Time information prediction method and device, storage medium and program product

Abstract

The embodiment of the invention provides a time information prediction method and device, a storage medium and a program product. In the embodiment of the invention, by constructing the space-time heterogeneous graph, organically fusing the time information associated with the commodity information and the address information associated with commodity receiving and sending, and capturing the complex dynamic change relationship in the logistics network, rare data can be supplemented, and meanwhile, the influence range and program of aging fluctuation can be predicted; therefore, the problems of data sparsity and aging volatility are effectively solved, and the prediction accuracy is improved; inverse probability distribution prediction is carried out on the space-time heterogeneous graph through a time information prediction model, inverse cumulative probability distribution of predicted arrival time is learned, and dependence on distribution hypothesis is reduced. Besides, the inverse cumulative probability distribution can express the possibility of the predicted arrival time (such as an abnormal value and an extreme value) under different probabilities instead of the predicted arrival time of a single point, so that the flexibility and robustness of predicting the predicted arrival time are improved.

Description

Technical Field

[0001] This application relates to the field of Internet technologies, and in particular, to a method, device, storage medium, and program product for predicting time information. Background Art

[0002] In the field of e-commerce, accurate prediction of the estimated time of arrival (ETA) of logistics is a key technology for enhancing user experience, optimizing supply chain management, and strengthening the competitiveness of the platform. Accurate ETA prediction can significantly improve user satisfaction, build trust and word-of-mouth, and increase user conversion rates.

[0003] Currently, some methods assume that ETA follows a normal distribution or an exponential distribution. During the process of a user browsing a product, the current time information, the address where the user will receive the product, and the address where the merchant ships the product are obtained. Based on this time information and relevant address information, the arrival time of the product is predicted through a normal distribution or an exponential distribution and transmitted transparently to the product browsing page. Such methods rely on assumptions about specific distributions, and the accuracy of the prediction results is relatively low, affecting the user experience on the e-commerce platform. Summary of the Invention

[0004] Embodiments of this application provide a method, device, storage medium, and program product for predicting time information, so as to improve the accuracy of time information prediction and enhance the user experience on the e-commerce platform.

[0005] An embodiment of this application provides a method for predicting time information, including: obtaining the time information when a user performs a target interaction operation on target product information, and a first address and a second address associated with the target product information; constructing a spatio-temporal heterogeneous graph, where the spatio-temporal heterogeneous graph includes the logistics state information of the logistics network from the first address to the second address in multiple time dimensions, and the multiple time dimensions are determined according to the time information; according to the spatio-temporal heterogeneous graph, invoking a time information prediction model to perform an inverse probability distribution prediction on the estimated delivery time of the target product information, so as to obtain the estimated arrival time of the target product information under multiple known distribution probabilities; according to the estimated arrival times under multiple known distribution probabilities, determining a target estimated arrival time, and displaying the target estimated arrival time on the page where the target product information is located.

[0006] An embodiment of this application further provides an electronic device, including: a memory and a processor; the memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program to implement the steps in the method for predicting time information provided by the embodiments of this application.

[0007] The embodiments of the present application further provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to implement the steps in the time information prediction method provided by the embodiments of the present application.

[0008] The embodiments of the present application further provide a computer program product, including a computer program / instructions, which, when executed by a processor, causes the processor to implement the steps in the time information prediction method provided by the embodiments of the present application.

[0009] In the embodiments of the present application, by constructing a spatio-temporal heterogeneous graph, the time information associated with commodity information and relevant address information are organically integrated to capture the complex dynamic change relationships in the logistics network, which can supplement rare data and predict the influence scope and process of timeliness fluctuations, thus effectively coping with the problems of data sparsity and timeliness volatility and improving the prediction accuracy; through the time information prediction model, inverse probability distribution prediction is performed on the spatio-temporal heterogeneous graph to learn the inverse cumulative probability distribution of the expected arrival time, reducing the dependence on distribution assumptions. In addition, the inverse cumulative probability distribution can express the possibility of the expected arrival time under different probabilities (such as including outliers and extreme values), rather than the expected arrival time of a single point, thereby improving the flexibility and robustness of predicting the expected arrival time.

[0010] Furthermore, during the model training process, frequent samples and rare samples in the entire dataset are considered, and the attention to rare samples is increased through sample density weights, improving the recognition ability for rare samples, thus better handling the problem of data imbalance and improving the accuracy of predicting time information. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings:

[0012] Figure 1 is a schematic flowchart of a time information prediction method provided by an exemplary embodiment of the present application;

[0013] Figure 2 is a schematic framework diagram of a time information prediction method provided by an exemplary embodiment of the present application;

[0014] Figure 3 is a schematic diagram of model training provided by an exemplary embodiment of the present application;

[0015] Figure 4 is a result schematic diagram of a time information prediction device provided by an exemplary embodiment of the present application;

[0016] Figure 5 A schematic structural diagram of an electronic device provided for another exemplary embodiment of the present application. Detailed implementation manners

[0017] To make the objectives, technical solutions, and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in conjunction with specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without making creative efforts shall fall within the protection scope of the present application.

[0018] It should be noted that in the case where the embodiments of the present application involve user information, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the embodiments of the present application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards of relevant countries and regions, and corresponding operation entrances are provided for users to select authorization or rejection. In addition, various models (including but not limited to language models or large models) involved in the present application comply with relevant laws and standards.

[0019] It is necessary for e-commerce platforms to predict logistics ETA. The current ETA prediction methods include the following several types, which will be introduced and explained separately.

[0020] A method for predicting ETA is a regression analysis method based on point prediction. The principle of this method is as follows: Obtain multiple sample data, where the sample data includes order placement time information, address information of the merchant's shipped goods, address information of the user's received goods, and actual arrival time; perform model training on the initial regression model through multiple sample data to obtain a target regression model. Among them, the regression model can include but not be limited to: linear regression, support vector regression, polynomial regression, etc. Subsequently, the time when the user browses the goods, the address information of the merchant's shipped goods, and the address information of the user's received goods can be input into the target regression model, and the target regression model outputs the estimated arrival time of the goods. This method can provide a single point prediction value (i.e., the estimated arrival time), does not model the probability distribution of ETA, and lacks the ability to handle uncertainties and outliers. In addition, data sparsity and data complexity cause the model to be easily affected by outliers, reducing the robustness and accuracy of the prediction.

[0021] Another method for predicting ETA: Assume that ETA conforms to a fixed probability distribution, such as a normal distribution or an exponential distribution, and is modeled through a parametric method. Parametric modeling means assuming that the sample data follows a known probability distribution and fitting the sample data by adjusting a finite number of parameters to obtain an ETA prediction model. When the probability distribution of the actual data deviates from the assumed probability distribution, the accuracy of the ETA prediction model will be significantly reduced. In addition, this method lacks in considering the problem of data imbalance, resulting in poor performance of the model in processing sparse time and low accuracy in predicting sparse events.

[0022] Another method for predicting ETA: Train a machine learning model (such as, random forest or gradient boosting) through sample data, and the machine learning model is modeled based on non-linear feature relationships. However, this method lacks in dealing with the volatility of data failure and has low prediction accuracy.

[0023] The above various methods may be affected by various factors such as the sparsity of order data, the volatility of logistics timeliness, data imbalance, outliers, and uncertainty when predicting ETA, resulting in the problem of low accuracy in predicting ETA.

[0024] Among them, 1) The sparsity of order data means that the order volume in some regions or time periods is small, resulting in sparse training data, making it difficult for the model to capture patterns and having low accuracy in predicting ETA. 2) The volatility of logistics timeliness means that affected by various factors, such as holidays, weather, or traffic, etc., the timeliness fluctuates greatly, and the volatility increases the difficulty of the model predicting ETA and it is difficult to adapt to sudden changes. 3) The order distribution in different regions, logistics companies, or time periods is uneven, resulting in data imbalance. The model may be biased towards the categories with larger quantities and ignore the categories with smaller quantities, affecting the prediction accuracy. 4) There may be outliers in the logistics data, such as abnormal timeliness caused by extreme weather or system errors. The outliers will interfere with the model training and reduce the prediction accuracy.

[0025] To address the problem of low accuracy in predicting ETA, the embodiments of the present application provide a time information prediction method. By constructing a spatio-temporal heterogeneous graph, the time information associated with commodity information and the address information related to the receipt and delivery of commodities are organically integrated to capture the complex dynamic change relationships in the logistics network, which can supplement rare data and at the same time predict the influence range and procedure of timeliness fluctuations, thereby effectively addressing the problems of data sparsity and timeliness volatility and improving the prediction accuracy; Through the time information prediction model, an inverse probability distribution prediction is performed on the spatio-temporal heterogeneous graph to learn the inverse cumulative probability distribution of the expected arrival time, reducing the dependence on distribution assumptions. In addition, the inverse cumulative probability distribution can express the possibility of the expected arrival time under different probabilities (such as, including outliers and extreme values), rather than the expected arrival time at a single point, thereby improving the flexibility and robustness of predicting the expected arrival time.

[0026] Further, during the model training process, frequent samples and rare samples in the entire dataset are considered, and the attention to rare samples is increased through sample density weights, improving the recognition ability for rare samples, thereby better handling the data imbalance problem and improving the accuracy of predicting time information.

[0027] It should be noted that an outlier refers to an expected arrival time in the dataset that significantly deviates from other expected arrival times and may be a real extreme situation. An extreme value refers to the values at both ends of the data distribution. For example, the maximum or minimum expected arrival time. A frequent sample refers to the majority-class samples in the dataset, and frequent samples occupy a large proportion of the dataset. A rare sample refers to a sample with a low occurrence frequency in the dataset, which may represent a small-probability event or an unconventional situation, etc. Rare samples can include outliers and extreme values.

[0028] The following will detail the technical solutions provided by each embodiment of the present application in conjunction with the accompanying drawings.

[0029] Figure 1 It is a schematic flowchart of a method for predicting time information provided by an exemplary embodiment of the present application. As Figure 1 shown, the method includes:

[0030] 101. Obtain the time information of the user's execution of the target interaction operation for the target commodity information, as well as the first address and the second address associated with the target commodity information;

[0031] 102. Construct a spatio-temporal heterogeneous graph, where the spatio-temporal heterogeneous graph includes the logistics state information of the logistics network from the first address to the second address in multiple time dimensions, and the multiple time dimensions are determined according to the time information;

[0032] 103. According to the spatio-temporal heterogeneous graph, call the time information prediction model to perform an inverse probability distribution prediction on the expected delivery time of the target commodity information, so as to obtain the expected arrival time of the target commodity information under multiple known distribution probabilities;

[0033] 104. Determine the target expected arrival time according to the expected arrival times under multiple known distribution probabilities, and display the target expected arrival time on the page where the target commodity information is located.

[0034] In this embodiment, the product information can be any product information under any category displayed on the e-commerce platform. For the convenience of distinction and description, subsequent descriptions will be made by taking the target product information as an example. The target interaction operation refers to the user operation performed on the target product information. From the dimension of the application scenario, the target interaction operation can include, but is not limited to: browsing operation, placing an order operation, favoriting operation, adding to cart operation, liking operation, sharing operation, etc. From the dimension of the implementation method, the target interaction operation can include, but is not limited to: interaction operations in various ways such as touch operation, gesture operation, voice operation, head movement operation, and eye movement operation; among them, the touch operation includes, but is not limited to: click operation, double-click operation, long-press operation, sliding operation, pinching operation, or mouse hovering operation, etc. The sliding operation includes, but is not limited to: linear sliding or curved sliding, etc.

[0035] In this embodiment, the user can perform a target interaction operation on the target product information on the e-commerce platform, obtain the time information of the user performing the target interaction operation, and obtain the first address and the second address associated with the target product information. For example, the first address can be the address information of the target product corresponding to the target product information sent by the merchant (which can be simply referred to as the shipping address), and the second address can be the address information provided by the user for receiving the target product (which can be simply referred to as the receiving address).

[0036] In this embodiment, a Spatio-Temporal Heterogeneous Graph (ST-HG) is constructed based on the time information, the first address, and the second address. The Spatio-Temporal Heterogeneous Graph is a logistics network structure that can reflect complex spatio-temporal dynamic change relationships. The Spatio-Temporal Heterogeneous Graph can include the logistics state information of the logistics network from the first address to the second address in multiple time dimensions. This logistics state information can reflect various factors affecting the logistics network, for example, the distance between the first address and the second address, the time consumption during transportation, the changes in traffic conditions in different time periods, etc.

[0037] Since the graph structure of the Spatio-Temporal Heterogeneous Graph contains different types of nodes (such as shipping points or receiving points, etc.) and edges (representing the association relationships between different nodes), the Spatio-Temporal Heterogeneous Graph can capture richer interaction patterns. Even for a specific node or edge, the data may be relatively sparse, but the information in the overall network can complement each other, thus alleviating the problem of local data sparsity.

[0038] Time-dependent volatility refers to the unpredictable changes in various metrics (such as transit time) in the logistics network over time. Such changes may be caused by external factors (such as weather conditions and traffic situations). Spatiotemporal heterogeneous graphs are adept at learning dynamic changes in time series. By encoding the node states at different times and the relationships between them, the model can learn the evolving trends and periodic patterns over time, which helps better understand and predict time-dependent volatility. There may be spatial dependencies and propagation effects in the logistics network. For example, a traffic jam in a certain area may spread and affect the transportation efficiency in adjacent areas. Spatiotemporal heterogeneous graphs can effectively capture such spatial correlations and combine the information in the time dimension to more accurately predict the scope and degree of the impact of time-dependent volatility.

[0039] It should be noted that different time information, different first addresses, and different second addresses may all result in different spatiotemporal heterogeneous graphs. At the application level, the constructed spatiotemporal heterogeneous graph will also vary if any one of the user, target interaction operation, time of executing the target interaction operation, and commodity information is different.

[0040] Among them, multiple time dimensions can be determined based on the time information of the executed target interaction operation. For example, preset the number of time dimensions, starting from this time information, determine a time dimension every set time unit until the required number of time dimensions is obtained. For instance, if the time of executing the target interaction operation is 15:00, the time unit is 1 hour, and the number of time dimensions is 24, then multiple time dimensions can be realized as the current 16:00, 17:00, …, 23:00, and 0:00, …, 13:00, 14:00 of the next day. Another example is to determine multiple time dimensions within a specific time period before or after the time information of the executed target interaction operation. For example, if the time of executing the target interaction operation is 15:00 in the afternoon, multiple time dimensions may include, but are not limited to, 11:00, 12:00, 13:00, 18:00, 19:00, 21:00, 22:00, 23:00 of the current day, and 0:00, 6:00, 7:00 of the next day, etc.

[0041] In this embodiment, a time information prediction model is provided. The implementation manner of the time information prediction model is not limited, and any model that can perform inverse probability distribution prediction is applicable to the embodiments of the present application. For example, in terms of the model scale, the time information prediction model in the embodiments of the present application can be a model with a relatively large parameter scale, such as a large language model (LLM), or a traditional model with a relatively small parameter scale, which is not limited herein. In terms of the implementation manner of the model, the time information prediction model can include, but is not limited to: Generative Adversarial Networks (GAN), Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN), Deep Neural Networks (DNN), and Residual Network, etc.

[0042] Among them, the time information prediction model does not need to determine the expected arrival time of the target commodity information according to a preset probability distribution function, but can predict the inverse cumulative probability distribution of the target commodity information through inverse probability distribution, reducing the dependence on distribution assumptions. The time information prediction model can predict the inverse cumulative probability distribution, and the inverse cumulative probability distribution can express the possibility of the expected arrival time (which can include outliers and extreme values) under different probabilities, rather than outputting a single-point expected arrival time, thereby improving the flexibility and robustness of predicting the expected arrival time.

[0043] In this embodiment, according to the spatio-temporal heterogeneous graph, the time information prediction model is called to perform inverse probability distribution prediction on the expected delivery time of the target commodity information, so as to obtain the expected arrival time of the target commodity information under multiple known distribution probabilities.

[0044] For example, a spatio-temporal heterogeneous graph can be input into a time information prediction model, and the inverse probability distribution prediction of the estimated delivery time of the target commodity information can be performed according to the spatio-temporal heterogeneous graph to obtain an inverse cumulative probability distribution. The inverse cumulative probability distribution can be represented by an Inverse Cumulative Distribution Function (ICDF). It can be considered that the inverse cumulative probability distribution includes the estimated arrival times under multiple known distribution probabilities. The multiple known distribution probabilities are not limited. For example, from 0% to 100%, a known distribution probability can be determined every 1%, and the multiple known distribution probabilities include 0%, 1%, …, 99% and 100%. Or, a known distribution probability can be determined every 10%, and the multiple known distribution probabilities include 0%, 10%, …, 90% and 100%. Or, the interval between the multiple known distribution probabilities is non-uniform, and the multiple known distribution probabilities can include: 90%, 95%, 98% and 100%. There is no limitation on this.

[0045] Among them, multiple known distribution probabilities can be input into the time information prediction model in advance, or multiple known distribution probabilities can be built into the time information prediction model as model parameters. There is no limitation on this.

[0046] In this embodiment, the target estimated arrival time can be determined according to the estimated arrival times under multiple known distribution probabilities. Among them, the implementation manner of determining the target estimated arrival time is not limited. For example, the estimated arrival times under the known distribution probabilities that exceed the set probability threshold among the multiple known distribution probabilities can be used as the target estimated arrival time. The probability threshold can include but is not limited to: 95%, 98% or 99%. Or, at least two candidate known distribution probabilities that exceed the set probability threshold among the multiple known distribution probabilities can be determined, and the weighted average of the estimated arrival times under the at least two candidate known distribution probabilities can be calculated as the target estimated arrival time. The weights can be adjusted based on the importance or historical performance of each candidate known distribution probability.

[0047] In this embodiment, the target estimated arrival time is displayed on the page where the target commodity information is located. Among them, the display form of the target estimated arrival time can include but is not limited to: plain text, rich text, static image, dynamic image, audio, or video, etc. The display form of the target estimated arrival time can include but is not limited to: embedded in the page, pop-up window, drop-down menu, card, and button, etc.

[0048] In one implementation, the user can perform a target interaction operation on the first page, where the target product information can be displayed on the first page, and the target estimated arrival time can also be displayed on the first page. For example, the first page can be the product details page, the target interaction operation can be the browsing operation of the user on the product details page, and the target estimated arrival time can also be displayed on the product details page. For example, the target estimated arrival time is displayed in text form on the first page. Exemplarily, one target estimated arrival time is "Place an order now, estimated to be delivered at 5 pm".

[0049] In another implementation, the user can perform a target interaction operation on the first page and jump to the second page, where the target product information is displayed on the second page. Additionally, the target estimated arrival time of the target product information can be displayed on the second page. For example, the target interaction operation is the order placement operation for the target product information, the first page is the order placement page, and the second page is the purchased page, which includes the target product information and its target estimated arrival time.

[0050] It should be noted that in the case where the embodiments of the present application involve the jump between the first page and the second page, the jump methods involved in the embodiments of the present application include but are not limited to: directly jumping from the first page to the second page, first jumping from the first page to the task page and then jumping to the second page when the corresponding task operation is completed on the task interface; the completion of the corresponding task operation on the task interface includes but is not limited to: when the task interface is implemented as a game interface, completing the game operation on the game interface; when the task interface is implemented as an identity authentication interface, completing the identity authentication on the identity authentication interface; when the task interface is implemented as a recharge interface, completing the recharge operation on the recharge interface; and so on.

[0051] In an optional embodiment, the implementation manner of constructing the spatio-temporal heterogeneous graph is not limited. An implementation manner is exemplarily provided below. A spatio-temporal heterogeneous graph is constructed through a spatio-temporal heterogeneous graph construction model. The implementation manner of the spatio-temporal heterogeneous graph construction model is not limited, and any model that can construct a spatio-temporal heterogeneous graph is applicable to the embodiments of the present application. For example, from the perspective of the model scale, the spatio-temporal heterogeneous graph construction model in the embodiments of the present application can be a model with a relatively large parameter scale, such as a large language model (LLM), or a traditional model with a relatively small parameter scale, which is not limited in this regard. From the perspective of the implementation manner of the model architecture, the spatio-temporal heterogeneous graph construction model can include, but is not limited to: Generative Adversarial Networks (GAN), Convolutional Neural Networks (CNN), Recurrent Neural Network (RNN), Deep Neural Networks (DNN), Residual Network, Encoder-Only architecture, Decoder-Only architecture, Encoder-Decoder architecture, architecture based on self-attention mechanism, etc.

[0052] Among them, the internal structure of the spatio-temporal heterogeneous graph construction model is not limited. Exemplarily, taking the spatio-temporal heterogeneous graph construction model including a graph convolutional network and an encoding network as an example for illustration. A Graph Convolutional Network (GCN) is a neural network for processing graph structures. It updates the representation of each node by aggregating the features of the node itself and its neighbor nodes, and can be regarded as a local smoothing operation on the graph, enabling adjacent nodes to share information. In the embodiments of the present application, the first address and the second address, as well as the adjacent addresses between the first address and the second address, are used as nodes in the spatio-temporal heterogeneous graph. The adjacent address is address information on the logistics route between the first address and the second address. There can be multiple logistics routes between the first address and the second address, and the number of adjacent addresses can be one or more. The encoding network is used to encode information for the time dimension and the space dimension (such as the first address and the second address).

[0053] Among them, in addition to the graph convolutional network, convolution aggregation can also be performed through a graph sampling and aggregation network or a graph convolutional network based on the attention mechanism, which is not limited in this regard. Among them, the connection relationship between nodes is weighted through the attention mechanism, making the model more flexible and capable of handling imbalance and noise in the graph.

[0054] Among them, the encoding network can also be implemented using a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU). Among them, the LSTM network is suitable for capturing long-distance dependencies in time series and is widely used for time series data prediction, but it does not handle spatial changes and dependencies sufficiently. Compared with LSTM, GRU has higher computational efficiency and shows good prediction performance in time series data.

[0055] Specifically, the first address and the second address are input into the graph convolutional network of the spatio-temporal heterogeneous graph construction model for convolutional aggregation to obtain a spatial feature map. The spatial feature map includes: a logistics network composed of the first address, the second address, and at least one adjacent address; the spatial feature map and the time information are input into the encoding network of the spatio-temporal heterogeneous graph construction model for encoding in the time dimension to obtain a spatio-temporal heterogeneous graph. The spatio-temporal heterogeneous graph includes: physical state information of the logistics network composed of the first address, the second address, and at least one adjacent address in multiple time dimensions. It should be noted that before constructing the spatio-temporal heterogeneous graph, the time information, the first address, and the second address can be encoded to obtain the time features corresponding to the time information, and the spatial features corresponding to the first address and the second address respectively; based on the time features and the above spatial features, the spatio-temporal heterogeneous graph is constructed.

[0056] Among them, the convolution in convolutional aggregation refers to the Graph Convolution operation. Graph convolution can update the representation of each node by aggregating the information of the node and its adjacent nodes. Aggregation refers to the process of collecting and integrating information from adjacent nodes. It usually involves some form of summarization of the features of the node itself and its adjacent nodes. Aggregation methods can include but are not limited to: average aggregation, pooling aggregation, sum aggregation, and attention mechanism-based aggregation, etc. Among them, in the graph convolutional network, the spatial relationship between nodes is defined by the adjacency matrix. When updating the node features, the features of the node itself and its adjacent nodes are weighted to obtain the updated node features. The weights for weighting can be defined by the adjacency matrix. The effect of this is that each node can "learn" useful information from its adjacent nodes, thereby better capturing the local structural information in the graph.

[0057] The structure of the spatio-temporal heterogeneous graph is not limited. For example, taking the first address, the second address, and the adjacent address as the shipping point, the receiving point, and the adjacent point respectively, Figure 2In this example, the spatio-temporal heterogeneous graph is shown with one adjacent node and three time dimensions, but it is not limited to this. Among them, the edges between the shipping point and the receiving point, and the edges between the shipping point and the adjacent nodes are called shipping edges, the edges between the receiving point and the adjacent nodes are called adjacent edges, and the edges between the shipping points, the edges between the receiving points, and the edges between the adjacent nodes are called time edges.

[0058] Among them, by performing convolutional aggregation on the logistics network information composed of the first address, the second address, and the adjacent address, the spatial dependencies and complex relationships in the logistics network can be captured. Further, by combining the physical state information in the time dimension, the status of the package at different times and locations can be predicted more accurately, thereby improving the predictability of the entire logistics process.

[0059] In an alternative embodiment, according to the spatio-temporal heterogeneous graph, the implementation manner of calling the time information prediction model to perform inverse probability distribution prediction on the expected delivery time of the target commodity information to obtain the expected arrival time of the target commodity information under multiple known distribution probabilities is not limited. The following is an exemplary description. The spatio-temporal heterogeneous graph is input into the time information prediction model. Multiple known distribution probabilities are built into the time information prediction model. Multiple intermediate features are generated according to the number of multiple known distribution probabilities and the spatio-temporal heterogeneous graph. Inverse probability distribution prediction is performed on the multiple intermediate features to obtain the expected arrival time of the target commodity information under multiple known distribution probabilities. Among them, the intermediate features contain information about the nodes involved in the logistics network and their mutual relationships, and can be used to predict the expected arrival time of the target commodity information under different known distribution probabilities. The number of intermediate features can change with the number of known distribution probabilities. For example, if there are 100 multiple known distribution probabilities, then the number of intermediate features is 100. Subsequently, inverse probability distribution prediction can be performed based on the multiple intermediate features to obtain the expected arrival time under multiple known distribution probabilities. Among them, the intermediate features can reflect rich logistics status information, which helps the model to more accurately predict the expected arrival time of the target commodity.

[0060] Optionally, the implementation manner of generating multiple intermediate features according to the number of multiple known distribution probabilities and the spatio-temporal heterogeneous graph is not limited. Multiple intermediate features can be generated by the feature cross network in the time information prediction model according to the number of multiple known distribution probabilities and the spatio-temporal heterogeneous graph. Any network that can perform feature crossing is applicable to the embodiments of the present application. For example, generative adversarial networks, convolutional neural networks, recurrent neural networks, deep neural networks (DNN), and residual networks, etc.

[0061] Specifically, an implementation method for generating multiple intermediate features based on the quantity of multiple known distribution probabilities and a spatio-temporal heterogeneous graph includes: generating a first spliced feature according to the spatio-temporal heterogeneous graph and the attribute information of the target commodity; inputting the first spliced feature into a feature cross network for feature crossing to obtain a second spliced feature; discretizing the second spliced feature according to the quantity of multiple known distribution probabilities to generate multiple intermediate features. Among them, the discretization can be a vector representation of the second spliced feature.

[0062] The attribute information of the target commodity may include but is not limited to: commodity category, size, weight, shelf life, inventory, function, and rating, as well as the merchant's level, delivery speed, and business scope, etc. For example, feature extraction can be performed on the attribute information of the target commodity to obtain commodity features, and the commodity features and the spatio-temporal heterogeneous graph are spliced to obtain the first spliced feature.

[0063] Among them, the internal structure of the feature cross network is not limited. For example, the feature cross network includes a cross module and a depth module. The cross module can explicitly construct interaction features, and the depth module is used to learn non-linear feature relationships. The combination of the two modules can enhance the processing ability of high-dimensional data.

[0064] Before inputting the first spliced feature into the feature cross network, it also includes inputting the first spliced feature into the feature embedding module of the feature cross network to perform feature embedding on the first spliced feature, and converting high-dimensional sparse categorical variables into low-dimensional dense vectors. As Figure 3 shown, x0 represents the first spliced feature passing through the feature embedding module, and the feature embedding layer is not shown in the figure. In Figure 3 , the cross module includes L layers, and the output of each layer can be expressed as x1, x2…x L1 . Taking the first layer as an example, the output of the first layer is expressed as x1 = x0x0 T w c,0 +b c,0 +x0, where the output of the second layer and the L1 layer is similar, and it can combine the information of the previous layer or the previous several layers. The weights w c,0 and b c,0 are different. The depth module includes L layers, and the output of each layer can be expressed as The output of the first layer is expressed as h1 = ReLu(w h,0 x0 + b h,0 ), the output of the second layer and the L1 layer is similar, and it can combine the information of the previous layer or the previous several layers. The weights w h,0 and b h,0 are different. The output result of the cross module and the output result of the depth module are spliced to obtain the second spliced feature.

[0065] Among them, by splicing the spatio-temporal heterogeneous graph and the attribute information of the target commodity to generate the first splicing feature, it is possible to effectively integrate multi-source information (such as time, space, commodity characteristics, etc.), enrich the information volume of the input data. This helps the model to more comprehensively understand the complex structure of the input data, thereby improving the prediction accuracy. Using the feature cross network to process the first splicing feature to obtain the second splicing feature can automatically mine and construct high-order feature interactions while retaining the original feature information, capturing more complex patterns and relationships.

[0066] Optionally, the implementation manner of predicting the inverse probability distribution of multiple intermediate features to obtain the estimated arrival time of the target commodity information under multiple known distribution probabilities is not limited. The time information prediction model includes an inverse probability distribution prediction network for performing inverse probability distribution prediction. The implementation manner of the inverse probability distribution prediction network is not limited, and generative adversarial networks, convolutional neural networks, recurrent neural networks, deep neural networks, and residual networks can be used.

[0067] The following provides an example. Input multiple intermediate features into the inverse probability distribution prediction network of the time information prediction model; perform feature transformation on the multiple intermediate features to obtain multiple time values corresponding to multiple known distribution probabilities; for any known distribution probability, accumulate the time values corresponding to other known distribution probabilities less than or equal to any known distribution probability to obtain the estimated arrival time under any known distribution probability. Among them, by accumulating the time values corresponding to other known distribution probabilities less than or equal to any known distribution probability to obtain the estimated arrival time under this probability, this method allows the model to output results with clear probability meanings. This way improves the transparency and interpretability of the model, facilitating users to understand the logic behind the prediction.

[0068] For example, if the 3 known distribution probabilities are 0.1, 0.8, and 0.9 respectively, and the 3 time values are 10 hours, 8 hours, and 6 hours respectively, then the estimated arrival time corresponding to the known distribution probability 0.1 is 10 hours, the estimated arrival time corresponding to the known distribution probability 0.8 is 10 + 8 = 18 hours, and the estimated arrival time corresponding to the known distribution probability 0.9 is 10 + 8 + 6 = 24 hours.

[0069] Such as Figure 3As shown, an example is illustrated where the inverse probability distribution prediction network includes a feature transformation module and an accumulation module. Any module that can transform intermediate features into time values can be used in the embodiments of the present application. An exemplary feature transformation module can use an activation function to transform multiple intermediate features into multiple time values corresponding to multiple known probability distributions. For example, the activation function can include, but is not limited to: rectified linear unit (relu), softplus function, sigmoid function (such as, Sigmoid), and softmax function. For example, the accumulation module can include, but is not limited to: cumulative sum (cumsum) module, accumulator module, and aggregate functions, etc.

[0070] In an alternative embodiment, considering that probability distribution prediction can provide multiple predicted arrival times corresponding to commodity information and their corresponding probability values. Assume that the probability distribution of the predicted arrival time is predicted through a model to obtain the cumulative probability distribution of the predicted arrival time, and this cumulative probability distribution can be represented by the Cumulative Distribution Function (CDF). There are some problems in predicting the cumulative probability distribution. That is, during the model training process, a loss function needs to be constructed based on the difference between the cumulative probability distribution predicted by the model and the true cumulative probability distribution, so as to realize model training based on the constructed loss function.

[0071] However, constructing the true cumulative probability distribution requires counting the predicted arrival times in the sample dataset. Constructing the true cumulative probability distribution based on the counted data may result in information loss, leading to a low accuracy of the model in predicting the predicted arrival time.

[0072] In the embodiments of the present application, to solve the problems that occur during the probability distribution prediction process, inverse probability distribution prediction is performed based on the time information prediction model to obtain the inverse cumulative probability distribution of the predicted arrival time. The inverse cumulative probability distribution can be represented by the Inverse Cumulative Distribution Function (ICDF). Among them, inverse probability distribution prediction refers to the inverse operation of probability distribution prediction. The inverse cumulative probability distribution includes: multiple known distribution probabilities and the predicted arrival times under multiple known distribution probabilities. During the model training process, a loss function is constructed based on the difference between the inverse cumulative probability distribution and the sample true predicted arrival time to realize model training.

[0073] Based on this, the training process of the time information prediction model includes: obtaining a plurality of sample data, where the sample data includes: the sample time information for performing an interaction operation on the sample commodity information, and the first sample address, the second sample address associated with the sample commodity information, and the actual arrival time of the sample commodity information; constructing a sample spatio-temporal heterogeneous graph, where the sample spatio-temporal heterogeneous graph includes the logistics state information of the logistics network from the first sample address to the second sample address in multiple time dimensions, and the multiple time dimensions are determined according to the sample time information; according to the sample spatio-temporal heterogeneous graph, calling the initial time information prediction model to perform an inverse probability distribution prediction on the estimated delivery time of the sample commodity information, so as to obtain the estimated arrival time of the sample commodity information under multiple known distribution probabilities; constructing an objective loss function according to the difference between the estimated arrival time and the actual arrival time under multiple known distribution probabilities; aiming at the objective that the loss function satisfies the iteration termination condition, performing iterative training on the initial time information prediction model to obtain the target time information prediction model.

[0074] Among them, for the introduction of the sample time information, the first sample address, the second sample address, the sample spatio-temporal heterogeneous graph, and the inverse probability distribution prediction, reference can be made to the relevant introductions of the time information, the first address, the second address, the spatio-temporal heterogeneous graph, and the inverse probability distribution prediction in the foregoing embodiments, which will not be elaborated here. Among them, the iteration termination conditions include but are not limited to: reaching the set number of iterations, reaching the set iteration time, or the value of the loss function being less than the set loss threshold, and the loss threshold can include but is not limited to: 0.05, 0.01, and 0.001, etc.

[0075] Among them, constructing an objective loss function based on the difference between the estimated arrival time and the actual arrival time under multiple known distribution probabilities can help more accurately measure the model prediction error and make targeted optimization adjustments. The design of the loss function directly affects the learning direction and final performance of the model. Therefore, a reasonable setting of the loss function is the key to improving the model performance.

[0076] Optionally, the implementation manner of constructing the target loss function according to the difference between the predicted arrival time and the actual arrival time under multiple known distribution probabilities is not limited. Exemplarily, an implementation manner of constructing the target loss function according to the difference between the predicted arrival time and the actual arrival time under multiple known distribution probabilities includes: generating a function curve corresponding to the actual arrival time based on a preset first value and second value, where the horizontal axis and the vertical axis of the function curve are the predicted arrival time and the known distribution probability respectively. In the function curve, the known distribution probability corresponding to the predicted arrival time greater than or equal to the actual arrival time is the first value, and the known distribution probability corresponding to the predicted arrival time less than the actual arrival time is the second value, and the first value is greater than the second value; constructing the target loss function according to the difference information between the probability values of the predicted arrival times under multiple known distribution probabilities on the function curve, where the probability value is the first value or the second value. In the above manner, it is not necessary to statistically calculate the true cumulative distribution probability, reducing information loss and improving the model training accuracy.

[0077] Among them, the implementation form of the function curve corresponding to the actual arrival time is not limited. For example, it can be implemented as a step function curve. Among them, the first value and the second value are not limited. For example, the first value and the second value are values between 0 and 1, and the first value is greater than the second value. For example, the first value can be 1 and the second value can be 0. Or, the first value can be 0.9 and the second value can be 0.1, and this is not limited. The horizontal axis and the vertical axis of the function curve corresponding to the actual arrival time are the predicted arrival time and the known distribution probability respectively. The value range of the predicted arrival time is from 0 to positive infinity, and the value range of the known distribution probability on the vertical axis is from 0 to 1. Among them, the known distribution probability corresponding to the predicted arrival time greater than or equal to the actual arrival time on the vertical axis can be set as the first value (such as 1), and the known distribution probability corresponding to the predicted arrival time less than the actual arrival time on the vertical axis can be set as the second value (such as 0).

[0078] Exemplarily, a target loss function is: Among them, CRPS is the Continuous Ranked Probability Score, representing the probability distribution loss function, y is the actual arrival time, {F -1 (q1), F -1 (q2), …, F -1 (q N )} is the inverse probability distribution predicted by the time information prediction model. q i is the known distribution probability, N is the number of known distribution probabilities, H is the step function, F is the cumulative probability distribution function (CDF), and F -1 represents the inverse cumulative probability distribution function (ICDF). Among them, at Figure 3Among them, for the inverse probability distribution {F -1 (q1), F -1 (q2), …, F -1 (q N )}, it is shown.

[0079] Further optionally, the implementation manner of constructing the target loss function according to the difference information between the probability values of the expected arrival times on the function curve under multiple known distribution probabilities and the multiple known distribution probabilities is not limited. Considering that during the process of model training, the learned model parameters tend to be biased towards frequent samples and easily ignore rare samples (which may be as important as frequent samples). To solve this problem, sample density weights can be introduced to weight the loss function, improving the accuracy of the model's prediction of rare events, thereby better reducing the problems brought by data imbalance.

[0080] Specifically, an initial loss function is constructed according to the difference information between the probability values of the expected arrival times on the function curve under multiple known distribution probabilities and the multiple known distribution probabilities; according to the actual arrival times included in multiple sample data, the sample density weights of the multiple sample data are determined respectively, where the sample density value of the sample data is negatively correlated with the sample density weight; according to the sample density weights of the multiple sample data, the initial loss function is weighted to construct the target loss function.

[0081] For example, according to the actual arrival times included in multiple sample data, the sample density values of the multiple sample data can be determined respectively, and based on the sample density values of the multiple sample data, the sample density weights of the multiple sample data are determined respectively. Among them, the kernel density estimation method can be used to calculate the density values of each sample data in the sample space. The kernel density estimation method is a non-parametric method for estimating the probability density of sample data. By setting a kernel function at each sample data, each kernel function can generate a smooth density contribution around it, and the density contributions generated by all kernel functions can be superimposed to form an overall density estimate, that is, the probability density function. By inputting the actual arrival time of the sample data into this probability density function, the sample density of the sample data can be obtained.

[0082] Exemplarily, the above CRPS function can be used as the initial loss function. On this basis, the target loss function can be implemented as where W d ’ (y) represents the sample density weight, and d represents the sample density.

[0083] It should be noted that, in addition to the above method of introducing sample density weights to solve the imbalance problem, the following methods can also be used: 1) Sample resampling. Oversampling (such as Synthetic Minority Over-sampling Technique, SMOTE) and undersampling (such as random undersampling) methods are used to adjust the data distribution to balance the classes. 2) Cost-sensitive learning. Class weights are introduced during the model training process, especially higher weights are assigned to rare events to reduce the negative impact of imbalanced data on the model. 3) Ensemble learning methods. Such as Adaboost and random forest, which enhance the processing effect of imbalanced problems by combining multiple weak classifiers.

[0084] In the embodiments of the present application, the model including the time information prediction model and the spatio-temporal heterogeneous graph construction model can be referred to as the Density-Sensitive Spatio-Temporal Graph Inverse Cumulative Probability Forecasting (DSTG-ICPF) model.

[0085] The method provided by the embodiments of the present application can bring the following technical effects:

[0086] 1) Significantly improve the prediction accuracy of ETA:

[0087] In the actual application of the e-commerce platform, the DSTG-ICPF model significantly reduces the error of ETA prediction. Some tests were conducted in the set application scenarios and it was found that: the recall rate of high-quality timeliness (such as next-day delivery) can be increased by about 20%, for example, it can be increased from 10% to 35%, and the accuracy can be increased by about 5%, for example, it can be increased from 85% to 90%. The improvement degree will vary depending on the scenario.

[0088] 2) Enhance the adaptability in complex logistics environments:

[0089] By leveraging the deep capture of spatio-temporal dependencies through spatio-temporal heterogeneous graph (ST-HG) modeling and the optimization of data imbalance by density-sensitive continuous ranked probability score (DenseCRPS), the model demonstrates stronger adaptability in complex logistics networks. This enables the model to flexibly handle various complex scenarios such as large-scale promotions, maintaining efficient prediction while supporting the smooth operation of logistics.

[0090] 3) Facilitate the operation decision-making of the e-commerce platform:

[0091] Provide a more comprehensive probability distribution prediction, support the formulation and implementation of refined operation strategies (such as the next-day delivery display ratio under different achievement rates), improve the response speed and fluency of the overall supply chain, and achieve higher operation efficiency.

[0092] Through the optimization of the above technical effects, the DSTG-ICPF model not only improves the prediction accuracy, but also brings substantial improvements in aspects such as user experience and operation efficiency, providing strong support for logistics management and decision-making.

[0093] The time information prediction method provided by the embodiments of the present application can be implemented on a terminal device or can be implemented by the cooperation of a terminal device and a server device. No matter which implementation manner, the following will introduce the detailed steps of the time information prediction method.

[0094] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can also be executed by different devices as the execution subject. For example, the execution subject of steps 101 to 103 can be device A; for another example, the execution subject of steps 101 and 102 can be device A, and the execution subject of step 103 can be device B; and so on.

[0095] In addition, in some processes described in the above embodiments and the accompanying drawings, there are multiple operations that appear in a specific order. However, it should be clearly understood that these operations can be executed not in the order in which they appear in this article or in parallel. The operation numbers such as 101, 102, etc. are only used to distinguish different operations, and the numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and these operations can be executed in order or in parallel. It should be noted that the descriptions such as "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0096] Figure 4 It is a schematic structural diagram of a process parameter processing device provided by another exemplary embodiment of the present application. As Figure 4 shown, the device includes: an acquisition module 41, a construction module 42, a call module 43, a determination module 44, and a display module 45.

[0097] The acquisition module 41 is used to acquire the time information of the user's execution of the target interaction operation for the target commodity information, as well as the first address and the second address associated with the target commodity information;

[0098] The construction module 42 is used to construct a spatio-temporal heterogeneous graph, and the spatio-temporal heterogeneous graph includes the logistics state information of the logistics network from the first address to the second address in multiple time dimensions, and the multiple time dimensions are determined according to the time information;

[0099] A calling module 43 is used to call a time information prediction model according to a spatio-temporal heterogeneous graph to perform an inverse probability distribution prediction on the expected delivery time of the target commodity information, so as to obtain the expected arrival times of the target commodity information under multiple known distribution probabilities.

[0100] A determining module 44 is used to determine a target expected arrival time according to the expected arrival times under multiple known distribution probabilities, and a display module 45 is used to display the target expected arrival time on the page where the target commodity information is located.

[0101] In an alternative embodiment, the construction module is specifically configured to: input the first address and the second address into the graph convolutional network of the spatio-temporal heterogeneous graph construction model for convolutional aggregation to obtain a spatial feature map, where the spatial feature map includes a logistics network composed of the first address, the second address, and at least one adjacent address; input the spatial feature map and the time information into the encoding network of the spatio-temporal heterogeneous graph construction model for encoding in the time dimension to obtain a spatio-temporal heterogeneous graph, where the spatio-temporal heterogeneous graph includes the physical state information of the logistics network composed of the first address, the second address, and at least one adjacent address in multiple time dimensions.

[0102] In an alternative embodiment, the calling module is specifically configured to: input the spatio-temporal heterogeneous graph into the time information prediction model, and generate multiple intermediate features according to the number of multiple known distribution probabilities and the spatio-temporal heterogeneous graph, where the multiple known distribution probabilities are built in the time information prediction model; perform an inverse probability distribution prediction on the multiple intermediate features to obtain the expected arrival times of the target commodity information under multiple known distribution probabilities.

[0103] Optionally, the calling module is specifically configured to: generate a first concatenated feature according to the spatio-temporal heterogeneous graph and the attribute information of the target commodity; input the first concatenated feature into the feature cross network of the time information prediction model for feature crossing to obtain a second concatenated feature; discretize the second concatenated feature according to the number of multiple known distribution probabilities to generate multiple intermediate features.

[0104] Optionally, the calling module is specifically configured to: input the multiple intermediate features into the inverse probability distribution prediction network of the time information prediction model; perform feature transformation on the multiple intermediate features to obtain multiple time values corresponding to multiple known probability distributions; for any known distribution probability, accumulate the time values corresponding to other known distribution probabilities that are less than or equal to the any known distribution probability to obtain the expected arrival time under the any known distribution probability.

[0105] In an optional embodiment, the device further includes a training module configured to obtain a plurality of sample data, where the sample data includes: sample time information for performing an interaction operation on sample commodity information, and a first sample address, a second sample address associated with the sample commodity information, and the actual arrival time of the sample commodity information; construct a sample spatio-temporal heterogeneous graph, where the sample spatio-temporal heterogeneous graph includes logistics state information of a logistics network from the first sample address to the second sample address in multiple time dimensions, and the multiple time dimensions are determined according to the sample time information; according to the sample spatio-temporal heterogeneous graph, call an initial time information prediction model to perform an inverse probability distribution prediction on the estimated delivery time of the sample commodity information, so as to obtain the estimated arrival time of the sample commodity information under multiple known distribution probabilities; construct a target loss function according to the difference between the estimated arrival time under multiple known distribution probabilities and the actual arrival time; and take the satisfaction of the iterative termination condition of the loss function as the goal, and perform iterative training on the initial time information prediction model to obtain a target time information prediction model.

[0106] Optionally, the training module is specifically configured to: generate a function curve corresponding to the actual arrival time based on a preset first value and a second value, where the horizontal axis and the vertical axis of the function curve are the estimated arrival time and the known distribution probability respectively. In the function curve, the known distribution probability corresponding to the estimated arrival time greater than or equal to the actual arrival time is the first value, and the known distribution probability corresponding to the estimated arrival time less than the actual arrival time is the second value, and the first value is greater than the second value; construct a target loss function according to the difference information between the probability values of the estimated arrival times under multiple known distribution probabilities on the function curve and the multiple known distribution probabilities, and the probability value is the first value or the second value.

[0107] Further optionally, the training module is specifically configured to: construct an initial loss function according to the difference information between the probability values of the estimated arrival times under multiple known distribution probabilities on the function curve and the multiple known distribution probabilities; determine the sample density weights of the multiple sample data according to the actual arrival times included in the multiple sample data, where the sample density of the sample data is negatively correlated with the sample density weight; and weight the initial loss function according to the sample density weights of the multiple sample data to construct a target loss function.

[0108] Regarding the detailed implementation manners and beneficial effects of each step in the Figure 4 device provided in the embodiments of the present application have been described in detail in the foregoing embodiments, and will not be elaborated herein.

[0109] Figure 5 This is a schematic structural diagram of an electronic device provided in an embodiment of the present application. As Figure 5 shown, in practice, the electronic device includes: a memory 54 and a processor 55.

[0110] A memory 54 for storing computer programs and configurable to store various other data to support operations on the electronic device. Examples of such data include instructions for any application or method for operating on the electronic device, data structures.

[0111] A processor 55, coupled to the memory 54, for executing the computer programs in the memory 54 to perform the time information prediction method described in the foregoing embodiments. For details, please refer to the descriptions of the foregoing embodiments and will not be elaborated herein.

[0112] Regarding what is provided in the embodiments of the present application Figure 5 The detailed implementation manners and beneficial effects of the steps in the device shown have been described in detail in the foregoing embodiments and will not be elaborated herein.

[0113] Furthermore, as Figure 5 shown, the electronic device further includes: other components such as a communication component 56, a display 57, a power supply component 58, an audio component 59, etc. Figure 5 Only some components are schematically shown in, which does not mean that the electronic device only includes Figure 5 the components shown in. Additionally, Figure 5 the components within the dashed box in are optional components, rather than mandatory components, and may vary depending on the product form of the electronic device. The electronic device of this embodiment can be implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, or an IOT device, or as a server device such as a conventional server, a cloud server, or a server array. If the electronic device of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, etc., it may include Figure 5 the components within the dashed box in; if the electronic device of this embodiment is implemented as a server device such as a conventional server, a cloud server, or a server array, it may not include Figure 5 the components within the dashed box in.

[0114] The above-mentioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0115] The above-mentioned communication component is configured to facilitate communication between the device where the communication component is located and other devices in a wired or wireless manner. The device where the communication component is located can access a wireless network based on a communication standard, such as a mobile communication network like 2G, 3G, 4G / LTE, 5G, etc., or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel.

[0116] The above-mentioned display includes a screen, and the screen can include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a touch panel, the screen can be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can not only sense the boundaries of touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operation.

[0117] The above-mentioned power supply component provides power for various components of the device where the power supply component is located. The power supply component can include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device where the power supply component is located.

[0118] The above-mentioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), and when the device where the audio component is located is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive external audio signals. The received audio signals can be further stored in the memory or transmitted via the communication component. In some embodiments, the audio component further includes a speaker for outputting audio signals.

[0119] Accordingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor is enabled to implement the steps in the above method embodiments. Among them, the computer-readable storage medium can be implemented by volatile or non-volatile or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium

[0120] Accordingly, an embodiment of the present application further provides a computer program product. The computer program product includes a computer program or instructions. When the computer program or instructions are executed by a processor, the processor is enabled to implement the steps in the above method embodiments. It should be understood that each process or a combination of multiple processes in the above method flow can be implemented by the computer program or instructions. In addition, these computer programs or instructions can be applied to the processors of general-purpose computers, special-purpose computers, embedded processors or other programmable data processing devices, so that the processors of general-purpose computers, special-purpose computers, embedded processors or other programmable data processing devices can be used as devices to implement the corresponding functions in the above method embodiments.

[0121] It should also be noted that the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, commodity or device including the element.

[0122] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various modifications and changes can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A time information prediction method, characterized in that: include: Acquire time information when a user performs a target interactive operation on target product information, and a first address and a second address associated with the target product information; Constructing a spatiotemporal heterogeneous graph, wherein the spatiotemporal heterogeneous graph includes logistics status information of a logistics network from the first address to the second address in multiple time dimensions, wherein the multiple time dimensions are determined according to the time information; According to the spatiotemporal heterogeneous graph, a time information prediction model is called to perform inverse probability distribution prediction on the estimated delivery time of the target product information, so as to obtain the estimated arrival time of the target product information under multiple known distribution probabilities; A target estimated arrival time is determined according to the estimated arrival times under the multiple known distribution probabilities, and the target estimated arrival time is displayed on the page where the target product information is located.

2. The method according to claim 1, characterized in that Construct a spatiotemporal heterogeneous graph, including: Inputting the first address and the second address into a graph convolution network of a spatiotemporal heterogeneous graph construction model for convolution aggregation to obtain a spatial feature graph, wherein the spatial feature graph includes: a logistics network consisting of the first address, the second address, and at least one adjacent address; The spatial feature map and the time information are input into the encoding network of the spatiotemporal heterogeneous graph construction model to encode the time dimension and obtain a spatiotemporal heterogeneous graph, wherein the spatiotemporal heterogeneous graph includes: physical state information of the logistics network composed of the first address, the second address and at least one adjacent address in multiple time dimensions.

3. The method according to claim 1, characterized in that According to the spatiotemporal heterogeneous graph, a time information prediction model is called to perform inverse probability distribution prediction on the estimated delivery time of the target product information to obtain the estimated arrival time of the target product information under multiple known distribution probabilities, including: Inputting the spatiotemporal heterogeneous graph into a time information prediction model, generating a plurality of intermediate features according to the number of a plurality of known distribution probabilities and the spatiotemporal heterogeneous graph, wherein the plurality of known distribution probabilities are built into the time information prediction model; An inverse probability distribution prediction is performed on the multiple intermediate features to obtain an estimated arrival time of the target product information under multiple known distribution probabilities.

4. The method according to claim 3, characterized in that According to the number of multiple known distribution probabilities and the spatiotemporal heterogeneous graph, multiple intermediate features are generated, including: Generate a first splicing feature according to the spatiotemporal heterogeneous graph and the attribute information of the target product; Inputting the first splicing feature into the feature crossover network of the time information prediction model to perform feature crossover to obtain a second splicing feature; The second concatenated feature is discretized according to the number of the plurality of known distribution probabilities to generate a plurality of intermediate features.

5. The method according to claim 3, characterized in that: Performing inverse probability distribution prediction on the multiple intermediate features to obtain the estimated arrival time of the target product information under multiple known distribution probabilities includes: Inputting the plurality of intermediate features into an inverse probability distribution prediction network of the temporal information prediction model; Performing feature conversion on the multiple intermediate features to obtain multiple time values ​​corresponding to multiple known probability distributions; For any known distribution probability, time values ​​corresponding to other known distribution probabilities that are less than or equal to the known distribution probability are accumulated to obtain the estimated arrival time under the known distribution probability.

6. The method according to any one of claims 1 to 5, characterized in that: Also includes: Acquire multiple sample data, the sample data including: sample time information for performing interactive operations on sample commodity information, and a first sample address and a second sample address associated with the sample commodity information, and an actual arrival time of the sample commodity information; Constructing a sample spatiotemporal heterogeneous graph, wherein the sample spatiotemporal heterogeneous graph includes logistics state information of a logistics network from the first sample address to the second sample address in multiple time dimensions, wherein the multiple time dimensions are determined according to the sample time information; According to the sample spatiotemporal heterogeneous graph, the initial time information prediction model is called to perform inverse probability distribution prediction on the estimated delivery time of the sample product information to obtain the estimated arrival time of the sample product information under multiple known distribution probabilities; Constructing a target loss function according to the difference between the estimated arrival time under the multiple known distribution probabilities and the actual arrival time; With the loss function satisfying the iteration termination condition as the goal, the initial time information prediction model is iteratively trained to obtain the target time information prediction model.

7. The method according to claim 6, characterized in that According to the difference between the estimated arrival time under the multiple known distribution probabilities and the actual arrival time, a target loss function is constructed, including: Based on a preset first value and a second value, a function curve corresponding to the actual arrival time is generated, wherein the horizontal axis and the vertical axis of the function curve are the estimated arrival time and the known distribution probability, respectively, and in the function curve, the known distribution probability corresponding to the estimated arrival time greater than or equal to the actual arrival time is the first value, and the known distribution probability corresponding to the estimated arrival time less than the actual arrival time is the second value, and the first value is greater than the second value; A target loss function is constructed according to the probability value of the estimated arrival time under the multiple known distribution probabilities on the function curve and the difference information between the multiple known distribution probabilities, and the probability value is the first value or the second value.

8. The method according to claim 7, characterized in that Constructing a target loss function according to the probability value of the estimated arrival time under the multiple known distribution probabilities on the function curve and the difference information between the multiple known distribution probabilities, including: constructing an initial loss function according to the probability value of the estimated arrival time under the multiple known distribution probabilities on the function curve and the difference information between the multiple known distribution probabilities; Determine, according to the actual arrival time included in the plurality of sample data, a sample density weight of each of the plurality of sample data, wherein the sample density of the sample data is negatively correlated with the sample density weight; The initial loss function is weighted according to the sample density weights of each of the plurality of sample data to construct a target loss function.

9. An electronic device, characterized in that: include: Memory and processor; The memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program to implement the steps in the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to implement the steps in the method according to any one of claims 1 to 8.

11. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the processor is caused to implement the steps in any one of the methods of claims 1-8.

Images