Commodity demand prediction method and system based on multi-modal machine learning

Abstract

The invention provides a commodity demand prediction method and system based on multi-modal machine learning, and relates to the technical field of commodity demand prediction. Comprising the steps that commodity data are collected and preprocessed, commodity features of the preprocessed commodity data in different scenes are extracted, e-commerce commodity comprehensive evaluation indexes are constructed, and the features of the preprocessed commodity data are divided based on a climb consumption algorithm; a DeepAR model is used to construct a commodity operation management initial model according to the feature division commodity data, an FTRL algorithm is used to analyze the feature division commodity data, and parameters of the commodity operation management initial model are adjusted according to newly generated competition data; clustering competitive product evaluation texts through a BERTopic model and constructing competitive influence indexes; and a commodity association weight matrix is constructed according to commodity sales data trend prediction, and perfect commodity demand prediction is obtained. According to the invention, prediction is carried out according to the commodity demand by constructing the commodity operation management initial model, and the timeliness and accuracy of prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of commodity demand forecasting technology, and in particular to a commodity demand forecasting method and system based on multimodal machine learning. Background Technology

[0002] With the booming development of the e-commerce industry and the increasing diversification of the consumer market, product demand is becoming increasingly complex and volatile. Most existing product demand forecasting methods rely solely on single sales or user behavior data, failing to fully integrate multimodal information such as text, images, and videos. E-commerce platforms frequently conduct various promotional activities, such as major shopping festivals like Singles' Day, Double Twelve, and 618, as well as daily limited-time discounts and spending-based reductions. These promotional activities have a significant impact on product sales, causing sales data to exhibit obvious volatility and irregularities, increasing the difficulty of forecasting. In contrast, product sales in other industries may be less affected by such large-scale, high-frequency promotional activities. The e-commerce market is highly competitive, with new products constantly emerging and consumers having numerous choices, potentially shortening the product lifecycle. A product may experience a short-term sales boom but quickly be replaced by new competitors, requiring continuous monitoring of market dynamics and competitor strategies.

[0003] The market environment is difficult to predict due to the volatile nature of promotions, competition, and consumer behavior. Predictive models also have limitations such as poor adaptability, high computational resource requirements, and insufficient multimodal fusion. Summary of the Invention

[0004] This invention provides a product demand forecasting method and system based on multimodal machine learning, which addresses the shortcomings of existing technologies in predicting market demand due to the volatile nature of promotions, competition, and consumer behavior.

[0005] On the one hand, this invention provides a commodity demand forecasting method based on multimodal machine learning, including:

[0006] S1: Collect product data from e-commerce platforms and preprocess it to obtain preprocessed product data.

[0007] S2: Extract product features from preprocessed product data under different scenarios, construct comprehensive evaluation indicators for e-commerce products, and classify the features of preprocessed product data based on the comparative consumption algorithm to obtain feature product data classification.

[0008] S3: Use the DeepAR model to build the initial model for product operation management, use the FTRL algorithm to analyze features and segment product data, and adjust the parameters of the initial model for product operation management based on newly generated competitive data.

[0009] S4: Cluster competitor evaluation texts using the BERTopic model and construct a competitive impact index to capture real-time market dynamics and obtain trend predictions for product sales data.

[0010] S5: Construct a product association weight matrix based on product sales data trend forecasts, quantify the market competition correlation and sales complementarity correlation between product categories, and obtain a more complete product demand forecast.

[0011] S6: Dynamically adjust the basic operational strategies of each product based on the product association weight matrix to generate the optimal product operation management plan that fits the product demand forecast.

[0012] According to the product demand forecasting method based on multimodal machine learning provided by this invention, the specific steps for constructing a comprehensive evaluation index system for e-commerce products are as follows:

[0013] We define product quality, user experience, sales performance, and market competitiveness as the dimensions for evaluation indicators.

[0014] We selected quality rating (product quality dimension), positive user review rate, page loading speed (user experience dimension), and after-sales service satisfaction as evaluation indicators.

[0015] The evaluation indicators are weighted and determined using the analytic hierarchy process (AHP).

[0016] The comprehensive evaluation index for the product is calculated based on the evaluation indicators and their weights.

[0017] According to the commodity demand forecasting method based on multimodal machine learning provided by the present invention, the specific steps in step S3 of constructing the initial commodity operation management model using the DeepAR model based on the feature category classification results are as follows:

[0018] The preprocessed commodity data features are divided into static feature data, time series feature data, and multimodal feature data. The static feature data is one-hot encoded, the numerical time series feature data is standardized, and the multimodal features are adapted to the DeepAR model.

[0019] Static feature data is input into the static layer for processing, while temporal features are input into the time series layer. By combining global and local attention mechanisms, a comprehensive feature vector is output.

[0020] The product feature vectors are divided into training and validation sets according to the time sequence.

[0021] Based on the objective of the commodity operation management model, mean squared error is selected as the loss function. The Adam optimizer is used and its hyperparameters are configured. The model is trained in multiple rounds using the training set, and the prediction results and loss values ​​of each round are output.

[0022] The prediction results are evaluated using a validation set to determine whether the loss value is greater than the preset range. If so, the hyperparameters are adjusted to obtain the initial model for commodity operation management.

[0023] According to the commodity demand forecasting method based on multimodal machine learning provided by the present invention, the specific steps for adjusting the initial model parameters of commodity operation management in step S3 are as follows:

[0024] By incorporating newly generated competitive data into the preprocessed product data features, a comprehensive feature dataset is constructed by combining competitor information with various product features.

[0025] Set the learning rate and regularization for the FTRL variant to update the parameters of the initial model for product operations management.

[0026] The initial model is associated with the FTRL algorithm and the objective function, so that the algorithm is guided by optimizing the objective function and uses integrated data to update the parameters of the initial model.

[0027] The comprehensive feature dataset is input into the initial model. After forward propagation to obtain the prediction results, the gradient is calculated. The FTRL algorithm is used to update the model parameters according to the gradient. The model is iterated multiple times to adapt to the changes in competitive data until the model is greater than the preset threshold on the validation set, thus obtaining the commodity operation management model.

[0028] According to the commodity demand forecasting method based on multimodal machine learning provided by the present invention, the specific steps for obtaining the commodity sales data trend forecast in step S4 are as follows:

[0029] Collect user review text data of competitors, preprocess it, and obtain a clean text dataset.

[0030] The BERTopic model was used to perform cluster analysis on the preprocessed competitor review texts to obtain different review topic clusters.

[0031] Based on the content characteristics of the evaluation topic clusters and combined with market feedback data, key factors affecting product competitiveness are extracted, and weights are set for these key factors.

[0032] The competitive impact index of different competing products is calculated based on the size of the evaluation topic cluster, sentiment tendency, and the weight of key factors.

[0033] Establish a correlation model between the competition impact index and product sales data, and dynamically update the competition impact index by monitoring the updates of competitor evaluation texts in real time.

[0034] The updated competition impact index and market dynamics data are input into the correlation model, and the future trend of commodity sales data is predicted by analyzing historical sales data trends.

[0035] According to the commodity demand forecasting method based on multimodal machine learning provided by the present invention, the specific steps for improving commodity demand forecasting in step S5 are as follows:

[0036] S51: Construct indicators based on market share changes and correlation values ​​to determine the degree of competitive correlation.

[0037] S52: Construct complementary correlation indicators based on sales data trends of different product categories and products that are frequently purchased simultaneously.

[0038] S53: Construct a product association weight matrix based on the indicators of competitive association degree and complementary association degree.

[0039] S54: Input the product association weight matrix into the product operation management model to improve product demand forecasting.

[0040] According to the commodity demand forecasting method based on multimodal machine learning provided by the present invention, the specific steps for analyzing market competition correlation in step S51 are as follows:

[0041] For each product category, its market share at different times is calculated based on sales data trends. By analyzing the changes in market share of each product category over time, the competitive landscape between categories can be determined.

[0042] Spearman correlation coefficient was used to conduct correlation analysis on the sales of different product categories, and the correlation values ​​between each pair of product categories were calculated.

[0043] Based on changes in market share and correlation values, specific market competition correlation indicators are constructed to obtain the degree of market competition correlation between each pair of product categories.

[0044] According to the commodity demand forecasting method based on multimodal machine learning provided by the present invention, the specific steps for constructing the commodity association weight matrix in step S53 are as follows: organize the market competition association degree and sales complementarity association degree values ​​between each pair of commodity categories into a matrix, wherein the rows and columns of the matrix correspond to different commodity categories, and the value in each cell is the association degree value between the corresponding pair of commodity categories, thereby obtaining the commodity association weight matrix.

[0045] According to the commodity demand forecasting method based on multimodal machine learning provided by the present invention, the specific steps in step S6 for generating the optimal commodity operation management plan that matches the commodity demand forecast are as follows:

[0046] Quantify and assign initial weights to improve sales performance, optimize inventory turnover, and enhance customer satisfaction.

[0047] Using a multi-objective optimization algorithm, the relationship characteristics and constraints are clarified based on the product association weight matrix, and the parameter range of each product's operation strategy is determined.

[0048] Generate basic operational strategy plans, calculate fitness values, and iteratively optimize parameters through selection, crossover, and mutation operations.

[0049] By introducing real-time sales data feedback, adjusting weights and the strength of constraints, and selecting the comprehensive optimal solution from the set of optimal solutions, an optimal product operation and management plan is formed.

[0050] On the other hand, the present invention also provides a commodity demand forecasting system based on multimodal machine learning, comprising:

[0051] The data acquisition module is used to collect product data from e-commerce platforms and preprocess it to obtain preprocessed product data.

[0052] The product feature extraction module is used to extract product features from preprocessed product data in different scenarios, construct comprehensive evaluation indicators for e-commerce products, and classify the features of preprocessed product data based on the comparative consumption algorithm to obtain feature product data classification.

[0053] The model building module is used to construct an initial model for commodity operation management based on the feature-based segmentation of commodity data using the DeepAR model, analyze the feature-based segmentation of commodity data using the FTRL algorithm, and adjust the parameters of the initial model for commodity operation management based on newly generated competitive data.

[0054] The sales forecasting module is used to cluster competitor evaluation texts using the BERTopic model and construct a competitive impact index to capture real-time market dynamics and obtain trend forecasts for product sales data.

[0055] Improve the demand module to construct a product association weight matrix based on product sales data trend forecasts, quantify the market competition correlation and sales complementarity correlation between different product categories, and obtain a more complete product demand forecast.

[0056] The product operation module is used to dynamically adjust the basic operation strategy of each product based on the product association weight matrix, and generate the optimal product operation management plan that fits the product demand forecast.

[0057] The product demand forecasting method and system based on multimodal machine learning provided by this invention forecasts product demand by constructing an initial product operation management model that considers competitive data, and then refines and adjusts it. The beneficial effects achieved are as follows:

[0058] This invention integrates data from multiple modalities, including promotional feature data, multimodal product data, competitive environment data, and user behavior data, enabling the model to comprehensively understand the market performance and potential demand of products from multiple perspectives.

[0059] This invention utilizes various models, including the DeepAR model, the FTRL algorithm, and the BERTopic model, each playing a unique role at different stages. The DeepAR model is used to construct the initial model for product operation management, capturing relevant characteristics and trends. The FTRL algorithm optimizes and adjusts model parameters based on competitive data, enabling it to better adapt to market changes. The BERTopic model focuses on competitor evaluation text analysis, constructing a competitive impact index to assist in grasping market dynamics. The collaborative work of multiple models enhances the scientific rigor and effectiveness of the entire prediction process. The FTRL algorithm fine-tunes the initial model parameters and incorporates newly generated competitive data in real time, allowing the model to adapt promptly to market changes, such as changes in competitor sales, price adjustments, and new promotional activities. The model can adjust prediction results based on the latest market dynamics, thereby better reflecting the actual changing trends in product demand and improving the timeliness and accuracy of predictions.

[0060] This invention focuses on multiple optimization objectives: improving sales performance, optimizing inventory turnover, and enhancing customer satisfaction. These objectives are transformed into quantifiable expressions and weighted accordingly. Simultaneously, the relational features extracted from the product association weight matrix are converted into constraints. The resulting operational management solution considers multiple core business needs, avoiding overemphasis on any single aspect while neglecting other important factors. Throughout the process, dynamic adjustments are continuously made based on real-time data feedback. Both model parameters and the operational management solution can flexibly adapt to market changes, ensuring that the final product operational management solution aligns with product demand forecasts and reflects actual market dynamics. This contributes to improving the overall competitiveness and operational efficiency of products in the market.

[0061] Based on user behavior data and algorithms for tracking and consuming in unrelated sectors, businesses can provide consumers with more personalized shopping experiences. By monitoring market dynamics and consumer feedback in real time, businesses can promptly understand changes in customer needs and problems. The results of product demand forecasting and the implementation of operational management plans can drive businesses to continuously optimize their products and services. Based on market demand and consumer feedback, businesses can continuously improve product features, enhance product quality, and optimize after-sales service, thereby better meeting customer needs and increasing customer satisfaction. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0063] Figure 1 This is a flowchart illustrating the commodity demand forecasting method based on multimodal machine learning provided in an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the process for obtaining trend predictions of commodity sales data provided in an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram of the modules of the commodity demand prediction system based on multimodal machine learning provided in an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0067] The following is combined with Figures 1-3 This invention describes a commodity demand forecasting method and system based on multimodal machine learning.

[0068] Figure 1 This is a flowchart illustrating the commodity demand forecasting method based on multimodal machine learning provided in this embodiment of the invention.

[0069] like Figure 1 As shown in the embodiments of the present invention, the commodity demand forecasting method and system based on multimodal machine learning is executed by a commodity demand forecasting method based on multimodal machine learning, which includes:

[0070] S1: Collect product data from e-commerce platforms and preprocess it to obtain preprocessed product data.

[0071] Product data includes: promotional feature data, product multimodal data, competitive environment data, and user behavior data.

[0072] Promotional data includes activity type (spend-and-refund / discount / flash sale), activity strength (discount rate / spend-and-refund threshold), activity duration, and promotional channels. Product multimodal data includes product title / details / promotional copy, main image / product display image, and price fluctuation series. Competitive environment data includes competitor price changes, competitor promotional strategies, user sentiment in reviews, and social media mentions.

[0073] User behavior data includes the sequence of clicks / add-to-cart / favorites during the event, historical promotional participation, and purchase conversion rate.

[0074] Data preprocessing includes: extracting promotional keywords from text data using TF-IDF and constructing sentiment polarity features; extracting visual features of promotional posters from image data using a pre-trained ResNet model; standardizing sales / price series data for time-series data and labeling them with promotional event timestamps; constructing a fusion feature matrix, aligning the features of each modality according to timestamps, and forming a multidimensional dataset.

[0075] S2: Extract the correlation features of preprocessed product data in different scenarios, construct a comprehensive evaluation index system for e-commerce products, and classify the features of preprocessed product data based on the comparative consumption algorithm to obtain the feature product data classification.

[0076] The specific steps for constructing a comprehensive evaluation index system for e-commerce products are as follows:

[0077] Determine the dimensions of the evaluation indicators, such as product quality, user experience, sales performance, and market competitiveness.

[0078] Product quality is a fundamental dimension for evaluating products, including durability, functional completeness, and design. User experience refers to the user's feelings during the purchase and use of the product, including interface friendliness, ease of operation, and after-sales service. Sales performance reflects the market acceptance of the product, including sales revenue, sales volume, and inventory turnover. Market competitiveness assesses the product's position in the market, including market share, brand influence, and price competitiveness.

[0079] For each dimension, we selected quality score (product quality), positive user review rate, page loading speed (user experience), and after-sales service satisfaction as evaluation indicators.

[0080] The evaluation indicators are weighted and the weight of each indicator is determined by the analytic hierarchy process.

[0081] A hierarchical model is constructed, with a comprehensive evaluation of the target e-commerce products at the highest level, product quality, user experience, sales performance, and market competitiveness at the middle level, and specific indicators at the lowest level. Pairwise comparisons are performed on each indicator to construct a judgment matrix, and consistency checks are conducted to determine the weight of each indicator.

[0082] The values ​​of each indicator are multiplied by their respective weights and then summed to obtain the comprehensive evaluation index system for e-commerce products.

[0083] Competitive consumption refers to the phenomenon where consumers are influenced by the consumption levels of those around them when making purchases, and tend to buy goods that are at a higher level than their own or comparable to those of others.

[0084] The specific steps for feature segmentation based on the algorithm for competitive consumption are as follows:

[0085] Extract key features of price, brand, and quality from product data and analyze how these features are related to competitive consumption behavior.

[0086] The algorithm for categorizing pre-processed product data uses a comparative consumption approach. Based on the product's price, brand, quality characteristics, and the influence of consumers' competitive mentality, the algorithm divides products into different categories, each corresponding to different comparative consumption characteristics.

[0087] The resulting product categories are evaluated to check their rationality and accuracy. Based on actual sales data and user feedback, the parameters of the conspicuous consumption algorithm are adjusted to improve the accuracy and usability of the classification.

[0088] Based on the evaluation results, the classification process is continuously iterated and improved to ensure that the classification results can better reflect the characteristics of conspicuous consumption behavior.

[0089] S3: Use the DeepAR model to construct an initial model for product operation management based on the feature classification results, use the FTRL algorithm to analyze the feature-classified product data, and fine-tune the parameters of the initial model for product operation management based on the newly generated competitive data.

[0090] The product data is sorted out according to different categories of features, such as distinguishing static features (e.g., product category, brand), time-series features (e.g., historical sales volume, price fluctuation data over time), and other multimodal features (e.g., promotional copy text, poster image related features), to ensure that the data of each feature is in an accurate and complete format and that there are no missing or outlier values ​​that may interfere with subsequent processing.

[0091] One-hot encoding is used to transform categorical features into vector representations, enabling them to be effectively processed by the model. Numerical time-series features are standardized or normalized. Appropriate deep learning techniques are used to extract and encode features from multimodal features such as text and images. For promotional copy, word vector models are used to transform the text into a sequence of word vectors, followed by pooling and other operations to obtain fixed-dimensional text feature vectors. For image features, convolutional neural networks are used to extract key features, transforming the image into a feature vector representation. Finally, these feature vectors from different modalities are integrated to adapt their dimensions and format to the input requirements of the DeepAR model.

[0092] In the DeepAR model, a dedicated static feature processing layer is set up. The encoded static features are input into the static feature processing layer, and these fixed attributes are initially mapped and fused to learn their inherent correlations and output the processed static features.

[0093] A time series processing layer based on a long short-term memory network is constructed. The standardized time series features are input into the time series processing module step by step. The sequential dependency relationship of the time series data is captured by a recurrent neural network, and the trends and periodicity of sales over time are learned. The encoded time series feature vector is output, which contains rich time series information and reflects the status changes of the product at different time points.

[0094] The global attention mechanism is used to identify different promotion types, specifically long-term impact patterns such as 618 and Double 11 promotions and daily discounts. By calculating the attention weights of relevant features at different time points and for different promotional activities, it highlights key factors that have a long-term impact on product operations, such as the impact of certain large-scale promotional activities on the sales trend of products in the following months. These key features are then weighted and fused before being output.

[0095] The local attention mechanism focuses on capturing the short-term sales impulse effect during the event period, specifically the hour before the flash sale and the limited-time purchase period. It can accurately capture the instantaneous stimulating effect of factors such as promotional activities and consumer behavior on product sales within a specific short period of time. By calculating attention weights, the features of the corresponding time period are weighted and integrated to output a feature vector that reflects key information about short-term sales fluctuations.

[0096] A multimodal fusion gating component is created, which takes the previously processed multimodal features such as promotional copy attractiveness-related features and poster click-through rates, as well as the output vectors of static and temporal features after processing by an attention mechanism, as input. The influence weights of each modality feature are dynamically adjusted by calculating gating weights based on the Sigmoid function, so as to achieve adaptive fusion of different modal features in different scenarios. Finally, a comprehensive feature vector that integrates multiple key information is output, which fully reflects the comprehensive influence of various factors on the product during operation.

[0097] The prepared encoded feature data is divided into training and validation sets to ensure that the data distribution of the training and validation sets is similar and can represent the characteristics and patterns of the overall data. For example, the division can be carried out in chronological order to avoid data leakage. This allows the model to learn data features reasonably during training and evaluate its performance through the validation set.

[0098] Based on the objectives of the product operation management model, select an appropriate loss function. For example, for numerical objectives such as predicting product sales, use mean squared error as the loss function to measure the error between the predicted value and the actual value.

[0099] The Adam optimizer is used to update model parameters. Based on factors such as data characteristics and model complexity, the learner's hyperparameters, such as learning rate and momentum, are configured and adjusted appropriately according to the training results.

[0100] The DeepAR model is trained multiple times using the training set. In each training round, the data is input into the model, the prediction result is obtained through forward propagation, the loss value is calculated through the loss function, and then the optimizer is used to backpropagate and update the model parameters based on the loss value. This process is repeated continuously to make the model gradually fit the data features and reduce the loss value.

[0101] After each training round, the model is evaluated using a validation set. The changes in the model's loss value and the performance of other evaluation metrics on the validation set are observed to determine whether the model is overfitting or underfitting. Based on the evaluation results, the model's hyperparameters are adjusted in a timely manner to optimize the model's performance until the model achieves satisfactory results on the validation set. This completes the training and initialization of the initial model for commodity operation management, resulting in an initial model capable of making commodity operation-related predictions based on input features.

[0102] The specific steps for constructing an initial model for commodity operation management based on feature category segmentation results using the DeepAR model are as follows:

[0103] Collect newly generated competitive data daily, including information such as competitor sales, price adjustments, and newly launched promotional activities. Integrate this data with existing product feature data in an appropriate manner, construct comparative features by comparing competitor sales data with your own product sales data, and incorporate competitor promotional activities into multimodal features in the form of category codes to form a comprehensive feature dataset containing competitive information.

[0104] Define key parameters for the FTRL algorithm, such as the learning rate and regularization parameters. Based on the objectives of product operations management, such as accurately predicting product sales and optimizing profits, construct a corresponding objective function. If predicting sales volume, the sum of squared differences between predicted and actual sales volume can be used as the objective function. The objective function should reasonably measure the gap between the model's prediction and the actual expected result.

[0105] The previously constructed initial model for product operations management is linked to the FTRL algorithm and objective function, enabling the FTRL algorithm to adjust the parameters of the initial model based on the objective function. The parameters of the initial model are set as the variables to be optimized by the FTRL algorithm, allowing the algorithm to update these parameters by analyzing the product data and competitor data after feature segmentation, in order to minimize the value of the objective function.

[0106] The comprehensive feature dataset containing competitive data is input into the initial model associated with the FTRL algorithm in batches of 100 data records each. Forward propagation calculation is performed to obtain the prediction results corresponding to each batch of data, such as predicted product sales and profits.

[0107] Based on the prediction results obtained from forward propagation and the constructed objective function, the gradients of the model parameters with respect to the objective function are calculated. Then, the FTRL algorithm is used to update the parameters of the initial model based on these gradients. The magnitude of the adjustment is determined by both the parameters of the FTRL algorithm and the magnitude of the gradients. This process is repeated to iteratively update the model parameters, enabling the model to better adapt to market changes after incorporating new competitive data, thereby improving prediction accuracy and operational management effectiveness.

[0108] S4: Cluster competitor evaluation texts using the BERTopic model and construct a competitive impact index to capture real-time market dynamics and obtain trend predictions for product sales data.

[0109] Collect user review text data of competitors, clean and preprocess it, including removing stop words and punctuation marks, and performing word form restoration, to obtain a clean text dataset.

[0110] The BERTopic model was used to perform cluster analysis on the preprocessed competitor review texts. The model automatically identified potential topics in the texts and grouped reviews that expressed similar views or discussed the same content into one category, resulting in different review topic clusters.

[0111] Based on the content characteristics of each thematic cluster and combined with market feedback data, key factors affecting product competitiveness are extracted, and corresponding weights are assigned to each key factor.

[0112] Based on the size, sentiment, and weight of key factors of each thematic cluster, the competitive influence index of different competing products is calculated. The index value reflects the size of the competitive influence of the competing product in the market.

[0113] Establish a correlation model between the competition impact index and product sales data. By monitoring the updates of competitor evaluation texts in real time, the competition impact index is dynamically updated to capture dynamic information such as changes in consumer preferences and shifts in competitors' advantages and disadvantages in the market.

[0114] The updated competition impact index and market dynamics data are input into the correlation model, and combined with historical sales data trends, analysis and calculations are performed to obtain future trend predictions for commodity sales data.

[0115] S5: Construct a product association weight matrix based on product sales data trend forecasts, quantify the market competition correlation and sales complementarity correlation between product categories, and obtain a more complete product demand forecast.

[0116] S51: Construct indicators based on market share changes and correlation values ​​to determine the degree of competitive correlation.

[0117] For each product category, its market share at different times is calculated based on sales data trends. The changes in market share of each product category over time are observed, and it is analyzed which categories show an upward trend in market share and which show a downward trend, thereby making a preliminary judgment on the competitive situation between them.

[0118] Spearman's correlation coefficient is used to analyze the correlation between sales volume or sales revenue data of different product categories, calculating the correlation values ​​between each pair of product categories. A high positive correlation indicates that they may have a strong substitution relationship in market competition and are direct competitors. A negative correlation or a weak correlation may indicate that the competitive relationship is not obvious.

[0119] Taking into account changes in market share and the results of correlation analysis, a specific market competition correlation index can be constructed. For example, a weighted comprehensive calculation formula can be set to add or multiply the correlation coefficient and the change in market share according to certain weights, so as to obtain a value that can quantitatively reflect the degree of market competition correlation between each pair of product categories. The value can be set between 0 and 1, with a larger value indicating a stronger competitive correlation.

[0120] S52: By examining the sales data trends of different product categories, we can identify sales volume and then use association rule mining algorithms to discover sets of product categories that are frequently purchased simultaneously, thus constructing a complementary association index.

[0121] Carefully examine the sales data trends of different product categories to identify combinations of product categories whose sales volume or sales revenue shows synchronous increases or decreases. These combinations often indicate complementary sales relationships. For example, when purchasing an electronic product, it's common to also buy related accessories, and their sales data tends to fluctuate synchronously. Use association rule mining algorithms to analyze sales transaction records and other data to identify sets of product categories that are frequently purchased together. Further determine the complementary relationships between product categories from a purchasing behavior perspective and analyze the frequency of sales pairings. Combine the changes in sales synergy with the association rule mining results to construct a sales complementarity correlation index. For example, weights can be assigned based on factors such as the frequency of simultaneous purchases and the degree of synchronous changes in sales data to calculate a value that reflects the degree of complementary sales correlation between product categories. The value can be set between 0 and 1, with a higher value indicating a stronger complementary correlation.

[0122] S53: Organize the calculated market competition correlation and sales complementarity correlation values ​​between each product category into a matrix. The rows and columns of the matrix correspond to different product categories, and the value in each cell is the correlation value between the corresponding pair of product categories. This forms the product correlation weight matrix.

[0123] S54: Integrate the constructed product association weight matrix into the existing product demand forecasting model. Add the relevant values ​​from the association weight matrix as new independent variables, or in a machine learning forecasting model, input them as part of the feature vector. This allows the model to utilize the association information between product categories for more accurate predictions. Use the forecasting model containing the association weight matrix information to simulate product demand forecasting, then compare the results with actual sales data to verify and evaluate whether the prediction accuracy has improved. If the accuracy improves significantly, it indicates that quantifying the association degree by constructing the product association weight matrix has effectively improved product demand forecasting. If the effect is not ideal, further analysis and adjustments to the construction of association degree indicators are needed to optimize the entire process until a satisfactory and improved product demand forecasting result is obtained.

[0124] S6: Dynamically adjust the basic operational strategies of each product based on the product association weight matrix to generate the optimal product operation management plan that fits the product demand forecast.

[0125] The core objective function for multi-objective optimization is determined, and sales performance improvement, inventory turnover optimization, and customer satisfaction improvement are transformed into quantifiable mathematical expressions. Initial weights are then assigned to each objective based on business priorities.

[0126] Based on the product association weight matrix, the competitive and complementary relationship characteristics between product categories are extracted and transformed into constraints. For example, products with high competitive correlation should avoid direct conflict in promotional strategies, while products with high complementary correlation should coordinate the adjustment of inventory levels and pricing strategies. At the same time, combined with the product demand forecast results, the range of basic operational strategy parameters for each product is clarified.

[0127] Then, a suitable multi-objective optimization algorithm framework is selected, and the objective function, constraints, and strategy parameters are input into the algorithm. A set of basic operational strategy schemes is generated by initializing the population. The fitness values ​​of each scheme on the three objectives are calculated, and selection, crossover, and mutation operations are performed based on non-dominated sorting and crowding distance to iteratively optimize the strategy parameters.

[0128] During the iteration process, real-time sales data and feedback on demand forecast deviations are dynamically introduced to adjust the weights of each target and the strength of constraints. For example, when the actual sales of a certain type of product are lower than the forecast, the weight of its sales performance target is increased, while the constraint threshold of inventory turnover is relaxed to ensure that the optimization direction is in line with the actual market.

[0129] From the Pareto optimal solution set obtained through optimization, the solution with the best overall performance is selected based on business scenario requirements. Specific operational strategy adjustment suggestions are output to form the final product operation management solution. The effectiveness of the solution is verified through small-scale pilot testing, and after fine-tuning based on feedback, it is fully implemented.

[0130] In summary, this embodiment provides a commodity demand forecasting method based on multimodal machine learning. By constructing an initial commodity operation management model considering competitive data to forecast commodity demand, and then refining and adjusting it, the beneficial effects achieved are as follows:

[0131] This invention integrates data from multiple modalities, including promotional feature data, multimodal product data, competitive environment data, and user behavior data, enabling the model to comprehensively understand the market performance and potential demand of products from multiple perspectives.

[0132] This invention utilizes various models, including the DeepAR model, the FTRL algorithm, and the BERTopic model, each playing a unique role at different stages. The DeepAR model is used to construct the initial model for product operation management, capturing relevant characteristics and trends. The FTRL algorithm optimizes and adjusts model parameters based on competitive data, enabling it to better adapt to market changes. The BERTopic model focuses on competitor evaluation text analysis, constructing a competitive influence index to assist in grasping market dynamics. The collaborative efforts of multiple models enhance the scientific rigor and effectiveness of the entire prediction process.

[0133] Based on the same general inventive concept, this invention also protects a big data-based reverse express delivery status identification and tracking management system. The big data-based reverse express delivery status identification and tracking management system provided by this invention will be described below. The big data-based reverse express delivery status identification and tracking management system described below can be referred to in correspondence with the AI-based turtle polypeptide library construction method described above.

[0134] Figure 3 This is a schematic diagram of the modules of the commodity demand prediction system based on multimodal machine learning provided in an embodiment of the present invention.

[0135] like Figure 3 As shown, the commodity demand forecasting system based on multimodal machine learning includes:

[0136] The data acquisition module is used to collect product data from e-commerce platforms and preprocess it to obtain preprocessed product data.

[0137] The product feature extraction module is used to extract product features from preprocessed product data in different scenarios, construct comprehensive evaluation indicators for e-commerce products, and classify the features of preprocessed product data based on the comparative consumption algorithm to obtain feature product data classification.

[0138] The model building module is used to construct an initial model for commodity operation management based on the feature-based segmentation of commodity data using the DeepAR model, analyze the feature-based segmentation of commodity data using the FTRL algorithm, and adjust the parameters of the initial model for commodity operation management based on newly generated competitive data.

[0139] The sales forecasting module is used to cluster competitor evaluation texts using the BERTopic model and construct a competitive impact index to capture real-time market dynamics and obtain trend forecasts for product sales data.

[0140] Improve the demand module to construct a product association weight matrix based on product sales data trend forecasts, quantify the market competition correlation and sales complementarity correlation between different product categories, and obtain a more complete product demand forecast.

[0141] The product operation module is used to dynamically adjust the basic operation strategy of each product based on the product association weight matrix, and generate the optimal product operation management plan that fits the product demand forecast.

[0142] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, are embodied in the form of a software product. This computer software product is stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A commodity demand forecasting method based on multimodal machine learning, characterized in that, include: S1: Collect product data from e-commerce platforms and preprocess it to obtain preprocessed product data; S2: Extract product features from preprocessed product data under different scenarios, construct comprehensive evaluation indicators for e-commerce products, and classify the features of preprocessed product data based on the comparative consumption algorithm to obtain feature product data classification; S3: Use the DeepAR model to build the initial model for product operation management, use the FTRL algorithm to analyze features and segment product data, and adjust the parameters of the initial model for product operation management based on newly generated competitive data; S4: Cluster competitor evaluation texts using the BERTopic model and construct a competitive impact index to capture real-time market dynamics and obtain trend predictions for product sales data; S5: Construct a product association weight matrix based on the product sales data trend prediction, quantify the market competition association and sales complementarity association between product categories, and obtain a complete product demand prediction. S6: Based on the product association weight matrix, dynamically adjust the basic operation strategy of each product to generate the optimal product operation management plan that fits the product demand forecast.

2. The commodity demand forecasting method based on multimodal machine learning according to claim 1, characterized in that, In step S2, the specific steps for constructing a comprehensive evaluation index system for e-commerce products are as follows: The dimensions of the evaluation indicators are determined as product quality, user experience, sales performance, and market competitiveness. We selected quality rating under the product quality dimension, positive user review rate, page loading speed under the user experience dimension, and after-sales service satisfaction as evaluation indicators. The evaluation indicators are weighted and the weights of the evaluation indicators are determined by the analytic hierarchy process. The comprehensive evaluation index for the product is calculated based on the evaluation indicators and their weights.

3. The commodity demand forecasting method based on multimodal machine learning according to claim 1, characterized in that, In step S3, the specific steps for constructing the initial model for commodity operation management using the DeepAR model based on the feature category classification results are as follows: The preprocessed commodity data features are divided into static feature data, time series feature data, and multimodal feature data. The static feature data is encoded using one-hot encoding, the numerical time series feature data is standardized, and the multimodal features are adapted to the DeepAR model. The static feature data is input into the static layer for processing, and the temporal features are input into the time series layer. Combining global and local attention mechanisms, a comprehensive feature vector is output. The product feature vectors are divided into training and validation sets according to the time sequence. The mean squared error is selected as the loss function according to the objective of the commodity operation management model. The Adam optimizer is used and its hyperparameters are configured. The model is trained in multiple rounds using the training set, and the prediction results and loss values ​​of each round are output. The prediction results are evaluated using a validation set to determine whether the loss value is greater than a preset range. If so, the hyperparameters are adjusted to obtain the initial model for commodity operation management.

4. The commodity demand forecasting method based on multimodal machine learning according to claim 1, characterized in that, In step S3, the specific steps for adjusting the initial model parameters of commodity operation management are as follows: By incorporating newly generated competitive data into the preprocessed product data features, a comprehensive feature dataset is constructed by combining competitor information with various product features. Set the learning rate and regularization for the FTRL variant to update the parameters of the initial model for product operations management; The initial model is associated with the FTRL algorithm and the objective function, so that the algorithm is guided by optimizing the objective function and uses integrated data to update the parameters of the initial model; The comprehensive feature dataset is input into the initial model. After forward propagation to obtain the prediction results, the gradient is calculated. The FTRL algorithm is used to update the model parameters according to the gradient. The model is iterated multiple times to adapt to the changes in competitive data until the model is greater than the preset threshold on the validation set, thus obtaining the commodity operation management model.

5. The commodity demand forecasting method based on multimodal machine learning according to claim 1, characterized in that, In step S4, the specific steps for obtaining the trend forecast of commodity sales data are as follows: Collect user review text data of competitors, preprocess it, and obtain a clean text dataset; The BERTopic model was used to perform cluster analysis on the preprocessed competitor review texts to obtain different review topic clusters; Based on the content characteristics of the evaluation topic clusters and combined with market feedback data, key factors affecting product competitiveness are extracted, and weights are set for these key factors. The competitive impact index of different competing products is calculated based on the size, sentiment, and weight of key factors of the evaluation topic cluster. Establish a correlation model between the competition impact index and product sales data, and dynamically update the competition impact index by monitoring the updates of competitor evaluation texts in real time; The updated competition impact index and market dynamics data are input into the correlation model, and the future trend of commodity sales data is predicted by analyzing historical sales data trends.

6. The commodity demand forecasting method based on multimodal machine learning according to claim 1, characterized in that, In step S5, the specific steps for improving commodity demand forecasting are as follows: S51: Construct indicators based on market share changes and correlation values ​​to determine the degree of competitive correlation; S52: Construct complementary correlation indicators based on sales data trends of different product categories and products that are frequently purchased simultaneously; S53: Construct a product association weight matrix based on the aforementioned competitive association degree and complementary association degree indicators; S54: Input the product association weight matrix into the product operation management model to improve product demand forecasting.

7. The commodity demand forecasting method based on multimodal machine learning according to claim 6, characterized in that, In step S51, the specific steps for analyzing the correlation between market competition and economics are as follows: For each product category, calculate its market share at different times based on sales data trends. Analyze the changes in market share of each product category over time to determine the competitive landscape between them. The Spearman correlation coefficient was used to conduct a correlation analysis on the sales of different product categories, and the correlation values ​​between each pair of product categories were calculated. Based on changes in market share and correlation values, specific market competition correlation indicators are constructed to obtain the degree of market competition correlation between each pair of product categories.

8. The commodity demand forecasting method based on multimodal machine learning according to claim 6, characterized in that, In step S53, the specific steps for constructing the product association weight matrix are as follows: organize the market competition association degree and sales complementarity association degree values ​​between each pair of product categories into a matrix. The rows and columns of the matrix correspond to different product categories, and the value in each cell is the association degree value between the corresponding pair of product categories, thus obtaining the product association weight matrix.

9. The commodity demand forecasting method based on multimodal machine learning according to claim 1, characterized in that, In step S6, the specific steps for generating the optimal product operation management plan that aligns with product demand forecasts are as follows: Quantify and assign initial weights to improve sales performance, optimize inventory turnover, and enhance customer satisfaction; Using a multi-objective optimization algorithm, the relationship characteristics and constraints are clarified based on the product association weight matrix, and the range of operational strategy parameters for each product is determined. Generate basic operational strategy plans, calculate fitness values, and iteratively optimize parameters through selection, crossover, and mutation operations; By introducing real-time sales data feedback, adjusting weights and the strength of constraints, and selecting the comprehensive optimal solution from the set of optimal solutions, an optimal product operation and management plan is formed.

10. A commodity demand forecasting system based on multimodal machine learning, applied to the commodity demand forecasting method based on multimodal machine learning as described in any one of claims 1 to 9, characterized in that, The commodity demand forecasting system includes: The data acquisition module is used to collect product data from e-commerce platforms and preprocess it to obtain preprocessed product data. The product feature extraction module is used to extract product features from preprocessed product data under different scenarios, construct comprehensive evaluation indicators for e-commerce products, and classify the features of preprocessed product data based on the comparative consumption algorithm to obtain feature product data classification. The model building module is used to build an initial model for commodity operation management by using the DeepAR model to divide commodity data according to features, using the FTRL algorithm to analyze features to divide commodity data, and adjusting the parameters of the initial model for commodity operation management based on newly generated competitive data. The sales forecasting module is used to cluster competitor evaluation texts using the BERTopic model and construct a competitive impact index to capture real-time market dynamics and obtain trend forecasts for product sales data. The demand module is improved to construct a product association weight matrix based on the product sales data trend prediction, quantify the market competition association and sales complementarity association between product categories, and obtain improved product demand prediction. The product operation module is used to dynamically adjust the basic operation strategy of each product based on the product association weight matrix, and generate the optimal product operation management plan that fits the product demand forecast.

Images