Sales prediction system based on machine learning

Abstract

The invention discloses a sales prediction system based on machine learning, and relates to the technical field of sales prediction, and the system comprises a data integration module which collects commodity display images and customer behavior image data in a shop by means of an image collection device. Comprising commodity display images in shops, customer behavior images and structured business data in an enterprise sales management system, comprehensive and accurate insight of sales activities is realized, the data integration module effectively fuses the images and the structured data, the quality and availability of the data are improved, a solid foundation is provided for subsequent analysis, and the system is suitable for popularization and application. The product association prediction module uses an advanced association rule algorithm to accurately judge the complementation and substitution relationship between products, and uses a multi-product joint prediction model to carry out sales prediction, so that the prediction accuracy and reliability are significantly improved, enterprises are helped to better understand the market demand, and the market competitiveness is improved. And an enterprise can be guided to optimize product combination and formulate an accurate marketing strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of sales forecasting, and particularly to a sales forecasting system based on machine learning. Background Art

[0002] The technical background of the present invention relates to sales forecasting systems and their applications in the retail industry. With the rapid development of big data and artificial intelligence technologies, the retail industry is gradually transforming towards intelligence and digitization. Sales forecasting, as a key link in retail management, is of great significance for guiding product portfolio, optimizing marketing strategies, and managing the supply chain. By collecting and analyzing information on sales data and customer behavior data, enterprises can more accurately grasp market trends, formulate effective sales strategies, and thus improve operational efficiency and profitability.

[0003] However, traditional sales forecasting technologies have many deficiencies. On the one hand, traditional technologies mainly rely on structured data for analysis, such as sales records and customer purchase records, while ignoring the rich information contained in unstructured data such as images. This results in prediction results that are often not comprehensive and accurate enough. On the other hand, traditional technologies usually only conduct sales forecasting for a single product and fail to fully consider the correlation relationships between products, such as complementary relationships and substitution relationships, thus limiting the depth and breadth of the forecasting. In addition, traditional technologies also have deficiencies in real-time monitoring and risk warning, and it is difficult to detect and respond to abnormal fluctuations and risk events in sales data in a timely manner.

[0004] Therefore, the development of a sales forecasting system based on machine learning will provide enterprises with more accurate and comprehensive sales forecasting and risk management services, which is helpful for enterprises to optimize product portfolio, formulate accurate marketing strategies, manage the supply chain, and enhance the competitiveness and profitability of enterprises. Summary of the Invention

[0005] The purpose of the present invention is to make up for the deficiencies of the existing technology and provide a sales forecasting system based on machine learning. By integrating multi-modal data, including images and structured business data, this system achieves a comprehensive and accurate insight into sales activities. Using advanced association rule algorithms and multi-product joint forecasting models, the system can accurately judge the correlation relationships between products and conduct sales forecasting, significantly improving the accuracy and reliability of the forecasting. At the same time, the application of the real-time sales monitoring module and the risk assessment and warning module enhances the enterprise's real-time monitoring and risk warning capabilities.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: A sales forecasting system based on machine learning, which includes:

[0007] Data Integration Module: Collect product display images and customer behavior image data in the store with the help of image acquisition devices, obtain structured business data such as sales data and customer purchase records from the enterprise sales management system, denoise, crop, and extract features from the image data, clean, deduplicate, and normalize the structured business data, and fuse the images with the structured business data to form a multi-modal feature vector set;

[0008] Product Association Prediction Module: Receive the data from the Data Integration Module, calculate the product association strength index using the association rule algorithm, judge the complementary and substitution relationships between products, construct a multi-product joint prediction model using the multi-product joint prediction formula, and train and optimize the model;

[0009] Real-time Sales Monitoring Module: Collect sales data in real time, perform preprocessing, calculate the sales anomaly deviation index in combination with the prediction data, and when the anomaly deviation index exceeds the preset threshold, determine that the sales data is abnormal and immediately issue an alarm;

[0010] Model Dynamic Update Module: According to the anomaly situation and market change information feedback by the Real-time Sales Monitoring Module, dynamically adjust the parameters of the multi-product joint prediction model with the help of the model parameter adjustment algorithm;

[0011] Risk Assessment and Early Warning Module: Integrate product association prediction and real-time sales monitoring data, calculate the comprehensive sales risk index in combination with raw material price information, and when the comprehensive sales risk index exceeds the preset risk threshold, issue a warning signal and provide coping strategy suggestions for adjusting the product portfolio, optimizing the marketing strategy, and managing the supply chain.

[0012] Furthermore, in the Data Integration Module, the data fusion algorithm is used to fuse the images with the structured business data. Let the image feature vector be the structured business data vector be the fused multi-modal feature vector be F multi-modal , and the calculation formula is: where λ and μ are the fusion weights of the image features and the structured business data, which are determined by historical data according to the actual scenario.

[0013] Even further, in the Product Association Prediction Module, the association rule algorithm is used to calculate the product association strength index. Let the product association strength index be RAI, and the calculation formula is: where S ij is the support degree of product i and product j, which is calculated by counting the number of transactions count ij where product i and product j are purchased simultaneously in the historical sales data and the total number of transactions count total , and the formula is: C ijis the confidence level of product i and product j, representing the probability of purchasing product j when product i is purchased. The formula is: count ij|i is the number of transactions where product j is purchased simultaneously in the transactions of purchasing product i. count i is the number of transactions of purchasing product i. L ij is the reciprocal of the lift of product i and product j. The lift calculation formula is:

[0014] Furthermore, the judgment of complementary and substitution relationships in the product association prediction module:

[0015] The judgment of the complementary relationship: When RAI ij > T RAI , and C ij is close to 1, it indicates that product i and product j are in a complementary relationship, where T RAI is the threshold of the association strength index, determined by the historical data quantile or industry standard value;

[0016] The judgment of the substitution relationship: If RAI ij < T RAI , when the sales volume of product i increases, the sales volume of product j shows an obvious downward trend, and at the same time C ij < 0.3, it indicates that product i and product j are in a substitution relationship, where T RAI is the threshold of the association strength index, determined according to industry requirements and data distribution.

[0017] Furthermore, the product association prediction module uses the multi-product joint prediction formula to construct a prediction model. Suppose there are n products. The predicted sales value of product i at time t is Y i，t . The calculation formula is: where RAI ij is the association strength index of product i and product j, α ij is the association weight between product i and product j, determined by historical sales data, Y j，t-1 is the actual sales quantity of product j in the previous time period t, X i，t is the self-characteristic vector of product i at time t, including factors such as product price, promotion activity intensity, and inventory level affecting sales, β i is the weight vector of the self-characteristics of product i, γ i is the constant term, used to correct the deviation of the model prediction.

[0018] Furthermore, for the training of the multi-product joint prediction model in the product association prediction module, it receives the multi-modal feature dataset from the data integration module, first divides it into a training set, a validation set, and a test set, initializes the parameters of the multi-product joint prediction model and defines the loss function, iteratively trains the model using the training set, updates the parameters according to the error, and adjusts the hyperparameters based on the evaluation using the validation set after each iteration. When the model performs stably on the training set and the validation set, it conducts a final evaluation using the test set, analyzes the model performance, optimizes and improves the model according to the evaluation results, and predicts the sales trends of different product combinations based on the model.

[0019] Furthermore, in the real-time sales monitoring module, the sales anomaly deviation index is calculated through the anomaly detection formula. Let the sales prediction data at the current time point be Y i，t , and the sales anomaly deviation index SAEI, with the calculation formula as follows: where S actual is the actual sales amount, σ S is the historical standard deviation of the sales amount, δ is the market environment sensitivity coefficient, determined according to market news and industry reports, and E t is the market environment index at the current time point, obtained by quantifying factors such as market activity and the intensity of competitors' promotional activities. By comparing with the preset threshold τ, if SAEI > τ, it is determined that the sales data is abnormal, and an alarm is immediately triggered, notifying relevant personnel in the form of text messages, emails, and system pop-ups.

[0020] Furthermore, in the model dynamic update module, the parameters of the multi-product joint prediction model are dynamically adjusted by means of the model parameter adjustment algorithm. Let the original model parameters be θ = (θ1, θ2,..., θ n ), and according to the abnormal situation and market changes, the calculation formula for the new model parameters θ' is as follows: where η is the learning rate, used to control the step size of parameter update, set through cross-validation or empirical values, ΔF multi-modeal is the change amount of the multi-modal feature vector, and ΔSAEI is the change amount of the sales anomaly deviation index, obtained by subtracting the sales anomaly deviation index at the previous time point from that at the current time point.

[0021] Furthermore, in the risk assessment and early warning module, the comprehensive sales risk index is calculated using the risk quantification formula. Let the comprehensive sales risk index be CSRI, and the calculation formula is as follows: where SAEI is the sales anomaly deviation index, and RAI max is the maximum value of the product association strength index, ΔP raw is the change amount of the raw material price, and P rawis the original price of raw materials, ρ is the weight coefficient of product association and sales anomaly deviation factors, ω is the weight coefficient of raw material price change factors. When CSRI exceeds the preset risk threshold T CSRI An alarm is triggered.

[0022] Compared with the prior art, the sales prediction system based on machine learning has the following beneficial effects:

[0023] First, by integrating multi-modal data, including in-store product display images, customer behavior images, and structured business data in the enterprise sales management system, the present invention realizes a comprehensive and accurate insight into sales activities. The data integration module effectively integrates images and structured data, improving the quality and usability of the data and providing a solid foundation for subsequent analysis. The product association prediction module uses advanced association rule algorithms to accurately judge the complementary and substitution relationships between products and applies a multi-product joint prediction model for sales prediction, significantly improving the accuracy and reliability of the prediction. This not only helps enterprises better understand market demand but also guides them to optimize product portfolios and formulate precise marketing strategies, thereby enhancing market competitiveness.

[0024] Second, through the real-time sales monitoring module and the risk assessment and early warning module, the present invention realizes dynamic monitoring and risk early warning of sales activities. The real-time sales monitoring module can immediately detect abnormal fluctuations in sales data and issue an alarm in a timely manner through a preset threshold judgment mechanism, ensuring that enterprises can quickly respond to market changes. The risk assessment and early warning module integrates various data, including product association prediction results, real-time sales data, and raw material price information, calculates a comprehensive sales risk index using a risk quantification formula, and when the risk index exceeds the preset threshold, the system automatically triggers an early warning and provides strategic suggestions for enterprises to cope with risks. This not only enhances the enterprise's risk management ability but also provides a scientific basis for enterprise decision-making and helps the enterprise develop steadily.

[0025] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on an investigation and study of the following text, or can be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings without creative efforts based on these drawings.

[0027] Figure 1Flowchart of a machine learning-based sales prediction system;

[0028] Figure 2 Framework diagram of a machine learning-based sales prediction system. Specific implementation manners

[0029] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in conjunction with the accompanying drawings and preferred embodiments, describe in detail the specific implementation manners, structures, features, and effects of the present invention as follows.

[0030] Embodiment 1:

[0031] Sales prediction for an electronics specialty store.

[0032] A large electronics specialty store mainly sells various electronics such as mobile phones, tablets, headphones, and chargers, has a certain influence in the market, and faces fierce competition. To more accurately grasp market demand and improve sales performance, the specialty store introduced a machine learning-based sales prediction system.

[0033] Data integration module: The specialty store installs high-definition cameras at various key positions in the store as image acquisition devices to collect product display images and customer behavior image data in the store. From the professional sales management system used by the store, detailed sales data and structured business data of customer purchase records are obtained, such as Figure 2 .

[0034] Denoise the image data to remove noise generated by equipment or environmental factors and improve image quality; then perform cropping to focus on key areas, such as product display areas and areas where customers are active frequently; finally, perform feature extraction, for example, extract the layout features of product displays, the staying time and trajectories of customers in different product areas.

[0035] Clean the structured business data to remove incorrect or incomplete data; perform deduplication to avoid interference caused by duplicate data to analysis; then perform normalization processing to unify data of different magnitudes to the same scale range.

[0036] Use a data fusion algorithm to fuse the images and structured business data. According to the past data research and experience of the specialty store, let the image feature vector be The structured business data vector be The fused multi-modal feature vector be F multi-modal , and the calculation formula is: Where the image feature fusion weight is λ, and the structured business data fusion weight is μ, forming a multi-modal feature data set containing product image features and sales business features, such as Figure 1 .

[0037] Product Association Prediction Module: Receives the data processed by the Data Integration Module, and uses the association rule algorithm to calculate the product association strength index. Taking mobile phones and headphones as an example, by counting the number of transactions where mobile phones and headphones are purchased simultaneously in historical sales data as count ij times, and the total number of transactions as count total times, the support formula is calculated as: Among the transactions where mobile phones are purchased, the number of transactions where headphones are purchased simultaneously is count ij|i times, and the number of transactions where mobile phones are purchased is count i times, the confidence formula is calculated as: The lift formula is calculated as: Furthermore, the reciprocal of the lift is obtained Finally, the product association strength index RAI is calculated, and the calculation formula is: By comparing with the association strength index threshold T RAI When RAI ij <T RAI , and C ij is close to 1, it is determined that mobile phones and headphones are in a complementary relationship.

[0038] Use the multi-product joint prediction formula to construct a multi-product joint prediction model. Suppose there are n products including mobile phones, headphones, and chargers. The sales prediction value of mobile phones at time t is Y i，t , and the calculation formula is: where RAI ij is the association strength index between product i and product j, a ij is the association weight between product i and product j, determined by historical sales data, Y j，t-1 is the actual sales quantity of product j in the previous time period t1, X i，t is the self-characteristic vector of product i at time t, including factors such as product price, promotion activity intensity, and inventory level that affect sales, β i is the weight vector of the self-characteristics of product i, γ i is a constant term. Divide the multi-modal feature dataset into a training set, a validation set, and a test set. Initialize the parameters of the multi-product joint prediction model, use the training set to iteratively train the model, update the parameters through an optimization algorithm according to the error generated in each iteration. After each iteration, use the validation set to evaluate the model performance, adjust the learning rate and regularization parameters according to the evaluation results. When the loss values of the model on the training set and the validation set tend to be stable and no longer decrease significantly, use the test set for the final evaluation, analyze the accuracy, recall rate and other performance indicators of the model, further optimize and improve the model according to the evaluation results, and finally predict the sales trends of different product combinations based on the optimized model.

[0039] Real-time Sales Monitoring Module: The sales system of the exclusive store collects the sales data of each product in real time, and performs preprocessing to remove outliers and duplicate data. Combining with the predicted data, it calculates the sales anomaly deviation index. Suppose the predicted sales data of mobile phones at the current time point is Y i，t units, and the sales quantity corresponding to the actual sales amount is S actual units, and the historical standard deviation of the sales amount is σ S , through the analysis of market news and industry reports, the market environment sensitivity coefficient is determined to be δ, and the market environment index at the current time point is obtained by quantifying factors such as market activity and the intensity of competitors' promotional activities as E t , according to the formula, the sales anomaly deviation index SAEI is calculated. The calculated sales anomaly deviation index SAEI is compared with the preset threshold τ. Since SAEI > τ, it is determined that the mobile phone sales data is abnormal, and the system immediately notifies the store manager, sales manager, and purchaser in the form of text messages, emails, and system pop-ups, reminding them to pay attention to the mobile phone sales situation.

[0040] Model Dynamic Update Module: Based on the abnormal situation and market change information fed back by the real-time sales monitoring module, such as a certain brand launching a competitive new mobile phone, resulting in a change in market demand, the parameters of the multi-product joint prediction model are dynamically adjusted by means of a model parameter adjustment algorithm. Suppose the original model parameters are θ = (θ1, θ2,..., θ n ), and the change amount of the multi-modal feature vector is ΔF multi-modal , obtained by comparing the multi-modal feature vectors of the new data and the old data, and the change amount of the sales anomaly deviation index ΔSAEI, obtained by subtracting the sales anomaly deviation index of the previous moment from the sales anomaly deviation index of the current moment. Set the learning rate to η, and calculate the new model parameters θ' according to the formula. The calculation formula is: So that the model can better adapt to market changes and improve the prediction accuracy.

[0041] Risk Assessment and Early Warning Module: Integrate the product association prediction and real-time sales monitoring data, and calculate the comprehensive sales risk index in combination with the raw material price information. It is known that the maximum value of the mobile phone product association strength index is RAI max , and the change amount of the raw material price is ΔP raw , the original price of the raw material is P raw , the weight coefficient of the product association and sales anomaly deviation factors is ρ, and the weight coefficient of the raw material price change factor is ω. Calculate the comprehensive sales risk index CSRI according to the formula. The calculation formula is: Among them, SAEI is the sales anomaly deviation index. When CSRI exceeds the preset risk threshold T CSRIWhen this happens, the system issues a warning signal. At the same time, it provides suggestions for coping strategies: adjust the product mix, reduce the purchase volume of mobile phone models that are more affected by the rising raw material prices, increase the purchase ratio of headphone and charger accessory products with higher profit margins, optimize the marketing strategy, launch combined promotional activities for mobile phones and headphones, such as giving away headphones when purchasing a designated mobile phone, to increase the overall sales volume of products, manage the supply chain, renegotiate prices with raw material suppliers, seek long-term stable supply contracts, or look for alternative raw materials to reduce cost risks.

[0042] In summary, the sales prediction system based on machine learning has achieved remarkable results in electronics specialty stores. The data integration module integrates multi-source data, laying a foundation for subsequent analysis; the product association prediction module accurately grasps product relationships, and the constructed model effectively predicts sales trends; the real-time sales monitoring module can detect anomalies in a timely manner to ensure stable sales; the model dynamic update module optimizes the model according to market changes to ensure accurate prediction; the risk assessment and warning module comprehensively considers multiple factors and provides practical coping strategies. This system comprehensively improves the scientific nature and foresight of sales management in specialty stores, enhances its market competitiveness, and helps achieve sustainable profitability and steady development in the highly competitive electronics market.

[0043] Example 2:

[0044] Sales prediction in a chain supermarket A chain supermarket has multiple stores in several cities, selling various products such as food, daily necessities, and fresh produce. It faces a large amount of sales data and complex market demand changes every day. To achieve refined operations and improve sales efficiency, this chain supermarket adopts a sales prediction system based on machine learning.

[0045] Data integration module: Install image acquisition devices in the shelf areas, aisles, and cashier positions of each store to collect images of product displays and images of customer behavior in the store. Obtain structured business data such as sales data, customer purchase records, and membership information of each store from the supermarket's central sales management system. Denoise the image data, extract the key areas of product displays and effective images of customer behavior through image cropping technology, and perform feature extraction, such as the way products are placed and the selection actions of customers in front of different shelves. Clean the structured business data, check and correct error values and missing values in the data, remove duplicate records, and through normalization processing, unify the data of different stores and different time periods to the same magnitude range. Use a data fusion algorithm to fuse the images and structured business data. The formula is: Form a multi-modal feature data set containing product image features, sales business features, and customer information features, providing comprehensive data support for subsequent predictive analysis.

[0046] Product Association Prediction Module: Receives the data processed by the Data Integration Module, and calculates the product association strength index using the association rule algorithm. Taking bread and milk as an example, the historical sales data of a store in a month is statistically analyzed. The number of transactions where both bread and milk are purchased is count ij , and the total number of transactions is count total . Calculate the support degree, and the formula is: Among the transactions where bread is purchased, the number of transactions where milk is also purchased is count ij|i , and the number of transactions where bread is purchased is count i . Calculate the confidence degree, and the formula is: Assume the support degree of milk is S ij . Calculate the lift degree, and the formula is: Furthermore, obtain the reciprocal of the lift degree Finally, calculate the product association strength index, and the formula is: By comparing with the association strength index threshold T RAI , and C ij is relatively high, it is determined that bread and milk are complementary relationships. Then, taking carbonated drinks and fruit juice as an example, after calculating their association strength index RAI ij < T RAI , and when the sales volume of carbonated drinks increases, the sales volume of fruit juice shows an obvious downward trend, and at the same time C ij < 0.3, it is determined that carbonated drinks and fruit juice are substitute relationships.

[0047] Construct a multi-product joint prediction model using the multi-product joint prediction formula. The Product Association Prediction Module constructs and trains a multi-product joint prediction model: Suppose there are n products including bread, milk, carbonated drinks, and fruit juice. For the sales prediction value Y of bread at time t i，t , the calculation formula is: Among them, determine the association weight α between bread and milk according to historical sales data ij , the weight vector β of the bread's own characteristics (such as price, promotion activity intensity, inventory level) i , and initialize the constant term γ i, divide the multi-modal feature dataset into a training set, a validation set, and a test set, initialize the parameters of the multi-product joint prediction model, and measure the difference between the predicted values and the true values of the model. Use the training set to iteratively train the model. During each training process, update the model parameters according to the error. After each iteration, use the validation set to evaluate the model and adjust the hyperparameters according to the evaluation results. When the loss values of the model on the training set and the validation set tend to be stable and the accuracy metric no longer improves significantly, it indicates that the model has achieved a good training effect. At this time, use the test set to conduct a final evaluation of the model. By analyzing various performance metrics of the model on the test set, further optimize and improve the model. The optimized model can more accurately predict the sales trends of different product combinations in the future time period, providing strong support for the supermarket's procurement, inventory management, and marketing strategy formulation.

[0048] Real-time sales monitoring module: The chain supermarket collects the sales data of various commodities in real time through the sales systems of each store. Immediately after data collection, preprocess the data to remove outliers and duplicate data caused by system failures or human operation errors, ensuring the accuracy and effectiveness of the data. Calculate the sales anomaly deviation index in combination with the predicted data. Assume that the predicted sales data of bread in a certain store at the current time point is Y i，t pieces, and the actual sales volume corresponding to the actual sales amount is S actual pieces. Through the statistical analysis of the bread sales amount data of this store in the past month, the historical standard deviation of the sales amount is obtained as σ S , the supermarket arranges special personnel to pay attention to market news, industry reports, and combines its own research on the local market to determine that the market environment sensitivity coefficient is δ. The market environment index at the current time point is obtained by quantifying factors such as market activity and the intensity of competitors' promotional activities as E t , and the sales anomaly deviation index of the bread in this store is calculated according to the formula as SAEI. The formula is: Compare the calculated sales anomaly deviation index SAEI with the preset threshold τ. Since SAEI > τ, it is determined that the sales data of bread in this store is abnormal. The system immediately notifies the store manager, sales supervisor, and purchasing specialist of this store via text message, and at the same time pops up a warning window in the supermarket's internal management system to remind relevant personnel to pay attention to the bread sales situation in a timely manner, analyze the reasons for the anomaly, and take corresponding measures.

[0049] Model dynamic update module: Based on the anomaly situation and market change information feedback by the real-time sales monitoring module, such as a well-known bread brand launching a new flavor of bread, attracting a large number of consumers, resulting in changes in market demand, or a large-scale event being held locally, and the demand structure of the surrounding residents for food being adjusted. Use the model parameter adjustment algorithm to dynamically adjust the parameters of the multi-product joint prediction model. Let the original model parameters be θ = (θ1, θ2,..., θ n), ΔF multi-modal The change in the multi-modal feature vector, ΔSAEI is obtained by comparing the multi-modal feature vectors of new data and old data. The change in the sales anomaly deviation index is obtained by subtracting the sales anomaly deviation index at the previous moment from the sales anomaly deviation index at the current moment. Set the learning rate to η, and calculate the new model parameters according to the formula: By dynamically adjusting the model parameters, the model can quickly adapt to the dynamic changes in the market, continuously maintain a high prediction accuracy, and provide a more reliable basis for the operation decision-making of the supermarket.

[0050] Risk assessment and early warning module: Integrate product association prediction and real-time sales monitoring data, and calculate the comprehensive sales risk index in combination with raw material price information. The maximum value of the bread product association strength index is RAI max , the price of flour raw materials has recently increased, and the change in raw material price is ΔP raw , the original price of the raw material is P raw , calculate the comprehensive sales risk index according to the formula. The formula is: When the comprehensive sales risk index CSRI exceeds the preset risk threshold T CSRI , the system issues a warning signal. At the same time, the following coping strategy suggestions are provided for the supermarket: Adjust the product mix: Appropriately reduce the purchase quantity of ordinary bread, increase the procurement ratio of new flavor bread or other bakery products less affected by the increase in raw material prices, optimize the product structure to meet the diverse needs of consumers, and reduce cost risks. Optimize the marketing strategy: In view of the complementary relationship between bread and milk, launch combined promotional activities, such as enjoying a discount on milk when purchasing a certain quantity of bread, to increase the overall sales volume of related products; for carbonated beverages and juice substitute products, adjust the display position and promotional efforts according to different seasons and consumer preferences to attract consumers to purchase. Manage the supply chain: Negotiate with flour suppliers to strive for more favorable purchase prices and stable supply contracts; at the same time, actively seek other high-quality raw material suppliers, expand the supply channels, and reduce the risks brought by raw material price fluctuations.

[0051] In summary, after the chain supermarket applies the sales prediction system based on machine learning, the operation management is optimized. The data integration module integrates multi-modal data and provides rich information; the product association prediction module clarifies product associations and optimizes the product mix; the real-time sales monitoring module quickly discovers sales anomalies and facilitates timely intervention; the model dynamic update module enables the model to keep up with market changes; the risk assessment and early warning module assesses risks and gives coping strategies to reduce operation risks. This system helps the supermarket reasonably arrange procurement and inventory, formulate marketing strategies, improve operation efficiency, meet customer needs, enhance the adaptability and profitability of the chain supermarket in the retail market, and promote the intelligent upgrade of the industry.

[0052] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments by using the above-disclosed technical content within the scope of the technical solution of the present invention. However, as long as it does not depart from the content of the technical solution of the present invention, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A sales prediction system based on machine learning, characterized in that, The system includes: Data integration module: Collect product display images and customer behavior image data in the store with the help of image acquisition devices, obtain structured business data such as sales data and customer purchase records from the enterprise sales management system, denoise, crop and extract features from the image data, clean, deduplicate and normalize the structured business data, and fuse the images with the structured business data to form a multi-modal feature vector set; Product association prediction module: Receive the data from the data integration module, calculate the product association strength index using the association rule algorithm, judge the complementary and substitution relationships between products, construct a multi-product joint prediction model using the multi-product joint prediction formula, and train and optimize the model; Real-time sales monitoring module: Real-time collect sales data, perform preprocessing, calculate the sales anomaly deviation index in combination with the prediction data, and when the anomaly deviation index exceeds the preset threshold, determine that the sales data is abnormal and immediately issue an alarm; Model dynamic update module: According to the abnormal situation and market change information feedback by the real-time sales monitoring module, dynamically adjust the parameters of the multi-product joint prediction model with the help of the model parameter adjustment algorithm; Risk assessment and early warning module: Integrate product association prediction and real-time sales monitoring data, calculate the comprehensive sales risk index in combination with raw material price information, and when the comprehensive sales risk index exceeds the preset risk threshold, issue a warning signal and provide countermeasure suggestions for adjusting the product portfolio, optimizing the marketing strategy, and managing the supply chain.

2. The sales prediction system based on machine learning according to claim 1, wherein In the data integration module, a data fusion algorithm is used to fuse images with structured business data. Let the image feature vector be and the structured business data vector be The fused multi-modal feature vector is F multi-modal , and the calculation formula is: where λ and μ are the fusion weights of the image features and the structured business data, which are determined according to the actual scenario through historical data.

3. The sales prediction system based on machine learning according to claim 1, characterized in that, In the product association prediction module, the association rule algorithm is used to calculate the product association strength index. Let the product association strength index be RAI, and the calculation formula is as follows: Among them, S ij is the support degree of product i and product j, which is calculated by counting the number of transactions count that purchase both product i and product j in the historical sales data ij and the total number of transactions count total The calculation formula is: C ij is the confidence degree of product i and product j, indicating the probability of purchasing product j when purchasing product i. The formula is: count ij|i is the number of transactions that purchase both product j and product i in the transactions of purchasing product i, and count i is the number of transactions of purchasing product i, and L ij is the reciprocal of the lift of product i and product j. The lift calculation formula is:

4. The sales prediction system based on machine learning according to claim 3, wherein, Judgment of complementary and substitution relationships in the product association prediction module: The complementary relationship judgment: When RAI ij >T RAI , and C ij is close to 1, it indicates that product i and product j are in a complementary relationship, where T RAI is the threshold of the association strength index, which is determined by the historical data quantile or the industry standard value; The substitution relationship judgment: If RAI ij <T RAI , when the sales volume of product i increases, the sales volume of product j shows an obvious downward trend, and at the same time C ij <0.3, it indicates that product i and product j are in a substitution relationship, where T RAI is the correlation strength index threshold, which is determined according to industry requirements and data distribution.

5. The sales prediction system based on machine learning according to claim 3, wherein In the product association prediction module, a prediction model is constructed using the multi-product joint prediction formula. Suppose there are n products. The predicted sales value of product i at time t is Y i,t , and the calculation formula is: where RAI ij is the association strength index between product i and product j, a ij is the association weight between product i and product j, determined by historical sales data, Y j，t-1 is the actual sales quantity of product j in the previous time period t, X i,t is the self-characteristic vector of product i at time t, including factors such as product price, promotion activity intensity, and inventory level that affect sales, β i is the weight vector of the self-characteristics of product i, γ i is a constant term used to correct the deviation of model prediction.

6. The sales prediction system based on machine learning according to claim 1, characterized in that, Training of the multi-product joint prediction model in the product association prediction module: Receive the multi-modal feature data set from the data integration module, first divide it into a training set, a validation set and a test set, initialize the parameters of the multi-product joint prediction model and define the loss function, use the training set to iteratively train the model, update the parameters according to the error, evaluate and adjust the hyperparameters accordingly using the validation set after each iteration, when the model performs stably on the training set and the validation set, use the test set for the final evaluation, analyze the model performance, optimize and improve the model according to the evaluation results, and predict the sales trends of different product combinations based on the model.

7. The machine learning-based sales prediction system according to claim 5, wherein In the real-time sales monitoring module, the sales anomaly deviation index is calculated through an anomaly detection formula. Let the sales forecast data at the current time point be Y i,t , and the sales anomaly deviation index SAEI. The calculation formula is as follows: where S actual is the actual sales amount, σ S is the historical standard deviation of the sales amount, δ is the market environment sensitivity coefficient, which is determined according to market news and industry reports, and E t is the market environment index at the current time point, which is obtained by quantifying factors such as market activity and the intensity of competitors' promotional activities. Compare it with the preset threshold τ. If SAEI > τ, it is determined that the sales data is abnormal, and an alarm is immediately triggered to notify relevant personnel in the form of text messages, emails, and system pop-ups.

8. The sales prediction system based on machine learning according to claim 2, wherein In the model dynamic update module, the parameters of the multi-product joint prediction model are dynamically adjusted by means of a model parameter adjustment algorithm. Let the original model parameters be θ = (θ1, θ2, …, θ n ). According to the abnormal situation and market changes, the calculation formula for the new model parameters θ ′ is as follows: Among them, η is the learning rate, which is used to control the step size of parameter update and is set through cross-validation or empirical values. ΔF multi-modal is the change amount of the multi-modal feature vector, and ΔSAEI is the change amount of the sales anomaly deviation index, which is obtained by subtracting the sales anomaly deviation index at the previous moment from the sales anomaly deviation index at the current moment.

9. The machine learning-based sales prediction system according to claim 7, wherein In the risk assessment and early warning module, a risk quantification formula is used to calculate the comprehensive sales risk index. Let the comprehensive sales risk index be CSRI, and the calculation formula is as follows: Among them, SAEI is the sales anomaly deviation index, and RAI max is the maximum value of the product correlation strength index, and ΔP raw is the change in the raw material price, P raw is the original price of the raw material, ρ is the weight coefficient of the product correlation and sales anomaly deviation factors, ω is the weight coefficient of the raw material price change factor. When CSRI exceeds the preset risk threshold T CSRI an alarm is triggered.

Images