0.5 to 0.5. A negative value indicates a decrease in spreadability with increased dosage. The environment-film stability correlation is the strength of the correlation between the increase in protection failure complaint rate in high humidity (>80%) environments and the crosslinking density of the film-forming agent or the stability process parameters of the emulsion system. The environmental film stability correlation is the rate of change in the product's film stability score when temperature / humidity increases. The environmental film stability correlation = (high temperature / high humidity film-forming score - normal temperature / normal humidity film-forming score) / normal temperature / normal humidity film-forming score × 100%, with a threshold range of -20% to 5%. A negative value indicates a decrease in film stability due to environmental degradation.

[0043] For example, the dosage-spreadability sensitivity coefficient: Using 100,000 user usage records extracted from a historical standard behavior dataset, when the user's single dosage increased from 0.5g to 0.6g, the spreadability score decreased from 3.2 to 2.8; the dosage-spreadability sensitivity coefficient for sunscreen = (3.2-2.8) / (0.6-0.5) = 4. The environment-film-forming stability correlation: The score was 4.2 at normal temperature and humidity (25℃, 50% humidity) and 2.5 at high temperature and humidity (35℃, 80% humidity). The environment-film-forming stability correlation = -40.5%, a negative value indicating that high temperature and humidity environments significantly reduce film-forming stability, suggesting that the current film-forming agent has insufficient weather resistance and requires adjustment of the film-forming agent compounding process.

[0044] Multi-factor coupling impact indexes quantify the nonlinear impact on a certain process dimension when two or more user behaviors or environmental factors work together.

[0045] The layering-pilling coupling index is the probability of pilling complaints when used in combination with a makeup primer in a low-temperature, dry environment. Compared to the multiplication factor of a single factor, it directly relates to powder surface treatment processes and polymer compatibility. It represents the probability of pilling feedback when sunscreen and foundation are used together. The layering-pilling coupling index = (Number of pilling complaints from layering / Total number of complaints from layering) × 100%, with a threshold range of 0-30%. The high-frequency reapplication-tolerance decay gradient is the slope of the curve showing the decrease in the rate of stinging sensation as the penetration rate of alcohol solvents in the formula changes when users reapplication more than 3 times per day. The high-frequency reapplication-tolerance decay gradient = (Number of high-frequency reapplication tolerance issues / Number of users who reapplication frequently) - (Number of low-frequency reapplication tolerance issues / Number of users who reapplication frequently), with a threshold range of 0-25%.

[0046] For example, the pilling coupling index for layered use: Based on 50,000 user records of layered use, the pilling feedback rate was 8% when using sunscreen alone, rising to 25% after layering foundation, and 22% after layering primer. The pilling coupling index for layered use was 23.5%, indicating insufficient surface treatment in the current formula. The tolerance decay gradient for high-frequency reapplication: Based on 30,000 user records of high-frequency reapplication (≥3 times daily), the feedback rate of skin redness / stirring was statistically analyzed: the tolerance problem feedback rate was 5% for users who reapplication once daily, rising to 15% for users who reapplication three times daily, and 22% for users who reapplication five times daily. Calculated using the formula ((number of high-frequency reapplication tolerance problems / number of high-frequency reapplication users) - (number of low-frequency reapplication tolerance problems / number of low-frequency reapplication users)), the tolerance decay gradient for high-frequency reapplication was 17%, indicating that high-frequency reapplication significantly exacerbates skin tolerance problems, suggesting that the current formula has an excessively high alcohol content.

[0047] Process window robustness metrics measure the ability of user experience metrics to remain stable within the allowable fluctuation window of current production process parameters, including homogeneous shear strength-skin feel consistency bandwidth and filling accuracy-dosage perception tolerance.

[0048] Homogeneous Shear Strength-Skin Feel Consistency Bandwidth: The standard deviation of user ratings for freshness within the allowable range of homogenization speed fluctuations. A narrower bandwidth indicates higher process robustness. Homogeneous Shear Strength-Skin Feel Consistency Bandwidth = Maximum Skin Feel Score - Minimum Skin Feel Score, with a threshold range of 0-1. Filling Accuracy-Perceived Dosage Tolerance: The range within which the actual volume difference of a single product due to filling accuracy will not trigger significant negative user feedback regarding product durability. Filling Accuracy-Perceived Dosage Tolerance is the proportion of user feedback regarding insufficient / excessive dosage when filling accuracy deviates. Filling Accuracy-Perceived Dosage Tolerance = (Number of Accuracy Deviation Feedback / Total Number of Feedback) × 100%, with a threshold range of 0-10%.

[0049] For example, the homogenization shear strength-skin feel consistency bandwidth: User skin feel scores (1-5 points) at different homogenization shear strengths of 3000 r / min, 4000 r / min, 5000 r / min, and 6000 r / min were statistically analyzed: 3.0 points at 3000 r / min, 3.8 points at 4000 r / min, 4.2 points at 5000 r / min, and 3.5 points at 6000 r / min. The homogenization shear strength-skin feel consistency bandwidth is 1.2 points; a larger bandwidth value indicates a narrower shear strength window for the current homogenization process. Filling accuracy-dosage perception error rate: User dosage perception feedback at different filling accuracy deviations of ±0.05g, ±0.1g, ±0.15g, and ±0.2g was statistically analyzed: the incomplete dosage feedback rate was 3% for ±0.05g, 8% for ±0.1g, 15% for ±0.15g, and 25% for ±0.2g. The filling accuracy-dosage perception tolerance rate is 25% (±0.2g). The high tolerance rate indicates that the filling accuracy deviation has a significant impact on user perception, and the filling equipment needs to be upgraded.

[0050] In this embodiment, by clearly defining multi-channel data collection paths, the comprehensiveness and authenticity of data sources are ensured, thus solving the problem of lagging traditional survey data. By training a graph neural network with an attention mechanism through predefined improved correlation indicators, features are extracted from the cleaned standard data, and the feature set is intelligently filtered to remove noisy data. By training a BP neural network to predict the basic adjustment cycle and combining it with external market demand forecasts for dynamic compensation, intelligent transformation is achieved, providing a data foundation for subsequent accurate cycle compensation and dynamic optimization.

[0051] S20: Determine the optimal process adjustment time period based on the historical product improvement association feature set and the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range; In this embodiment, the preset time range is a pre-defined process adjustment decision coverage period, which is the time boundary between demand forecasting and process adjustment; the predicted product demand growth rate is the average growth rate of product demand within the preset time range predicted by a Long Short-Term Memory (LSTM) network based on historical demand data and improvement correlation characteristics; the predicted product demand fluctuation coefficient is the coefficient of variation of the predicted month-on-month growth rate of demand within the preset time range, reflecting the severity of demand fluctuations; the optimal process adjustment time period is the process adjustment interval determined by combining historical improvement correlation characteristics, predicted demand growth rate and fluctuation coefficient, predicted by a BP neural network and then compensated and calibrated, which is the optimal period for balancing response speed and production stability.

[0052] For a specific sunscreen, historical survey records of similar sunscreens are retrieved to collect a set of sample product improvement-related features and a corresponding set of sample process adjustment time cycle labels, generating a process adjustment time cycle predictor. The historical improvement-related feature set of the sunscreen is input into the predictor to obtain the initial predicted process adjustment time cycle. A comprehensive compensation coefficient is calculated to compensate for the initial predicted process adjustment time cycle, yielding the final optimal process adjustment time cycle.

[0053] Step S20 in the method provided in this application embodiment includes: Using the product attributes of the target product as constraints, historical survey records of similar products are retrieved to collect several sample product improvement association feature sets and several sample process adjustment time cycle label sets; Using the aforementioned sample product improvement association feature set as input and the aforementioned sample process adjustment time cycle label set as supervision, a BP neural network is trained until convergence to generate a process adjustment cycle predictor. Using the process adjustment cycle predictor, the predicted process adjustment time cycle is obtained based on the analysis of the historical product improvement correlation feature set; The predicted process adjustment time period is compensated based on the predicted product demand growth rate and the predicted product demand fluctuation coefficient within a preset time range to obtain the optimal process adjustment time period.

[0054] In this embodiment, the product attributes of the target product are used as constraints. Historical survey records of similar products are retrieved to collect several sample product improvement-related feature sets and several sample process adjustment time period label sets. Among them, product attributes are the core characteristics of the target sunscreen product, including product type, sun protection index, efficacy positioning, price range, etc., which are the core basis for screening similar products and ensuring the relevance and comparability of sample data. Similar products are marketed sunscreen products that are highly matched with the target product in terms of product attributes, and must meet the requirements of consistency or similarity in key attributes such as sun protection index, efficacy positioning, and price range. Historical survey records are structured data such as user behavior data, process adjustment records, and market feedback data of similar products.

[0055] Based on the target product, filter conditions with product attribute constraints are set, and historical research records of similar products meeting the conditions are retrieved from the brand database. The historical research records of each similar product are analyzed: a set of sample product improvement-related features is extracted, including six features: dosage-spreadability sensitivity coefficient, environment-film formation stability correlation, layered use-pilling coupling index, homogeneous shear strength-skin feel consistency bandwidth, etc. Then, the corresponding sample process adjustment time period label set is extracted. Finally, a matching sample feature set and label set covering the adjustment period are collected.

[0056] Secondly, using a set of improved features from several sample products as input and a set of process adjustment time cycle labels from several samples as supervision, a backpropagation (BP) neural network is trained until convergence to generate a process adjustment cycle predictor. The BP neural network is a multi-layer feedforward neural network trained using the backpropagation algorithm, consisting of an input layer, hidden layers, and an output layer. It can learn the nonlinear mapping relationship between input and output, making it suitable for continuous value prediction tasks such as process adjustment cycles.

[0057] The specific steps for constructing a process adjustment cycle predictor using a BP neural network are as follows: Model Structure: The main structure of the process adjustment cycle predictor includes an input layer, a hidden layer, and an output layer. The input layer is a neural network layer that receives the feature set associated with sample product improvements; the number of nodes is consistent with the feature dimension. The hidden layer is an intermediate layer that performs a nonlinear transformation on the input features and uses the ReLU activation function to enhance the model's fitting ability. The output layer is a neural network layer that outputs the predicted process adjustment time cycle, using a linear activation function to output continuous predicted values.

[0058] Model training involves inputting the sample feature set into the neural network, setting the initial learning rate and weights, assigning weights, calculating the error between the prediction result and the sample parameter configuration set using the mean squared error function, adjusting the weights and calculating the error, iterating and repeating until the error is minimized, calculating the prediction period through forward propagation, comparing it with the sample label period to calculate the loss, updating the network weights through backpropagation, and iterating until the loss value stabilizes. After each training epoch, the performance is evaluated using a validation set to avoid overfitting. When the MSE loss on the training set decreases by less than 1e-5 for five consecutive epochs and the MSE loss on the validation set stabilizes below 0.01, the model is considered converged, and the process adjustment period predictor is obtained.

[0059] Next, the process adjustment cycle predictor is used to obtain the predicted process adjustment time cycle based on the analysis of historical product improvement correlation feature sets. The predicted process adjustment time cycle is a preliminary process adjustment cycle suggestion output by the predictor based on the historical feature set of the target product, and serves as the basis for subsequent compensation calibration. The historical product improvement correlation feature set of the target product is standardized and then input into the process adjustment cycle predictor.

[0060] The process adjustment cycle predictor internally calculates the weights of each feature, performs a nonlinear transformation in the hidden layer, and then calculates the prediction cycle in the output layer. The final output is the predicted process adjustment time cycle for the target product, reflecting preliminary adjustment suggestions based on historical product improvement correlation features.

[0061] For example, the historical product improvement associated feature set of the target product—including dosage-spreadability sensitivity coefficient -4, environment-film stability correlation degree -40%, superimposed use-pilling coupling index 23%, high-frequency touch-up coating-tolerance attenuation gradient 17%, homogeneous shear strength-skin feel consistency bandwidth 1.2 points, and filling accuracy-dosage perception fault rate 25%—is standardized and then input into the process adjustment cycle predictor. The final output is a predicted process adjustment time cycle of 42 days for the target product.

[0062] Finally, the optimal process adjustment time period is obtained by compensating for the predicted product demand growth rate and the predicted product demand volatility coefficient within a preset time range. Here, the preset time range is the coverage period for the target product's process adjustment decision; the predicted product demand growth rate is the average monthly growth rate of the target product's demand within the preset time range, reflecting the market demand growth trend; the predicted product demand volatility coefficient is the coefficient of variation of the target product's month-on-month demand growth rate within the preset time range, reflecting the degree of market demand fluctuation—the higher the coefficient, the greater the fluctuation; and the optimal process adjustment time period is the final process adjustment period after dynamic demand compensation, which matches the product's own characteristics and adapts to changes in market demand, serving as the optimal decision-making basis for process adjustments.

[0063] Obtain demand forecast data, predict the growth rate of target product demand within a preset time range, predict the product demand fluctuation coefficient, calculate the compensation coefficient, and use the compensation coefficient to compensate for the predicted process adjustment time cycle.

[0064] In step S20 of the method provided in this application embodiment, obtaining the predicted product demand growth rate and the predicted product demand fluctuation coefficient within a preset time range includes: Obtain the historical demand month-on-month growth rate sequence of the target product within a preset time granularity over a historical time range; By calling up the historical operation records of similar products, and using the time span of the preset time range as a constraint, several sample product improvement related feature sets and several sample demand month-on-month growth rate sequences are collected. The historical demand month-on-month growth rate sequences recorded in the historical time range of different sample product improvement related feature sets and sample demand month-on-month growth rate sequences are statistically analyzed as sample predicted demand month-on-month growth rate sequences, thus obtaining several sample predicted demand month-on-month growth rate sequences. Using the aforementioned sample product improved association feature set and the aforementioned sample demand month-on-month growth rate sequence as input, and using the aforementioned sample predicted demand month-on-month growth rate sequence as supervision, a long short-term memory network is trained until convergence to generate a demand growth prediction plugin. Using the aforementioned demand growth forecasting plugin, based on the historical product improvement related feature set and the historical demand month-on-month growth rate sequence, a predicted demand month-on-month growth rate sequence within a preset time range is obtained. The mean of the predicted demand month-on-month growth rate series is calculated to obtain the predicted product demand growth rate. Calculate the coefficient of variation of the predicted demand month-on-month growth rate series as the predicted product demand volatility coefficient.

[0065] In this embodiment, the historical demand month-on-month growth rate sequence of the target product within a preset time granularity over a historical time range is first obtained. Here, the historical time range is the complete sales cycle of the target sunscreen product since its market launch, serving as the time boundary for demand data collection; the preset time granularity is the statistical unit for demand data, ensuring that the data granularity matches the subsequent forecast period; the historical demand month-on-month growth rate is the percentage increase in product demand within two adjacent time granularities, reflecting the monthly trend of demand changes; and the historical demand month-on-month growth rate sequence is a set of historical demand month-on-month growth rates arranged chronologically, serving as the foundational data for subsequent demand forecasting.

[0066] Historical sales data are extracted from the brand sales database using the target product as an index, and the month-on-month growth rate is calculated. The historical demand month-on-month growth rate is calculated as (this month's demand - last month's demand) / last month's demand × 100%. The historical demand month-on-month growth rate is calculated for all preset time granularities within the historical time range, and the historical demand month-on-month growth rate sequence is generated by arranging them in chronological order to cover the demand change trend within the historical time range.

[0067] For example, taking sunscreen as an example, monthly sales data from June 2022 to December 2024 are extracted from the brand's sales database: 10,000 bottles sold in June 2022, 10,500 bottles in July, and so on. The month-on-month growth rate is calculated: July 2022: (10,500-10,000) / 10,000×100%=5%; August 2022: (11,760-10,500) / 10,500×100%=12%..., resulting in 30 monthly month-on-month growth rate values. A sequence is generated: arranging the data chronologically into a historical demand month-on-month growth rate sequence: [5%, 12%, 8%, 15%, 10%, 5%, 3%, 8%, 12%, ...], covering the demand trend from July 2022 to December 2024.

[0068] Secondly, historical operational records of similar products are retrieved, constrained by a preset time span. Several sample product improvement-related feature sets and several sample demand month-on-month growth rate sequences are collected. The historical demand month-on-month growth rate sequences recorded within the historical time span for different sample product improvement-related feature sets and sample demand month-on-month growth rate sequences are then used as sample predicted demand month-on-month growth rate sequences, resulting in several sample predicted demand month-on-month growth rate sequences. The historical operational records of similar products are complete operational data of sunscreen products matching the target product attributes in the brand database, including product improvement-related features, monthly demand month-on-month growth rate, market environment data, etc. The preset time span is the length of the target prediction period, used to constrain the time span of the sample data.

[0069] The sample product improvement correlation feature set is the improvement correlation features extracted from the operation records of similar products that are consistent with the feature dimensions of the target product; the sample demand month-on-month growth rate sequence is the demand month-on-month growth rate sequence of similar products over a preset time span in history; the sample predicted demand month-on-month growth rate sequence is the actual demand month-on-month growth rate sequence after the preset time span corresponding to the sample feature set and the sample sequence.

[0070] The system retrieves data on similar products, including operational records of similar sunscreen products from a brand database. Data is filtered by a preset time span, and the month-on-month growth rate sequence of sample demand within that time span is extracted. The system also extracts the product improvement correlation feature set for the corresponding time period, and then extracts the actual month-on-month growth rate sequence for subsequent preset time spans as the sample prediction sequence. Finally, sample data is collected, with each group containing a sample feature set, a sample sequence, and a sample prediction sequence, covering demand change patterns under different market environments and product characteristics.

[0071] A training sample set is constructed using historical data of similar products, linking product characteristics, historical demand sequences, and future demand sequences to provide a learning basis for the LSTM model. The constraint of a preset time range ensures that the sample and the prediction target dimension are consistent, enabling the model to learn the patterns of demand forecasting within the preset time range. If the sample time span is inconsistent with the prediction target, the model will be unable to learn long-term demand trends.

[0072] For example, the operational records of 10 similar sunscreen products from 2018 to 2024 were retrieved from the brand database. Data was filtered by a 6-month time span: the month-on-month growth rate of sample demand was extracted for 6 consecutive months, for product A from January to June 2022: [8%, 12%, 15%, 18%, 15%, 12%]; the improved correlation feature set of sample products was extracted, for product A from January to June 2022: usage-spreadability sensitivity coefficient -3.5, pilling coupling index 20%; the sample prediction sequence was extracted, for product A from July to December 2022: [18%, 20%, 15%, 10%, 8%, 5%]. Finally, 200 sets of sample data were collected.

[0073] Furthermore, using several sample product improvement feature sets and several sample demand month-on-month growth rate sequences as input, and several sample predicted demand month-on-month growth rate sequences as supervision, a Long Short-Term Memory (LSTM) network is trained until convergence to generate a demand growth prediction plugin. The LSTM network is a special type of recurrent neural network (RNN) that uses a gating mechanism to solve the gradient vanishing problem of traditional RNNs, enabling it to capture long-term dependencies in time series data and making it suitable for time series tasks such as demand forecasting.

[0074] For example, the steps to construct a demand growth prediction plugin using a Long Short-Term Memory (LSTM) network are as follows: Data preparation: The feature set of the sample data is standardized and normalized to the 0-1 interval. The month-on-month growth rate sequence of the sample is standardized and converted into a dimensionless value of -1 to 1. The sample data is then divided into training set and validation set in an 8:2 ratio.

[0075] Model Construction: The Long Short-Term Memory (LSTM) network mainly consists of an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer contains product feature nodes and historical sequence nodes; the LSTM layer uses the tanh activation function; the fully connected layer maps the output of the LSTM layer; the output layer outputs a standardized predicted month-on-month growth rate sequence, which is then de-standardized to the actual growth rate.

[0076] Model Training: Set the learning rate to 0.001, the number of iterations to 200, and the loss function to MSE. Input the training set into the model and perform forward propagation to calculate the predicted sequence. Update the parameters through backpropagation. Evaluate the model performance on the validation set after each training epoch. Training is terminated when the validation set loss no longer decreases for 10 consecutive epochs. After training, perform a final evaluation on the test set. If the loss value stabilizes at 0.8 and the validation set mean absolute error is 0.8%, the convergence condition is met, and training is stopped. Finally, save the converged model parameters and generate the demand growth prediction plugin.

[0077] Furthermore, using the demand growth forecasting plugin, based on the historical product improvement correlation feature set and the historical demand month-on-month growth rate sequence, a predicted demand month-on-month growth rate sequence is obtained within a preset time range. This predicted demand month-on-month growth rate sequence is the set of predicted demand month-on-month growth rate values ​​output by the model within the preset time range, and serves as the direct basis for calculating the demand growth rate and volatility coefficient.

[0078] For example, the historical product improvement correlation feature set (e.g., usage-expansion sensitivity coefficient -4, pilling coupling index 23%, film-forming stability correlation -40%), and the historical demand month-on-month growth rate sequence (actual sequence from July to December 2024 [15%, 18%, 15%, 12%, 10%, 8%]) are standardized and input into the demand growth forecasting plugin. The correlation between the features and the historical sequence is analyzed through an LSTM layer, outputting a standardized forecast sequence. After inverse standardization, the predicted demand month-on-month growth rate sequence for January to June 2025 is obtained: [5%, 8%, 12%, 15%, 18%, 15%].

[0079] Furthermore, the mean of the predicted demand month-on-month growth rate series is calculated to obtain the predicted product demand growth rate. The mean calculation involves taking the arithmetic mean of all values ​​in the predicted demand month-on-month growth rate series; the predicted product demand growth rate is the average monthly growth rate of demand within the prediction period, reflecting the overall growth trend of demand and serving as a core parameter for process adjustment cycle compensation.

[0080] The forecasted product demand growth rate is obtained by calculating the mean of the predicted month-on-month growth rate series. The predicted product demand growth rate is calculated as (r1 + r2 + ... + rn) / n (n = integer). If the monthly peak (18%) is used directly as the growth basis, the demand growth trend will be overestimated, resulting in an excessively short adjustment period. The 12.17% calculated by the mean objectively reflects the overall growth level of the forecast period, including seasonal peaks, balancing the growth levels of the off-season and peak season, and making the compensation for the process adjustment period more reasonable.

[0081] For example, the predicted month-on-month growth rate of demand is [5%, 8%, 12%, 15%, 18%, 15%]. The average value is calculated using the formula: (5%+8%+12%+15%+18%+15%) / 6=73% / 6≈12.17%, which is rounded down to 12%. Therefore, the predicted product demand growth rate is 12.17% / month.

[0082] Finally, the coefficient of variation (COP) of the predicted month-on-month demand growth rate series is calculated as the predicted product demand volatility coefficient. The COP is a relative indicator measuring the dispersion of data, eliminating the influence of data units and suitable for comparing the volatility of different series. The standard deviation measures the deviation between the predicted series value and the mean, reflecting the monthly fluctuation range of the month-on-month demand growth rate. The predicted product demand volatility coefficient represents the relative dispersion of the month-on-month demand growth rate within the prediction period; a higher coefficient indicates more drastic demand fluctuations. The mean and standard deviation of the predicted month-on-month demand growth rate series are calculated, and then the COP is calculated, where the COP = (standard deviation / mean). The COP is then used as the predicted product demand volatility coefficient.

[0083] For example, for the target product sunscreen, the average of the squared differences of the predicted demand month-on-month growth rate series within a preset time range is approximately 0.1217; the standard deviation is approximately 0.0447, and the coefficient of variation is approximately (0.0447 / 0.1217) ≈ 0.367, that is, the predicted product demand fluctuation coefficient is 0.367.

[0084] In step S20 of the method provided in this application embodiment, the predicted process adjustment time period is compensated based on the predicted product demand growth rate and the predicted product demand fluctuation coefficient within a preset time range to obtain the optimal process adjustment time period, including: Access historical operational records of similar products to collect the average historical product demand growth rate and the average historical product demand fluctuation coefficient within a historical time range; The ratio of the predicted product demand growth rate to the average historical product demand growth rate is used as the first period compensation coefficient. The ratio of the average historical product demand fluctuation coefficient to the predicted product demand fluctuation coefficient is used as the second period compensation coefficient. The comprehensive periodic compensation coefficient is calculated by weighting the first periodic compensation coefficient and the second periodic compensation coefficient. The product of the comprehensive cycle compensation coefficient and the predicted process adjustment time period is taken as the optimal process adjustment time period.

[0085] In this embodiment, the ratio of the predicted product demand growth rate to the historical average product demand growth rate is first used as the first period compensation coefficient.

[0086] First, retrieve historical operational records of similar sunscreen and whitening products from the brand database to obtain data on similar products. Extract the demand growth rate and demand fluctuation coefficient for each product in each year, and calculate the average historical demand growth rate and the average historical demand fluctuation coefficient.

[0087] For example, the operational records of 10 similar sunscreen products from 2018 to 2024 are retrieved from the brand database. The demand growth rate for each product in each year is extracted: Product A: 12% in 2018, 10% in 2019, 8% in 2020, etc.; Demand fluctuation coefficient: Product A: 0.32 in 2018, 0.30 in 2019, 0.28 in 2020, etc. The historical average demand growth rate is calculated to be 10% / month; the historical average demand fluctuation coefficient is 0.30.

[0088] Secondly, the ratio of the historical average product demand fluctuation coefficient to the predicted product demand fluctuation coefficient is used as the second-period compensation coefficient. Here, the predicted product demand growth rate is the demand growth rate of the target product within a preset time frame; the first-period compensation coefficient reflects the multiple relationship between the predicted growth rate of the target product and the industry benchmark. The first-period compensation coefficient is the ratio of the predicted growth rate to the historical average growth rate. If the first-period compensation coefficient > 1, it indicates that the predicted growth rate is higher than the industry level, and the process adjustment cycle needs to be shortened to improve response speed; if the first-period compensation coefficient < 1, the opposite is true.

[0089] For example, the predicted product demand growth rate is 0.1217 / month, the historical average product demand growth rate is 0.1 / month, and the first-cycle compensation coefficient is 0.1217 / 0.1≈1.22.

[0090] Next, a comprehensive periodic compensation coefficient is calculated by weighting the first-period compensation coefficient and the second-period compensation coefficient. The predicted product demand volatility coefficient is the target product demand volatility coefficient within a preset time range. The second-period compensation coefficient reflects the ratio of the predicted volatility coefficient of the target product to the industry benchmark. The second-period compensation coefficient is the ratio of the historical average product demand volatility coefficient to the predicted product demand volatility coefficient. A second-period compensation coefficient < 1 indicates that the predicted volatility coefficient is higher than the industry level, requiring an extended process adjustment cycle to reduce production risk; conversely, a second-period compensation coefficient > 1 indicates the opposite. A larger predicted volatility coefficient indicates an unstable growth rate, and thus a longer process adjustment cycle.

[0091] When the predicted product demand growth rate is higher than the industry's historical average (>1), it indicates that the market expansion window has opened, and the process adjustment cycle needs to be shortened to quickly respond to opportunities and meet new demands. When the predicted product demand volatility coefficient is higher than the industry's historical average (<1), it indicates that the market has entered a high uncertainty stage, and the process adjustment cycle needs to be extended to avoid the operational risks of production and demand mismatch and ensure the stability of the supply chain and quality.

[0092] For example, the historical average product demand volatility coefficient is 0.30, the predicted product demand volatility coefficient is 0.367, and the second-cycle compensation coefficient is approximately 0.30 / 0.367.

[0093] Ultimately, the product of the comprehensive cycle compensation coefficient and the predicted process adjustment time period is taken as the optimal process adjustment time period. The comprehensive cycle compensation coefficient is the final compensation coefficient that integrates both growth rate and volatility factors; the optimal process adjustment time period is the final process adjustment period calibrated by the comprehensive compensation coefficient, representing the best decision result that balances market responsiveness and production stability.

[0094] The comprehensive cycle compensation coefficient = first cycle compensation coefficient × growth weight + second cycle compensation coefficient × volatility weight, comprehensively reflecting the balance between response speed and risk control. The weights are set based on industry experience; if the influence of the growth rate is higher, the growth rate weight is higher. The optimal process adjustment time period = comprehensive cycle compensation coefficient × predicted process adjustment time period. Adjusting only based on the first cycle compensation coefficient would ignore high volatility risk; adjusting only based on the second cycle compensation coefficient would miss growth opportunities. The weighted calculation reflects a slight shortening of the growth period while also taking into account a slight extension of the volatility period, achieving a dual balance.

[0095] For example, the comprehensive compensation coefficient = (first coefficient × 0.6 + second coefficient × 0.4) = 1.22 × 0.6 + 0.82 × 0.4 ≈ 1.06. The optimal process adjustment time cycle = prediction cycle × comprehensive compensation coefficient 42 × 1.06 = 44.52 days, rounded to 45 days. This cycle was shortened due to rapid demand growth and appropriately extended due to large fluctuations, achieving a balance between response speed and stability.

[0096] In this embodiment, an LSTM forecasting model is constructed to predict future demand growth rates and their volatility coefficients, reflecting market risks. By comparing the current predicted growth rate and volatility coefficient with historical industry benchmarks, a comprehensive compensation coefficient is obtained to ensure production stability and the effectiveness of adjustments. Ultimately, the optimal process adjustment time cycle is determined, achieving intelligent synchronization between adjustment rhythms and market rhythms.

[0097] S30: Collect user behavior datasets of the target product within the optimal process adjustment time period through multiple channels, perform data preprocessing and product quality correlation feature extraction, and determine product improvement indicators based on the product quality correlation feature set; In this embodiment, the user behavior dataset within the optimal process adjustment time period is user behavior data collected in real time through multiple channels within the optimal process adjustment time period; product quality correlation feature extraction is the process of extracting features directly related to product quality from the preprocessed user behavior data within the period; the product quality correlation feature set is a set of features obtained after extraction that reflects the correlation between user behavior and product quality performance within the period; and the product improvement indicators are specific goals that process optimization needs to achieve, determined based on the product quality correlation feature set.

[0098] By continuously collecting user behavior data through multiple channels, and after preprocessing, using associated features as constraints, features are extracted to form a product quality associated feature set, and product improvement indicators are determined.

[0099] Step S30 in the method provided in this application embodiment includes: Configure product quality related indicators, wherein the product quality related indicators include at least the scenario-based sufficient usage compliance rate, negative feedback behavior pattern specificity index, efficacy perception-behavioral input efficiency ratio, experience consistency entropy, and silent churn risk score. Using the product quality correlation index as a constraint, several sample user behavior datasets and several sample product quality correlation feature label sets are collected, and a graph neural network with an attention mechanism is trained to construct a quality correlation feature extractor. The quality correlation feature extractor is used to extract product quality correlation features to obtain a product quality correlation feature set, and product improvement indicators are determined based on the product quality correlation feature set.

[0100] In this embodiment of the application, product quality-related indicators are first configured, including at least the scenario-based sufficient usage compliance rate, the negative feedback behavior pattern specificity index, the efficacy perception-behavioral input efficiency ratio, the experience consistency entropy, and the silent churn risk score.

[0101] Contextualized adequate usage compliance rate (SCR) is the percentage of times a user achieves the theoretically adequate usage standard in a single application within a specific usage scenario, out of the total number of uses in that scenario. It can be estimated using smart packaging data or a usage-purchase cycle model. A low SCR indicates strong process-related issues, such as product spreadability, stickiness, or sweat compatibility defects in the current scenario. It is the core feature for extracting context-skin feel-process correlation characteristics.

[0102] The Behavioral Pattern Specificity Index (NFSI) measures whether a negative feedback is strongly associated with one or more specific, definable prior user behaviors. A higher index indicates a more clearly defined behavioral trigger for the quality issue. The co-occurrence strength and confidence of negative evaluations and specific behavioral sequences can be calculated through association rule mining or sequence pattern recognition. Strong process associations, manifested as high NFSI negative feedback, provide valuable clues for extracting deep features such as formulation incompatibilities or process parameters incompatible with specific usage conditions. This can guide the model to discover complex patterns where process parameters Y lead to problem Z under behavior X.

[0103] The Efficacy Perceived-Effectiveness Ratio (EPER) is the ratio of the degree of positive skin changes reported by users to the behavioral costs borne by users in achieving these changes. It can be calculated using user self-rating scores and behavioral log data. A low EPER indicates a strong correlation with the product's process, suggesting poor delivery efficiency of the product's active ingredients or ineffective final efficacy. This can directly guide feature extraction models to identify user behavior patterns associated with process parameters such as ingredient penetration rate, sustained-release technology, and formulation bioavailability.

[0104] Experience Consistency Entropy (ECE) quantifies the degree of fluctuation in a product's impact on a user's experience by analyzing the disorder (entropy value) of a user's long-term behavioral and skin condition sequences. Ultimately, the entropy values ​​of the user group are aggregated. The entropy formula from information theory can be used to calculate user time-series data. Strong process correlations manifest as high ECE, reflecting batch-to-batch inconsistencies in the production process or inherent instability in the formula itself, allowing feature extraction models to capture production fluctuations that lead to unstable user experiences. Silent Churn Risk Score (SCRS) is based on user behavior decay patterns and is used to predict the probability of ceasing use and not repurchasing within a future period. Prediction can be made using survival analysis models or machine learning classification models. When strongly correlated with product processes, the behavioral characteristics of users with high SCRS represent the general characteristics of the product. By analyzing common user behaviors, global process optimization features related to basic skin feel, long-term satisfaction, and overall competitiveness can be extracted.

[0105] Configure indicators for target products: scenario-based sufficient usage compliance rate, negative feedback behavior pattern specificity index, efficacy perception-behavioral input efficiency ratio, experience consistency entropy, and silent churn risk score, and construct a quality indicator system to provide clear standards for subsequent feature extraction and improvement direction positioning.

[0106] Secondly, using product quality correlation indicators as constraints, several sample user behavior datasets and several sample product quality correlation feature label sets were collected. A graph neural network with an attention mechanism was trained to construct a quality correlation feature extractor. The sample user behavior datasets are raw data extracted from historical user behavior data of the target product and matched with the quality correlation indicators. The sample product quality correlation feature label sets are feature labels calculated based on the quality correlation indicators and corresponding to the sample behavior datasets. The graph neural network with an attention mechanism (GAT) introduces an attention mechanism into the graph neural network, which can automatically focus on user behavior features strongly correlated with quality indicators, improving the accuracy of feature extraction.

[0107] For example, the steps to construct a quality correlation feature extractor using a graph neural network that incorporates an attention mechanism are as follows: Model data: Collect sample user behavior datasets that include user usage scenarios, feedback content, repeat purchase records, etc., and calculate the corresponding sample feature label set based on quality correlation indicators.

[0108] Model Structure: A GAT model is constructed, with user behavior features as the input layer, an attention mechanism introduced in the hidden layer, weights assigned to key features, and quality-related feature labels as the output layer. Based on a multi-layer stacked graph attention layer, each layer first performs a linear transformation on the node features through a shared weight matrix, then calculates the non-normalized attention coefficients between node pairs through the attention mechanism (the dot product of the concatenated features activated by LeakyReLU and the learnable attention vector), then normalizes the attention coefficients of neighboring nodes using softmax, and finally weights and aggregates the neighboring features with the normalized coefficients as weights and outputs a new node representation after non-linear activation.

[0109] Model Training: The sample set was divided into training and validation sets in a 7:3 ratio, with a learning rate of 0.001. After obtaining the node prediction results through forward propagation, the cross-entropy loss with L2 regularization was calculated. Backpropagation and the Adam optimizer were used to iteratively update the weight matrix and attention parameters. Dropout and early stopping strategies were employed during training to prevent overfitting. The training run consisted of 200 iterations, with the loss function being MSE. At the 150th iteration, the average error between the model's predicted labels and the true labels was less than 3%, meeting the convergence condition, and training was stopped. Ultimately, the model learned adaptive weights and inter-node dependencies to improve performance in tasks such as node classification. The converged model parameters were saved, and a quality correlation feature extractor was generated.

[0110] Finally, a quality correlation feature extractor is used to extract product quality correlation features, obtaining a product quality correlation feature set. Based on this feature set analysis, product improvement indicators are determined. The product quality correlation feature set is a collection of features strongly correlated with quality indicators extracted by the quality correlation feature extractor from the latest user behavior data of the target product. Product improvement indicators are specific improvement directions and quantifiable targets determined based on the gap between the quality correlation feature set and the indicator target threshold.

[0111] Input user behavior data of the target product into the quality correlation feature extractor to obtain the product quality correlation feature set; compare with the target threshold and analyze the gap of each feature: sufficient use compliance rate, negative feedback index of rubbing off mud, and experience consistency entropy, and finally determine the product improvement indicators.

[0112] For example, user behavior data (5000 records) of the target product from January to March 2025 were input into the quality correlation feature extractor to obtain the product quality correlation feature set: Seaside sufficient usage compliance rate 55% (target 60%), pilling negative feedback index 25% (target 20%), efficacy-input ratio 1.0 (target 1.2), experience consistency entropy 0.4 (target 0.3), and silent churn score 35% (target 30%). The differences between the features were analyzed: the low seaside sufficient usage compliance rate indicates a need to reduce the thickener ratio due to the product's thick texture; the high pilling negative feedback index requires optimization of the powder modification process; and the high experience consistency entropy requires optimization of homogenization parameters. The final product improvement indicators were determined as follows: seaside sufficient usage compliance rate increased to 60%; pilling negative feedback index decreased to 20%; and experience consistency entropy decreased to 0.3.

[0113] In this embodiment, by configuring deeply correlated indicators of user experience and product quality, and by using graph neural networks for feature extraction, the system can analyze behavioral data within a cycle to identify user behavior signals caused by product process defects, optimize product improvement indicators, clarify process optimization goals, and ensure that the optimization direction is always consistent with user experience.

[0114] S40: With the goal of approximating the product improvement indicators, optimize the production process of the target product to obtain the optimal process adjustment scheme, and within the preset time range, perform iterative optimization of the production process based on user behavior data collected in real time within the optimal process adjustment time period.

[0115] In this embodiment, the optimization objective is to approximate the product improvement indicators as the core objective, while minimizing the consumption of process resources. Production process optimization is the process of predicting the effects of different process schemes and selecting the optimal scheme within the production process parameter adjustment threshold through a deep learning model. The optimal process adjustment scheme is the combination of process parameters that meets the product improvement indicators and has the lowest process resource consumption.

[0116] For the optimization target, a first process adjustment scheme is generated within the process parameter adjustment threshold, the effect of the first process adjustment scheme is predicted, the quality coefficient of the first scheme is calculated, and schemes are randomly generated again for iterative optimization to determine the scheme as the optimal process adjustment scheme.

[0117] Step S40 in the method provided in this application embodiment includes: A first process adjustment scheme is randomly generated within the adjustment threshold of the production process parameters of the target product. A processing result predictor is constructed using deep learning, wherein the input of the processing result predictor is a process adjustment plan, and the output is product quality indicators and process resource consumption; Using the processing result predictor, the product processing result is predicted according to the first process adjustment plan, and the first predicted quality index and the first predicted process resource consumption are output. With the goal of approximating the product improvement index and minimizing process resource consumption, the first scheme quality coefficient is obtained by evaluating the first predicted quality index and the first predicted process resource consumption. Continue to randomly select and iteratively evaluate solutions within the production process parameter adjustment threshold until the preset number of optimization convergence times is reached. Output the process adjustment solution corresponding to the maximum solution quality coefficient during the optimization process as the optimal process adjustment solution.

[0118] In this embodiment, firstly, a first process adjustment scheme is randomly generated within the adjustment threshold of the production process parameters of the target product. The production process parameter adjustment threshold is the adjustable range of each process parameter of the target sunscreen product; the threshold setting ensures that the process adjustment does not exceed the safety and production feasibility boundaries. The process adjustment scheme is a specific combination of process parameters, with each parameter corresponding to a definite value within the threshold. Random generation involves determining parameter values ​​through random sampling within each process parameter threshold and combining them to form the initial process adjustment scheme, ensuring that the scheme covers the parameter space within the threshold.

[0119] Based on the core process parameters of the target product, determine the adjustment threshold, and within each process parameter threshold, determine the parameter value by random sampling, and combine them to generate the first process adjustment scheme.

[0120] For example, taking the core process parameters of the target product (SPF50+ waterproof sunscreen) as an example, the adjustment thresholds are determined as follows: thickener addition ratio: 0.5%-2.0%; titanium dioxide powder modification temperature: 80℃-120℃; homogenization speed: 3000r / min-8000r / min; sunscreen compounding ratio: 15%-25%. A first scheme is randomly generated: thickener 1.2%, powder modification temperature 95℃, homogenization speed 6000r / min, and sunscreen compounding ratio 20%, which is combined into the first process adjustment scheme.

[0121] Secondly, a processing result predictor is constructed using deep learning. The input to the processing result predictor is a process adjustment plan, and the output is product quality indicators and process resource consumption. Specifically, the processing result predictor is a prediction model built on deep learning, with input being combinations of process parameters and output being corresponding product quality indicators. The product quality indicator output predicts the quality performance of the product under the given process plan; the process resource consumption output predicts the production resource input for the given process plan, including material consumption and energy consumption.

[0122] For example, the specific steps for constructing a processing result predictor using a CNN feature extraction and fully connected layer prediction architecture are as follows: Data collection: Collect historical process schemes and corresponding product quality inspection data and resource consumption data. Use historical product quality inspection data and historical resource consumption data as input and divide them into training set and validation set in an 8:2 ratio.

[0123] Model construction: A CNN feature extraction and fully connected layer prediction architecture is adopted. The CNN layer extracts the nonlinear correlation features of process parameters, and the fully connected layer outputs product quality indicators and process resource consumption.

[0124] Model training: The learning rate is set to 0.001, and the iteration is 300 rounds. The loss functions for product quality indicators and process resource consumption are the mean squared error function (MSE) and the mean absolute error function (MAE), respectively. The training is completed when the loss value stabilizes at 0.02, and a processing result predictor is generated.

[0125] Next, using the processing result predictor, the product processing result is predicted based on the first process adjustment plan, outputting the first predicted quality index and the first predicted process resource consumption. The first predicted quality index is the quality performance data output by the predictor for the first process adjustment plan; the first predicted process resource consumption is the resource input data output by the predictor for the first process adjustment plan. The first process adjustment plan is input into the processing result predictor, which outputs the first predicted quality index and the first predicted resource consumption.

[0126] For example, the first process adjustment scheme: thickener 1.2%, powder modification temperature 95℃, homogenization speed 6000r / min, sunscreen 20%, is input into the processing result predictor, and the first predicted quality index is output: the compliance rate of sufficient use at the beach is 58%, the negative feedback index of pilling is 23%, and the entropy of experience consistency is 0.35. None of these meet the improvement indicators: compliance rate ≥60%, pilling index ≤20%, entropy ≤0.3; the first predicted resource consumption is output: unit energy consumption 1.2kWh, raw material cost 8.5 yuan, and production time 45 minutes / batch.

[0127] Simultaneously, with the goals of approximating product improvement indicators and minimizing process resource consumption, a first scheme quality coefficient is obtained based on the first predicted quality indicator and the first predicted process resource consumption. The scheme quality coefficient is a quantitative indicator that comprehensively measures the degree of quality compliance and resource consumption optimization of the process scheme; approximating improvement indicators: the higher the quality compliance score, the closer the scheme is to the improvement indicators; minimizing resource consumption: the higher the resource consumption optimization score, the more resource-efficient the scheme.

[0128] The first scheme's quality coefficient = first predicted quality index × first predicted process resource consumption evaluation, where the quality compliance score is the weighted sum of the compliance rates of each quality index. The first predicted quality index is obtained based on the weighted sum of the compliance rates of each quality index, and the first predicted process resource consumption evaluation is the weighted sum of process resources.

[0129] For example, if the first predicted quality indicator is compliance rate (0.58 / 0.6=0.967), rubbing index (20% / 23%=87%), and experience consistency entropy (0.3 / 0.35=85.7%), then the first predicted quality indicator = 96.7%×0.4+87%×0.4+85.7%×0.2=90.9%; the first predicted process resource consumption evaluation = (benchmark consumption - predicted consumption) / benchmark consumption×100%, where the benchmark consumption is the traditional process consumption: energy consumption 1.5kWh, cost 9 yuan, and the first predicted process resource consumption evaluation = (1.5-1.2) / 1.5+(9-8.5) / 9) / 2≈0.139, then the first scheme quality coefficient = 0.967×0.139≈0.134.

[0130] Finally, within the production process parameter adjustment threshold, random selection and iterative evaluation of solutions continue until the preset number of optimization convergence attempts is reached. The process adjustment solution corresponding to the highest quality coefficient during the optimization process is output as the optimal process adjustment solution. The preset number of optimization convergence attempts is a set upper limit for iterations to ensure the optimization process is completed within a reasonable time. The optimal process adjustment solution is the one with the highest quality coefficient among all iterative solutions, representing the optimal combination of process parameters that satisfies quality standards and resource consumption.

[0131] After iteration, multiple process schemes are generated and coefficients are calculated. The scheme with the largest quality coefficient among all schemes is selected as the optimal process adjustment scheme.

[0132] For example, after 1000 iterations, multiple process schemes are generated and coefficients are calculated. Among them, the 56th scheme parameter has a scheme quality coefficient of 81.3% corresponding to the prediction result, which is the highest among all schemes, and is regarded as the optimal process adjustment scheme.

[0133] In this embodiment, a processing result predictor is constructed using a deep learning model to establish a mapping relationship between process parameter adjustment schemes and predicted product quality and resource consumption. The optimal solution is sought by taking both the approximation of product improvement indicators and the minimization of process resource consumption as optimization objectives to ensure the adaptability of the scheme. Through an iterative optimization mechanism, the optimal process adjustment scheme that achieves the maximum user experience at the lowest cost is intelligently explored in the process parameter space to ensure the optimization effect.

[0134] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, firstly, by clearly defining multi-channel data collection paths, the comprehensiveness and authenticity of data sources are ensured, thus solving the problem of lagging traditional survey data. Secondly, by training a graph neural network with an attention mechanism through predefined improved correlation indicators, features are extracted from the cleaned standard data, and the feature set is intelligently filtered to remove noisy data. Thirdly, by training a BP neural network to predict the basic adjustment cycle and combining it with external market demand forecasts for dynamic compensation, intelligent transformation is achieved, providing a data foundation for subsequent accurate cycle compensation and dynamic optimization.

[0135] Secondly, by constructing an LSTM forecasting model, future demand growth rates and volatility coefficients are predicted to reflect market risks. By comparing the current predicted growth rate and volatility coefficient with historical industry benchmarks, a comprehensive compensation coefficient is obtained to ensure production stability and the effectiveness of adjustments. Ultimately, the optimal process adjustment time cycle is determined, achieving intelligent synchronization between adjustment rhythms and market rhythms.

[0136] Furthermore, by configuring deeply correlated indicators of real user experience and product quality, and by using graph neural networks for feature extraction, we can achieve in-depth quality problem analysis from behavioral data within the cycle, identify behavioral signals generated by users due to product process defects, optimize product improvement indicators, clarify the goals of process optimization, and ensure that the optimization direction is always consistent with the user experience.

[0137] Finally, a processing result predictor is constructed using a deep learning model, establishing a mapping relationship between process parameter adjustment schemes and predicted product quality and resource consumption. Approximating product improvement indicators and minimizing process resource consumption are used as optimization objectives to seek the optimal solution and ensure the suitability of the scheme. Through an iterative optimization mechanism, the optimal process adjustment scheme that achieves the maximum user experience at the lowest cost is intelligently explored in the process parameter space to ensure the optimization effect.

[0138] Example 2, as Figure 2 As shown, based on the same inventive concept as the personal care product process optimization method based on user behavior data provided in Embodiment 1, this embodiment of the invention also provides a personal care product process optimization system based on user behavior data, including: Data acquisition module 11 is used to collect historical user behavior datasets within a historical time range since the target product was launched on the market through multiple channels, perform data preprocessing and product improvement related feature extraction on the historical user behavior datasets, and obtain a historical product improvement related feature set. The time cycle prediction module 12 is used to determine the optimal process adjustment time cycle based on the historical product improvement association feature set and the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range. The quality feature extraction module 13 is used to collect user behavior datasets of the target product within the optimal process adjustment time period through multiple channels, perform data preprocessing and product quality correlation feature extraction, and determine product improvement indicators based on the product quality correlation feature set. The process optimization module 14 is used to optimize the production process of the target product with the goal of approximating the product improvement indicators to obtain the optimal process adjustment scheme, and to iteratively optimize the production process based on user behavior data collected in real time within the preset time range and the optimal process adjustment time period. In one embodiment, the data acquisition module 11 is used for: Configure multiple data collection paths, wherein the data collection paths include at least the brand's official application, e-commerce transaction platform, social media content platform and user research interviews; Using the aforementioned multi-channel data collection path, historical user behavior datasets are collected within a historical time range since the target product was launched into the market. The target product is a sun protection personal care product, and the types of user behavior data include at least behavioral dimensions, scenario dimensions, efficacy experience dimensions, emotional tendency dimensions, and spatiotemporal correlation dimensions.

[0139] Specifically, the historical user behavior dataset undergoes data preprocessing and product improvement-related feature extraction to obtain a historical product improvement-related feature set, including: The historical user behavior dataset is processed for outlier handling, missing value imputation, and multi-source data alignment to generate a historical standard behavior dataset. Predefine product improvement-related indicators for the target product; Using the product improvement correlation index as a constraint, several sample user behavior datasets and several sample product improvement correlation feature label sets are collected, and a graph neural network with an attention mechanism is trained to construct an improvement correlation feature extractor. Using the improved correlation feature extractor, product improvement correlation features are extracted from the historical standard behavior dataset to obtain a historical product improvement correlation feature set.

[0140] The product improvement correlation indicators include at least process sensitivity indicators, multi-factor coupling influence indicators, and process window robustness indicators. The process sensitivity indicators include at least the dosage-spreadability sensitivity coefficient and the environment-film formation stability correlation. The multi-factor coupling influence indicators include at least the superimposed use-pilling coupling index and the high-frequency recoating-tolerance attenuation gradient. The process window robustness indicators include at least the homogeneous shear strength-skin feel consistency bandwidth and the filling accuracy-dosage perception tolerance rate.

[0141] In one embodiment, the time period prediction module 12 is used for: Using the product attributes of the target product as constraints, historical survey records of similar products are retrieved to collect several sample product improvement association feature sets and several sample process adjustment time cycle label sets; Using the aforementioned sample product improvement association feature set as input and the aforementioned sample process adjustment time cycle label set as supervision, a BP neural network is trained until convergence to generate a process adjustment cycle predictor. Using the process adjustment cycle predictor, the predicted process adjustment time cycle is obtained based on the analysis of the historical product improvement correlation feature set; The predicted process adjustment time period is compensated based on the predicted product demand growth rate and the predicted product demand fluctuation coefficient within a preset time range to obtain the optimal process adjustment time period.

[0142] This includes obtaining the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range, including: Obtain the historical demand month-on-month growth rate sequence of the target product within a preset time granularity over a historical time range; By calling up the historical operation records of similar products, and using the time span of the preset time range as a constraint, several sample product improvement related feature sets and several sample demand month-on-month growth rate sequences are collected. The historical demand month-on-month growth rate sequences recorded in the historical time range of different sample product improvement related feature sets and sample demand month-on-month growth rate sequences are statistically analyzed as sample predicted demand month-on-month growth rate sequences, thus obtaining several sample predicted demand month-on-month growth rate sequences. Using the aforementioned sample product improved association feature set and the aforementioned sample demand month-on-month growth rate sequence as input, and using the aforementioned sample predicted demand month-on-month growth rate sequence as supervision, a long short-term memory network is trained until convergence to generate a demand growth prediction plugin. Using the aforementioned demand growth forecasting plugin, based on the historical product improvement related feature set and the historical demand month-on-month growth rate sequence, a predicted demand month-on-month growth rate sequence within a preset time range is obtained. The mean of the predicted demand month-on-month growth rate series is calculated to obtain the predicted product demand growth rate. Calculate the coefficient of variation of the predicted demand month-on-month growth rate series as the predicted product demand volatility coefficient.

[0143] The optimal process adjustment time period is obtained by compensating the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range, including: Access historical operational records of similar products to collect the average historical product demand growth rate and the average historical product demand fluctuation coefficient within a historical time range; The ratio of the predicted product demand growth rate to the average historical product demand growth rate is used as the first period compensation coefficient. The ratio of the average historical product demand fluctuation coefficient to the predicted product demand fluctuation coefficient is used as the second period compensation coefficient. The comprehensive periodic compensation coefficient is calculated by weighting the first periodic compensation coefficient and the second periodic compensation coefficient. The product of the comprehensive cycle compensation coefficient and the predicted process adjustment time period is taken as the optimal process adjustment time period.

[0144] In one embodiment, the quality feature extraction module 13 is used for: Configure product quality related indicators, wherein the product quality related indicators include at least the scenario-based sufficient usage compliance rate, negative feedback behavior pattern specificity index, efficacy perception-behavioral input efficiency ratio, experience consistency entropy, and silent churn risk score. Using the product quality correlation index as a constraint, several sample user behavior datasets and several sample product quality correlation feature label sets are collected, and a graph neural network with an attention mechanism is trained to construct a quality correlation feature extractor. The quality correlation feature extractor is used to extract product quality correlation features to obtain a product quality correlation feature set, and product improvement indicators are determined based on the product quality correlation feature set.

[0145] In one embodiment, the process optimization module 14 is used for: A first process adjustment scheme is randomly generated within the adjustment threshold of the production process parameters of the target product. A processing result predictor is constructed using deep learning, wherein the input of the processing result predictor is a process adjustment plan, and the output is product quality indicators and process resource consumption; Using the processing result predictor, the product processing result is predicted according to the first process adjustment plan, and the first predicted quality index and the first predicted process resource consumption are output. With the goal of approximating the product improvement index and minimizing process resource consumption, the first scheme quality coefficient is obtained by evaluating the first predicted quality index and the first predicted process resource consumption. Continue to randomly select and iteratively evaluate solutions within the production process parameter adjustment threshold until the preset number of optimization convergence times is reached. Output the process adjustment solution corresponding to the maximum solution quality coefficient during the optimization process as the optimal process adjustment solution.

[0146] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, the data acquisition module 11 first clarifies the data acquisition path through multiple channels, ensuring the comprehensiveness and authenticity of the data sources and solving the problem of lagging traditional survey data. Then, a graph neural network with an attention mechanism is trained using predefined improved correlation indicators to extract features from the cleaned standard data, intelligently filtering feature sets and eliminating noisy data. Finally, a backpropagation (BP) neural network is trained to predict the basic adjustment cycle, and dynamic compensation is performed in conjunction with external market demand forecasts, achieving intelligent transformation and providing a data foundation for subsequent precise cycle compensation and dynamic optimization.

[0147] Secondly, an LSTM forecasting model is constructed using the time cycle forecasting module 12 to predict future demand growth rates and their volatility coefficients, reflecting market risks. By comparing the current predicted growth rate and volatility coefficient with historical industry benchmarks, a comprehensive compensation coefficient is obtained to ensure production stability and the effectiveness of adjustments. Ultimately, the optimal process adjustment time cycle is determined, achieving intelligent synchronization between the adjustment rhythm and market rhythms.

[0148] Furthermore, through the quality feature extraction module 13, the correlation indicators between the user's real experience and product quality are configured, and the graph neural network is used again for feature extraction. This enables the analysis of in-depth quality problems from behavioral data within the cycle, identifies the behavioral signals generated by users due to product process defects, optimizes product improvement indicators, clarifies the goals of process optimization, and ensures that the optimization direction is always consistent with the user experience.

[0149] Finally, through the process optimization module 14, a processing result predictor is constructed using a deep learning model to establish a mapping relationship between process parameter adjustment schemes and predicted product quality and resource consumption. Approximating product improvement indicators and minimizing process resource consumption are used as optimization objectives to seek the optimal solution and ensure the adaptability of the scheme. Through an iterative optimization mechanism, the optimal process adjustment scheme that achieves the maximum user experience at the lowest cost is intelligently explored in the process parameter space to ensure the optimization effect.

[0150] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0151] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0152] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for optimizing the manufacturing process of personal care products based on user behavior data, characterized in that, The methods include: By collecting historical user behavior datasets from the target product within a historical time range since its market launch through multiple channels, and performing data preprocessing and product improvement-related feature extraction on the historical user behavior datasets, a historical product improvement-related feature set is obtained. The optimal process adjustment time period is determined based on the historical product improvement correlation feature set and the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range. By collecting user behavior datasets of the target product within the optimal process adjustment time period through multiple channels, and performing data preprocessing and product quality correlation feature extraction, product improvement indicators are determined based on the product quality correlation feature set. With the goal of approximating the product improvement indicators, the production process of the target product is optimized to obtain the optimal process adjustment scheme. Within the preset time range, the production process is iteratively optimized based on user behavior data collected in real time within the optimal process adjustment time period.

2. The method for optimizing the process of personal care products based on user behavior data according to claim 1, characterized in that, Historical user behavior datasets were collected through multiple channels over a historical period since the target product was launched on the market, including: Configure multiple data collection paths, wherein the data collection paths include at least the brand's official application, e-commerce transaction platform, social media content platform and user research interviews; Using the aforementioned multi-channel data collection path, historical user behavior datasets are collected within a historical time range since the target product was launched into the market. The target product is a sun protection personal care product, and the types of user behavior data include at least behavioral dimensions, scenario dimensions, efficacy experience dimensions, emotional tendency dimensions, and spatiotemporal correlation dimensions.

3. The method for optimizing the process of personal care products based on user behavior data according to claim 1, characterized in that, The historical user behavior dataset is preprocessed and product improvement-related feature extraction is performed to obtain a historical product improvement-related feature set, including: The historical user behavior dataset is processed for outlier handling, missing value imputation, and multi-source data alignment to generate a historical standard behavior dataset. Predefine product improvement-related indicators for the target product; Using the product improvement correlation index as a constraint, several sample user behavior datasets and several sample product improvement correlation feature label sets are collected, and a graph neural network with an attention mechanism is trained to construct an improvement correlation feature extractor. Using the improved correlation feature extractor, product improvement correlation features are extracted from the historical standard behavior dataset to obtain a historical product improvement correlation feature set.

4. The personal care product process optimization method based on user behavior data according to claim 3, characterized in that, The types of product improvement correlation indicators include at least process sensitivity indicators, multi-factor coupling influence indicators, and process window robustness indicators. Among them, the process sensitivity indicators include at least the dosage-spreadability sensitivity coefficient and the environment-film formation stability correlation degree. The multi-factor coupling influence indicators include at least the superimposed use-pilling coupling index and the high-frequency recoating-tolerance attenuation gradient. The process window robustness indicators include at least the homogeneous shear strength-skin feel consistency bandwidth and the filling accuracy-dosage perception error rate.

5. The method for optimizing the process of personal care products based on user behavior data according to claim 1, characterized in that, The optimal process adjustment time period is determined based on the historical product improvement correlation feature set and the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range, including: Using the product attributes of the target product as constraints, historical survey records of similar products are retrieved to collect several sample product improvement association feature sets and several sample process adjustment time cycle label sets; Using the aforementioned sample product improvement association feature set as input and the aforementioned sample process adjustment time cycle label set as supervision, a BP neural network is trained until convergence to generate a process adjustment cycle predictor. Using the process adjustment cycle predictor, the predicted process adjustment time cycle is obtained based on the analysis of the historical product improvement correlation feature set; The predicted process adjustment time period is compensated based on the predicted product demand growth rate and the predicted product demand fluctuation coefficient within a preset time range to obtain the optimal process adjustment time period.

6. The method for optimizing the process of personal care products based on user behavior data according to claim 5, characterized in that, Obtain the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range, including: Obtain the historical demand month-on-month growth rate sequence of the target product within a preset time granularity over a historical time range; By calling up the historical operation records of similar products, and using the time span of the preset time range as a constraint, several sample product improvement related feature sets and several sample demand month-on-month growth rate sequences are collected. The historical demand month-on-month growth rate sequences recorded in the historical time range of different sample product improvement related feature sets and sample demand month-on-month growth rate sequences are statistically analyzed as sample predicted demand month-on-month growth rate sequences, thus obtaining several sample predicted demand month-on-month growth rate sequences. Using the aforementioned sample product improved association feature set and the aforementioned sample demand month-on-month growth rate sequence as input, and using the aforementioned sample predicted demand month-on-month growth rate sequence as supervision, a long short-term memory network is trained until convergence to generate a demand growth prediction plugin. Using the aforementioned demand growth forecasting plugin, based on the historical product improvement related feature set and the historical demand month-on-month growth rate sequence, a predicted demand month-on-month growth rate sequence within a preset time range is obtained. The mean of the predicted demand month-on-month growth rate series is calculated to obtain the predicted product demand growth rate. Calculate the coefficient of variation of the predicted demand month-on-month growth rate series as the predicted product demand volatility coefficient.

7. The method for optimizing the process of personal care products based on user behavior data according to claim 6, characterized in that, The predicted process adjustment time period is compensated based on the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range to obtain the optimal process adjustment time period, including: Access historical operational records of similar products to collect the average historical product demand growth rate and the average historical product demand fluctuation coefficient within a historical time range; The ratio of the predicted product demand growth rate to the average historical product demand growth rate is used as the first period compensation coefficient. The ratio of the average historical product demand fluctuation coefficient to the predicted product demand fluctuation coefficient is used as the second period compensation coefficient. The comprehensive periodic compensation coefficient is calculated by weighting the first periodic compensation coefficient and the second periodic compensation coefficient. The product of the comprehensive cycle compensation coefficient and the predicted process adjustment time period is taken as the optimal process adjustment time period.

8. The method for optimizing the process of personal care products based on user behavior data according to claim 1, characterized in that, Data preprocessing and product quality correlation feature extraction are performed. Based on the product quality correlation feature set, product improvement indicators are determined, including: Configure product quality related indicators, wherein the product quality related indicators include at least the scenario-based sufficient usage compliance rate, negative feedback behavior pattern specificity index, efficacy perception-behavioral input efficiency ratio, experience consistency entropy, and silent churn risk score. Using the product quality correlation index as a constraint, several sample user behavior datasets and several sample product quality correlation feature label sets are collected, and a graph neural network with an attention mechanism is trained to construct a quality correlation feature extractor. The quality correlation feature extractor is used to extract product quality correlation features to obtain a product quality correlation feature set, and product improvement indicators are determined based on the product quality correlation feature set.

9. The method for optimizing the process of personal care products based on user behavior data according to claim 1, characterized in that, With the goal of approximating the product improvement indicators, the production process of the target product is optimized to obtain the optimal process adjustment scheme, including: A first process adjustment scheme is randomly generated within the adjustment threshold of the production process parameters of the target product. A processing result predictor is constructed using deep learning, wherein the input of the processing result predictor is a process adjustment plan, and the output is product quality indicators and process resource consumption; Using the processing result predictor, the product processing result is predicted according to the first process adjustment plan, and the first predicted quality index and the first predicted process resource consumption are output. With the goal of approximating the product improvement index and minimizing process resource consumption, the first scheme quality coefficient is obtained by evaluating the first predicted quality index and the first predicted process resource consumption. Continue to randomly select and iteratively evaluate solutions within the production process parameter adjustment threshold until the preset number of optimization convergence times is reached. Output the process adjustment solution corresponding to the maximum solution quality coefficient during the optimization process as the optimal process adjustment solution.

10. A personal care product process optimization system based on user behavior data, characterized in that, The system is used to implement the personal care product process optimization method based on user behavior data according to any one of claims 1-9, the system comprising: The data acquisition module is used to collect historical user behavior datasets within a historical time range since the target product was launched on the market through multiple channels, perform data preprocessing and product improvement related feature extraction on the historical user behavior datasets, and obtain a historical product improvement related feature set. The time cycle prediction module is used to determine the optimal process adjustment time cycle based on the historical product improvement correlation feature set and the predicted product demand growth rate and predicted product demand fluctuation coefficient within a preset time range. The quality feature extraction module is used to collect user behavior datasets of the target product within the optimal process adjustment time period through multiple channels, perform data preprocessing and product quality correlation feature extraction, and determine product improvement indicators based on the product quality correlation feature set. The process optimization module is used to optimize the production process of the target product with the goal of approximating the product improvement indicators, obtain the optimal process adjustment scheme, and perform iterative optimization of the production process based on user behavior data collected in real time within the preset time range and the optimal process adjustment time period.

Images