Commodity sales prediction method and system based on improved neural network

Abstract

The invention discloses a commodity sales volume prediction method and system based on an improved neural network, and relates to the technical field of artificial intelligence, and the method comprises the steps: constructing a dynamic heterogeneous graph network, calculating the time sequence dynamic association strength, and quantifying the correlation between a node relation and sales volume; based on a self-adaptive space-time attention mechanism, a seasonal mode and a sudden mode in the sales volume data are identified, and the influence of emergencies on the sales volume is quantified; applying a bidirectional causal analysis framework to evaluate the direct, indirect and interactive influence of each node on the sales volume; generating a cross-domain commodity embedding representation based on an evaluation result, and establishing an association bridging mechanism of long-tail commodities and hot-sell commodities; and carrying out sales volume prediction through a multi-level cooperation mechanism, and outputting a sales volume prediction result of the target commodity. According to the method, a time sensing gating mechanism and a heterogeneous graph network are combined, modeling is carried out through a commodity-user-promotion-event quaternary relationship, a dynamic change rule among sales volume influence factors is effectively captured, and the adaptability of a prediction model to market fluctuation is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for predicting commodity sales based on an improved neural network. Background Technology

[0002] Existing methods for predicting commodity sales primarily rely on single-structure neural network models, which are ineffective at handling multi-source heterogeneous data, leading to insufficient identification of sales patterns in complex market environments. Traditional methods have significant limitations in dealing with seasonal fluctuations and the impact of unforeseen events, struggling to distinguish the direct and indirect effects of various factors on sales, and lacking reasonable causal explanation mechanisms. Furthermore, current technologies use a uniform model framework for best-selling and long-tail products, ignoring the correlations and differences between different types of products, and failing to fully utilize the sales patterns of best-selling products to guide the prediction of long-tail products.

[0003] Meanwhile, current methods for predicting commodity sales have significant shortcomings in areas such as multi-dimensional data collaborative analysis, cross-domain commodity knowledge sharing, and the integration of multi-level sales influencing factors, making it difficult to adapt to the complex sales models and market fluctuations in the actual retail environment. Summary of the Invention

[0004] This invention provides a product sales forecasting method based on an improved neural network, which addresses the technical problems of existing technologies in multi-source heterogeneous data processing, time-series pattern recognition, causal relationship analysis, and long-tail product forecasting.

[0005] In view of this, the first aspect of the present invention provides a product sales forecasting method based on an improved neural network, comprising: constructing a dynamic heterogeneous graph network including product nodes, user nodes, promotional activity nodes, and external event nodes, and calculating the temporal dynamic correlation strength between nodes to quantify the correlation between node relationships and sales; identifying seasonal and sudden patterns in sales data based on an adaptive spatiotemporal attention mechanism to quantify the impact of sudden events on sales; applying a bidirectional causal analysis framework to simultaneously perform forward prediction and backward attribution to evaluate the direct, indirect, and interactive effects of each node on sales; generating a cross-domain product embedding representation based on the evaluation results to establish a bridging mechanism between long-tail products and best-selling products; and using the product embedding representation and the bridging mechanism to perform sales forecasting through a multi-level collaborative mechanism to output the sales forecast result of the target product.

[0006] Optionally, the construction process of the dynamic heterogeneous graph network includes: constructing a four-element heterogeneous graph network of product-user-promotion-event, and optimizing the connection strength between nodes through a dynamic weight adjustment algorithm; mapping node entities to a low-dimensional semantic space, and using a multi-head attention mechanism to capture the interaction relationships between different types of nodes; establishing a temporal dependency model between nodes, and quantifying the dynamic changes in association strength through time-aware gated recursive units; performing association strength evaluation based on causal inference, and calculating the direct and indirect contribution values ​​of each node to sales through counterfactual analysis.

[0007] Optionally, identifying seasonal and sudden patterns in sales data includes: constructing a multi-scale time series decomposition network to split sales data into trend, seasonal, and residual components; designing an adaptive time window mechanism to dynamically adjust the observation time span and capture seasonal patterns in different periods; using anomaly detection algorithms to identify sudden fluctuations in sales data and determining the occurrence time of sudden events based on statistical significance tests; constructing an event-sales response model to analyze the magnitude, duration, and recovery pattern of sales changes before and after sudden events; designing a dual-channel attention network, with one channel focusing on seasonal patterns and the other on sudden patterns, and dynamically balancing the importance weights of the two patterns through a gating mechanism; and constructing a sales impact response function to quantify the short-term and long-term impacts of different types of sudden events on sales, forming a scoring system for the impact of sudden events.

[0008] Optionally, applying a two-way causal analysis framework includes: establishing a causal graph structure for sales forecasting and constructing a neural network architecture for two-way information flow; employing causal discovery algorithms to identify potential causal paths and hidden interaction effects; building a counterfactual prediction model to evaluate the independent impact of changes in a single node on sales through intervention testing; performing path analysis methods to decompose and quantify the direct and indirect effects of each node on sales through different paths; and employing an interaction term detection mechanism to identify synergistic or inhibitory effects generated by combinations of multiple nodes.

[0009] Optionally, establishing a bridging mechanism between long-tail products and best-selling products includes: constructing product feature vectors, which contain the product's intrinsic attributes, historical sales data, and causal analysis evaluation results; designing a cross-domain representation learning network to map products with different sales performances to the same semantic space; employing a contrastive learning method to reduce the distance between best-selling products and long-tail products with similar features in the embedding space; and constructing a product similarity calculation matrix to quantify the degree of association between products based on multi-dimensional feature weighting.

[0010] Optionally, establishing a bridging mechanism between long-tail products and best-selling products also includes: performing cluster analysis to identify product groups with similar sales patterns and influencing factors; establishing a knowledge transfer channel from best-selling products to long-tail products; generating a product association network, representing the bridging relationship between best-selling products and long-tail products through a graph structure, and calculating the bridging strength.

[0011] Optionally, sales forecasting through a multi-level collaborative mechanism includes: constructing an integrated commodity sales forecasting framework, comprising two core components: a commodity feature layer and a market association layer; in the commodity feature layer, training a basic commodity sales forecasting model based on commodity embedding representations and historical sales data; and in the market association layer, integrating category information, market environment factors, and inter-commodity relationships to construct a sales adjustment model.

[0012] The second aspect of this invention provides a product sales forecasting system based on an improved neural network, comprising: a dynamic graph construction module for constructing a dynamic heterogeneous graph network containing product nodes, user nodes, promotional activity nodes, and external event nodes, and calculating the temporal dynamic correlation strength between nodes to quantify the correlation between node relationships and sales; a temporal pattern recognition module for identifying seasonal and sudden patterns in sales data based on an adaptive spatiotemporal attention mechanism to quantify the impact of sudden events on sales; a bidirectional causal analysis module for applying a bidirectional causal analysis framework to simultaneously perform forward prediction and backward attribution to evaluate the direct, indirect, and interactive impacts of each node on sales; a product representation learning module for generating cross-domain product embedding representations based on the evaluation results and establishing a bridging mechanism between long-tail products and best-selling products; and a sales forecasting module for using product embedding representations and the bridging mechanism to perform sales forecasting through a multi-level collaborative mechanism and outputting the sales forecast results for the target product.

[0013] The beneficial effects of this invention are as follows: Firstly, it innovatively combines a time-aware gating mechanism with heterogeneous graph networks. By modeling a four-element relationship between product, user, promotion, and event, it effectively captures the dynamic changing patterns among factors influencing sales, enhancing the predictive model's adaptability to market fluctuations. Secondly, it innovatively designs a dual-channel attention architecture for seasonality and sudden events, enhancing the model's ability to identify both regular and sudden sales fluctuations through adaptive gating to balance the weights of the two modes. Thirdly, it proposes a bidirectional causal inference method combining counterfactual prediction and path analysis, improving the model's explanatory power for sales impact mechanisms by quantitatively evaluating direct and indirect effects. Finally, it designs a cross-domain knowledge transfer mechanism from best-selling products to long-tail products, achieving product feature mapping through comparative learning to overcome the data sparsity problem of long-tail products. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart outlining a framework for a product sales forecasting method based on an improved neural network.

[0016] Figure 2 This is a flowchart illustrating the construction process of a dynamic heterogeneous graph network for a product sales prediction method based on an improved neural network.

[0017] Figure 3 This is a flowchart of a bidirectional causal analysis framework for a product sales forecasting method based on an improved neural network. Detailed Implementation

[0018] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. 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.

[0019] Example 1, referring to Figures 1-3 This is the first embodiment of the present invention, which provides a product sales prediction method based on an improved neural network. By constructing a dynamic heterogeneous graph network, introducing an adaptive spatiotemporal attention mechanism, applying a bidirectional causal analysis framework, and designing a multi-level collaborative mechanism, it achieves high-precision prediction of product sales. The framework flowchart is shown below. Figure 1 As shown, the details are as follows: S1: Construct a dynamic heterogeneous graph network and calculate the temporal dynamic correlation strength between nodes to quantify the correlation between node relationships and sales.

[0020] In a preferred embodiment of the present invention, the flowchart for constructing a dynamic heterogeneous graph network is as follows: Figure 2 As shown, it includes the following steps: S1.1: Collect basic product attributes, sales trajectory, user purchase behavior, promotional activity information, and external event records to form a multi-dimensional dataset.

[0021] S1.2: Perform adaptive sliding time window segmentation on the historical time series data of each node to identify correlation patterns at different time granularities.

[0022] Specifically, a multi-level time granularity library including hourly, daily, weekly, monthly, and quarterly levels is established, and an initial window length is set for each time granularity. An entropy fluctuation detection mechanism is designed, which calculates the information entropy within each time window in real time. When the entropy value change exceeds the entropy fluctuation threshold, the window is adjusted. This entropy fluctuation threshold is dynamically calculated based on the entropy value distribution characteristics of historical data. Periodic patterns, trend patterns, and sudden patterns are extracted under different time windows. The periodic pattern represents the recurring pattern of data, the trend pattern represents the overall direction of data change, and the sudden pattern represents the abnormal fluctuation characteristics of data.

[0023] S1.3: Construct a quaternary heterogeneous graph network of product-user-promotion-event, and optimize the connection strength between nodes through a dynamic weight adjustment algorithm.

[0024] In this step, a heterogeneous graph structure is first designed, mapping four types of nodes—products, users, promotional activities, and external events—and their attributes to different types of vertices and edges, defining the basic attribute characteristics of the nodes. Next, a time-decaying edge weight calculation formula is constructed, so that interactions that occur more recently have higher initial weights, and the weights decay with time intervals, with the decay rate set according to the relationship between different node types. Then, the basic association weights between nodes are calculated, including the intensity of user purchase behavior, the degree of influence of promotional activities, and the degree of association of external events, forming a complete heterogeneous graph network structure.

[0025] S1.4: Map node entities to a low-dimensional semantic space and use a multi-head attention mechanism to capture the interaction relationships between different types of nodes.

[0026] Specifically, the attribute features of different types of nodes are converted into vector representations of a unified dimension. An attention mechanism is used to adaptively weight the correlation strength between different types of nodes, thereby optimizing the feature extraction effect and improving the discriminativeness of heterogeneous node representations.

[0027] S1.5: Establish a temporal dependency model between nodes and quantify the dynamic changes in association strength through time-aware gated recursive units.

[0028] Furthermore, a dual-layer temporal memory structure is designed, comprising a short-term memory module and a long-term memory module, to capture node association patterns at different time scales. A time-interval-aware memory decay mechanism is constructed, adaptively adjusting the memory retention rate based on the time interval of node interactions, enabling the effective processing of non-uniformly sampled temporal data. A basic gating mechanism is implemented, including update and forget gates, whose parameters are dynamically adjusted through a time interval function to improve the memory capacity for long-interval events. A node interaction frequency adaptation mechanism is constructed, with optimization strategies designed for high-frequency and low-frequency interaction nodes respectively, to prevent high-frequency nodes from dominating the model learning process and to enhance sensitivity to rare but important interactions.

[0029] S1.6: Perform a correlation strength assessment based on causal inference, and calculate the direct and indirect contribution values ​​of each node to sales through counterfactual analysis.

[0030] Furthermore, a time-series processing do operator is constructed to separate direct and indirect effects by implementing intervention operations on specific nodes and generating interaction behavior sequences. Instrumental variable methods and matching methods are used to eliminate confounding factors, and a causal inference model is established to improve inference reliability. A Bayesian network path analysis model is designed to track the propagation path of influence and quantify mediating effects by establishing conditional probability relationships between nodes. The difference-in-differences method is used to evaluate the effectiveness of policy interventions and the impact of emergencies, generating a dynamic change sequence of correlation strength to achieve attribution analysis of changes in correlation strength. This method achieves accurate calculation of the sales contribution of each node through counterfactual analysis, effectively assessing the degree of direct and indirect influence.

[0031] Preferably, this invention, through a dynamic heterogeneous graph network structure and adaptive time window analysis, can effectively handle multi-source heterogeneous data and adapt to changes in sales patterns at different time scales. The quaternary heterogeneous graph network combined with a time decay weight mechanism improves the model's ability to model complex product sales relationships. The two-layer temporal memory structure solves the problem of processing non-uniformly sampled time-series data, enhancing the ability to capture long-term and short-term sales patterns. The association strength assessment method based on causal inference can distinguish the direct and indirect impacts of various factors on sales volume, improving the interpretability of prediction results and providing a more reliable decision-making basis for sales strategy optimization.

[0032] S2: Based on an adaptive spatiotemporal attention mechanism, it identifies seasonal and sudden patterns in sales data and quantifies the impact of sudden events on sales.

[0033] In a preferred embodiment of the present invention, the following steps are included: S2.1: Construct a multi-scale time series decomposition network to split sales data into trend components, seasonal components, and residual components.

[0034] Specifically, a multi-scale decomposition unit is designed, comprising deformable convolutional layers, attention-gated layers, and residual connections. It captures temporal patterns of different frequencies through parallel convolutional kernels of varying sizes. A non-parametric Bayesian structure is applied to dynamically determine the number of periodic patterns in the sales sequence through variational inference, automatically identifying multiple superimposed seasonal cycles. A trend extraction module based on causal convolution is designed, using dilated convolution to increase the receptive field and adaptive regularization to constrain the smoothness of trend components. An end-to-end component reconstruction mechanism is constructed to ensure that each decomposed component can accurately reconstruct the original sales sequence, and the information content allocation of each component is optimized through information bottleneck theory.

[0035] S2.2: Design an adaptive time window mechanism to dynamically adjust the observation time span and capture seasonal patterns of different cycles.

[0036] Specifically, an adaptive threshold-based time window generator is designed, and the optimal observation window length is determined through information entropy and autocorrelation analysis. A multi-scale sliding window mechanism is introduced to maintain observation windows of different lengths to capture seasonal patterns of different cycles such as daily, weekly, monthly, and quarterly. A window optimization module is constructed to dynamically adjust window parameters, including window length, sliding step size, and overlap ratio, based on sales fluctuation characteristics and prediction accuracy. A window attention mechanism is designed to assign different weights to different time windows to highlight important seasonal features. An adaptive sampling strategy is applied to increase sampling density at key time nodes to improve the accuracy of characterizing seasonal transition periods.

[0037] S2.3: Employ an anomaly detection algorithm to identify sudden fluctuations in sales data and determine the timing of the sudden event based on statistical significance testing.

[0038] Sudden fluctuations are identified through moving average differences and standard deviation thresholds. When the difference between actual sales and predicted sales exceeds a set multiple of the standard deviation, it is judged as an anomaly. The statistical significance test adopts the improved Mann-Whitney U test. By comparing the difference in sales distribution before and after the sudden event, the point where the sudden event occurs is confirmed when the p-value is less than the significance level (e.g., 0.05).

[0039] S2.4: Construct an event-sales response model to analyze the magnitude, duration, and recovery pattern of sales changes before and after a sudden event.

[0040] Specifically, a classification system for emergencies is established, hierarchically encoding emergencies based on event type, scope of impact, duration, and severity; a sales response network is designed, including an input layer, hidden layer, impact response layer, recovery dynamic layer, and output layer, to simulate the impact and recovery characteristics of different types of emergencies on sales; a sales elasticity assessment module is constructed, which analyzes the sensitivity differences of different product categories to the same event by calculating elasticity coefficients, forming a product-event sensitivity matrix, where the assessment dimensions include product category, price range, and sales channel differences; an adaptive recovery trajectory predictor is applied to dynamically generate sales recovery curves based on event characteristics and product attributes, predicting the complete process from impact to recovery; and a multi-event superposition effect analysis mechanism is implemented, modeling the composite impact of multiple simultaneous or consecutive emergencies on sales through nonlinear combination functions.

[0041] S2.5: Design a dual-channel attention network, with one channel focusing on seasonal patterns and the other on sudden patterns, and dynamically balance the importance weights of the two patterns through a gating mechanism.

[0042] Furthermore, a seasonal pattern recognition channel is constructed, employing a multi-head time-aware self-attention mechanism to capture periodic features at different time granularities, and extracting dominant periodic patterns through frequency-selective convolutional layers. A sudden event pattern recognition channel is also constructed, applying a jump recursive unit enhanced with change point detection to focus on discontinuous changes in sales sequences, and highlighting significant short-term fluctuations through an anomaly amplification mechanism. A cross-channel interaction mechanism is designed to allow information exchange between the seasonal and sudden event channels, capturing the perturbation effect of sudden events on seasonal patterns and the modulation effect of seasonal background on the intensity of sudden event impact. A context-based adaptive gating unit is designed to dynamically adjust the weight contributions of the two channels based on the current market environment, event intensity, and historical sales fluctuation patterns. A meta-learning mechanism is introduced, enabling the gating unit to learn the optimal weight allocation strategy from historical data and automatically adjust the weight calculation function for different product categories and market stages.

[0043] S2.6: Construct a sales impact response function to quantify the short-term and long-term impacts of different types of emergencies on sales and form an emergency impact rating system.

[0044] Furthermore, a family of parameterized impulse response functions is designed, including instantaneous impact functions, delayed response functions, and oscillating decay functions. These functions are combined and optimized to fit the sales change patterns caused by different types of sudden events. Piecewise functions and decay factors are used to capture the changing patterns of event impact over time, distinguishing between short-term impact effects and long-term persistent effects. A product-event coupling matrix is ​​introduced to quantify the sensitivity coefficients of different product categories to various sudden events, and an adaptive learning algorithm is used to dynamically update the coefficients. An impact decomposition framework is designed to decompose the overall impact of sudden events into three dimensions: direct impact, supply chain transmission impact, and consumer psychological impact, and the contribution weights of these three dimensions are quantified.

[0045] Preferably, through multi-scale time series decomposition and an adaptive time window mechanism, this invention can accurately identify seasonal and sudden patterns in sales data. A dual-channel attention network achieves a balanced capture of periodic changes and abnormal fluctuations, improving the adaptability of sales forecasting in complex market environments. The constructed sales shock response function and event-sales response model effectively quantify the short-term and long-term impacts of sudden events on the sales of different products, providing a quantitative basis for assessing the impact of sudden events and offering decision support for enterprises to formulate response strategies when facing market fluctuations.

[0046] S3: Apply a two-way causal analysis framework to simultaneously perform forward forecasting and backward attribution, and evaluate the direct, indirect, and interactive impacts of each node on sales.

[0047] In a preferred embodiment of the present invention, the flowchart of the bidirectional causal analysis framework is as follows: Figure 3 As shown, it includes the following steps: S3.1: Establish the cause-effect graph structure for sales forecasting, and define the causal relationships and conditional independence assumptions between nodes.

[0048] S3.2: Construct a neural network architecture for bidirectional information flow, including forward propagation from factors to sales and backward derivation from sales to factors.

[0049] Specifically, a symmetric encoder-decoder network is designed. The encoder compresses multi-dimensional factor features into latent representations, and the decoder reconstructs sales data through these latent representations. An attention-guided information selection mechanism is constructed, using cross-entropy and mutual information metrics to filter key factor nodes in the forward propagation. A gradient reversal layer is used to connect the forward and backward paths, making the optimization objective of the backward path antagonistic to that of the forward path. A dual loss function is designed, where the forward propagation loss is proportional to the prediction accuracy, and the backward derivation loss is proportional to the attribution residual. A residual recurrent network structure is introduced, feeding back the forward prediction residual and the backward attribution bias to the network parameters after each iteration. A sparse connection learning method based on Gumbel-Softmax sampling is applied to dynamically prune redundant connections between factors and sales. Through an alternating training strategy, the forward and backward paths are frozen sequentially in each training batch to achieve collaborative optimization of the two paths.

[0050] S3.3: Employ causal discovery algorithms to identify potential causal paths and hidden interaction effects.

[0051] Specifically, the method identifies potential causal paths between nodes through conditional independence tests, calculates path confidence scores, and retains paths with significance exceeding a threshold. A node combination test is used to discover hidden interaction effects; when the combined effect significantly deviates from the sum of the effects of individual nodes, it is labeled as a synergistic or inhibitory effect. Bootstrap sampling is applied to verify the stability of the discovered paths and interaction effects, eliminating results that are unstable during resampling. This path and interaction effect analysis method can accurately identify implicit causal paths and node interaction patterns in the data, providing a reliable structural basis for subsequent sales forecasting and influencing factor analysis.

[0052] S3.4: Construct a counterfactual prediction model and evaluate the independent impact of changes at a single node on sales through intervention testing.

[0053] Specifically, a node intervention graph generator is constructed to create multiple sets of counterfactual graph instances by replacing the feature representations or connection weights of target nodes; a difference comparison framework is designed to calculate the deviation distribution function of the original graph and the intervention graph in the sales forecast results; a hierarchical sampling strategy is adopted to perform intervention tests under different node types, different time windows, and different intervention intensities; an intervention sensitivity curve is established to quantify the response gradient of the sales forecast results to changes in node features; a Bayesian inference mechanism is used to calculate the posterior influence probability of each node under different intervention conditions; a backpropagation enhanced causal inference algorithm is applied to track the propagation path and decay mode of the intervention influence; and an intervention influence scoring system is established to generate a comprehensive evaluation index of the independent influence of nodes by combining the influence intensity, influence duration, and influence scope.

[0054] S3.5: Implement path analysis methods to decompose and quantify the direct and indirect effects of each node on sales through different paths.

[0055] Furthermore, a causal path tracker is constructed to identify all possible paths from the source node to the sales node and classify them according to path length and type. A structural equation model decomposition algorithm is applied to break down the total node effect into direct and indirect effects transmitted along different paths. A path-specific activation function is designed to distinguish the independent contribution of a single path by controlling the on / off state of path weights. Monte Carlo path sampling technology is used to randomly select a subset of paths in the causal graph to evaluate the synergistic effect of path combinations. A path importance ranking mechanism is established, calculating path weight scores based on effect size, stability, and universality. A path comparison analysis framework is introduced to reveal the key determinants of path selection by comparing the effect differences of similar nodes under different paths. A path effect visualization matrix is ​​developed to map direct and indirect effects into heatmaps, visually displaying the influence distribution of each node on different paths.

[0056] S3.6: Employ an interaction term detection mechanism to identify synergistic or inhibitory effects generated by multi-node combinations.

[0057] Furthermore, based on the causal paths identified in S3.5, the joint occurrence frequency and conditional probability distribution of node combinations are extracted; the conditional mutual information between node pairs on different paths is calculated to quantify the additional impact beyond the sum of direct and indirect effects; the importance ranking of node combinations is performed to identify the key interaction terms that contribute the most to the accuracy of sales forecasting; an interaction term intervention experiment is constructed to compare the effect differences between individual interventions on each node and simultaneous interventions on multiple nodes; the interaction effect coefficient is calculated, with positive values ​​indicating synergistic effects (exceeding the sum of individual effects) and negative values ​​indicating inhibitory effects (below the sum of individual effects); node combinations with interaction gains exceeding the statistical significance threshold are selected to form the minimum necessary interaction term set; and an interaction network topology graph is constructed to visualize the strength of synergistic and inhibitory relationships between nodes.

[0058] Preferably, this invention achieves unified processing of sales forecasting and attribution of influencing factors by constructing a bidirectional causal analysis framework. Based on a causal graph structure and a neural network architecture with bidirectional information flow, the model can simultaneously perform accurate forward prediction and reliable backward attribution. The application of counterfactual prediction models and path analysis methods enables the system to accurately assess the direct and indirect effects of each factor on sales, and to identify synergistic and inhibitory effects among multiple nodes through an interaction term detection mechanism. This comprehensive causal analysis capability improves the interpretability and decision support value of the sales forecasting model, enabling enterprises to more clearly understand the key factors affecting sales and their mechanisms of action.

[0059] S4: Generate cross-domain product embedding representations based on the evaluation results, and establish a bridging mechanism between long-tail products and best-selling products.

[0060] In a preferred embodiment of the present invention, the following steps are included: S4.1: Construct a product feature vector, which includes the product's intrinsic attributes, historical sales data, and causal analysis evaluation results.

[0061] S4.2: Design a cross-domain representation learning network to map products with different sales performances to the same semantic space.

[0062] Specifically, a dual-encoder network structure based on graph attention mechanism is constructed, in which the first encoder processes the features of best-selling products and the second encoder processes the features of long-tail products. The two encoders share some weight parameters. A cross-domain alignment loss function is designed, which includes three sub-objectives: structure preservation term, domain invariant term, and sales relevance term. An adaptive feature importance evaluation module is introduced to dynamically adjust the weight contributions of product intrinsic attributes, historical sales patterns, and external influencing factors in the representation learning process.

[0063] Optionally, a comparative learning framework based on sales performance can be constructed. By using a positive and negative sample pair selection strategy, cross-domain products with similar sales patterns can be placed closer together in the embedding space. A domain-adaptive regularization mechanism can be designed to reduce the distribution differences between the best-selling product domain and the long-tail product domain, thereby improving the consistency of the representation space.

[0064] S4.3: Employ a contrastive learning approach to reduce the distance between best-selling products and long-tail products with similar characteristics in the embedding space.

[0065] S4.4: Construct a product similarity calculation matrix and quantify the degree of association between products based on multi-dimensional feature weighting.

[0066] S4.5: Perform cluster analysis to identify product groups with similar sales patterns and influencing factors.

[0067] Among them, the cluster analysis adopts the spectral clustering algorithm combined with the dynamic kernel function, determines the optimal number of clusters through the adaptive similarity threshold, and introduces the sales fluctuation pattern as the cluster weight factor to construct a hierarchical product grouping system in the embedding space.

[0068] S4.6: Establish a knowledge transfer channel from best-selling products to long-tail products.

[0069] Specifically, feature patterns are extracted from the sales sequence of best-selling products through a similarity weighting mechanism, and knowledge transfer of sales fluctuation patterns and influencing factors is achieved by combining the attributes of long-tail products.

[0070] S4.7: Generate a product association network, represent the bridging relationship between best-selling products and long-tail products using a graph structure, and calculate the bridging strength.

[0071] Furthermore, a heterogeneous association graph with products as nodes is constructed, and the connection between best-selling products and long-tail products is established based on a multi-dimensional similarity matrix; a bridging strength calculation function is designed, which comprehensively considers the weighted scores of three dimensions: sales pattern similarity, attribute matching degree, and market position relevance; a two-way information flow mechanism is introduced, and the directionality and effectiveness of knowledge transfer are represented by directed edge weights; a random walk algorithm is used to calculate the reachability between nodes in the network and quantify the bridging effectiveness of different paths; and the strength coefficient of bridging relationships is dynamically adjusted based on time series change analysis to adapt to the market fluctuation environment.

[0072] Preferably, by constructing cross-domain product embedding representations and association bridging mechanisms, this invention achieves knowledge sharing and transfer between best-selling products and long-tail products. A dual-encoder network based on graph attention maps products with different sales performances to a unified semantic space, enabling the model to effectively capture the common features of similar products under different sales states. The application of contrastive learning methods reduces the distance between cross-class products with similar features in the embedding space, improving the accuracy of identifying inter-product associations. The established knowledge transfer channel from best-selling products to long-tail products promotes experience sharing of sales patterns and influencing factors, providing data support for sales forecasting and marketing strategy optimization for long-tail products, and helping enterprises tap into the market potential of long-tail products.

[0073] S5: Utilizes product embedding representation and association bridging mechanism to perform sales forecasting through a multi-level collaboration mechanism, and outputs the sales forecast results for the target product.

[0074] In a preferred embodiment of the present invention, the following steps are included: S5.1: Construct an integrated commodity sales forecasting framework, which includes two core components: the commodity feature layer and the market association layer.

[0075] S5.2: In the product feature layer, a basic prediction model for product sales is trained based on product embedding representation and historical sales data.

[0076] Specifically, a multi-dimensional feature fusion structure is constructed, and a parameterized gating mechanism is used to achieve dynamic weight coupling between the product embedding representation vector and historical sales time-series data. A dual attention mechanism is designed, in which the time attention component adaptively selects key time points in the historical sales data, and the feature attention component dynamically adjusts the importance weights of different dimensions in the embedding representation. A multi-scale recursive structure based on a long short-term memory network is constructed to simultaneously capture short-term fluctuation patterns and long-term evolution trends in sales. A nonlinear feature interaction module is designed to efficiently calculate the high-order interaction relationships of latent feature dimensions in the product embedding vector using a low-rank tensor decomposition method. A time-varying parameter update mechanism is applied to automatically adjust the model parameters according to the prediction error of different time windows, thereby improving the model's adaptability to market changes.

[0077] Optionally, an adaptive prediction framework for sales distribution can be constructed, which dynamically selects the prediction objective function based on the statistical characteristics of historical sales data, effectively addressing the heterogeneity problem of sales distribution for different products.

[0078] S5.3: In the market association layer, integrate category information, market environment factors, and inter-product relationships to construct a sales adjustment model.

[0079] Furthermore, a multi-level heterogeneous graph structure is constructed, using node type encoding and edge attribute weighting to represent internal product relationships, inter-category relationships, and market environment influence channels. Local topological feature extraction is performed, utilizing parameterized neighborhood aggregation functions to capture the synergistic and competitive effects of sales within product categories. Market environment factors are vectorized by transforming macroeconomic indicators, seasonal factors, and regional consumption trends into low-dimensional representation vectors through nonlinear projection. A bidirectional information channel is established between the product feature layer and the market relationship layer, using basic prediction results as conditional inputs to the adjustment model while simultaneously feeding back market relationship features for feature enhancement of the basic model. An inter-layer residual learning mechanism is constructed, enabling the adjustment model to focus on learning the correction amount of the basic prediction results, and integrating the prediction results of the two layers through an additive model to generate the final sales prediction value.

[0080] S5.4: Design an inter-layer information fusion mechanism to effectively integrate the prediction information of the two layers through feature sharing and attention mechanisms.

[0081] S5.5: Apply the correlation bridging transmission algorithm, based on the correlation bridging mechanism, to transmit the sales patterns of best-selling products to the sales forecasting process of long-tail products.

[0082] Specifically, a similarity matrix between best-selling products and long-tail products is constructed, and multi-dimensional feature similarity scores are calculated based on product embedding representation. A sales pattern extractor is designed to identify key temporal patterns and seasonal features from historical sales data of best-selling products. A sales pattern adaptation function is established to adjust the sales patterns of best-selling products to reference patterns suitable for the characteristics of long-tail products through parameterized transformation. An attention-based pattern transfer mechanism is implemented to selectively integrate the sales pattern features of best-selling products into the prediction process of long-tail products according to similarity weights. A sales difference compensation mechanism is constructed to automatically adjust the scale difference and feature shift in the pattern transfer process through residual learning. A progressive transmission strategy is designed to gradually transfer the sales pattern from high-selling products to extremely low-selling products according to the similarity level, forming a complete transmission link.

[0083] S5.6: Construct a multi-objective optimization mechanism to balance prediction accuracy and model generalization ability, and generate the final sales forecast results for the target product.

[0084] Furthermore, a composite loss function is designed to comprehensively consider three optimization objectives: point prediction error, distribution prediction bias, and prediction uncertainty. A product importance weighting mechanism is implemented, dynamically adjusting the loss weights of different products during the optimization process based on product value and sales volume. A prediction stability regularization term is constructed to penalize prediction results sensitive to input perturbations, improving model robustness. Model complexity constraints are applied, controlling model complexity through structured sparsity and parameter norm limitations to prevent overfitting. A balanced training strategy for best-selling and long-tail products is designed, ensuring sufficient optimization attention for long-tail products through gradient adjustment. Multi-granularity prediction error evaluation is implemented, simultaneously optimizing daily, weekly, and monthly prediction accuracy to improve the model's generalization ability across different time scales. An adaptive learning rate scheduling mechanism is constructed to dynamically adjust the parameter update step size based on the convergence status of different product types.

[0085] Preferably, by constructing an integrated commodity sales forecasting framework, this invention achieves collaborative forecasting between the commodity feature layer and the market association layer. The multi-level fusion mechanism fully utilizes commodity characteristics and market environment information, improving the comprehensiveness and accuracy of the forecast. The association bridging and transmission algorithm effectively transmits the sales patterns of best-selling commodities to long-tail commodities, solving the forecasting difficulty caused by the sparsity of long-tail commodity data. The application of a multi-objective optimization mechanism balances forecast accuracy and model generalization ability, improving the model's adaptability to different commodity types and sales environments. The overall architecture ensures both the forecast accuracy of best-selling commodities and the forecast reliability of long-tail commodities, providing enterprises with more comprehensive sales forecasting support.

[0086] Furthermore, this embodiment also provides a product sales forecasting system based on an improved neural network, including: a dynamic graph construction module, used to construct a dynamic heterogeneous graph network containing product nodes, user nodes, promotional activity nodes, and external event nodes, and calculate the temporal dynamic correlation strength between nodes to quantify the correlation between node relationships and sales; a temporal pattern recognition module, used to identify seasonal and sudden patterns in sales data based on an adaptive spatiotemporal attention mechanism, and quantify the impact of sudden events on sales; a bidirectional causal analysis module, used to apply a bidirectional causal analysis framework, simultaneously performing forward prediction and backward attribution, to evaluate the direct, indirect, and interactive impacts of each node on sales; a product representation learning module, used to generate cross-domain product embedding representations based on the evaluation results, and establish a bridging mechanism between long-tail products and best-selling products; and a sales forecasting module, used to use product embedding representations and the bridging mechanism to perform sales forecasting through a multi-level collaborative mechanism, and output the sales forecast results of the target product.

[0087] In summary, this invention innovatively combines a time-aware gating mechanism with heterogeneous graph networks. By modeling a four-element relationship between product, user, promotion, and event, it effectively captures the dynamic changes among factors influencing sales, enhancing the predictive model's adaptability to market fluctuations. It innovatively designs a dual-channel attention architecture for seasonality and sudden events, using adaptive gating to balance the weights of the two modes, thus enhancing the model's ability to identify both regular and sudden sales fluctuations. It proposes a bidirectional causal inference method combining counterfactual prediction and path analysis, improving the model's explanatory power for sales impact mechanisms by quantitatively evaluating direct and indirect effects. Finally, it designs a cross-domain knowledge transfer mechanism from best-selling products to long-tail products, achieving product feature mapping through comparative learning to overcome the data sparsity problem of long-tail products.

[0088] Example 2, refer to Figures 1-3 This is the second embodiment of the present invention, which provides a commodity sales forecasting method based on an improved neural network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0089] To verify the effectiveness of the improved neural network-based sales prediction method in real-world scenarios, this study selected sales data from a large retail chain enterprise from January 2020 to December 2023 as the experimental dataset. This dataset contains over 50,000 SKUs, covering multiple categories including food, clothing, electronics, and daily necessities, as well as approximately 15 million user purchase records, 35,000 promotional activity records, and sales-related external event data. The experimental dataset was divided into training, validation, and test sets in a 7:2:1 ratio, with data from the most recent three months (October to December 2023) used as the test set to evaluate the model's actual predictive performance.

[0090] In the construction phase of the dynamic heterogeneous graph network, the system first preprocessed the collected multidimensional data and segmented the time-series data based on the adaptive sliding time window technique. The entropy fluctuation threshold was set to 0.15; when the information entropy change exceeded this threshold, window adjustment was triggered. Five initial window lengths for different time granularities were set in the experiment: hourly (24 hours), daily (7 days), weekly (4 weeks), monthly (3 months), and quarterly (3 quarters). During the construction of the quaternary heterogeneous graph network, the edge weight decay rate was set according to the node type: 0.05 / day for user-product interaction, 0.03 / day for promotion-product association, and 0.02 / day for external event-product association. Through experimental optimization, the time window for the short-term memory module in the time-series dependency model was set to 14 days, and the time window for the long-term memory module was set to 90 days. The association strength assessment based on causal inference shows that, among best-selling electronic product categories, promotional activities directly contribute an average of 35.2% to sales, user repurchase behavior indirectly contributes 28.7%, external events (such as holidays) contribute 18.5%, and the remainder contributes to basic sales volume. Through this stage of analysis, the system successfully identified 73.6% of the key influencing factor nodes, an improvement of 21.4 percentage points compared to traditional association rule mining methods.

[0091] In the seasonal and sudden pattern recognition stage based on an adaptive spatiotemporal attention mechanism, the system designs a multi-scale time series decomposition network containing three deformable convolutional layers with different kernel sizes (3, 7, and 15). In the adaptive time window mechanism, the information entropy threshold is set to 0.12, and the autocorrelation coefficient threshold is set to 0.65, forming five main time windows: daily (24 hours), weekly (7 days), monthly (30 days), quarterly (90 days), and annual (365 days). In the sudden fluctuation recognition, the standard deviation threshold is set to 2.5, and the statistical significance level is set to 0.05. Experimental analysis shows that the system can identify sudden fluctuations in sales during 11 major promotional periods, such as Singles' Day and Spring Festival, in the food category, with an average lead time of 2.3 days and an accuracy rate of 91.5%. Event-sales response model analysis reveals significant differences in elasticity coefficients among different product categories to sudden events: the average elasticity coefficient for electronics is 2.13, for food it is 1.87, for clothing it is 1.62, and for daily necessities it is 1.45. Dynamic weight adjustment of the dual-channel attention network shows that during normal sales periods, the seasonal mode channel weight averages 0.73, while the sudden event mode channel weight averages 0.27; during large-scale promotional events, the seasonal mode channel weight decreases to 0.34, while the sudden event mode channel weight increases to 0.66. This dynamic balancing mechanism improves the model's predictive accuracy by 17.8% in complex market environments.

[0092] In the application of the bidirectional causal analysis framework, the system constructed a causal graph structure for sales prediction containing 297 nodes, including 156 product nodes, 85 user node groups, 35 promotional activity nodes, and 21 external event nodes. In the bidirectional information flow neural network architecture, the encoder-decoder hidden layer dimension was set to 256, the number of attention heads to 8, and the inversion coefficient λ of the gradient inversion layer to 0.1. The counterfactual prediction model evaluated the independent impact of each node on sales through 10,000 Monte Carlo intervention experiments. The results showed that the average intervention sensitivity of price factors was 0.72, promotional intensity was 0.68, seasonality factors were 0.57, and brand influence was 0.45. Path analysis identified 468 valid causal paths, including 189 direct effect paths and 279 indirect effect paths. In the best-selling mobile phone category, price promotions increased sales by an average of 23.5% through direct effects and by an average of 16.8% through indirect effects of user referrals. The interaction detection mechanism identified 124 significant interactions, of which 43 were synergistic effects (interaction effect coefficients greater than 1) and 81 were inhibitory effects (interaction effect coefficients less than 1). For example, the synergistic effect coefficient between price promotions and holidays was 1.37, indicating that price promotions during holidays are 37% more effective than during ordinary periods; while the interaction effect coefficient between certain high-end brands and significant price discounts was 0.82, indicating that significant price discounts may have a negative impact on the image of high-end brands and reduce the effectiveness of promotions.

[0093] In the construction phase of the association bridging mechanism based on cross-domain product embedding representation, the system divides 50,000 SKUs into best-selling products (approximately 5,000 products, accounting for 80% of total sales) and long-tail products (approximately 45,000 products, accounting for 20% of total sales) according to sales volume distribution. In the dual-encoder network structure, the encoder hidden layer dimension is set to 128, and the shared weight ratio is 40%. In the cross-domain alignment loss function, the weight of the structure preservation term is set to 0.4, the weight of the domain invariant term is set to 0.3, and the weight of the sales relevance term is set to 0.3. In contrastive learning, the selection threshold for positive sample pairs is set to 0.85, and the selection threshold for negative sample pairs is set to 0.35. Cosine similarity index is used for product similarity calculation, and spectral clustering algorithm is used for cluster analysis, ultimately forming 427 product groups. After constructing the knowledge transfer channel from best-selling products to long-tail products, the system calculates 15,723 high-weight bridging relationships, with an average bridging strength of 0.63. Experimental results show that through knowledge transfer, the accuracy of sales forecasting for long-tail products has improved significantly: MAPE (mean absolute percentage error) decreased from 42.7% to 27.5%, and RMSE (root mean square error) decreased by 32.8%. In particular, the improvement in forecasting accuracy is more significant in long-tail product categories with strong seasonality and large sales fluctuations (such as specialty clothing and special holiday foods), with MAPE decreasing by 19.6 percentage points.

[0094] In the multi-level collaborative prediction stage, the product feature layer adopts an LSTM network structure with a hidden layer dimension of 256, and both temporal attention and feature attention heads are 8. The market association layer constructs a 3-layer graph neural network with 64 neurons in each layer, using the LeakyReLU activation function. In the association bridging and propagation algorithm, the similarity threshold is set to 0.7, the initial value of the pattern propagation weight is 0.6, and it is dynamically adjusted according to the prediction error. In the multi-objective optimization mechanism, the weight of point prediction error is 0.5, the weight of distribution prediction bias is 0.3, and the weight of prediction uncertainty is 0.2. The overall performance of the model on the entire test set is as follows: MAPE is 12.3%, MAE (mean absolute error) is 6.8 units, and RMSE is 9.2 units. Among them, the MAPE of best-selling products is 8.7%, and the MAPE of long-tail products is 19.6%. Time granularity analysis shows that the daily prediction MAPE is 15.2%, the weekly prediction MAPE is 10.8%, and the monthly prediction MAPE is 8.9%. As shown in Table 1, the commodity sales forecasting method based on the improved neural network proposed in this invention performs excellently in all performance indicators compared with mainstream forecasting methods in the industry.

[0095] Table 1 Comparison of the performance of the present invention with existing technologies. method Overall MAPE (%) Best-selling product MAPE (%) MAPE (%) for long-tail products Accuracy rate of emergency identification (%) Forecast lead time (days) ARIMA 28.5 19.7 46.2 63.5 0 LSTM 19.6 12.3 36.4 76.4 1.5 Graph Neural Networks 16.3 10.9 31.2 81.7 1.8 This invention 12.3 8.7 19.6 91.5 2.3 As shown in Table 1, the method of this invention significantly improves upon existing technologies in all indicators. The overall MAPE is 4.0 percentage points lower and 24.5% higher than the closest graph neural network method. In long-tail product forecasting, the MAPE is 11.6 percentage points lower and 37.2% higher, demonstrating the advantages of this invention in handling long-tail product forecasting problems. Regarding the accuracy of identifying sudden events, the method of this invention reaches 91.5%, 9.8 percentage points higher than the graph neural network method. The forecast lead time reaches 2.3 days, providing enterprises with more time to adjust inventory and marketing strategies.

[0096] These data fully demonstrate that the present invention has made significant progress in the field of commodity sales forecasting through innovative technologies such as dynamic heterogeneous graph networks, adaptive spatiotemporal attention mechanisms, bidirectional causal analysis frameworks, and cross-domain knowledge transfer, especially in solving the two difficult problems of long-tail commodity forecasting and market fluctuation response.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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 method for predicting commodity sales based on an improved neural network, characterized in that, include: Construct a dynamic heterogeneous graph network containing product nodes, user nodes, promotional activity nodes, and external event nodes, and calculate the temporal dynamic correlation strength between nodes to quantify the correlation between node relationships and sales. Based on an adaptive spatiotemporal attention mechanism, seasonal and sudden patterns in sales data are identified, and the impact of sudden events on sales is quantified. By applying a two-way causal analysis framework, forward forecasting and backward attribution are performed simultaneously to assess the direct, indirect, and interactive impacts of each node on sales. Based on the evaluation results, a cross-domain product embedding representation is generated, and a bridging mechanism for the association between long-tail products and best-selling products is established. By utilizing the product embedding representation and association bridging mechanism, sales forecasting is performed through a multi-level collaborative mechanism, and the sales forecast results of the target product are output.

2. The commodity sales forecasting method based on an improved neural network according to claim 1, characterized in that, The construction process of the dynamic heterogeneous graph network includes: Construct a quaternary heterogeneous graph network of product-user-promotion-event, and optimize the connection strength between nodes through a dynamic weight adjustment algorithm; Map node entities to a low-dimensional semantic space and use a multi-head attention mechanism to capture the interaction relationships between different types of nodes; Establish a temporal dependency model between nodes and quantify the dynamic changes in association strength through time-aware gated recursive units; Perform a correlation strength assessment based on causal inference, and calculate the direct and indirect contribution values ​​of each node to sales through counterfactual analysis.

3. The commodity sales forecasting method based on an improved neural network according to claim 1, characterized in that, The identification of seasonal and sudden patterns in sales data includes: Construct a multi-scale time series decomposition network to split sales data into trend components, seasonal components, and residual components; An adaptive time window mechanism is designed to dynamically adjust the observation time span and capture seasonal patterns of different cycles; Anomaly detection algorithms are used to identify sudden fluctuations in sales data, and the timing of the sudden events is determined based on statistical significance tests. Construct an event-sales response model to analyze the magnitude, duration, and recovery pattern of sales changes before and after a sudden event; Design a dual-channel attention network, with one channel focusing on seasonal patterns and the other on sudden patterns, and dynamically balance the importance weights of the two patterns through a gating mechanism; Construct a sales impact response function to quantify the short-term and long-term impacts of different types of emergencies on sales, and form an emergency impact rating system.

4. The commodity sales forecasting method based on an improved neural network according to claim 1, characterized in that, The application of the bidirectional causal analysis framework includes: Establish a causal graph structure for sales forecasting and construct a neural network architecture for bidirectional information flow; A causal discovery algorithm is used to identify potential causal paths and hidden interaction effects; Construct a counterfactual prediction model and evaluate the independent impact of changes at a single node on sales through intervention testing; The path analysis method is used to decompose and quantify the direct and indirect effects of each node on sales through different paths. An interaction term detection mechanism is used to identify the synergistic or inhibitory effects generated by the combination of multiple nodes.

5. The commodity sales forecasting method based on an improved neural network according to claim 1, characterized in that, The mechanism for establishing a connection between long-tail products and best-selling products includes: Construct a product feature vector, which includes the product's intrinsic attributes, historical sales data, and causal analysis and evaluation results; Design a cross-domain representation learning network to map products with different sales performance to the same semantic space; By employing a contrastive learning approach, the distance between best-selling products and long-tail products with similar characteristics in the embedding space is reduced; Construct a product similarity calculation matrix and quantify the degree of association between products based on multi-dimensional feature weighting.

6. The commodity sales forecasting method based on an improved neural network according to claim 1, characterized in that, The mechanism for establishing a connection between long-tail products and best-selling products also includes: Perform cluster analysis to identify product groups with similar sales patterns and influencing factors; Establish a knowledge transfer channel from best-selling products to long-tail products; Generate a product association network, represent the bridging relationship between best-selling products and long-tail products using a graph structure, and calculate the bridging strength.

7. The commodity sales forecasting method based on an improved neural network according to claim 1, characterized in that, The sales forecasting through a multi-level collaborative mechanism includes: An integrated commodity sales forecasting framework is constructed, comprising two core components: a commodity feature layer and a market association layer. At the product feature layer, a basic prediction model for product sales is trained based on product embedding representations and historical sales data. At the market association layer, category information, market environment factors, and inter-product relationships are integrated to construct a sales adjustment model.

8. A commodity sales forecasting system based on an improved neural network, characterized in that, include: The dynamic graph construction module is used to construct a dynamic heterogeneous graph network containing product nodes, user nodes, promotional activity nodes, and external event nodes, and to calculate the temporal dynamic correlation strength between nodes and quantify the correlation between node relationships and sales. The temporal pattern recognition module is used to identify seasonal and sudden patterns in sales data based on an adaptive spatiotemporal attention mechanism, and to quantify the impact of sudden events on sales. The two-way causal analysis module is used to apply the two-way causal analysis framework, simultaneously performing forward forecasting and backward attribution to evaluate the direct, indirect, and interactive impacts of each node on sales. The product representation learning module is used to generate cross-domain product embedding representations based on the evaluation results and establish a bridging mechanism between long-tail products and best-selling products. The sales forecasting module is used to forecast sales through a multi-level collaborative mechanism by utilizing the product embedding representation and association bridging mechanism, and output the sales forecast results of the target product.

Images