Data generation method and device, model training method and device and information prediction method and device

Abstract

The invention provides a data generation method and device, a model training method and device, and an information prediction method and device, and the data method comprises the steps: obtaining first product data of a historical product and second product data of a target product, the historical product being the same as the target product in type; performing multi-view construction based on the first product data and the second product data to obtain a first multi-view representation of a historical product and a second multi-view representation of a target product; constructing an in-view local graph structure and a cross-view global graph structure according to the first multi-view representation; performing covariant enhancement on the second multi-view representation based on the in-view local graph structure and the cross-view global graph structure to obtain first enhanced covariant data of the target product; generating second enhanced sample data of the target product based on the first product data and the second product data; wherein the first enhanced covariable data and the second enhanced sample data are used for training a prediction model, and the prediction model is used for determining prediction information of the target product.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and more specifically, to a method and apparatus for data generation, model training, and information prediction in the field of computer technology. Background Technology

[0002] In various application scenarios, leveraging existing product data to obtain predictive information for the future can aid in relevant decision-making. However, in the early stages of a product's lifecycle, existing product data often exhibits characteristics of sparsity or instability, making accurate predictions difficult to achieve based on limited data. Therefore, improving the accuracy of product prediction information has become a pressing technical problem to be solved. Summary of the Invention

[0003] This specification provides a data generation, model training, and information prediction method and apparatus, which can improve the accuracy of product prediction information.

[0004] Firstly, a data generation method is provided, including: Obtain the first product data of historical products and the second product data of the target product, where the historical products and the target products are of the same type. Multi-views are constructed based on the first product data and the second product data respectively, resulting in a first multi-view representation of historical products and a second multi-view representation of the target product; Construct a local graph structure within a view and a global graph structure across views based on the first multi-view representation; Based on the local graph structure within the view and the global graph structure across the view, covariate enhancement is performed on the second multi-view representation to obtain the first enhanced covariate data of the target product. Based on the first product data and the second product data, generate the second enhanced sample data for the target product; The first enhanced covariate data and the second enhanced sample data are used to train the prediction model, which is then used to determine the prediction information for the target product.

[0005] In conjunction with the first aspect, in some possible implementations, multiple views are constructed based on the first product data and the second product data respectively, resulting in a first multi-view representation of historical products and a second multi-view representation of the target product, including: Based on the first product data, multiple preset context windows, and multiple preset time scales, historical product observation data is generated under multiple time views. Based on the second product data, multiple context windows, and multiple time scales, generate observation data of the target product in multiple time views; The preset time-series feature extraction model is invoked to model the observation data of historical products under multiple time views and the observation data of target products under multiple time views, respectively, to obtain the first multi-view representation of historical products and the second multi-view representation of target products.

[0006] Combining the first aspect and the above implementation methods, in some possible implementation methods, a local graph structure within a view and a global graph structure across views are constructed based on the first multi-view representation, including: The first multi-view representation is mapped based on a preset first mapping matrix to obtain the in-view spatial representation; The inner product is calculated based on the spatial representation within the view to obtain the local graph structure within the view; The first multi-view representation is mapped based on a preset second mapping matrix to obtain a cross-view space representation. The inner product is calculated based on the cross-view space representation to obtain the cross-view global graph structure.

[0007] Combining the first aspect and the above implementation methods, in some possible implementation methods, based on the local graph structure within the view and the global graph structure across views, covariate enhancement is performed on the second multi-view representation to obtain the first enhanced covariate data of the target product, including: The system retrieves the set of neighboring products within the view associated with the target product based on the second multi-view representation and the local graph structure within the view, and retrieves the set of neighboring products across the view associated with the target product based on the second multi-view representation and the global graph structure across the view. A dual-path multi-view diffusion is performed on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information; Sales forecast results for the target product are obtained based on aggregated information. The sales forecast results are used to characterize the sales volume of the target product in a future preset time period. The first enhancing covariate data for the target product is determined based on sales forecast results.

[0008] Combining the first aspect and the above implementation methods, in some possible implementation methods, retrieving the set of intra-view neighbor products associated with the target product based on the second multi-view representation and the intra-view local graph structure, and retrieving the set of cross-view neighbor products associated with the target product based on the second multi-view representation and the cross-view global graph structure, includes: For each view in the second multi-view representation, the in-view similarity is calculated based on the in-view representation of the target product under the view and the in-view representation of the historical products under the view; Based on in-view similarity, retrieve the set of one-hop neighbor products in the view that are associated with the target product; Calculate cross-view similarity based on the cross-view representation of the target product in the view and the cross-view representation of historical products in the view; Based on cross-view similarity, retrieve cross-view similarities associated with the target product. Figure 1 Jump to neighbor product collection; Based on the local graph structure within the view, retrieve the set of two-hop neighbor products within the view for each product in the set of one-hop neighbor products within the view. Based on the cross-view global graph structure, for cross-view Figure 1 Each product in the neighbor product collection can be retrieved across views. Figure 2 Jump to neighbor product collection; Combine the set of one-hop neighbor products within the view and the set of two-hop neighbor products within the view into the set of neighbor products within the view associated with the target product, and combine the cross-view neighbor products. Figure 1 Jump Neighbor product collection, cross-view Figure 2 The jump neighbor product set combination is the cross-view neighbor product set associated with the target product.

[0009] Combining the first aspect and the above implementation methods, in some possible implementations, a two-path multi-view diffusion is performed on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information, including: For each first-hop neighbor product in the set of first-hop neighbor products in the view, based on the local graph structure in the view, the information of each product in the set of second-hop neighbor products in the view corresponding to the first-hop neighbor product is aggregated to obtain the updated representation of the first-hop neighbor product. Based on the updated first-hop neighbor representation and the in-view representation of the target product in the view, the similarity between the target product and each first-hop neighbor product is calculated as the propagation weight, and the updated first-hop neighbor representation is weighted and summed according to the propagation weight to obtain the in-view aggregation information of the view. For cross-view Figure 1 For each second-hop neighbor product in the hop neighbor product set, based on the cross-view global graph structure, the cross-view corresponding cross-hop neighbor products are aggregated. Figure 2 The information of each product in the next-hop neighbor product set is used to obtain the updated representation of the second-hop neighbor product; Based on the updated second-hop neighbor representation and the cross-view representation of the target product in the view, the similarity between the target product and each second-hop neighbor product is calculated as the propagation weight, and the updated second-hop neighbor representation is weighted and summed according to the propagation weight to obtain the cross-view aggregation information of the view. The view's intra-view aggregation information and cross-view aggregation information are merged to obtain the view's merged aggregation information; The aggregated information is obtained by splicing together the fused and aggregated information of all views in the second multi-view representation.

[0010] Combining the first aspect and the above implementation methods, in some possible implementation methods, based on the first product data and the second product data, second enhanced sample data of the target product is generated, including: Identify similar historical products from historical products that have similar multi-view representations to the target product; The product sequence of similar historical products is determined from the first product data, and the product sequence of the target product is determined from the second product data; The preset sequence decomposition model is invoked to decompose the product sequence of similar historical products into the trend, seasonal and residual terms of similar historical products, and to decompose the product sequence of the target product into the trend, seasonal and residual terms of the target product. The trend terms of the target product and the trend terms of similar historical products, the seasonal terms of the target product and the seasonal terms of similar historical products, and the residual terms of the target product and the residual terms of similar historical products are merged to obtain the merged trend terms, merged seasonal terms, and merged residual terms. The trend term, seasonal term, and residual term are reconstructed to obtain the second enhanced sample data of the target product.

[0011] Combining the first aspect and the above implementation methods, in some possible implementation methods, the trend items of the target product and the trend items of similar historical products, the seasonal items of the target product and the seasonal items of similar historical products, and the residual items of the target product and the residual items of similar historical products are respectively fused to obtain fused trend items, fused seasonal items, and fused residual items, including: The trend items of the target product and the trend items of similar historical products are weighted and merged to obtain the merged trend item; After phase alignment, the seasonal items of the target product and the seasonal items of similar historical products are weighted and fused to obtain the fused seasonal item. Generative fusion is performed on the residuals of the target product and the residuals of similar historical products to obtain the fused residuals.

[0012] Combining the first aspect and the above implementation methods, in some possible implementation methods, the residual terms of the target product and the residual terms of similar historical products are generatively fused to obtain fused residual terms, including: Based on preset weight parameters, a preset generative adversarial network model is invoked to perform generative fusion of the residual terms of the target product and the residual terms of similar historical products to obtain the fused residual terms.

[0013] Secondly, a model training method is provided, including: Acquire first enhanced covariate data and second enhanced sample data, wherein the first enhanced covariate data and the second enhanced sample data are generated based on the data generation method of the first aspect; The preset initial model is trained based on the first enhanced covariate data and the second enhanced sample data until the preset convergence condition is reached. The parameter-updated initial model is then determined as the prediction model. The prediction model is used to determine the prediction information for the target product.

[0014] Thirdly, an information prediction method is provided, including: Obtain third-party product data for the target product; Based on the third product data of the target product, a preset prediction model is invoked to perform information prediction and obtain the prediction information of the target product. The prediction model is trained using first enhanced covariate data and second enhanced sample data generated based on the data generation method of the first aspect.

[0015] Fourthly, a data generation apparatus is provided, comprising: The acquisition unit is used to acquire the first product data of historical products and the second product data of the target product, where the historical products and the target products are of the same type. The first construction unit is used to construct multiple views based on the first product data and the second product data respectively, to obtain the first multi-view representation of the historical product and the second multi-view representation of the target product; The second building unit is used to build a local graph structure within a view and a global graph structure across views based on the first multi-view representation; The enhancement unit is used to enhance the covariates of the second multi-view representation based on the local graph structure within the view and the global graph structure across the view, so as to obtain the first enhanced covariate data of the target product. The generation unit is used to generate second enhanced sample data of the target product based on the first product data and the second product data; The first enhanced covariate data and the second enhanced sample data are used to train the prediction model, which is then used to determine the prediction information for the target product.

[0016] Fifthly, a model training device is provided, comprising: The acquisition unit is used to acquire first enhanced covariate data and second enhanced sample data, wherein the first enhanced covariate data and the second enhanced sample data are generated based on the data generation method of the first aspect. The training unit is used to train the preset initial model based on the first augmented covariate data and the second augmented sample data until the preset convergence condition is reached, and the parameter-updated initial model is determined as the prediction model. The prediction model is used to determine the prediction information for the target product.

[0017] Sixthly, an information prediction device is provided, comprising: The acquisition unit is used to acquire third-party product data of the target product. The prediction unit is used to call a preset prediction model to predict information based on the third product data of the target product, and obtain the prediction information of the target product. The prediction model is trained using first enhanced covariate data and second enhanced sample data generated based on the data generation method of the first aspect.

[0018] In a seventh aspect, a computer-readable storage medium is provided that stores computer program code, which, when executed, implements the steps of the above-described method.

[0019] Eighthly, an electronic device is provided, comprising: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the steps of the above-described method.

[0020] Ninthly, a computer program product is provided, which stores at least one instruction, the steps of the above method being implemented when the at least one instruction is executed by a processor.

[0021] The main solution proposed in this application includes: firstly, acquiring first product data of historical products of the same type and second product data of the target product; then, constructing a first multi-view representation of the historical products and a second multi-view representation of the target products through multi-view construction, and constructing an intra-view local graph structure and a cross-view global graph structure based on the first multi-view representation, thereby establishing a multi-dimensional feature description and structured association between products; on this basis, covariate enhancement is performed based on the intra-view local graph structure and the cross-view global graph structure, integrating the information of historical products into the target products to generate first enhanced covariate data, thereby enriching the feature information used for training; at the same time, the scale and diversity of the training data are expanded by generating second enhanced sample data; in summary, by generating first enhanced covariate data and second enhanced sample data, the quality and quantity of training data are improved, thereby improving the accuracy of the prediction information of the target products determined by the prediction model. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a data generation method provided in an embodiment of this specification; Figure 2 This is a flowchart illustrating a model training method provided in an embodiment of this specification; Figure 3 This is a flowchart illustrating an information prediction method provided in the embodiments of this specification; Figure 4 This is a schematic diagram of the structure of a data generation device provided in the embodiments of this specification; Figure 5 This is a schematic diagram of the structure of a model training device provided in the embodiments of this specification; Figure 6 This is a schematic diagram of the structure of an information prediction device provided in the embodiments of this specification; Figure 7 This is a schematic diagram of the structure of an electronic device provided in the embodiments of this specification. Detailed Implementation

[0023] The technical solutions in this specification will now be described clearly and in detail with reference to the accompanying drawings. In the description of the embodiments in this specification, unless otherwise stated, " / " indicates "or," for example, A / B can mean A or B. "And / or" in the text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of the embodiments in this specification, "multiple" refers to two or more than two.

[0024] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0025] In various application scenarios, existing product data can be used to obtain predictive information for the future, which can then be used to make relevant decisions. However, in the early stages of a product's lifecycle, the existing product data is often sparse or unstable, making it difficult to achieve accurate predictions based on limited product data.

[0026] Taking e-commerce application scenarios as an example, newly listed products lack sufficient historical user behavior data in the early stages of release, such as click-through rate, add-to-cart rate, and conversion rate. This makes it difficult for the platform to accurately predict its future sales trends or sales volume, which in turn affects the formulation of inventory management strategies and may lead to inventory backlog or stockouts. At the same time, the allocation of marketing resources is also limited due to the lack of data support.

[0027] Taking content recommendation application scenarios as an example, newly published articles or videos have highly sparse initial user interaction data (such as clicks, shares, and comments), making it difficult for the system to effectively assess their potential popularity trend and matching degree with user interests. Consequently, the recommendation system cannot achieve accurate personalized push during the cold start phase, reducing content distribution efficiency and user engagement.

[0028] Taking financial applications as an example, during the initial launch phase of wealth management products, the existing product data is characterized by sparsity and instability. On the one hand, limited by factors such as product types and launch duration, the sample size is limited, making it difficult to capture diverse market behavior patterns. On the other hand, raw data such as historical subscription records and yield rates are basic characteristic information and cannot fully reflect potential customer behavior and market dynamics. These two problems together lead to low accuracy in predicting product sales or scale growth during this phase, affecting the decision-making effectiveness of asset management institutions.

[0029] In summary, whether it's product sales, content distribution, or asset management services, the key technological challenge lies in how to effectively extract crucial information and build generalizable predictive models from limited and unstable existing product data to improve predictive performance in the early stages of a product's lifecycle. In other words, improving the accuracy of product predictions has become a pressing technical problem.

[0030] To address the aforementioned issues, the main solution proposed in this application includes: firstly, acquiring first product data of historical products of the same type and second product data of the target product; then, constructing a first multi-view representation of the historical products and a second multi-view representation of the target products through multi-view construction, and constructing an intra-view local graph structure and a cross-view global graph structure based on the first multi-view representation, thereby establishing a multi-dimensional feature description and structured association between products; based on this, covariate enhancement is performed based on the intra-view local graph structure and the cross-view global graph structure, integrating the information of historical products into the target products to generate first enhanced covariate data, thereby enriching the feature information used for training; simultaneously, the scale and diversity of the training data are expanded by generating second enhanced sample data; in summary, by generating first enhanced covariate data and second enhanced sample data, the quality and quantity of the training data are improved, thereby improving the accuracy of the prediction information of the target products determined by the prediction model.

[0031] The data generation method, model training method, and information prediction method provided in the embodiments of this specification will be described in detail below.

[0032] Please see Figure 1 This is a flowchart illustrating a data generation method provided in an embodiment of this specification. Figure 1 As shown, the method in the embodiments of this specification may include the following steps S102-S110.

[0033] S102, obtain the first product data of the historical product and the second product data of the target product, where the historical product and the target product are of the same type.

[0034] Specifically, considering the limited historical data and scarcity of key features faced by products at specific lifecycle stages, this embodiment proposes to alleviate the prediction difficulties at this stage through data augmentation technology. First, it is necessary to acquire first-stage product data for historical products and second-stage product data for the target product, where the historical and target products are of the same type. Historical products refer to similar products that have passed this specific lifecycle stage and have complete historical data; the target product refers to the product that needs to be predicted at this specific lifecycle stage; the first-stage product data for historical products refers to the complete sales volume, yield data, subscription records, and other time-series data of the historical product within its lifecycle; the second-stage product data for the target product refers to the limited time-series data accumulated by the target product during this specific lifecycle stage. It should be noted that the aforementioned specific lifecycle stage can refer to the initial stage of a product's launch or other data-sparse stages within its lifecycle, characterized by insufficient historical data accumulation, large fluctuations in sales or user behavior, limited sample size, and scarcity of key features; for example, the cold start stage, which refers to the initial period when a financial product or other product is first launched into the market. During this period, there is a lack of sufficient historical data to support conventional prediction models, making the prediction of future sales of the product face significant uncertainty.

[0035] Regarding this step, in some possible implementations, the original data of historical products and the original data of the target product can be obtained through data retrieval, and these can be used as the first product data of the historical products and the second product data of the target products, respectively. In some possible implementations, the original product data can be preprocessed, and the historical product-related data obtained after preprocessing can be used as the first product data of the historical products, and the target product-related data obtained after preprocessing can be used as the second product data of the target products.

[0036] S104, construct multiple views based on the first product data and the second product data respectively, to obtain the first multi-view representation of the historical products and the second multi-view representation of the target products.

[0037] Specifically, to capture the diverse change patterns of products across different time scales and context windows, it is necessary to construct multiple views based on the first and second product data respectively, resulting in a first multi-view representation of the historical product and a second multi-view representation of the target product. Here, multi-view construction refers to generating observation data of the product under different views through multiple preset context windows and time scales; the first multi-view representation of the historical product refers to the set of feature representations of the historical product across multiple time views; and the second multi-view representation of the target product refers to the set of feature representations of the target product across multiple time views.

[0038] Regarding this step, in some possible implementations, a first multi-view representation of the historical product and a second multi-view representation of the target product can be generated based on the first product data of the historical product and the second product data of the target product through multi-scale transformation and feature extraction. In some possible implementations, mathematical operations can be performed on the first product data of the historical product and the second product data of the target product to obtain the first multi-view representation of the historical product and the second multi-view representation of the target product.

[0039] S106, construct the local graph structure within the view and the global graph structure across views based on the first multi-view representation.

[0040] Specifically, in order to establish structured relationships between products, it is necessary to construct intra-view local graph structures and cross-view global graph structures based on the first multi-view representation. Intra-view local graph structures refer to the local relationship graphs between products constructed within a single view; cross-view global graph structures refer to the global relationship graphs between products constructed across different views.

[0041] Regarding this step, in some possible implementations, mapping and inner product calculations can be performed on the first multi-view representation of the historical product to construct an intra-view local graph structure and a cross-view global graph structure. Alternatively, mathematical operations can be performed on the first multi-view representation of the historical product to generate the intra-view local graph structure and the cross-view global graph structure.

[0042] S108, based on the local graph structure within the view and the global graph structure across views, perform covariate enhancement on the second multi-view representation to obtain the first enhanced covariate data of the target product.

[0043] Specifically, to enhance the feature extraction capability of the target product and incorporate relevant information from historical products, it is necessary to perform covariate enhancement on the second multi-view representation based on the local graph structure within the view and the global graph structure across views, thereby obtaining the first enhanced covariate data of the target product. Covariate enhancement refers to the process of enriching the features of the target product by retrieving and aggregating information from similar historical products; the first enhanced covariate data of the target product refers to the enhanced feature data containing historical product information generated through covariate enhancement.

[0044] Regarding this step, in some possible implementations, the first enhanced covariate data of the target product can be generated through information retrieval and aggregation based on the second multi-view representation, the intra-view local graph structure, and the cross-view global graph structure. In some possible implementations, mathematical operations can be performed on the second multi-view representation, the intra-view local graph structure, and the cross-view global graph structure of the target product to obtain the first enhanced covariate data of the target product.

[0045] S110, Based on the first product data and the second product data, generate the second enhanced sample data of the target product; wherein, the first enhanced covariate data and the second enhanced sample data are used to train the prediction model, and the prediction model is used to determine the prediction information of the target product.

[0046] Specifically, to expand the scale and diversity of the training samples, it is necessary to generate second augmented sample data for the target product based on the first and second product data. The first augmented covariate data and the second augmented sample data are used to train the prediction model, which is then used to determine the prediction information for the target product. The second augmented sample data for the target product refers to the augmented time-series data generated through sample augmentation techniques; the prediction model refers to the machine learning model used to predict the future sales of the target product; and the prediction information for the target product refers to the predicted sales volume of the target product within a predetermined future time period.

[0047] Regarding this step, in some possible implementations, second enhanced sample data for the target product can be generated based on the first product data of historical products and the second product data of the target product through data fusion and reconstruction. In other possible implementations, logical operations and mathematical transformations can be performed on the first product data of historical products and the second product data of the target product to produce the second enhanced sample data for the target product.

[0048] The training process of the prediction model can be described as follows: the first enhanced covariate data and the second enhanced sample data of the target product are used as training datasets and input into the preset initial prediction model for training. The parameters of the initial prediction model are continuously adjusted through the backpropagation algorithm until the preset convergence condition is reached. The trained initial prediction model is then used as the final prediction model.

[0049] The process of using a prediction model to determine the forecast information of a target product can be described as follows: acquiring real-time product data of the target product, inputting the real-time product data into a trained prediction model, and outputting the forecast information of the target product through the prediction model, wherein the forecast information includes the sales forecast value of the target product within a preset future time period.

[0050] In this embodiment, firstly, first product data of historical products of the same type and second product data of the target product are acquired. Then, a first multi-view representation of the historical products and a second multi-view representation of the target product are constructed through multi-view construction. Based on the first multi-view representation, an intra-view local graph structure and a cross-view global graph structure are constructed, thereby establishing a multi-dimensional feature description and structured association between products. On this basis, covariate augmentation is performed based on the intra-view local graph structure and the cross-view global graph structure, integrating information from historical products into the target product to generate first augmented covariate data, enriching the feature information used for training. Simultaneously, the scale and diversity of the training data are expanded by generating second augmented sample data. In summary, by generating first augmented covariate data and second augmented sample data, the quality and quantity of the training data are improved, thereby enhancing the accuracy of the prediction information of the target product determined by the prediction model.

[0051] In one embodiment, the step of "constructing multiple views based on the first product data and the second product data respectively to obtain a first multi-view representation of historical products and a second multi-view representation of the target product" can be further refined and may include the following steps: Based on the first product data, multiple preset context windows, and multiple preset time scales, historical product observation data is generated under multiple time views. Based on the second product data, multiple context windows, and multiple time scales, generate observation data of the target product in multiple time views; The preset time-series feature extraction model is invoked to model the observation data of historical products under multiple time views and the observation data of target products under multiple time views, respectively, to obtain the first multi-view representation of historical products and the second multi-view representation of target products.

[0052] Specifically, at a particular stage of a product's lifecycle, the product exhibits diverse change patterns under different time scales and context windows. A single scale and perspective are insufficient to fully capture its dynamic characteristics. This embodiment proposes a technical solution that utilizes multi-scale transformation and multi-view construction.

[0053] On the one hand, it is necessary to generate observation data of historical products under multiple time views based on the first product data, multiple preset context windows, and multiple preset time scales. Among them, the multiple preset context windows refer to sliding windows with different time lengths, used to capture the short-term fluctuations and long-term trends of the product; the multiple preset time scales refer to different time granularities, such as daily, weekly, and monthly, used to analyze product behavior from different time dimensions; the observation data of historical products under multiple time views refers to the multi-dimensional time-series feature matrix obtained after processing through the above context windows and time scales.

[0054] Regarding this step, in some possible implementations, a time series segmentation function can be used to extract the data of the first product to obtain subsequences under different context windows. Then, the subsequences can be converted into observation data at multiple time scales through resampling technology, and the converted observation data can be used as the observation data of historical products under multiple time views.

[0055] On the other hand, it is necessary to generate observation data for the target product across multiple time views based on the second product data, multiple context windows, and multiple time scales. Specifically, the observation data for the target product across multiple time views refers to a multi-dimensional time-series feature matrix generated using the same parameter configuration as historical products.

[0056] Regarding this step, in some possible implementations, the second product data can be processed using the same window and scale parameters as the historical products to obtain observation data of the target product under multiple time views, and this observation data can be directly used for subsequent modeling.

[0057] Next, a pre-defined temporal feature extraction model is invoked to model the observation data of the historical product under multiple time views and the observation data of the target product under multiple time views, respectively, to obtain the first multi-view representation of the historical product and the second multi-view representation of the target product. The pre-defined temporal feature extraction model refers to a Transformer-based temporal feature extraction model that captures dependencies under different time views through a multi-head self-attention mechanism. The first multi-view representation of the historical product refers to a feature tensor containing multiple views, each with a pre-defined feature dimension. The second multi-view representation of the target product refers to a feature tensor with the same number of views and feature dimensions as the historical product but containing only the feature tensor of a single target product.

[0058] Regarding this step, in some possible implementations, the observation data of historical products under multiple time views can be input into a preset time-series feature extraction model, and feature representations under each view can be generated through a feature encoding layer. The feature representations of all views can be combined to obtain the first multi-view representation of the historical product. At the same time, the same operation can be performed on the observation data of the target product under multiple time views to obtain the second multi-view representation of the target product.

[0059] For example, for historical product sequences extracted from the first product data First, for each preset context window and each preset time scale The historical product sequence is processed to generate observation data in a time view. This processing may include: according to the time scale. Historical product sequences are resampled to match the granularity of this time scale; then, based on the context window... A sliding window is applied to the resampled sequence to obtain multiple sub-sequences, which together constitute the observation data under this time view. After performing the above operation on historical products, an observation data set of the historical products under multiple time views can be obtained. Similarly, the same operation is performed on the target product sequence extracted from the second product data to obtain an observation data set of the target product under multiple time views. Next, the observation data set of historical products under multiple time views is input into the time series feature extraction model, i.e., the TSFM model. This model learns feature representations under different time views through a multi-head self-attention mechanism and outputs the first multi-view representation of the historical products. Simultaneously, the observation data sets of the target product under multiple time views are input into the same time-series feature extraction model, and a second multi-view representation of the target product is output. .

[0060] In this embodiment, firstly, based on first product data, multiple preset context windows, and multiple preset time scales, observation data of historical products under multiple time views is generated. Secondly, based on second product data, multiple context windows, and multiple time scales, observation data of the target product under multiple time views is generated. This step utilizes different context windows and time scales to process the product data, enabling the generated observation data to reflect the behavioral characteristics of the product under different time dimensions and local contexts. Subsequently, a preset temporal feature extraction model is invoked to model the observation data of historical products under multiple time views and the observation data of the target product under multiple time views, respectively, to obtain the first multi-view representation of the historical product and the second multi-view representation of the target product. This step uses the temporal feature extraction model to model the multi-view observation data, transforming the observation data under different views into a structured feature representation. Through the above steps, the first and second multi-view representations covering features of multiple time views can be extracted from the original product data, providing a data foundation for subsequently constructing local graph structures within views and global graph structures across views based on the first multi-view representation, thereby creating conditions for covariate augmentation based on graph structures.

[0061] In one embodiment, the step of "constructing a local graph structure within a view and a global graph structure across views based on the first multi-view representation" can be further refined and may include the following steps: The first multi-view representation is mapped based on a preset first mapping matrix to obtain the in-view spatial representation; The inner product is calculated based on the spatial representation within the view to obtain the local graph structure within the view; The first multi-view representation is mapped based on a preset second mapping matrix to obtain a cross-view space representation. The inner product is calculated based on the cross-view space representation to obtain the cross-view global graph structure.

[0062] Specifically, considering the differences in product association patterns within and between different views, the association construction under a single perspective cannot fully reflect the complex relationships between products. This embodiment proposes a technical solution to construct local graph structures within a view and global graph structures across views through dual-path mapping.

[0063] First, the first multi-view representation needs to be mapped based on a preset first mapping matrix to obtain the in-view spatial representation. Here, the preset first mapping matrix refers to the learnable parameter matrix used to project the first multi-view representation onto the in-view space; mapping refers to the mathematical process of converting a high-dimensional feature representation into a specific subspace representation through linear transformation; and the in-view spatial representation refers to the set of feature representations of historical products within a single view.

[0064] Regarding this step, in some possible implementations, the first multi-view representation can be multiplied by a preset first mapping matrix to obtain the in-view spatial representation, and the obtained in-view spatial representation can be used as the representation of historical products in the in-view spatial representation.

[0065] Then, inner product calculation is performed based on the spatial representation within the view to obtain the local graph structure within the view. Here, inner product calculation refers to the mathematical operation of measuring similarity by calculating the dot product between feature vectors.

[0066] Regarding this step, in some possible implementations, the inner product can be calculated for the feature vectors of any two historical products in the spatial representation within the view. The calculation result can be used as the edge weight of the corresponding node in the local graph structure within the view, and the matrix formed by the edge weights can be used as the local graph structure within the view.

[0067] Next, the first multi-view representation is mapped based on a preset second mapping matrix to obtain a cross-view space representation. Here, the preset second mapping matrix refers to the learnable parameter matrix used to project the first multi-view representation onto the cross-view space; the cross-view space representation refers to the set of feature representations of historical products after being fused across multiple views.

[0068] Regarding this step, in some possible implementations, the first multi-view representation can be multiplied by a preset second mapping matrix to obtain a cross-view space representation, and the obtained cross-view space representation can be used as the representation of historical products in the cross-view space.

[0069] Finally, the inner product is calculated based on the cross-view space representation to obtain the cross-view global graph structure.

[0070] Regarding this step, in some possible implementations, the inner product can be calculated for the feature vectors of any two historical products in the cross-view space representation. The calculation result can be used as the edge weight of the corresponding node in the cross-view global graph structure, and the matrix formed by the edge weights can be used as the cross-view global graph structure.

[0071] For example, firstly, a first multi-view representation of a historical product. Based on the preset first mapping matrix, a spatial representation within the view is obtained. This mapping process can be represented as a matrix multiplication operation, where... This is the preset first mapping matrix. Then, based on the spatial representation within the view... Perform inner product calculations to obtain the local diagram structure within the view. The calculation process can be represented as follows: Rather than transpose Multiplication operation, that is Each element in the resulting matrix Characterizes historical products With historical products Similarity within the view. Next, the first multi-view representation for historical products. Based on the preset second mapping matrix, a cross-view space representation is obtained. This mapping process can be represented as a matrix multiplication operation, where... This is the preset second mapping matrix. Finally, based on the cross-view space representation... Perform inner product calculations to obtain the cross-view global graph structure. The calculation process can be represented as follows: Rather than transpose Multiplication operation, that is Each element in the resulting matrix Characterizes historical products With historical products Global relevance across views.

[0072] In this embodiment, the intra-view local graph structure is constructed using a first mapping matrix and inner product calculation, representing the local similarity relationships between historical products within the same view; the cross-view global graph structure is constructed using a second mapping matrix and inner product calculation, representing the global association relationships between historical products in different views. By constructing the intra-view local graph structure and the cross-view global graph structure respectively, a structured data foundation is provided for subsequent steps to retrieve the intra-view neighbor product set and the cross-view neighbor product set for the target product, enabling the subsequent covariate enhancement process to utilize both local and global information for feature enhancement simultaneously.

[0073] In one embodiment, the step of "enhancing the second multi-view representation based on the local graph structure within the view and the global graph structure across views to obtain the first enhanced covariate data of the target product" can be further refined and may include the following steps: The system retrieves the set of neighboring products within the view associated with the target product based on the second multi-view representation and the local graph structure within the view, and retrieves the set of neighboring products across the view associated with the target product based on the second multi-view representation and the global graph structure across the view. A dual-path multi-view diffusion is performed on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information; Sales forecast results for the target product are obtained based on aggregated information. The sales forecast results are used to characterize the sales volume of the target product in a future preset time period. The first enhancing covariate data for the target product is determined based on sales forecast results.

[0074] Specifically, considering that product feature information is limited and dynamically changes in a specific life cycle stage, single-view association retrieval is difficult to fully capture the multi-dimensional similarity between products, and traditional diffusion methods cannot effectively integrate local and global information. This embodiment proposes to systematically enhance the feature representation of target products through dual-path neighbor retrieval and multi-view information diffusion mechanism.

[0075] To comprehensively cover the potential association information of the target product, it is necessary to retrieve the set of neighboring products associated with the target product within the view based on the second multi-view representation and the local graph structure within the view, and to retrieve the set of neighboring products associated with the target product across views based on the second multi-view representation and the global graph structure across views. The set of neighboring products within the view refers to the set of historical products that have high similarity to the target product in a single view, including one-hop neighbors and two-hop neighbors; the set of neighboring products across views refers to the set of historical products that have global association with the target product across different views, including one-hop neighbors and two-hop neighbors.

[0076] Regarding this step, in some possible implementations, the similarity between the target product's in-view representation and the historical product's in-view representation can be calculated to retrieve the set of one-hop neighbor products within the view from the in-view local graph structure. Furthermore, the set of two-hop neighbor products within the view can be retrieved by extending the retrieval through the adjacency relationships of the in-view local graph structure. Simultaneously, based on the similarity between the target product's cross-view representation and the historical product's cross-view representation, the cross-view global graph structure can be used to retrieve cross-view... Figure 1 Jump to neighboring product sets, and further expand the retrieval across views through adjacency relationships in the cross-view global graph structure. Figure 2 Jump-neighbor product set; merge the intra-view one-hop neighbor product set and the intra-view two-hop neighbor product set into an intra-view neighbor product set, and cross-view... Figure 1 Jump Neighbor Product Collection and Cross-View Figure 2 Jump neighbor product sets are merged into cross-view neighbor product sets.

[0077] Then, a dual-path multi-view diffusion is performed on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information. Here, dual-path multi-view diffusion refers to the process of hierarchical information aggregation along intra-view paths and cross-view paths respectively; aggregated information refers to the target product enhanced feature representation generated after fusing intra-view local information and cross-view global information.

[0078] Regarding this step, in some possible implementations, for each product in the in-view neighbor product set, the information of its associated two-hop neighbor products can be aggregated based on the in-view local graph structure, and the representation of the one-hop neighbor products can be updated; the similarity weights between the updated one-hop neighbor product representations and the in-view representations of the target product can be calculated, and the updated one-hop neighbor product representations can be weighted and summed to obtain the in-view aggregated information; the same diffusion operation can be performed on the cross-view neighbor product set to obtain the cross-view aggregated information; the in-view aggregated information and the cross-view aggregated information can be fused, and the fused information of all views can be concatenated to obtain the aggregated information.

[0079] Next, sales forecast results for the target product are obtained based on the aggregated information. These sales forecast results represent the sales volume of the target product over a predetermined future time period. Specifically, the sales forecast results for the target product refer to the sequence of predicted future sales values ​​output by the forecasting model after processing the aggregated information.

[0080] Regarding this step, in some possible implementations, the aggregated information can be input into a preset prediction model, and the aggregated information can be time-series modeled through a fully connected layer or a recurrent neural network to output a series of sales forecast values ​​for the target product in the future preset time period, and this series of sales forecast values ​​can be used as the sales forecast result.

[0081] Finally, the first enhancing covariate data for the target product is determined based on the sales forecast results.

[0082] Regarding this step, in some possible implementations, the predicted value of each time step in the sales forecast results can be used as a covariate feature to construct the first enhanced covariate data with the same dimensions as the original covariates; or the sales forecast results can be concatenated or weighted with the original covariates to generate the first enhanced covariate data containing the enhanced features.

[0083] In this embodiment, historical product information related to the target product is obtained from two dimensions by retrieving the intra-view neighbor product set and the cross-view neighbor product set based on the intra-view local graph structure and the cross-view global graph structure, respectively. Furthermore, a dual-path multi-view diffusion is performed on the intra-view neighbor product set and the cross-view neighbor product set to aggregate the information of two-hop neighbors to one-hop neighbors, and then to the target product, achieving hierarchical information transfer. This aggregated information is then used to obtain the sales forecast result for the target product, which quantifies the impact of historically similar products on future sales. Finally, based on the first enhanced covariate data determined by this sales forecast result, dynamic features predicted based on related products are introduced for the target product, providing training data for the subsequent prediction model and thus affecting prediction performance.

[0084] In one embodiment, the steps of "retrieving the set of neighboring products within the view associated with the target product based on the second multi-view representation and the local graph structure within the view, and retrieving the set of neighboring products across the view associated with the target product based on the second multi-view representation and the global graph structure across the view" can be further refined and may include the following steps: For each view in the second multi-view representation, the in-view similarity is calculated based on the in-view representation of the target product under the view and the in-view representation of the historical products under the view; Based on in-view similarity, retrieve the set of one-hop neighbor products in the view that are associated with the target product; Calculate cross-view similarity based on the cross-view representation of the target product in the view and the cross-view representation of historical products in the view; Based on cross-view similarity, retrieve cross-view similarities associated with the target product. Figure 1 Jump to neighbor product collection; Based on the local graph structure within the view, retrieve the set of two-hop neighbor products within the view for each product in the set of one-hop neighbor products within the view. Based on the cross-view global graph structure, for cross-view Figure 1 Each product in the neighbor product collection can be retrieved across views. Figure 2 Jump to neighbor product collection; Combine the set of one-hop neighbor products within the view and the set of two-hop neighbor products within the view into the set of neighbor products within the view associated with the target product, and combine the cross-view neighbor products. Figure 1 Jump Neighbor product collection, cross-view Figure 2 The jump neighbor product set combination is the cross-view neighbor product set associated with the target product.

[0085] Specifically, considering that product feature information is limited and dynamically changes in a specific life cycle stage, and that association retrieval in a single view is difficult to fully capture the multi-dimensional similarity between products, this embodiment proposes a technical solution to obtain related product information from both a local view within the view and a global view across the view through a dual-path neighbor retrieval mechanism.

[0086] First, for each view in the second multi-view representation, it is necessary to calculate the intra-view similarity based on the intra-view representations of the target product and historical products within the view. Intra-view similarity refers to a metric used to quantify the degree of association between the target product and historical products in the local feature space within a single view.

[0087] Regarding this step, in some possible implementations, cosine similarity or Euclidean distance can be used to calculate the similarity score between the in-view representation of the target product and the in-view representation of each historical product, and the calculated similarity score can be used as the in-view similarity.

[0088] Then, based on the in-view similarity, the set of in-view one-hop neighbor products associated with the target product is retrieved. This set of in-view one-hop neighbor products associated with the target product refers to a set of a predetermined number of historical products that have the highest local similarity to the target product within a single view, selected based on in-view similarity.

[0089] Regarding this step, in some possible implementations, the similarity within the view can be sorted from high to low, a preset number of historical products with the highest ranking can be selected, and the set of selected historical products can be determined as the set of one-hop neighbor products within the view associated with the target product.

[0090] Next, cross-view similarity is calculated based on the cross-view representations of the target product and historical products within the view. Cross-view similarity refers to a metric used to quantify the global correlation between the target product and historical products in the feature space fused from multiple views.

[0091] Regarding this step, in some possible implementations, the correlation strength between the cross-view representation of the target product and the cross-view representation of each historical product can be calculated through inner product operation or attention mechanism, and the calculated correlation strength can be used as the cross-view similarity.

[0092] Subsequently, based on cross-view similarity, cross-view similarities associated with the target product are retrieved. Figure 1 Jump to neighboring product sets. Among them, cross-views associated with the target product. Figure 1 The jump neighbor product set refers to a set of a preset number of historical products that have the highest similarity to the target product across global views, selected based on cross-view similarity.

[0093] Regarding this step, in some possible implementations, cross-view similarity can be sorted from high to low, and a preset number of historical products with the highest rankings can be selected. The set of selected historical products is then defined as the cross-view similarity associated with the target product. Figure 1 Jump to neighbor product collection.

[0094] Then, based on the local graph structure within the view, the set of two-hop neighbor products within the view is retrieved for each product in the set of one-hop neighbor products within the view. Here, the set of two-hop neighbor products within the view refers to the set of other historical products that are directly connected to products in the set of one-hop neighbor products within the view, within the local graph structure of the view.

[0095] Regarding this step, in some possible implementations, for each product in the set of one-hop neighbor products within the view, the adjacency matrix of the local graph structure within the view can be traversed to extract all nodes that have a direct connection relationship with the product, and the historical products corresponding to the extracted nodes can be combined into a set of two-hop neighbor products within the view.

[0096] Subsequently, based on the cross-view global graph structure, cross-view... Figure 1 Each product in the neighbor product collection can be retrieved across views. Figure 2 A collection of products related to skipping neighbors. Among them, cross-view... Figure 2 Jump neighbor product sets refer to products that, in a cross-view global graph structure, are related to cross-view... Figure 1 The set of other historical products that are directly connected to the products in the neighboring product set.

[0097] Regarding this step, among some possible implementations, it can be targeted at cross-view... Figure 1 For each product in the neighbor product set, traverse the adjacency matrix of the cross-view global graph structure, extract all nodes that have direct connections to that product, and form a cross-view structure with the historical products corresponding to the extracted nodes. Figure 2 Jump to neighbor product collection.

[0098] Finally, it is necessary to combine the set of one-hop neighbor products within the view and the set of two-hop neighbor products within the view into the set of neighbor products within the view associated with the target product, and to combine the cross-view neighbor products... Figure 1 Jump Neighbor product collection, cross-view Figure 2 The jump neighbor product set combination is the cross-view neighbor product set associated with the target product.

[0099] Regarding this step, in some possible implementations, a union operation can be used to merge the set of one-hop neighbor products within the view and the set of two-hop neighbor products within the view, and the merged result can be used as the set of neighbor products within the view associated with the target product; a union operation can also be used to merge cross-view neighbor products. Figure 1 Jump Neighbor Product Collection and Cross-View Figure 2 The jump neighbor product set is merged, and the merged result is used as the cross-view neighbor product set associated with the target product.

[0100] For example, for the second multi-view representation Each view First, based on the target product in the view The view below shows With historical products in view The view below shows Calculate intra-view similarity. Specifically, by calculating... and The inner product of these two elements yields a similarity vector, which can then be used to... The operation retrieves the top results with the highest similarity. These historical products constitute the set of one-hop neighboring products within the view associated with the target product. This process can be represented as: ; Next, based on the local graph structure within the view , is a set of neighboring products within the view. Each product in Retrieve the set of two-hop neighbor products within its view. This process can be represented as, for each one-hop neighbor... ,exist Find with The first directly connected ones with the highest connection weight 1 node, namely: ; Subsequently, based on the target product in the view Cross-view representation below Representation of historical products across views Calculate cross-view similarity. Specifically, by calculating... and The inner product of these two elements yields a cross-view similarity vector, which can then be obtained through... The operation retrieves the top results with the highest similarity. A historical product constitutes a cross-perspective related to the target product. Figure 1 Jump Neighbor Product Collection This process can be represented as: ; Subsequently, based on the cross-view global graph structure For cross-view Figure 1 Jump Neighbor Product Collection Each product in Search its cross-view Figure 2 Jump Neighbor Product Collection This process can be represented as, for each one-hop neighbor... ,exist Find with The first directly connected ones with the highest connection weight 1 node, namely: ; Finally, we will target the view. The retrieved set of one-hop neighbor products within the view and the corresponding set of two-hop neighbor products within the view. Combining, for example, through union operations. This yields the set of neighboring products within the view associated with the target product. Simultaneously, it will target the view... Cross-view retrieved Figure 1 Jump Neighbor Product Collection and corresponding cross-view Figure 2 Jump Neighbor Product Collection Combining, for example, through union operations. This yields the set of cross-view neighboring products associated with the target product. (This is related to the second multi-view representation.) The above operation is performed on each view.

[0101] In this embodiment, the intra-view one-hop neighbor product set obtained through intra-view similarity retrieval captures historical products with local features similar to the target product in a single time view. Further, the intra-view two-hop neighbor product set is retrieved based on the intra-view local graph structure, utilizing the intra-view topological relationships to expand the range of locally similar products. Cross-view similarity retrieval is then used to obtain cross-view similarity... Figure 1 The jump neighbor product set captures historical products that are similar to the target product's global features in the multi-view fusion feature space. Further cross-view global graph structure is used for retrieval. Figure 2 The hop neighbor product set leverages cross-view topological relationships to expand the scope of globally similar products. It combines the intra-view one-hop neighbor product set and the intra-view two-hop neighbor product set to extend the cross-view product range. Figure 1 Jump Neighbor Product Collection and Cross-View Figure 2 The jump neighbor product set combination provides neighbor product information containing direct similarity and indirect association for subsequent dual-path multi-view diffusion, enabling the information aggregation process to be based on more comprehensive historical data, thereby enhancing the integrity of feature representation.

[0102] In one embodiment, the step of "performing dual-path multi-view diffusion on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information" can be further refined and may include the following steps: For each first-hop neighbor product in the set of first-hop neighbor products in the view, based on the local graph structure in the view, the information of each product in the set of second-hop neighbor products in the view corresponding to the first-hop neighbor product is aggregated to obtain the updated representation of the first-hop neighbor product. Based on the updated first-hop neighbor representation and the in-view representation of the target product in the view, the similarity between the target product and each first-hop neighbor product is calculated as the propagation weight, and the updated first-hop neighbor representation is weighted and summed according to the propagation weight to obtain the in-view aggregation information of the view. For cross-view Figure 1 For each second-hop neighbor product in the hop neighbor product set, based on the cross-view global graph structure, the cross-view corresponding cross-hop neighbor products are aggregated. Figure 2 The information of each product in the next-hop neighbor product set is used to obtain the updated representation of the second-hop neighbor product; Based on the updated second-hop neighbor representation and the cross-view representation of the target product in the view, the similarity between the target product and each second-hop neighbor product is calculated as the propagation weight, and the updated second-hop neighbor representation is weighted and summed according to the propagation weight to obtain the cross-view aggregation information of the view. The view's intra-view aggregation information and cross-view aggregation information are merged to obtain the view's merged aggregation information; The aggregated information is obtained by splicing together the fused and aggregated information of all views in the second multi-view representation.

[0103] Specifically, considering that product feature information is limited and dynamically changes in a specific life cycle stage, traditional diffusion methods cannot effectively integrate local information within a view and global information across views. This embodiment proposes a dual-path multi-view diffusion mechanism through hierarchical aggregation and weighted fusion to integrate multi-dimensional related product information and enhance the robustness of feature representation.

[0104] First, for each first-hop neighbor product in the in-view set of first-hop neighbor products, based on the in-view local graph structure, it is necessary to aggregate the information of each product in the corresponding in-view set of second-hop neighbor products to obtain the updated representation of the first-hop neighbor product. The updated representation of the first-hop neighbor product refers to the new feature representation generated after weighted aggregation of the feature information of each product in the in-view set of second-hop neighbor products. This new feature representation incorporates information from indirectly related products, enriching the feature dimensions of the first-hop neighbor product.

[0105] Regarding this step, in some possible implementations, for each first-hop neighbor product in the in-view one-hop neighbor product set, the in-view two-hop neighbor product set corresponding to that first-hop neighbor product in the in-view local graph structure can be traversed. Based on the adjacency matrix of the in-view local graph structure, the connection weight between each product in the in-view two-hop neighbor product set and the first-hop neighbor product can be calculated. The feature information of each product in the in-view two-hop neighbor product set can be weighted and summed according to the connection weight to obtain the aggregated information of the first-hop neighbor product. Then, the aggregated information of the first-hop neighbor product can be concatenated or fused with the original feature information of the first-hop neighbor product to generate the updated representation of the first-hop neighbor product.

[0106] Then, based on the updated first-hop neighbor representations and the in-view representation of the target product within the view, the similarity between the target product and each first-hop neighbor product is calculated as a propagation weight. The updated first-hop neighbor representations are then weighted and summed according to these propagation weights to obtain the in-view aggregate information of the view. Here, the propagation weight refers to a metric used to quantify the degree of association between the target product and each first-hop neighbor product in the in-view feature space; the in-view aggregate information of the view refers to the aggregated feature representation that integrates local in-view information, generated by weighting and summing the updated first-hop neighbor representations.

[0107] Regarding this step, in some possible implementations, for each first-hop neighbor product in the in-view one-hop neighbor product set, the cosine similarity or inner product similarity between the in-view representation of the target product and the updated representation of the first-hop neighbor product can be calculated. The calculated similarity score is used as the propagation weight between the target product and the first-hop neighbor product. Then, based on the propagation weight, the updated representations of all first-hop neighbor products in the in-view one-hop neighbor product set are weighted and summed. The weighted sum is used as the in-view aggregation information of the view.

[0108] Next, regarding cross-view Figure 1 For each second-hop neighbor product in the hop neighbor product set, based on the cross-view global graph structure, the cross-view corresponding cross-hop neighbor products are aggregated. Figure 2 The information of each product in the next-hop neighbor product set is used to obtain the updated representation of the second-hop neighbor product. The updated representation of the second-hop neighbor product refers to the information obtained through cross-view... Figure 2 After weighted aggregation of the feature information of each product in the jump neighbor product set, a new feature representation is generated. This new feature representation integrates the information of indirectly related products across views and expands the feature space of the second jump neighbor products.

[0109] Regarding this step, among some possible implementations, it can be targeted at cross-view... Figure 1For each second-hop neighbor product in the hop neighbor product set, traverse the cross-view global graph structure to find the corresponding cross-view product. Figure 2 Jump neighbor product set, based on the adjacency matrix of the cross-view global graph structure, calculate cross-view Figure 2 The connection weights between each product in the hop neighbor product set and the second hop neighbor product are used to determine the cross-view... Figure 2 The feature information of each product in the jump neighbor product set is weighted and summed to obtain the aggregated information of the second jump neighbor product. Then, the aggregated information of the second jump neighbor product is concatenated or merged with the original feature information of the second jump neighbor product to generate the updated representation of the second jump neighbor product.

[0110] Subsequently, based on the updated second-hop neighbor representations and the cross-view representations of the target product within the view, the similarity between the target product and each second-hop neighbor product is calculated as a propagation weight. Then, the updated second-hop neighbor representations are weighted and summed according to these propagation weights to obtain the cross-view aggregation information of the view. Here, the cross-view aggregation information refers to the aggregated feature representation that incorporates global cross-view information, generated by weighting and summing the updated second-hop neighbor representations.

[0111] Regarding this step, among some possible implementations, it can be targeted at cross-view... Figure 1 For each second-hop neighbor product in the hop neighbor product set, calculate the cosine similarity or inner product similarity between the cross-view representation of the target product and the updated representation of the second-hop neighbor product. Use the calculated similarity score as the propagation weight between the target product and the second-hop neighbor product, and then apply the propagation weight to the cross-view... Figure 1 The update representations of all second-hop neighbor products in the hop neighbor product set are weighted and summed, and the weighted sum is used as the cross-view aggregation information for this view.

[0112] Subsequently, in order to integrate local information within the view and global information across views to form a unified view feature representation, it is necessary to fuse the view's intra-view aggregated information and cross-view aggregated information to obtain the view's fused aggregated information. Here, the view's fused aggregated information refers to the comprehensive feature representation generated by concatenating or weighting the intra-view aggregated information and cross-view aggregated information, resulting in a representation that includes both intra-view local features and cross-view global features.

[0113] Regarding this step, in some possible implementations, the intra-view aggregation information and cross-view aggregation information of the view can be concatenated to obtain the concatenated fused features. Then, the concatenated fused features can be fused through a preset fully connected layer or attention mechanism to generate the fused aggregation information of the view. Alternatively, the intra-view aggregation information and cross-view aggregation information of the view can be weighted and summed to obtain the fused aggregation information of the view.

[0114] Finally, in order to integrate the feature information of all views and form a global aggregated representation of the target product, it is necessary to stitch together the fusion and aggregation information of all views in the second multi-view representation to obtain the aggregated information.

[0115] Regarding this step, in some possible implementations, the fusion and aggregation information of all views in the second multi-view representation can be spliced ​​together according to the view order to obtain the spliced ​​global feature, and then the global feature can be used as the aggregation information; or, the fusion and aggregation information of all views in the second multi-view representation can be weighted and fused through a preset feature fusion network to generate the aggregation information.

[0116] For example, for the second multi-view representation Each view First, perform a two-path multi-view diffusion within the view. For a set of one-hop neighbor products within the view... Each first-hop neighbor product Based on the local graph structure within the view Aggregate first-hop neighbor products The corresponding set of two-hop neighbor products within the view Information about each product within the aggregation process. The aggregation process first calculates the join weights. Its formula is: ; in, Represents nodes in the local graph structure within the view. With nodes The connection weights, with the denominator based on all two-hop neighbors. Normalize the data. Then, based on this weight, normalize the set of two-hop neighbor products within the view. Feature information of each product Perform a weighted summation to obtain aggregated information. Its formula is: ; Next, we will introduce the first hop neighbor product. Original feature information With aggregated information The data is stitched together and fused using a multilayer perceptron (MLP) to obtain the first-hop neighbor product. The update indicates Its formula is: ; Then, based on the updated first-hop neighbor representation and target product in view The view below shows Calculate the target product and each of its first-hop neighbor products. Similarity between them as propagation weight Its formula is: ; in, For the first hop neighbor product The view is represented within the text. Finally, based on the propagation weights... Representing all updated first-hop neighbors Perform a weighted summation to obtain the view. Aggregated information within the view Its formula is: ; For cross-view paths, perform the same operations as for intra-view paths. Figure 1 Jump Neighbor Product Collection Each second-hop neighbor product Based on cross-view global graph structure Aggregate its corresponding cross-view Figure 2 Jump Neighbor Product Collection Information on each product in the text has been updated. Then, based on the updated second-hop neighbor representation... and target product in view Cross-view representation below Calculate the propagation weights And perform a weighted summation of all updated second-hop neighbor representations to obtain the view. Cross-view aggregation information .

[0117] Then, view Aggregated information within the view and cross-view aggregation information By merging, a view is obtained. Information fusion and aggregation This fusion process can be represented as a dimensionality reduction after splicing the two elements, i.e.: ; Downsample represents the dimensionality reduction operation.

[0118] Finally, the second multi-view representation All views in the middle (total) (individual) fused and aggregated information The data is then concatenated to obtain the final aggregated information. Its formula is: ; in, This represents the final aggregated information. This aggregated information integrates multi-level and multi-dimensional related product information, both local within the view and global across the view, providing high-quality feature input for subsequent sales forecasting.

[0119] In this embodiment, a dual-path multi-view diffusion mechanism is used to aggregate information from two-hop neighbor products within a view to one-hop neighbor products within the view, and then aggregate the updated representations of one-hop neighbor products within the view to the target product, thus obtaining aggregated information within the view; simultaneously, cross-view information is also aggregated. Figure 2 Information aggregation of neighbor products to cross-view Figure 1 Jump to neighboring products, then cross-view Figure 1 The update representation of skipping neighboring products is aggregated to the target product, resulting in cross-view aggregation information. Further, the intra-view aggregation information and cross-view aggregation information are fused to obtain fused aggregation information for the view. Finally, the fused aggregation information from all views is concatenated to obtain the aggregated information. This process, through hierarchical aggregation and weighted fusion, integrates information from multi-dimensional related products, improving the richness and accuracy of the target product's feature representation and providing high-quality feature input for subsequent predictions.

[0120] In one embodiment, the step of "generating second enhanced sample data of the target product based on the first product data and the second product data" can be further refined and may include the following steps: Identify similar historical products from historical products that have similar multi-view representations to the target product; The product sequence of similar historical products is determined from the first product data, and the product sequence of the target product is determined from the second product data; The preset sequence decomposition model is invoked to decompose the product sequence of similar historical products into the trend, seasonal and residual terms of similar historical products, and to decompose the product sequence of the target product into the trend, seasonal and residual terms of the target product. The trend terms of the target product and the trend terms of similar historical products, the seasonal terms of the target product and the seasonal terms of similar historical products, and the residual terms of the target product and the residual terms of similar historical products are merged to obtain the merged trend terms, merged seasonal terms, and merged residual terms. The trend term, seasonal term, and residual term are reconstructed to obtain the second enhanced sample data of the target product.

[0121] Specifically, the first step is to identify similar historical products from the historical products that have similar multi-view representations to the target product. Here, "similar multi-view representation" refers to the similarity between the multi-view representation of the historical product and the second multi-view representation of the target product in the feature space meeting a preset condition; "similar historical products" refers to historical products that are highly correlated with the target product in terms of multi-view features.

[0122] Regarding this step, in some possible implementations, the similarity score between the second multi-view representation of the target product and the first multi-view representation of each historical product can be calculated, and historical products whose similarity scores meet a preset threshold can be identified as similar historical products; or, according to the similarity scores sorted from high to low, a preset number of historical products ranked at the top can be selected as similar historical products.

[0123] Then, the product sequence of similar historical products is determined from the first product data, and the product sequence of the target product is determined from the second product data. The product sequence of similar historical products refers to the time-series data of similar historical products extracted from the first product data, which reflects the sales volume or scale changes of similar historical products throughout their complete lifecycle. The product sequence of the target product refers to the limited time-series data accumulated by the target product at a specific lifecycle stage, extracted from the second product data, which reflects the sales volume or scale changes of the target product at that stage.

[0124] Regarding this step, in some possible implementations, the corresponding time-series data can be retrieved from the first product data based on the product identifier of similar historical products, and the retrieved time-series data can be used as the product sequence of similar historical products; the corresponding time-series data can be retrieved from the second product data based on the product identifier of the target product, and the retrieved time-series data can be used as the product sequence of the target product.

[0125] Next, a preset sequence decomposition model is invoked to decompose the product series of similar historical products into trend, seasonal, and residual components, and the product series of the target product is also decomposed into trend, seasonal, and residual components. Here, the preset sequence decomposition model refers to a mathematical model used to decompose time-series data into trend, seasonal, and residual components, such as the STL decomposition model; the trend component of similar historical products refers to the long-term trend component in the product series of similar historical products; the seasonal component of similar historical products refers to the seasonal component that changes with a fixed period in the product series of similar historical products; the residual component of similar historical products refers to the random fluctuation component after removing the trend and seasonal components from the product series of similar historical products; the trend component of the target product refers to the long-term trend component in the product series of the target product; the seasonal component of the target product refers to the seasonal component that changes with a fixed period in the product series of the target product; and the residual component of the target product refers to the random fluctuation component after removing the trend and seasonal components from the product series of the target product.

[0126] Regarding this step, in some possible implementations, the product series of similar historical products can be input into a preset sequence decomposition model, which outputs the trend term, seasonal term, and residual term of the similar historical products; the product series of the target product can be input into a preset sequence decomposition model, which outputs the trend term, seasonal term, and residual term of the target product.

[0127] Subsequently, to generate diverse augmented samples while preserving the inherent structural characteristics of the time-series data, it is necessary to fuse the trend term of the target product with the trend terms of similar historical products, the seasonal term of the target product with the seasonal term of similar historical products, and the residual term of the target product with the residual term of similar historical products, respectively, to obtain the fused trend term, fused seasonal term, and fused residual term. The fused trend term refers to the new trend component generated after fusing the trend terms of the target product and similar historical products; the fused seasonal term refers to the new seasonal component generated after fusing the seasonal terms of the target product and similar historical products; and the fused residual term refers to the new random fluctuation component generated after fusing the residual terms of the target product and similar historical products.

[0128] Regarding this step, in some possible implementations, the trend terms of the target product and the trend terms of similar historical products can be weighted and fused to obtain the fused trend term; the seasonal terms of the target product and the seasonal terms of similar historical products can be phase-aligned and then weighted and fused to obtain the fused seasonal term; the residual terms of the target product and the residual terms of similar historical products can be fused through a generative model to obtain the fused residual term.

[0129] Finally, in order to construct complete enhanced time-series data, it is necessary to reconstruct the fused trend term, fused seasonal term, and fused residual term to obtain the second enhanced sample data of the target product.

[0130] Regarding this step, in some possible implementations, the fused trend term, fused seasonal term, and fused residual term can be added together to obtain the reconstructed product sequence, and this reconstructed product sequence can be used as the second enhanced sample data of the target product.

[0131] For example, suppose there is an existing target product sequence and similar historical products sequence First, a preset sequence decomposition model, such as the STL decomposition model, is invoked for the target product. The product sequence is decomposed to obtain the target product. Trend items Seasonal items and residuals This decomposition process, based on a time-series additive model, can be expressed as: ; Among them, trend item Extracted using the Locally Weighted Regression Smoothing (LOESS) method, the formula is as follows: ; Then, the trend term is removed from the original sequence to obtain the detrended sequence. : ; Next, for the detrended sequence Use the LOESS method again to extract the seasonal items. : ; Finally, the residual term The result is obtained by subtracting the trend term and seasonal term from the original sequence: ; For similar historical products Product Series Perform the same decomposition operation to obtain its trend term. Seasonal items and residuals .

[0132] Next, the decomposed items are merged separately. For the trend item, a weighted merging method is used to obtain the merged trend item. : ; in, These are preset weight parameters used to control the target product. and similar historical products The proportion of trend items in the fusion; Random noise is introduced to increase diversity. For the seasonal term, phase alignment is performed first, followed by weighted fusion to obtain the fused seasonal term. : ; Here, PhaseAlign is the phase alignment function used to adjust the phase shift between the two seasonal terms, and then based on the weight parameters. The residuals are then merged. A generative fusion approach is used to obtain the merged residuals. : ; GAN stands for Generative Adversarial Network, which uses two residual sequence terms and weight parameters. As input, generate a new residual sequence that incorporates the combined properties of both.

[0133] Finally, the trend items will be merged. , Integration of seasonal items and fusion residuals Reconstruction was performed to obtain the second enhanced sample data of the target product. : ; in, It is the final generated enhanced sample sequence, which introduces new temporal variation patterns while maintaining a similar structure to the original sequence.

[0134] In this embodiment, sequence decomposition deconstructs the product sequence of the target product and the product sequences of similar historical products into three time-series components of different granularities: trend, seasonality, and residual. This allows the fusion operation on the original sequences to be performed based on independent components, thereby generating a new time-series pattern after reconstruction. Since the product sequences of similar historical products are obtained based on similar multi-view representation filtering, the basic patterns of their components are related to the target product. Therefore, the second augmented sample data of the target product generated by fusion maintains usability in structure. At the same time, differentiated fusion operations are performed on the trend, seasonality, and residual components, introducing variation details different from the original sequences. This increases the diversity of training data while ensuring usability. In summary, this second augmented sample data expands the sample set used for model training, helps improve the generalization ability of the prediction model when faced with limited samples at a specific life cycle stage, and thus improves the accuracy of the target product prediction information.

[0135] In one embodiment, the steps of "merging the trend items of the target product and the trend items of similar historical products, the seasonal items of the target product and the seasonal items of similar historical products, and the residual items of the target product and the residual items of similar historical products to obtain the merged trend item, the merged seasonal item, and the merged residual item" can be further refined and may include the following steps: The trend items of the target product and the trend items of similar historical products are weighted and merged to obtain the merged trend item; After phase alignment, the seasonal items of the target product and the seasonal items of similar historical products are weighted and fused to obtain the fused seasonal item. Generative fusion is performed on the residuals of the target product and the residuals of similar historical products to obtain the fused residuals.

[0136] Specifically, considering that different components have different data characteristics during the time series data fusion process, such as the trend term reflecting long-term change patterns, the seasonal term having periodic phase shifts, and the residual term containing high-frequency random noise, it is difficult to simultaneously guarantee the diversity and accuracy of the generated samples if the same fusion method is used. This embodiment proposes a technical solution that adopts differentiated fusion mechanisms for the trend term, seasonal term, and residual term respectively.

[0137] On the one hand, it is necessary to perform weighted fusion of the trend items of the target product and the trend items of similar historical products to obtain a fused trend item. Here, weighted fusion refers to linearly combining the trend items of the target product and the trend items of similar historical products through preset weight parameters to fuse the information of the two in long-term change patterns.

[0138] Regarding this step, in some possible implementations, the trend items of the target product and the trend items of similar historical products can be linearly weighted based on preset weight parameters, and the result of the linear weighted calculation can be used as the fused trend item. The preset weight parameters can be fixed values ​​or learnable parameters, which are used to adjust the contribution ratio of the trend items of the target product and the trend items of similar historical products during fusion.

[0139] On the other hand, it is necessary to perform phase alignment and weighted fusion on the seasonal items of the target product and the seasonal items of similar historical products to obtain the fused seasonal item. Here, phase alignment and weighted fusion refers to adjusting the phase offset between the seasonal items of the target product and the seasonal items of similar historical products before fusion, so that the periodic features of the two are aligned in the time dimension, and then performing weighted fusion.

[0140] Regarding this step, in some possible implementations, the phase offset between the seasonal items of the target product and the seasonal items of similar historical products can be calculated through cross-correlation or dynamic time warping algorithms. Based on this phase offset, the seasonal items of similar historical products or the seasonal items of the target product are time-shifted to achieve phase alignment. Subsequently, the seasonal items of the phase-aligned target product and the seasonal items of similar historical products are weighted and fused, and the result of the weighted fusion is used as the fused seasonal item.

[0141] On the other hand, it is necessary to perform generative fusion on the residuals of the target product and the residuals of similar historical products to obtain fused residuals. Generative fusion refers to using a generative model to learn the joint distribution of the residuals of the target product and the residuals of similar historical products, and generating a new residual sequence that contains the mixed characteristics of the two.

[0142] Regarding this step, in some possible implementations, the residuals of the target product and the residuals of similar historical products can be used as input data. A pre-defined generative adversarial network model is then called to perform generative fusion to obtain the fused residuals. The generative adversarial network model learns the data distribution characteristics of the two residuals through adversarial training. The output fused residuals retain the statistical characteristics of the original residuals while introducing new random fluctuation patterns, thereby enhancing sample diversity.

[0143] For example, for the target product Trend items and similar historical products Trend items Weighted fusion is performed to obtain the fusion trend item. The formula for this fusion process is: ; in, These are preset weight parameters used to control the target product. The contribution ratio of trend items in the fusion, These are similar historical products. The contribution ratio of the trend item; The introduced random noise is used to increase the diversity of the trend terms after fusion.

[0144] For target products Seasonal items and similar historical products Seasonal items First, phase alignment is performed, followed by weighted fusion to obtain the fused seasonal term. The formula for this fusion process is: ; Here, PhaseAlign represents the phase alignment function, which adjusts the target product. Seasonal items Similar historical products Seasonal items The phase shift between them aligns their periodic characteristics in the time dimension; superscript This indicates that after alignment, based on the weight parameters... The two seasonal terms are weighted and merged.

[0145] For target products residuals and similar historical products residuals The fusion residual term is obtained through generative fusion. The formula for this fusion process is: ; Here, GAN stands for Generative Adversarial Network model, which uses the target product as a reference. residuals Similar historical products residuals and weight parameters As input, the joint data distribution of the two residual terms is learned through adversarial training, and a new residual sequence containing the mixed properties of the two is generated. This generative fusion method, while preserving the statistical properties of the original residuals, introduces a new random fluctuation pattern, thereby enhancing the detail diversity of the generated samples.

[0146] In this embodiment, by employing weighted fusion on the trend term, the basic direction of change in the sequence at the macro level is maintained; by using phase alignment followed by weighted fusion on the seasonal term, the problem of decreased fusion effect caused by time misalignment of periodic features is solved, ensuring the effective combination of periodic patterns; by employing generative fusion on the residual term, new random fluctuation patterns are generated while retaining the statistical characteristics of the original residual term, increasing the detail diversity of the generated samples. In summary, the second enhanced sample data of the target product obtained by reconstructing the target product by fusing the trend term, seasonal term, and residual term maintains structural consistency with the original time series data, while containing variation details different from the original sequence, thereby expanding the diversity of training data while ensuring sample usability.

[0147] In one embodiment, the step of "generatively fusing the residuals of the target product and the residuals of similar historical products to obtain the fused residuals" can be further refined and may include the following steps: Based on preset weight parameters, a preset generative adversarial network model is invoked to perform generative fusion of the residual terms of the target product and the residual terms of similar historical products to obtain the fused residual terms.

[0148] Specifically, considering that the residual terms contain high-frequency random noise during the time series data fusion process, traditional linear fusion methods are difficult to effectively preserve their statistical characteristics and generate diverse new patterns. This embodiment proposes a technical solution of learning the joint distribution of residual terms through generative adversarial networks.

[0149] Specifically, based on preset weight parameters, a preset generative adversarial network (GAN) model is invoked to generatively fuse the residuals of the target product and the residuals of similar historical products to obtain a fused residual. The preset weight parameters are learnable parameters used to control the contribution ratio of the residuals of the target product and the residuals of similar historical products in the generative fusion process. The preset GAN model is an adversarial training framework consisting of a generator and a discriminator. This framework learns the data distribution characteristics of the residuals of the target product and the residuals of similar historical products to generate a new residual sequence containing the hybrid characteristics of both.

[0150] Regarding this step, in some possible implementations, the residuals of the target product and similar historical products can be input as input data into the generator of a pre-defined generative adversarial network (GAN) model. Simultaneously, pre-defined weight parameters are input as conditional information into the generator. Guided by these pre-defined weight parameters, the generator generates candidate fusion residuals based on the feature distributions of the residuals of the target product and similar historical products. These candidate fusion residuals are then input into the discriminator of the pre-defined GAN model. The discriminator determines the authenticity of the candidate fusion residuals based on the statistical characteristics of the residuals of the target product and similar historical products. The generator parameters are optimized through adversarial training between the generator and the discriminator until a pre-defined convergence condition is met. Finally, the candidate fusion residuals output by the generator are used as the fusion residuals.

[0151] For example, the pre-defined generative adversarial network model includes a generator. and a discriminator First, the residuals of the target product are... Residual terms of similar historical products and preset weight parameters As conditional information, along with a random noise vector Common input to generator In the middle. Generator The goal is to learn the data distribution of the true residual term and generate a candidate fusion residual term that is indistinguishable from the true data. This process can be represented as: ; Then, the candidate fusion residual terms Compared with the actual residual (i.e., the residual of the target product) Residual terms of similar historical products () are input together into the discriminator Middle. Discriminator The goal is to determine as accurately as possible whether the input residual is real or generated by the generator, and its output is a probability value. During adversarial training, the discriminator... The goal is to maximize the probability of discrimination against the true residual term and minimize the probability of discrimination against the candidate fused residual term. Its loss function can be expressed as: ; in, Indicates from The true residual term of the sampled data. Meanwhile, the generator... The goal is to minimize the discriminator The discrimination accuracy, that is, to make the generated candidate fusion residuals as real as possible by the discriminator, can be expressed by the loss function as follows: ; Through generator With discriminator The two sides engage in adversarial training, with one gaining strength while the other loses strength, until the model converges and the generator... It can learn the statistical properties of the true residuals. Finally, the trained generator... The output candidate fusion residuals are used as the final fusion residuals. .

[0152] In this embodiment, a generative adversarial network model is used to generatively fuse the residuals of the target product and the residuals of similar historical products. This generates a new random fluctuation pattern while preserving the statistical characteristics of the original residuals, increasing the detail diversity of the generated samples and thus expanding the diversity of the training data while ensuring the availability of the samples.

[0153] Please see Figure 2 This is a flowchart illustrating a model training method provided in the embodiments of this specification. Figure 2 As shown, the method in the embodiments of this specification may include the following steps S202-S204.

[0154] S202, Obtain the first enhanced covariate data and the second enhanced sample data, wherein the first enhanced covariate data and the second enhanced sample data are generated based on the data generation method; S204, The preset initial model is trained based on the first enhanced covariate data and the second enhanced sample data until the preset convergence condition is reached, and the parameter-updated initial model is determined as the prediction model; wherein, the prediction model is used to determine the prediction information of the target product.

[0155] Specifically, considering that product feature information is limited and dynamically changes in a specific life cycle stage, it is difficult to achieve ideal prediction accuracy by directly using raw data to train the prediction model. This embodiment proposes a technical solution to improve model performance by using augmented data.

[0156] First, it is necessary to obtain the first augmented covariate data and the second augmented sample data, which are generated based on the data generation method. The specific details of the first augmented covariate data, the second augmented sample data, and the data generation method have been described in the above embodiments and will not be repeated here.

[0157] In some possible implementations, the above data generation method can be executed to generate first enhanced covariate data and second enhanced sample data; in some possible implementations, the first enhanced covariate data and second enhanced sample data generated based on the above data generation method can be read directly.

[0158] Then, based on the first augmented covariate data and the second augmented sample data, the preset initial model is trained until a preset convergence condition is met. The parameter-updated initial model is then determined as the prediction model. The prediction model is used to determine the prediction information of the target product. Here, the preset initial model refers to an untrained neural network model, such as a recurrent neural network, a long short-term memory network, or an attention mechanism model; the preset convergence condition refers to the model's loss function value on the validation set being less than a preset threshold, or the rate of change of the loss function value being less than a preset threshold over multiple consecutive training epochs, or the training reaching a preset maximum number of epochs; the parameter-updated initial model refers to a model that has been trained and meets the convergence condition; the prediction model refers to a machine learning model used to predict the future sales of the target product; and the prediction information of the target product refers to the predicted sales value of the target product within a preset future time period.

[0159] Regarding this step, in some possible implementations, the first augmented covariate data and the second augmented sample data can be used as the training dataset, and the actual sales data of the target product can be used as the label. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the parameters of the preset initial model are updated using the gradient descent method. In other possible implementations, an input feature sequence can be constructed based on the first augmented covariate data and the second augmented sample data. This input feature sequence is then fed into the preset initial model, and the model parameters are adjusted by minimizing the mean squared error or mean absolute error between the predicted sales data and the actual sales data until the preset convergence condition is met. During training, a learning rate decay strategy or regularization method can also be used to prevent overfitting.

[0160] In this embodiment, a pre-set initial model is trained using first enhanced covariate data and second enhanced sample data. The first enhanced covariate data introduces dynamic features predicted based on associated products, enriching the feature dimensions of the training data; the second enhanced sample data expands the scale and diversity of the training samples. During training, the model can learn more diverse temporal patterns and richer feature representations, thereby reducing the risk of overfitting to a single data source. This enables the trained prediction model to have stronger generalization ability when facing datasets with limited sample size and sparse features, thus enabling it to more accurately determine the prediction information of the target product.

[0161] Please see Figure 3 This is a flowchart illustrating an information prediction method provided in an embodiment of this specification. Figure 3 As shown, the method in the embodiments of this specification may include the following steps S302-S304.

[0162] S302, Obtain third-party product data for the target product.

[0163] S304, based on the third product data of the target product, call the preset prediction model to perform information prediction and obtain the prediction information of the target product; wherein, the prediction model is trained based on the first enhanced covariate data and the second enhanced sample data generated by the data generation method.

[0164] Specifically, considering that product data is sparse and dynamically changes in a specific life cycle stage, it is difficult to guarantee accuracy by directly using the original data for prediction. This embodiment proposes a technical solution to make predictions using a prediction model that has been enhanced and trained.

[0165] First, it is necessary to obtain the third-party product data of the target product. The target product has been described in the above embodiments and will not be repeated here. The third-party product data of the target product refers to the real-time or recent time-series data of the target product at a specific stage of its life cycle, which is used to input the prediction model to generate future predicted values.

[0166] Regarding this step, in some possible implementations, third-party product data of the target product can be acquired in real time using sensors or data acquisition systems. In other possible implementations, historical recent data of the target product can be retrieved from a database as third-party product data.

[0167] Then, in order to accurately predict the future state of the target product, it is necessary to call a preset prediction model based on the third product data of the target product to perform information prediction and obtain the predicted information of the target product. The prediction model is trained based on the first enhanced covariate data and the second enhanced sample data generated by the data generation method. The prediction model, data generation method, first enhanced covariate data, and second enhanced sample data have been described in the above embodiments and will not be repeated here.

[0168] Regarding this step, in some possible implementations, third-product data can be input into the prediction model, and the model can output prediction information through forward propagation. In other possible implementations, third-product data can be combined with external environmental data (such as market indicators) and input into the prediction model to improve prediction accuracy.

[0169] Taking e-commerce applications as an example, the target product can be a newly listed item, and the third-party product data can be real-time interactive data such as click-through rate and add-to-cart rate in the early stages of the product's launch. In this implementation plan, the predictive model outputs forecast information for the target product, which represents the predicted sales value for the product in the coming week, thus assisting in inventory management.

[0170] Taking content recommendation as an example, the target product can be a newly published article or video, and the third-party product data can be the initial click and share data after the content is published. In accordance with this implementation plan, the predictive model outputs predicted information about the target product, which is used to characterize the future popularity trend of the content and optimize the recommendation strategy.

[0171] Taking financial application scenarios as an example, the target product can be a newly issued wealth management product, and the third-party product data can be subscription records and yield fluctuation data in the early stages of the product's launch. In conjunction with this implementation plan, the predictive model outputs forecast information for the target product, which is used to characterize the product's future growth forecast and support asset management decisions.

[0172] In this embodiment, the training process of the prediction model utilizes first augmented covariate data and second augmented sample data obtained through data generation methods. The first augmented covariate data, through dual-path multi-view retrieval and diffusion, introduces information about similar historical products, enriching the feature representation of the target product. The second augmented sample data, through decomposition and multi-scale mixing, expands the diversity of training samples. Therefore, the prediction model learns richer temporal patterns and deeper correlations during the training phase. During the prediction phase, when third-party product data of the target product is input, the model can better capture potential patterns in sparse data scenarios based on its augmented generalization ability learned during training, thereby improving the accuracy of predictions for the target product.

[0173] The following will combine Figure 4 This specification provides a detailed description of the data generation apparatus provided in the embodiments. It should be noted that... Figure 4 The data generation device 1 in this specification is used to execute the data generation device 1 in this specification. Figure 1 The methods shown in the embodiments are illustrated for ease of explanation, showing only the parts related to the embodiments of this specification. For specific technical details not disclosed, please refer to this specification. Figure 1 The embodiment shown. Specifically, the data generation device 1 includes: Acquisition unit 11 is used to acquire the first product data of historical products and the second product data of target products, wherein the historical products and target products are of the same type. The first construction unit 12 is used to construct multiple views based on the first product data and the second product data respectively, to obtain the first multi-view representation of the historical product and the second multi-view representation of the target product; The second construction unit 13 is used to construct a local graph structure within a view and a global graph structure across views based on the first multi-view representation; Enhancement unit 14 is used to enhance the covariates of the second multi-view representation based on the local graph structure within the view and the global graph structure across the view, so as to obtain the first enhanced covariate data of the target product; Generation unit 15 is used to generate second enhanced sample data of the target product based on the first product data and the second product data; The first enhanced covariate data and the second enhanced sample data are used to train the prediction model, which is then used to determine the prediction information for the target product.

[0174] Optionally, the first building unit 12 is also used for: Based on the first product data, multiple preset context windows, and multiple preset time scales, historical product observation data is generated under multiple time views. Based on the second product data, multiple context windows, and multiple time scales, generate observation data of the target product in multiple time views; The preset time-series feature extraction model is invoked to model the observation data of historical products under multiple time views and the observation data of target products under multiple time views, respectively, to obtain the first multi-view representation of historical products and the second multi-view representation of target products.

[0175] Optionally, the first building unit 12 is also used for: The first multi-view representation is mapped based on a preset first mapping matrix to obtain the in-view spatial representation; The inner product is calculated based on the spatial representation within the view to obtain the local graph structure within the view; The first multi-view representation is mapped based on a preset second mapping matrix to obtain a cross-view space representation. The inner product is calculated based on the cross-view space representation to obtain the cross-view global graph structure.

[0176] Optionally, the enhancement unit 14 is also used for: The system retrieves the set of neighboring products within the view associated with the target product based on the second multi-view representation and the local graph structure within the view, and retrieves the set of neighboring products across the view associated with the target product based on the second multi-view representation and the global graph structure across the view. A dual-path multi-view diffusion is performed on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information; Sales forecast results for the target product are obtained based on aggregated information. The sales forecast results are used to characterize the sales volume of the target product in a future preset time period. The first enhancing covariate data for the target product is determined based on sales forecast results.

[0177] Optionally, the enhancement unit 14 is also used for: For each view in the second multi-view representation, the in-view similarity is calculated based on the in-view representation of the target product under the view and the in-view representation of the historical products under the view; Based on in-view similarity, retrieve the set of one-hop neighbor products in the view that are associated with the target product; Calculate cross-view similarity based on the cross-view representation of the target product in the view and the cross-view representation of historical products in the view; Based on cross-view similarity, retrieve cross-view similarities associated with the target product. Figure 1 Jump to neighbor product collection; Based on the local graph structure within the view, retrieve the set of two-hop neighbor products within the view for each product in the set of one-hop neighbor products within the view. Based on the cross-view global graph structure, for cross-view Figure 1 Each product in the neighbor product collection can be retrieved across views. Figure 2 Jump to neighbor product collection; Combine the set of one-hop neighbor products within the view and the set of two-hop neighbor products within the view into the set of neighbor products within the view associated with the target product, and combine the cross-view neighbor products. Figure 1 Jump Neighbor product collection, cross-view Figure 2 The jump neighbor product set combination is the cross-view neighbor product set associated with the target product.

[0178] Optionally, the enhancement unit 14 is also used for: For each first-hop neighbor product in the set of first-hop neighbor products in the view, based on the local graph structure in the view, the information of each product in the set of second-hop neighbor products in the view corresponding to the first-hop neighbor product is aggregated to obtain the updated representation of the first-hop neighbor product. Based on the updated first-hop neighbor representation and the in-view representation of the target product in the view, the similarity between the target product and each first-hop neighbor product is calculated as the propagation weight, and the updated first-hop neighbor representation is weighted and summed according to the propagation weight to obtain the in-view aggregation information of the view. For cross-view Figure 1 For each second-hop neighbor product in the hop neighbor product set, based on the cross-view global graph structure, the cross-view corresponding cross-hop neighbor products are aggregated. Figure 2 The information of each product in the next-hop neighbor product set is used to obtain the updated representation of the second-hop neighbor product; Based on the updated second-hop neighbor representation and the cross-view representation of the target product in the view, the similarity between the target product and each second-hop neighbor product is calculated as the propagation weight, and the updated second-hop neighbor representation is weighted and summed according to the propagation weight to obtain the cross-view aggregation information of the view. The view's intra-view aggregation information and cross-view aggregation information are merged to obtain the view's merged aggregation information; The aggregated information is obtained by splicing together the fused and aggregated information of all views in the second multi-view representation.

[0179] Optionally, the generating unit 15 is also used for: Identify similar historical products from historical products that have similar multi-view representations to the target product; The product sequence of similar historical products is determined from the first product data, and the product sequence of the target product is determined from the second product data; The preset sequence decomposition model is invoked to decompose the product sequence of similar historical products into the trend, seasonal and residual terms of similar historical products, and to decompose the product sequence of the target product into the trend, seasonal and residual terms of the target product. The trend terms of the target product and the trend terms of similar historical products, the seasonal terms of the target product and the seasonal terms of similar historical products, and the residual terms of the target product and the residual terms of similar historical products are merged to obtain the merged trend terms, merged seasonal terms, and merged residual terms. The trend term, seasonal term, and residual term are reconstructed to obtain the second enhanced sample data of the target product.

[0180] Optionally, the generating unit 15 is also used for: The trend items of the target product and the trend items of similar historical products are weighted and merged to obtain the merged trend item; After phase alignment, the seasonal items of the target product and the seasonal items of similar historical products are weighted and fused to obtain the fused seasonal item. Generative fusion is performed on the residuals of the target product and the residuals of similar historical products to obtain the fused residuals.

[0181] Optionally, the generating unit 15 is also used for: Based on preset weight parameters, a preset generative adversarial network model is invoked to perform generative fusion of the residual terms of the target product and the residual terms of similar historical products to obtain the fused residual terms.

[0182] For the effects achievable in this embodiment, please refer to the relevant embodiments of the above data generation method, which will not be repeated here.

[0183] The following will combine Figure 5 This specification provides a detailed description of the model training apparatus provided in the embodiments. It should be noted that... Figure 5 The model training device 2 in this specification is used to execute the instructions. Figure 2 The methods shown in the embodiments are illustrated for ease of explanation, showing only the parts related to the embodiments of this specification. For specific technical details not disclosed, please refer to this specification. Figure 2 The embodiment shown. Specifically, the model training device 2 includes: The acquisition unit 21 is used to acquire the first enhanced covariate data and the second enhanced sample data, wherein the first enhanced covariate data and the second enhanced sample data are generated based on the above data generation method; Training unit 22 is used to train a preset initial model based on the first augmented covariate data and the second augmented sample data until the preset convergence condition is reached, and to determine the parameter-updated initial model as the prediction model. The prediction model is used to determine the prediction information for the target product.

[0184] The effects achievable in this embodiment can be found in the relevant embodiments of the above model training method, and will not be repeated here.

[0185] The following will combine Figure 6 This specification provides a detailed description of the information prediction device provided in the embodiments. It should be noted that... Figure 6 The information prediction device 3 in the specification is used to perform the functions described herein. Figure 3 The methods shown in the embodiments are illustrated for ease of explanation, showing only the parts related to the embodiments of this specification. For specific technical details not disclosed, please refer to this specification. Figure 3 The embodiment shown. Specifically, the information prediction device 3 includes: Acquisition unit 31 is used to acquire third product data of the target product; Prediction unit 32 is used to call a preset prediction model to perform information prediction based on the third product data of the target product, and obtain the prediction information of the target product; The prediction model is trained using the first augmented covariate data and the second augmented sample data generated by the above data generation method.

[0186] For the effects achievable in this embodiment, please refer to the relevant embodiments of the information prediction method described above, which will not be repeated here.

[0187] Please see Figure 7 This document provides a schematic diagram of the structure of an electronic device as an embodiment of the present specification. Figure 7 As shown, the electronic device 1000 may include: at least one processor 1001, such as a CPU; at least one network interface 1004; an input / output interface 1003; a memory 1005; and at least one communication bus 1002. The communication bus 1002 is used to enable communication between these components. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk drive. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 7 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, an input / output interface module, and application programs.

[0188] exist Figure 7 In the electronic device 1000 shown, the input / output interface 1003 is mainly used to provide an input interface for the user and to obtain the user input data.

[0189] In one embodiment, processor 1001 can be used to invoke an application stored in memory 1005 and specifically perform the following operations: Obtain the first product data of historical products and the second product data of the target product, where the historical products and the target products are of the same type. Multi-views are constructed based on the first product data and the second product data respectively, resulting in a first multi-view representation of historical products and a second multi-view representation of the target product; Construct a local graph structure within a view and a global graph structure across views based on the first multi-view representation; Based on the local graph structure within the view and the global graph structure across the view, covariate enhancement is performed on the second multi-view representation to obtain the first enhanced covariate data of the target product. Based on the first product data and the second product data, generate the second enhanced sample data for the target product; The first enhanced covariate data and the second enhanced sample data are used to train the prediction model, which is then used to determine the prediction information for the target product.

[0190] Optionally, when the processor 1001 performs multi-view construction based on the first product data and the second product data respectively to obtain a first multi-view representation of the historical products and a second multi-view representation of the target product, it specifically performs the following operations: Based on the first product data, multiple preset context windows, and multiple preset time scales, historical product observation data is generated under multiple time views. Based on the second product data, multiple context windows, and multiple time scales, generate observation data of the target product in multiple time views; The preset time-series feature extraction model is invoked to model the observation data of historical products under multiple time views and the observation data of target products under multiple time views, respectively, to obtain the first multi-view representation of historical products and the second multi-view representation of target products.

[0191] Optionally, when the processor 1001 constructs the local graph structure within a view and the global graph structure across views based on the first multi-view representation, it specifically performs the following operations: The first multi-view representation is mapped based on a preset first mapping matrix to obtain the in-view spatial representation; The inner product is calculated based on the spatial representation within the view to obtain the local graph structure within the view; The first multi-view representation is mapped based on a preset second mapping matrix to obtain a cross-view space representation. The inner product is calculated based on the cross-view space representation to obtain the cross-view global graph structure.

[0192] Optionally, when processor 1001 performs covariate enhancement on the second multi-view representation based on the in-view local graph structure and the cross-view global graph structure to obtain the first enhanced covariate data of the target product, it specifically performs the following operations: The system retrieves the set of neighboring products within the view associated with the target product based on the second multi-view representation and the local graph structure within the view, and retrieves the set of neighboring products across the view associated with the target product based on the second multi-view representation and the global graph structure across the view. A dual-path multi-view diffusion is performed on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information; Sales forecast results for the target product are obtained based on aggregated information. The sales forecast results are used to characterize the sales volume of the target product in a future preset time period. The first enhancing covariate data for the target product is determined based on sales forecast results.

[0193] Optionally, when the processor 1001 performs the following operations: retrieving the set of neighboring products within the view associated with the target product based on the second multi-view representation and the local graph structure within the view, and retrieving the set of neighboring products across the view associated with the target product based on the second multi-view representation and the global graph structure across the view: For each view in the second multi-view representation, the in-view similarity is calculated based on the in-view representation of the target product under the view and the in-view representation of the historical products under the view; Based on in-view similarity, retrieve the set of one-hop neighbor products in the view that are associated with the target product; Calculate cross-view similarity based on the cross-view representation of the target product in the view and the cross-view representation of historical products in the view; Based on cross-view similarity, retrieve cross-view similarities associated with the target product. Figure 1 Jump to neighbor product collection; Based on the local graph structure within the view, retrieve the set of two-hop neighbor products within the view for each product in the set of one-hop neighbor products within the view. Based on the cross-view global graph structure, for cross-view Figure 1 Each product in the neighbor product collection can be retrieved across views. Figure 2 Jump to neighbor product collection; Combine the set of one-hop neighbor products within the view and the set of two-hop neighbor products within the view into the set of neighbor products within the view associated with the target product, and combine the cross-view neighbor products. Figure 1 Jump Neighbor product collection, cross-view Figure 2 The jump neighbor product set combination is the cross-view neighbor product set associated with the target product.

[0194] Optionally, when processor 1001 performs a two-path multi-view diffusion on the in-view neighbor product set and the cross-view neighbor product set to obtain aggregated information, it specifically performs the following operations: For each first-hop neighbor product in the set of first-hop neighbor products in the view, based on the local graph structure in the view, the information of each product in the set of second-hop neighbor products in the view corresponding to the first-hop neighbor product is aggregated to obtain the updated representation of the first-hop neighbor product. Based on the updated first-hop neighbor representation and the in-view representation of the target product in the view, the similarity between the target product and each first-hop neighbor product is calculated as the propagation weight, and the updated first-hop neighbor representation is weighted and summed according to the propagation weight to obtain the in-view aggregation information of the view. For cross-view Figure 1 For each second-hop neighbor product in the hop neighbor product set, based on the cross-view global graph structure, the cross-view corresponding cross-hop neighbor products are aggregated. Figure 2 The information of each product in the next-hop neighbor product set is used to obtain the updated representation of the second-hop neighbor product; Based on the updated second-hop neighbor representation and the cross-view representation of the target product in the view, the similarity between the target product and each second-hop neighbor product is calculated as the propagation weight, and the updated second-hop neighbor representation is weighted and summed according to the propagation weight to obtain the cross-view aggregation information of the view. The view's intra-view aggregation information and cross-view aggregation information are merged to obtain the view's merged aggregation information; The aggregated information is obtained by splicing together the fused and aggregated information of all views in the second multi-view representation.

[0195] Optionally, when the processor 1001 generates second enhanced sample data of the target product based on the first product data and the second product data, it specifically performs the following operations: Identify similar historical products from historical products that have similar multi-view representations to the target product; The product sequence of similar historical products is determined from the first product data, and the product sequence of the target product is determined from the second product data; The preset sequence decomposition model is invoked to decompose the product sequence of similar historical products into the trend, seasonal and residual terms of similar historical products, and to decompose the product sequence of the target product into the trend, seasonal and residual terms of the target product. The trend terms of the target product and the trend terms of similar historical products, the seasonal terms of the target product and the seasonal terms of similar historical products, and the residual terms of the target product and the residual terms of similar historical products are merged to obtain the merged trend terms, merged seasonal terms, and merged residual terms. The trend term, seasonal term, and residual term are reconstructed to obtain the second enhanced sample data of the target product.

[0196] Optionally, when the processor 1001 performs the following operations to fuse the trend items of the target product and the trend items of similar historical products, the seasonal items of the target product and the seasonal items of similar historical products, and the residual items of the target product and the residual items of similar historical products to obtain the fused trend item, fused seasonal item, and fused residual item: The trend items of the target product and the trend items of similar historical products are weighted and merged to obtain the merged trend item; After phase alignment, the seasonal items of the target product and the seasonal items of similar historical products are weighted and fused to obtain the fused seasonal item. Generative fusion is performed on the residuals of the target product and the residuals of similar historical products to obtain the fused residuals.

[0197] Optionally, when the processor 1001 performs generative fusion of the residuals of the target product and the residuals of similar historical products to obtain the fused residuals, it specifically performs the following operations: Based on preset weight parameters, a preset generative adversarial network model is invoked to perform generative fusion of the residual terms of the target product and the residual terms of similar historical products to obtain the fused residual terms.

[0198] In one embodiment, processor 1001 can be used to invoke an application stored in memory 1005 and specifically perform the following operations: Obtain the first augmented covariate data and the second augmented sample data, which are generated based on the above data generation method; The preset initial model is trained based on the first enhanced covariate data and the second enhanced sample data until the preset convergence condition is reached. The parameter-updated initial model is then determined as the prediction model. The prediction model is used to determine the prediction information for the target product.

[0199] In one embodiment, processor 1001 can be used to invoke an application stored in memory 1005 and specifically perform the following operations: Obtain third-party product data for the target product; Based on the third product data of the target product, a preset prediction model is invoked to perform information prediction and obtain the prediction information of the target product. The prediction model is trained using the first augmented covariate data and the second augmented sample data generated by the above data generation method.

[0200] For the effects achievable in this embodiment, please refer to the relevant embodiments of the above-mentioned data generation method, model training method, and information prediction method; they will not be repeated here.

[0201] This specification also provides a computer storage medium, which stores computer program code that, when executed, performs the above-described functions. Figure 1 - Figure 3 The data generation method, model training method, and information prediction method of the illustrated embodiments can be found in the following documentation for detailed execution processes: Figure 1 - Figure 3 The specific details of the illustrated embodiments will not be elaborated here.

[0202] This specification also provides an embodiment of a computer program product, which stores at least one instruction, which, when executed by a processor, implements the above-described... Figure 1 - Figure 3 The data generation method, model training method, and information prediction method of the illustrated embodiments can be found in the following documentation for detailed execution processes: Figure 1 - Figure 3 The specific details of the illustrated embodiments will not be elaborated here.

[0203] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0204] The above-disclosed embodiments are merely preferred embodiments of this specification and should not be construed as limiting the scope of this specification. Therefore, any equivalent variations made in accordance with the claims of this specification shall still fall within the scope of this specification.

Claims

1. A data generation method, comprising: Obtain first product data of historical products and second product data of target products, wherein the historical products are of the same type as the target products; Based on the first product data and the second product data, multi-views are constructed respectively to obtain a first multi-view representation of the historical product and a second multi-view representation of the target product; Construct a local graph structure within a view and a global graph structure across views based on the first multi-view representation; Based on the local graph structure within the view and the global graph structure across the view, the second multi-view representation is enhanced with covariates to obtain the first enhanced covariate data of the target product. Based on the first product data and the second product data, a second enhanced sample data for the target product is generated; The first enhanced covariate data and the second enhanced sample data are used to train the prediction model, which is used to determine the prediction information of the target product.

2. The method according to claim 1, wherein constructing multiple views based on the first product data and the second product data respectively to obtain a first multi-view representation of the historical product and a second multi-view representation of the target product includes: Based on the first product data, multiple preset context windows, and multiple preset time scales, observation data of the historical product under multiple time views is generated; Based on the second product data, the multiple context windows, and the multiple time scales, observation data of the target product under multiple time views is generated. A preset time-series feature extraction model is invoked to model the observation data of the historical product under multiple time views and the observation data of the target product under multiple time views, respectively, to obtain the first multi-view representation of the historical product and the second multi-view representation of the target product.

3. The method according to claim 1, wherein constructing the intra-view local graph structure and the cross-view global graph structure based on the first multi-view representation comprises: The first multi-view representation is mapped based on a preset first mapping matrix to obtain the in-view spatial representation; The inner product is calculated based on the spatial representation within the view to obtain the local graph structure within the view; The first multi-view representation is mapped based on a preset second mapping matrix to obtain a cross-view space representation; The inner product is calculated based on the cross-view space representation to obtain the cross-view global graph structure.

4. The method according to claim 1, wherein the step of performing covariate enhancement on the second multi-view representation based on the local graph structure within the view and the global graph structure across views to obtain the first enhanced covariate data of the target product includes: The system retrieves a set of neighboring products within the view associated with the target product based on the second multi-view representation and the in-view local graph structure, and retrieves a set of neighboring products across the view associated with the target product based on the second multi-view representation and the cross-view global graph structure. A dual-path multi-view diffusion is performed on the in-view neighbor product set and the cross-view neighbor product set to obtain aggregated information; Based on the aggregated information, a sales forecast result for the target product is obtained, and the sales forecast result is used to characterize the sales volume of the target product in a future preset time period; The first enhanced covariate data for the target product is determined based on the sales forecast results.

5. The method according to claim 4, wherein retrieving the set of intra-view neighbor products associated with the target product based on the second multi-view representation and the intra-view local graph structure, and retrieving the set of cross-view neighbor products associated with the target product based on the second multi-view representation and the cross-view global graph structure, comprises: For each view of the second multi-view representation, the in-view similarity is calculated based on the in-view representation of the target product under the view and the in-view representation of the historical product under the view; Based on the in-view similarity, retrieve the set of one-hop neighbor products in the view associated with the target product; Calculate cross-view similarity based on the cross-view representation of the target product in the view and the cross-view representation of the historical product in the view; Based on the cross-view similarity, retrieve the set of cross-view one-hop neighbor products associated with the target product; Based on the local graph structure within the view, retrieve the set of two-hop neighbor products within the view for each product in the set of one-hop neighbor products within the view. Based on the cross-view global graph structure, retrieve the cross-view second-hop neighbor product set for each product in the cross-view one-hop neighbor product set; The set of one-hop neighbor products within the view and the set of two-hop neighbor products within the view are combined into a set of neighbor products within the view associated with the target product, and the set of one-hop neighbor products across views and the set of two-hop neighbor products across views are combined into a set of neighbor products across views associated with the target product.

6. The method according to claim 5, wherein performing dual-path multi-view diffusion on the intra-view neighbor product set and the cross-view neighbor product set to obtain aggregated information includes: For each first-hop neighbor product in the set of first-hop neighbor products in the view, based on the local graph structure in the view, the information of each product in the set of second-hop neighbor products in the view corresponding to the first-hop neighbor product is aggregated to obtain the updated representation of the first-hop neighbor product. Based on the updated first-hop neighbor representation and the in-view representation of the target product in the view, the similarity between the target product and each first-hop neighbor product is calculated as a propagation weight, and the updated first-hop neighbor representation is weighted and summed according to the propagation weight to obtain the in-view aggregation information of the view. For each second first-hop neighbor product in the cross-view first-hop neighbor product set, based on the cross-view global graph structure, the information of each product in the cross-view second-hop neighbor product set corresponding to the second first-hop neighbor product is aggregated to obtain the updated representation of the second first-hop neighbor product. Based on the updated second-hop neighbor representation and the cross-view representation of the target product in the view, the similarity between the target product and each second-hop neighbor product is calculated as a propagation weight, and the updated second-hop neighbor representation is weighted and summed according to the propagation weight to obtain the cross-view aggregation information of the view. The view's intra-view aggregation information and cross-view aggregation information are fused together to obtain the view's fused aggregation information; The aggregated information is obtained by splicing together the fused and aggregated information of all views in the second multi-view representation.

7. The method according to claim 1, wherein generating second enhanced sample data of the target product based on the first product data and the second product data comprises: Identify similar historical products from the historical products that have a similar multi-view representation to the target product; The product sequence of the similar historical products is determined from the first product data, and the product sequence of the target product is determined from the second product data; The preset sequence decomposition model is invoked to decompose the product sequence of the similar historical products into the trend item, seasonal item and residual item of the similar historical products, and the product sequence of the target product is decomposed into the trend item, seasonal item and residual item of the target product. The trend terms of the target product and the trend terms of similar historical products, the seasonal terms of the target product and the seasonal terms of similar historical products, and the residual terms of the target product and the residual terms of similar historical products are fused to obtain a fused trend term, a fused seasonal term, and a fused residual term. The fusion trend term, the fusion seasonal term, and the fusion residual term are reconstructed to obtain the second enhanced sample data of the target product.

8. The method according to claim 7, wherein fusing the trend term of the target product and the trend term of similar historical products, the seasonal term of the target product and the seasonal term of similar historical products, and the residual term of the target product and the residual term of similar historical products to obtain a fused trend term, a fused seasonal term, and a fused residual term, comprises: The trend items of the target product and the trend items of similar historical products are weighted and fused to obtain a fused trend item; The seasonal items of the target product and the seasonal items of similar historical products are phase-aligned and then weighted and fused to obtain the fused seasonal item. Generative fusion is performed on the residual terms of the target product and the residual terms of the similar historical products to obtain the fused residual terms.

9. The method according to claim 8, wherein the generative fusion of the residuals of the target product and the residuals of the similar historical products to obtain the fused residuals comprises: Based on preset weight parameters, a preset generative adversarial network model is invoked to perform generative fusion on the residual terms of the target product and the residual terms of similar historical products to obtain fused residual terms.

10. A model training method, comprising: Obtain first enhanced covariate data and second enhanced sample data, wherein the first enhanced covariate data and the second enhanced sample data are generated based on the data generation method according to any one of claims 1 to 9; The preset initial model is trained based on the first enhanced covariate data and the second enhanced sample data until the preset convergence condition is reached, and the parameter-updated initial model is determined as the prediction model. The prediction model is used to determine the prediction information for the target product.

11. An information prediction method, comprising: Obtain third-party product data for the target product; Based on the third product data of the target product, a preset prediction model is invoked to perform information prediction, thereby obtaining the predicted information of the target product; The prediction model is trained based on the first enhanced covariate data and the second enhanced sample data generated by the data generation method of any one of claims 1 to 9.

12. A data generation apparatus, comprising: The acquisition unit is used to acquire first product data of historical products and second product data of target products, wherein the historical products are of the same type as the target products. The first construction unit is configured to construct multiple views based on the first product data and the second product data respectively, to obtain a first multi-view representation of the historical product and a second multi-view representation of the target product; The second construction unit is used to construct a local graph structure within a view and a global graph structure across views based on the first multi-view representation; The enhancement unit is used to perform covariate enhancement on the second multi-view representation based on the local graph structure within the view and the global graph structure across the view, to obtain the first enhanced covariate data of the target product. The generation unit is configured to generate second enhanced sample data of the target product based on the first product data and the second product data; The first enhanced covariate data and the second enhanced sample data are used to train the prediction model, which is used to determine the prediction information of the target product.

13. A model training device, comprising: An acquisition unit is configured to acquire first enhanced covariate data and second enhanced sample data, wherein the first enhanced covariate data and the second enhanced sample data are generated based on the data generation method according to any one of claims 1 to 9; The training unit is used to train a preset initial model based on the first augmented covariate data and the second augmented sample data until a preset convergence condition is reached, and to determine the parameter-updated initial model as the prediction model. The prediction model is used to determine the prediction information for the target product.

14. An information prediction device, comprising: The acquisition unit is used to acquire third-party product data of the target product. The prediction unit is used to call a preset prediction model to perform information prediction based on the third product data of the target product, and obtain the prediction information of the target product. The prediction model is trained based on the first enhanced covariate data and the second enhanced sample data generated by the data generation method of any one of claims 1 to 9.

15. A computer-readable storage medium storing computer program code that, when executed, implements the steps of the method as claimed in any one of claims 1 to 11.

16. An electronic device, the electronic device comprising: A processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the steps of the method as claimed in any one of claims 1 to 11.

17. A computer program product storing at least one instruction that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 11.

Images