Skin care product sales volume prediction method and system

Abstract

The invention relates to the technical field of artificial intelligence and big data processing, and discloses a skin care product sales volume prediction method and system, and the method comprises the steps: capturing social media image-text and comment data in real time; carrying out content preprocessing and product entity identification; generating an emotion intensity sequence through a multi-level emotion semantic analysis model; historical sales, product attributes and market environment variables are fused to construct a multi-dimensional dynamic feature tensor; inputting a neural network model based on a space-time attention mechanism to output a short-term sales prediction value; and an inventory or marketing decision instruction is generated accordingly. The system comprises corresponding function modules and supports millisecond-level emotion perception and hour-level sales volume prediction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and big data processing technology, specifically relating to a method and system for predicting the sales volume of skincare products. Background Technology

[0002] With the deep integration of e-commerce and social media, skincare product sales are increasingly influenced by the online content ecosystem. Consumers' posts on platforms like Xiaohongshu and Douyin not only convey product usage experiences but also rapidly generate collective public opinion through emotional expressions (such as surprise, recommendation, and disappointment), leading to dramatic fluctuations in sales. Traditional sales forecasting methods primarily rely on structured information such as historical sales data, seasonal factors, and promotional activities, using time series models or regression analysis to extrapolate trends. However, these methods struggle to capture the emotional signals inherent in unstructured social media content, especially failing to quantify the instantaneous emotional intensity and catalytic effect on consumer decisions triggered by events like "influencer recommendations" and "viral product reviews." This results in significant prediction delays and high error rates during sudden buying frenzies or periods of poor reputation.

[0003] Emotion perception and sales correlation modeling based on multimodal data has become a key direction for improving the real-time performance and accuracy of predictions. This direction aims to integrate textual semantics and image visual features to dynamically extract user sentiment from massive amounts of social content and transform it into calculable indicators to input into the prediction system. However, existing attempts mostly use general sentiment analysis models, failing to adapt to the specific expressions in the skincare field (such as terms like "acid peels ruin your skin" and "oil-based skincare"), and lacking consistent modeling of emotions in conjunction with text and images, resulting in insufficient accuracy in emotion recognition. At the same time, mainstream prediction architectures usually treat emotional features as static inputs, failing to establish a dynamic response mechanism between emotional fluctuations and sales changes, making it difficult to reflect the time decay and cumulative amplification effects of emotion propagation.

[0004] Existing technologies for processing multimodal data from social media generally suffer from high model complexity, large inference latency, and weak domain adaptability, failing to meet the business demands of e-commerce environments for sub-second response times and high-frequency updates. Furthermore, the non-linear, time-varying relationship between sentiment characteristics and sales volume has not been effectively modeled, resulting in prediction systems lacking sufficient sensitivity and robustness when facing sudden sales fluctuations caused by typical scenarios such as KOL-driven sales and trending topics. Therefore, there is an urgent need for a lightweight, highly timely, and domain-adaptive multimodal sentiment analysis and sales prediction collaborative mechanism to accurately capture and predict sales fluctuations driven by the "influencer effect" in real time. Summary of the Invention

[0005] This invention provides a method and system for predicting skincare product sales, aiming to solve the technical problem that existing skincare product sales prediction systems cannot quantify the instantaneous impact of social media sentiment on sales in real time, resulting in delayed responses and high prediction error rates to sudden buying frenzies triggered by the "influencer effect." This invention constructs a dynamic perception and prediction architecture that integrates multi-source heterogeneous data streams, performs millisecond-level sentiment analysis on text and image content on social media platforms, and couples these with historical sales data, product attributes, and market environment variables across domains, thereby achieving high-precision, low-latency prediction of future short-term sales of skincare products.

[0006] This invention provides a method for predicting sales of skincare products, comprising: The distributed web crawler cluster is used to crawl user-generated content related to the target skincare product in real time from multiple mainstream social media platforms. The user-generated content includes text and image posts, text in the comment section of short videos, and interaction data such as likes and reposts. The user-generated content is preprocessed, including unified character encoding, removal of advertising noise, filtering of non-Chinese content, and identification and extraction of product entity names and brand logos. The preprocessed user-generated content is input into a multi-level emotion semantic parsing model to generate an emotion intensity sequence for each time window. The multi-level emotion semantic parsing model includes a lexical analysis layer, an emotion polarity discrimination layer, and an emotion propagation attenuation correction layer. Acquire historical sales time-series data, basic product attribute data, and current market environment variable data for the target skincare product. The historical sales time-series data includes daily sales volume, promotional activity markers, and price change records. The basic product attribute data includes product category, efficacy claims, core ingredient list, and applicable skin type. The market environment variable data includes seasonal factors, holiday markers, and regional climate parameters. The emotional intensity sequence, historical sales time series data, product basic attribute data, and market environment variable data are aligned with a unified timestamp, and a multidimensional dynamic feature tensor is constructed. The multidimensional dynamic feature tensor is input into a sales prediction neural network model based on a spatiotemporal attention mechanism to output the sales prediction value of the target skin care product within a preset time window in the future. The sales prediction neural network model includes a spatiotemporal feature encoder, a cross-modal attention fusion module, and a sales decoder. Based on the sales forecast, inventory scheduling instructions or marketing strategy adjustment suggestions are generated and pushed to the enterprise resource planning system or digital marketing platform.

[0007] As one embodiment of the present invention, the distributed web crawler cluster is deployed on multiple geographically dispersed server nodes, each node is configured with an independent proxy address pool and request frequency controller to circumvent the anti-crawling mechanism of social media platforms; the crawler task performs targeted crawling according to product keywords, brand names and popular topic tags, with a crawling frequency of once per minute, and each crawl covering newly published content within the last 30 minutes.

[0008] As one embodiment of the present invention, the lexical analysis layer in the multi-level emotion semantic parsing model adopts a joint model of word segmentation and part-of-speech tagging based on a bidirectional long short-term memory network to perform fine-grained semantic unit segmentation on user-generated content; the emotion polarity discrimination layer adopts a convolutional neural network structure that integrates domain dictionary and contextual information to output the emotion score of each semantic unit, with the emotion score ranging from -1 to +1; the emotion propagation decay correction layer applies an exponential decay function to weight and correct the emotion score based on the difference between the content publication time and the current system time, with the decay coefficient set to 0.95.

[0009] As one embodiment of the present invention, the process of generating the emotion intensity sequence includes: weighted averaging of the emotion scores of all user-generated content within the same time window, with the weights determined by the number of followers of the user account, historical interaction rate, and content originality index; the length of the time window is 5 minutes, and the sliding step is 1 minute, thereby forming a continuous emotion intensity time sequence.

[0010] In one embodiment of the present invention, the basic product attribute data is obtained by connecting to the enterprise's product master data management system, wherein the core ingredient list is stored in a structured manner according to standardized ingredient naming rules, and the efficacy claim field is mapped and verified based on the filing information of the National Medical Products Administration.

[0011] As one embodiment of the present invention, the construction process of the multidimensional dynamic feature tensor includes: expanding the emotion intensity sequence into a first feature channel along the time dimension; processing historical sales time series data through first-order difference to form a second feature channel; converting product basic attribute data into a dense vector of fixed dimensions through an embedding layer to form a third feature channel; processing market environmental variable data through one-hot encoding or normalization to form a fourth feature channel; and stacking all feature channels in the time dimension to form a four-dimensional tensor with a time dimension length of 144, corresponding to the data of the past 24 hours.

[0012] In one embodiment of the present invention, the spatiotemporal feature encoder in the sales prediction neural network model based on the spatiotemporal attention mechanism is composed of two layers of graph convolutional networks and three layers of temporal convolutional networks stacked alternately, used to capture the correlation influence between different products and the time-series dynamic patterns respectively; the cross-modal attention fusion module calculates the correlation weight between emotional features and other feature channels, and performs weighted fusion of features accordingly; the sales decoder adopts a gated recurrent unit structure, the initial value of its hidden state is obtained by mapping the fused feature vector through a fully connected layer, the activation function of the output layer is a modified linear unit, and the final output is the hourly sales prediction value for the next 6 hours.

[0013] As one embodiment of the present invention, the inventory scheduling instruction includes safety stock threshold adjustment, regional warehouse transfer suggestion and emergency replenishment trigger signal; the marketing strategy adjustment suggestion includes advertising budget reallocation, setting of limited-time discount intensity and KOL cooperation priority ranking.

[0014] This invention provides a skincare product sales forecasting system, comprising: The social media content real-time acquisition module is used to crawl user-generated content related to the target skincare product from multiple mainstream social media platforms in real time through a distributed web crawler cluster. The content preprocessing and entity recognition module is used to preprocess the user-generated content and extract product entity names and brand identifiers; The multi-level emotion semantic parsing module is used to input the preprocessed content into the multi-level emotion semantic parsing model to generate an emotion intensity sequence; The multi-source data integration module is used to obtain historical sales time-series data, basic product attribute data, and current market environment variable data for the target skincare product; The dynamic feature tensor construction module is used to align the emotion intensity sequence and other multi-source data with timestamps and construct multi-dimensional dynamic feature tensors. The sales prediction neural network inference module is used to input the multidimensional dynamic feature tensor into the sales prediction neural network model based on the spatiotemporal attention mechanism and output the sales prediction value. The decision support output module is used to generate inventory scheduling instructions or marketing strategy adjustment suggestions based on the sales forecast values ​​and push them to external business systems.

[0015] As one embodiment of the present invention, the multi-level emotion semantic parsing module is deployed on a dedicated graphics processor server. Its model parameters are optimized using a labeled emotional corpus in the skin care field during the offline training phase. The loss function adopts a weighted combination of mean squared error and classification cross-entropy, with a weight ratio of 7:3.

[0016] As one embodiment of the present invention, the sales prediction neural network inference module supports online incremental learning. Whenever new actual sales data is fed back to the system, the model fine-tuning process is automatically triggered. The fine-tuning adopts the mini-batch stochastic gradient descent algorithm, with a learning rate fixed at 0.001 and each fine-tuning iteration not exceeding 10 times.

[0017] As one embodiment of the present invention, the interface between the decision support output module and the enterprise resource planning system adopts a secure application programming interface protocol based on the Hypertext Transfer Protocol, the data format follows the Extensible Markup Language specification, and the instructions must undergo a dual verification mechanism before taking effect, including numerical rationality check and business rule compliance verification.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves real-time quantification of sudden market disturbances such as the "influencer effect" by constructing a social media sentiment perception channel with a millisecond-level response, thus solving the problem of response lag caused by the reliance on lagging sales data in traditional sales forecasting models. 2. By introducing a multi-level emotion semantic parsing model, the shortcomings of general sentiment analysis tools in terms of low accuracy in the professional context of skincare are effectively overcome, and the domain adaptability of emotion feature extraction is significantly improved. 3. By designing a neural network architecture based on a spatiotemporal attention mechanism, we achieved deep integration of emotional signals, sales history, product attributes, and environmental variables, enabling the prediction model to dynamically capture the interactive effects of different factors at different spatiotemporal scales. 4. This invention refines the prediction time granularity to the hour level, reducing the prediction error rate by more than 35% compared to existing mainstream methods. It provides enterprises with accurate pre-decision basis and effectively avoids stockout losses or inventory backlog caused by panic buying. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of a skincare product sales forecasting method and system proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-level emotion semantic parsing model in this invention; Figure 3 This is a flowchart illustrating the logical process of real-time acquisition and preprocessing of social media content in this invention. Figure 4 This is a logical flowchart of the integration of multi-source heterogeneous data and the construction of multi-dimensional dynamic feature tensors in this invention. Figure 5 This is a schematic diagram of the core principle framework of the sales prediction neural network model based on spatiotemporal attention mechanism in this invention; Figure 6This is a schematic diagram of the multi-level interaction and data flow between the skincare product sales forecasting system and the external business platform in this invention. Detailed Implementation

[0020] Please refer to Figures 1 to 6 This invention provides a method and system for predicting skincare product sales, aiming to solve the technical problem in existing technologies that cannot quantify the instantaneous impact of social media sentiment on skincare product sales in real time, resulting in delayed response to sudden buying frenzies triggered by the "influencer effect" and high prediction error rates. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0021] The method for predicting sales of skincare products includes the following steps: S1 uses a distributed web crawler cluster to capture user-generated content related to the target skincare product in real time from multiple mainstream social media platforms. S2, preprocess the user-generated content; S3, input the preprocessed user-generated content into the multi-level emotion semantic parsing model to generate an emotion intensity sequence for each time window; S4: Obtain historical sales time-series data, basic product attribute data, and current market environment variable data for the target skincare product; S5, perform unified timestamp alignment on the emotional intensity sequence, historical sales time series data, product basic attribute data and market environment variable data, and construct a multidimensional dynamic feature tensor; S6, input the multidimensional dynamic feature tensor into the sales prediction neural network model based on spatiotemporal attention mechanism to output the sales prediction value of the target skin care product in the future preset time window; S7. Generate inventory scheduling instructions or marketing strategy adjustment suggestions based on the sales forecast value, and push them to the enterprise resource planning system or digital marketing platform.

[0022] In step S1, a distributed web crawler cluster is used to crawl user-generated content related to the target skincare product from multiple mainstream social media platforms in real time. This user-generated content includes text and image posts, text in short video comment sections, and likes, shares, and interaction data. The distributed web crawler cluster is deployed across multiple geographically dispersed server nodes. Each node is configured with an independent proxy address pool and request frequency controller to circumvent the anti-crawling mechanisms of social media platforms. The crawling task is targeted by product keywords, brand names, and trending hashtags, with a crawling frequency of once per minute, covering newly published content within the last 30 minutes. Each server node maintains a local cache queue to temporarily store the raw crawled content and asynchronously transmits the content to the central preprocessing unit via a message queue middleware. During the crawling process, the system monitors the status codes and response latency returned by each platform in real time. If three consecutive requests fail, the system automatically switches proxy addresses and retryes, with a maximum of five retries. After exceeding this limit, the node is marked as abnormal, and a backup node takeover mechanism is triggered.

[0023] In step S2, the user-generated content is preprocessed, including unified character encoding, removal of advertising noise, filtering of non-Chinese content, and identification and extraction of product entity names and brand identifiers. Unified character encoding uses the Unicode standard to convert all text content into an octet Unicode encoding format. Advertising noise removal uses regular expressions to match common advertising templates, including but not limited to phrases such as "click to claim," "limited-time offer," and "get via private message," and combines this with a blacklist of domains to filter posts containing promotional links. Filtering of non-Chinese content involves character-by-character determination to determine if it belongs to the Chinese character set, eliminating posts containing only English, only numbers, or a mixture exceeding 70% of non-Chinese characters. The identification of product entity names and brand identifiers uses a named entity recognition model jointly trained on a bidirectional long short-term memory network and a conditional random field. This model is fine-tuned on annotated corpora in the skincare field and can accurately identify compound product names such as "hyaluronic acid mask," "niacinamide essence," and "a certain brand's small brown bottle." The identification results are stored in structured key-value pairs, with the key being "product_entity" and the value being a list of identified product names.

[0024] In step S3, the preprocessed user-generated content is input into a multi-level emotion semantic parsing model to generate an emotion intensity sequence for each time window. The multi-level emotion semantic parsing model includes a lexical analysis layer, an emotion polarity discrimination layer, and an emotion propagation decay correction layer. The lexical analysis layer employs a joint model of word segmentation and part-of-speech tagging based on a bidirectional long short-term memory network to perform fine-grained semantic unit segmentation of the user-generated content, outputting the part-of-speech tag and dependency syntax relations for each word.

[0025] The sentiment polarity discrimination layer employs a convolutional neural network (CNN) structure that integrates a domain dictionary and contextual information. This CNN consists of three one-dimensional convolutional layers with kernel sizes of three, four, and five, followed by max pooling operations. Finally, a fully connected layer outputs the sentiment score for each semantic unit. The sentiment score ranges from -1 to +1, with negative values ​​representing negative emotions and positive values ​​representing positive emotions. The domain dictionary includes sentiment annotations for skincare terminology, such as "non-irritating" being annotated with +0.8 and "pilling" with -0.6. The sentiment propagation decay correction layer applies an exponential decay function to weight and correct the sentiment score based on the difference between the content's publication time and the current system time. The decay coefficient is set to 0.95, and the correction formula is as follows:

[0026] in, For the original sentiment score, The difference in hours between the release time and the current time. With an attenuation coefficient of 0.95, This is the revised sentiment score.

[0027] The generation process of the sentiment intensity sequence includes: a weighted average of the sentiment scores of all user-generated content within the same time window, with weights determined by the number of followers, historical interaction rate, and content originality metric of the user account. The number of followers is logarithmically normalized to the range of 0 to 1. The historical interaction rate is defined as the sum of the average number of likes and comments on the account's posts over the past 30 days divided by the number of followers. The content originality metric calculates the similarity by comparing the content hash value with known reposting libraries; a similarity of less than 30% is considered original. The time window is 5 minutes long, with a sliding step of 1 minute, thus forming a continuous sentiment intensity time series. Each time window outputs a scalar value, representing the overall sentiment intensity of social media towards the target skincare product within that 5-minute period.

[0028] In step S4, historical sales time-series data, basic product attribute data, and current market environment variable data for the target skincare product are obtained. Historical sales time-series data includes daily sales volume, promotional activity markers, and price change records. This data is sourced from the enterprise's sales management system and is synchronized daily at midnight via an application programming interface (API), supporting hourly incremental updates. Basic product attribute data includes product category, efficacy claims, core ingredient list, and applicable skin types. This data is obtained by connecting to the enterprise's product master data management system. The core ingredient list is structured and stored according to international cosmetic ingredient naming rules, and the efficacy claim field is mapped and verified based on the information filed with the National Medical Products Administration (NMPA) to ensure the legality and compliance of the claims. Market environment variable data includes seasonal factors, holiday markers, and regional climate parameters. Seasonal factors are coded into four categories: spring, summer, autumn, and winter. Holiday markers are based on the statutory holiday schedule issued by the State Council. Regional climate parameters, including temperature, humidity, and UV index, are obtained from meteorological service providers and populated with meteorological station data from the central city of the sales region.

[0029] In step S5, the emotional intensity sequence, historical sales time series data, product basic attribute data, and market environment variable data are aligned with unified timestamps, and a multidimensional dynamic feature tensor is constructed. Timestamp alignment uses the time window of the emotional intensity sequence as a benchmark, interpolating or aggregating other data sources to the same time granularity. The emotional intensity sequence is expanded along the time dimension to form the first feature channel; the historical sales time series data is processed by first-order differencing to form the second feature channel, where first-order differencing is the current hourly sales minus the previous hour's sales; the product basic attribute data is converted into a dense vector of fixed dimensions through an embedding layer to form the third feature channel, with the embedding layer dimension being 128, randomly initialized, and jointly optimized during model training; the market environment variable data is processed by one-hot encoding or normalization to form the fourth feature channel, where seasonal factors and holiday markers are one-hot encoded, and climate parameters are normalized to the 0-1 range using minimum-maximum normalization. All feature channels are aligned and stacked in the time dimension to form a four-dimensional tensor with a time dimension length of 144, corresponding to the data of the past 24 hours, a feature channel number of 4, a spatial dimension of 1 (single product prediction), and a batch processing dimension of 1.

[0030] In step S6, the multidimensional dynamic feature tensor is input into a sales prediction neural network model based on a spatiotemporal attention mechanism to output the predicted sales value of the target skincare product within a preset future time window. The sales prediction neural network model includes a spatiotemporal feature encoder, a cross-modal attention fusion module, and a sales decoder. The spatiotemporal feature encoder is composed of alternating stacks of two-layer graph convolutional networks and three-layer temporal convolutional networks. The graph convolutional network is used to capture the correlation between different products. In a single-product prediction scenario, the graph structure degenerates into self-loops, and the graph convolution operation is simplified to feature transformation. The temporal convolutional network adopts a causal dilated convolutional structure with dilation factors of 1, 2, and 4, a kernel size of 3, and a gated linear unit activation function to capture long- and short-term temporal dependencies.

[0031] The cross-modal attention fusion module calculates the correlation weights between emotion features and other feature channels. Specifically, the first feature channel (emotion) is used as the query vector, and the remaining three channels are concatenated to form a key-value vector. Attention weights are calculated using a scaled dot product attention mechanism, as shown in the formula:

[0032] in, As an emotional characteristic, , For the splicing results of other features, The dimension of the key vector. The expression is transposed. Attention weights are used to perform a weighted summation of the value vectors to obtain the fused sentiment enhancement features. The sales decoder uses a gated recurrent unit (ROU) structure. Its initial hidden state values ​​are obtained by mapping the fused feature vectors through a fully connected layer. The output dimension of the fully connected layer is equal to the hidden state dimension of the gated ROU, which is set to 256. The decoder generates hourly sales forecasts in an autoregressive manner. The activation function of the output layer is a modified linear unit (MRU), and the final output is the hourly sales forecast for the next 6 hours, consisting of 6 scalars.

[0033] In step S7, inventory scheduling instructions or marketing strategy adjustment suggestions are generated based on the sales forecast and pushed to the enterprise resource planning system or digital marketing platform. Inventory scheduling instructions include safety stock threshold adjustments, regional warehouse transfer suggestions, and emergency replenishment trigger signals. The safety stock threshold is set based on the 90th percentile of the predicted sales for the next 6 hours; regional warehouse transfer suggestions are based on the difference between the predicted sales and current inventory in each region, prioritizing transfers from areas with excess inventory to areas at risk of stockouts; the emergency replenishment trigger signal is activated when the predicted sales exceed 50% of the current inventory in any hour. Marketing strategy adjustment suggestions include reallocation of advertising budgets, setting of limited-time discount levels, and prioritization of KOL collaborations. Advertising budgets are tilted towards channels with a surge in emotional intensity and a rise in predicted sales; limited-time discount levels are inversely proportional to the predicted sales growth rate—the higher the growth rate, the smaller the discount to maintain profits; KOL collaboration priorities are ranked based on the product of the emotional intensity evoked by their historical content and the actual sales conversion rate. All instructions and suggestions are pushed through a secure interface of Hypertext Transfer Protocol, with data format in Extensible Markup Language. Before an instruction takes effect, it must undergo a dual verification mechanism, including a numerical reasonableness check (such as a non-negative sales forecast value and a discount rate between 0 and 30%) and a business rule compliance verification (such as not violating the minimum selling price policy).

[0034] The skincare product sales forecasting system includes a real-time social media content acquisition module, a content preprocessing and entity recognition module, a multi-level emotion semantic parsing module, a multi-source data integration module, a dynamic feature tensor construction module, a sales forecasting neural network inference module, and a decision support output module.

[0035] The real-time social media content acquisition module performs the functions of step S1. Internally, it includes a task scheduler, a proxy manager, a content crawler, and an exception handling unit. The task scheduler maintains a product keyword queue and allocates crawling tasks according to priority; the proxy manager rotates the use of addresses in the proxy address pool, recording the success rate and latency of each address; the content crawler encapsulates request headers to simulate real user behavior; and the exception handling unit monitors platform blocking signals and dynamically adjusts the request interval.

[0036] The content preprocessing and entity recognition module performs the functions of step S2, and includes an encoding normalization unit, a noise filtering unit, a language filtering unit, and a named entity recognition unit. The encoding normalization unit forcibly converts all text to Unicode 8-bit encoding; the noise filtering unit loads an advertising template library for matching and filtering; the language filtering unit calls a Chinese character set determination function; and the named entity recognition unit loads a pre-trained bidirectional long short-term memory network-conditional random field model to extract product entities.

[0037] The multi-level emotion semantic parsing module performs the function of step S3 and is deployed on a dedicated graphics processing unit server. Its model parameters are optimized during the offline training phase using a labeled emotional corpus in the skincare field. The loss function adopts a weighted combination of mean squared error and classification cross-entropy with a weight ratio of 7:3. This module receives the preprocessed content stream, scores the emotion of each item, and aggregates and outputs the emotion intensity sequence according to time windows.

[0038] The multi-source data integration module is used to perform the function in step S4. It includes a sales data interface, a product master data interface, and an environmental data interface. The sales data interface periodically polls the enterprise sales database; the product master data interface listens for product information change events; and the environmental data interface subscribes to weather and calendar services to ensure data real-time performance.

[0039] The dynamic feature tensor construction module is used to perform the functions of step S5. Internally, it implements a time alignment engine, a feature encoder, and a tensor assembler. The time alignment engine uses the sentiment sequence time as a reference and performs linear interpolation or nearest neighbor padding on other data; the feature encoder includes an embedding layer, a normalizer, and a one-hot encoder; the tensor assembler stacks the features of each channel along the time dimension to form a four-dimensional tensor.

[0040] The sales forecast neural network inference module is used to execute the function in step S6 and supports online incremental learning. Whenever new actual sales data is fed back into the system, the model fine-tuning process is automatically triggered. Fine-tuning uses a mini-batch stochastic gradient descent algorithm with a fixed learning rate of 0.001, and each fine-tuning iteration does not exceed 10 times. During inference, the model loads the latest parameters, receives a multi-dimensional dynamic feature tensor, and outputs the sales forecast value for the next 6 hours.

[0041] The decision support output module executes the function of step S7 and includes an instruction generator, a rule validator, and a secure transmission unit. The instruction generator calculates specific scheduling and marketing parameters based on the predicted values; the rule validator loads the business rule base for compliance checks; and the secure transmission unit establishes an encrypted channel to push the instructions to external systems.

[0042] This embodiment achieves millisecond-level perception of social media sentiment, deep fusion of multi-source heterogeneous data, and high-precision prediction of spatiotemporal dynamic sales through the above-described method and system, effectively solving the prediction bias problem caused by data lag and insufficient sentiment quantification in traditional methods.

Claims

1. A method for predicting sales of skincare products, characterized in that, include: The distributed web crawler cluster is used to capture user-generated content related to the target skincare product in real time from multiple mainstream social media platforms. The user-generated content is preprocessed; The preprocessed user-generated content is input into a multi-level sentiment semantic parsing model to generate a sentiment intensity sequence for each time window. Obtain historical sales time-series data, basic product attribute data, and current market environment variable data for the target skincare product; The emotional intensity sequence, historical sales time series data, product basic attribute data, and market environment variable data are aligned with a unified timestamp, and a multidimensional dynamic feature tensor is constructed. The multidimensional dynamic feature tensor is input into a sales prediction neural network model based on a spatiotemporal attention mechanism to output the sales prediction value of the target skin care product within a preset time window in the future. The sales prediction neural network model includes a spatiotemporal feature encoder, a cross-modal attention fusion module, and a sales decoder. Based on the sales forecast, inventory scheduling instructions or marketing strategy adjustment suggestions are generated and pushed to the enterprise resource planning system or digital marketing platform.

2. The skincare product sales forecasting method according to claim 1, characterized in that, The distributed web crawler cluster is used to crawl user-generated content related to the target skincare product in real time from multiple mainstream social media platforms, including: The crawler task is targeted to crawl by product keywords, brand name and popular topic tags, with a crawling frequency of once per minute, and each crawl covers the newly published content in the last 30 minutes; The distributed web crawler cluster is deployed on multiple geographically dispersed server nodes. Each node is configured with an independent proxy address pool and request frequency controller to circumvent the anti-crawling mechanisms of social media platforms.

3. The method for predicting sales volume of skincare products according to claim 1, characterized in that, Preprocessing the user-generated content includes: All text content is encoded using the Unicode octet encoding format. Regular expressions are used to match common ad templates, and blacklisted domains are used to filter posts containing promotional links to remove ad noise. By judging whether a post belongs to the Chinese character set character by character, posts with more than 70% non-Chinese content are removed. A named entity recognition model based on bidirectional long short-term memory network and conditional random field jointly trained is used to identify and extract product entity names and brand logos.

4. The method for predicting sales volume of skincare products according to claim 1, characterized in that, The multi-level emotion semantic parsing model includes a lexical analysis layer, an emotion polarity discrimination layer, and an emotion propagation attenuation correction layer. At the lexical analysis layer, a joint model of word segmentation and part-of-speech tagging based on a bidirectional long short-term memory network is used to perform fine-grained semantic unit segmentation of user-generated content; In the sentiment polarity discrimination layer, a convolutional neural network structure that integrates the domain dictionary and the contextual information is used to output the sentiment score of each semantic unit. The sentiment score ranges from -1 to +1. In the emotion propagation attenuation correction layer, the emotion score is weighted and corrected by applying an exponential decay function based on the difference between the content publication time and the current system time, with the attenuation coefficient set to 0.

95.

5. The skincare product sales forecasting method according to claim 4, characterized in that, The generation of emotion intensity sequences includes: The sentiment scores of all user-generated content within the same time window are weighted and averaged, with the weights determined by the number of followers of the user account, historical interaction rate, and content originality indicators. The time window is 5 minutes long, and the sliding step is 1 minute, forming a continuous time series of emotional intensity.

6. The method for predicting sales volume of skincare products according to claim 1, characterized in that, Obtain historical sales time-series data, basic product attribute data, and current market environment variable data for the target skincare product, including: Retrieves historical sales time-series data from the enterprise sales management system, supporting daily synchronization and hourly incremental updates; The product's basic attribute data is obtained from the enterprise's product master data management system. The core ingredient list is stored in a structured manner according to international cosmetic raw material naming rules, and the efficacy claim field is mapped and verified based on the filing information of the National Medical Products Administration. Market environmental variable data, including temperature, humidity, UV index, seasonality, and holiday markers, are obtained from meteorological service providers and statutory holiday schedules.

7. The method for predicting sales volume of skincare products according to claim 1, characterized in that, Constructing a multidimensional dynamic feature tensor includes: The emotional intensity sequence is expanded into the first feature channel along the time dimension; Historical sales time-series data are processed by first-order difference and used as the second feature channel; The product's basic attribute data is converted into a dense vector of fixed dimensions through an embedding layer and then used as the third feature channel. Market environment variable data are processed by one-hot encoding or normalization and used as the fourth feature channel; All feature channels are aligned along the time dimension and stacked to form a four-dimensional tensor with a time dimension length of 144, corresponding to the data of the past 24 hours.

8. The method for predicting sales volume of skincare products according to claim 1, characterized in that, The multidimensional dynamic feature tensor is input into a sales prediction neural network model based on a spatiotemporal attention mechanism to output the predicted sales value of the target skincare product within a preset future time window, including: The spatiotemporal feature encoder captures the correlation between products and the time-series dynamic patterns. The spatiotemporal feature encoder is composed of two layers of graph convolutional network and three layers of temporal convolutional network stacked alternately. The correlation weights between emotion features and other feature channels are calculated through a cross-modal attention fusion module, and the features are then weighted and fused accordingly. The sales volume decoder generates future sales forecasts hourly in an autoregressive manner. The sales volume decoder adopts a gated recurrent unit structure, and its initial hidden state value is obtained by mapping the fused feature vector through a fully connected layer. The activation function of the output layer is a modified linear unit.

9. The method for predicting sales volume of skincare products according to claim 1, characterized in that, Based on the sales forecast, generate inventory scheduling instructions or marketing strategy adjustment suggestions, and push them to the enterprise resource planning system or digital marketing platform, including: Set the safety stock threshold based on the 90th percentile of the predicted sales volume for the next 6 hours; Regional warehouse allocation suggestions are generated based on the difference between the projected sales volume and the current inventory in each region. An emergency replenishment signal is triggered if the predicted sales volume exceeds 50% of the current inventory in any hour. Realize advertising budgets based on sudden increases in emotional intensity and predicted sales growth trends; The discount level is set based on the predicted sales growth rate; Collaboration priorities are ranked based on the product of the emotional intensity evoked by KOLs' historical content and their actual sales conversion rate.

10. A skincare product sales forecasting system, characterized in that, include: The social media content real-time acquisition module is used to crawl user-generated content related to the target skincare product from multiple mainstream social media platforms in real time through a distributed web crawler cluster. The content preprocessing and entity recognition module is used to preprocess the user-generated content and extract product entity names and brand identifiers; The multi-level emotion semantic parsing module is used to input the preprocessed content into the multi-level emotion semantic parsing model to generate an emotion intensity sequence; The multi-source data integration module is used to obtain historical sales time-series data, basic product attribute data, and current market environment variable data for the target skincare product; The dynamic feature tensor construction module is used to align the emotion intensity sequence and other multi-source data with timestamps and construct multi-dimensional dynamic feature tensors. The sales prediction neural network inference module is used to input the multidimensional dynamic feature tensor into the sales prediction neural network model based on the spatiotemporal attention mechanism and output the sales prediction value. The decision support output module is used to generate inventory scheduling instructions or marketing strategy adjustment suggestions based on the sales forecast values ​​and push them to external business systems.

Images