User traffic prediction method, device, equipment, medium and program product

Abstract

The invention provides a user traffic prediction method which can be applied to the technical field of artificial intelligence. The method comprises the following steps: acquiring first user traffic time sequence data in a three-dimensional tensor form; based on the first user traffic time sequence data, executing local feature extraction and dimension reduction processing to obtain dimension reduction feature sequence data; based on the dimension reduction feature sequence data, executing multi-layer long time sequence dependence modeling processing to obtain depth time sequence feature sequence data; on the basis of the depth time sequence feature sequence data, feature weighting processing is executed to obtain context vector data; and based on the context vector data, performing prediction processing based on regression to obtain user traffic prediction data. The invention further provides a user traffic prediction device and equipment, a storage medium and a program product.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically to a user traffic prediction method, apparatus, device, medium, and program product. Background Technology

[0002] As the physical carriers of banking operations, bank branches are crucial for banks to assess service quality and customer satisfaction. However, branches commonly face resource misallocation issues due to fluctuating customer traffic, resulting in persistent long customer wait times. This not only impacts customer experience but also poses a risk of customer churn. Current technical solutions in this field have limitations. One popular approach is to use image recognition technology, such as acquiring surveillance video images from branches and then processing them by frame segmentation and display to count customer traffic. However, this method is overly complex and redundant, with a cumbersome process, and its predictive model is too simplistic to form an effective predictive model. Furthermore, other time-series-based prediction techniques, whether traditional statistical methods or conventional machine learning methods, often fail to meet the stability and accuracy requirements for refined operational scheduling when dealing with highly complex, nonlinear, and dynamically changing data such as bank customer traffic. Summary of the Invention

[0003] In view of the above problems, this application provides user traffic prediction methods, apparatus, devices, media and program products.

[0004] According to a first aspect of this application, a user traffic prediction method is provided, the method comprising: acquiring first user traffic time series data in three-dimensional tensor form; performing local feature extraction and dimensionality reduction processing based on the first user traffic time series data to obtain dimensionality-reduced feature sequence data; performing multi-layer long-time-series dependency modeling processing based on the dimensionality-reduced feature sequence data to obtain deep-time-series feature sequence data; performing feature weighting processing based on the deep-time-series feature sequence data to obtain context vector data; and performing regression-based prediction processing based on the context vector data to obtain user traffic prediction data.

[0005] According to an embodiment of this application, obtaining the first user traffic time series data in three-dimensional tensor form includes: obtaining the second user traffic time series data after conversion based on a proportional range; and performing a time series conversion process based on a sliding window based on the second user traffic time series data to obtain the first user traffic time series data in three-dimensional tensor form.

[0006] According to an embodiment of this application, obtaining the second user traffic time series data after proportional range conversion includes: obtaining the original user traffic time series data; performing missing value imputation processing on the original user traffic time series data to obtain complete user traffic time series data; and performing normalization processing based on a preset numerical range on the complete user traffic time series data to obtain the second user traffic time series data.

[0007] According to an embodiment of this application, the step of performing a sliding window-based time series transformation process based on the second user traffic time series data to obtain the first user traffic time series data in a three-dimensional tensor form includes: performing sliding window sampling processing along the time axis of the second user traffic time series data with a preset time step as the window size to obtain multiple input-output sample pairs, wherein the input of each input-output sample pair is a sequence of data with consecutive time steps, and the output of each input-output sample pair is data of a single time step immediately following the consecutive time steps; performing combination processing based on the input sequence data within the multiple input-output sample pairs to generate the first user traffic time series data in a three-dimensional tensor form, wherein the three dimensions of the three-dimensional tensor form represent the number of samples, the time step, and the number of features, respectively.

[0008] According to an embodiment of this application, the step of performing local feature extraction and dimensionality reduction processing based on the first user traffic time series data to obtain dimensionality-reduced feature sequence data includes: performing one-dimensional convolution processing on the first user traffic time series data to obtain local feature data representing local time patterns; performing nonlinear transformation processing on the local feature data to obtain nonlinearized feature data; and performing dimensionality reduction processing based on max pooling on the nonlinearized feature data to obtain the dimensionality-reduced feature sequence data that retains the time series dimension.

[0009] According to an embodiment of this application, the step of performing multi-layer long temporal dependency modeling processing based on the dimensionality-reduced feature sequence data to obtain deep temporal feature sequence data includes: performing a first long temporal dependency modeling process based on the dimensionality-reduced feature sequence data to obtain a first hidden state sequence data representing surface temporal dependencies; and performing a second long temporal dependency modeling process based on the first hidden state sequence data to obtain the deep temporal feature sequence data representing deep temporal dependencies.

[0010] According to an embodiment of this application, the step of performing feature weighting processing based on the deep temporal feature sequence data to obtain context vector data specifically includes: performing probability weight allocation processing on feature vectors at different time steps in the deep temporal feature sequence data to obtain the weight corresponding to each time step; and performing weighted summation processing on feature vectors at all time steps in the deep temporal feature sequence data based on the weights to fold the time dimension and obtain the context vector data representing the weighted information.

[0011] According to an embodiment of this application, the step of performing regression-based prediction processing based on the context vector data to obtain user traffic prediction data specifically includes: inputting the context vector data into a preset regression layer, wherein the regression layer is configured as a mapping structure with a single output node; and performing weighted summation and linear mapping processing on the context vector data based on the regression layer to obtain the user traffic prediction data.

[0012] According to an embodiment of this application, the method further includes: performing a reverse normalization transformation based on a preset numerical range on the user traffic prediction data to obtain actual user traffic prediction data.

[0013] A second aspect of this application provides a user traffic prediction device, the device comprising: a first acquisition module configured to acquire first user traffic time series data in three-dimensional tensor form; a first processing module configured to perform local feature extraction and dimensionality reduction processing based on the first user traffic time series data to obtain dimensionality-reduced feature sequence data; a second processing module configured to perform multi-layer long-term dependency modeling processing based on the dimensionality-reduced feature sequence data to obtain deep time series feature sequence data; a third processing module configured to perform feature weighting processing based on the deep time series feature sequence data to obtain context vector data; and a fourth processing module configured to perform regression-based prediction processing based on the context vector data to obtain user traffic prediction data.

[0014] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0015] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0017] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0018] Figure 1 The illustrations depict application scenarios of user traffic prediction methods, apparatuses, devices, media, and program products according to embodiments of this application.

[0019] Figure 2 A flowchart illustrating a user traffic prediction method according to an embodiment of this application is shown schematically.

[0020] Figure 3 A flowchart illustrating step S100 of the user traffic prediction method according to an embodiment of this application is shown schematically.

[0021] Figure 4 A flowchart illustrating step S110 of the user traffic prediction method according to an embodiment of this application is shown schematically.

[0022] Figure 5 A flowchart illustrating step S120 of the user traffic prediction method according to an embodiment of this application is shown schematically.

[0023] Figure 6 A flowchart illustrating step S200 of the user traffic prediction method according to an embodiment of this application is shown schematically.

[0024] Figure 7 A flowchart illustrating step S300 of the user traffic prediction method according to an embodiment of this application is shown schematically.

[0025] Figure 8 A flowchart illustrating step S400 of the user traffic prediction method according to an embodiment of this application is shown schematically.

[0026] Figure 9 A flowchart illustrating step S500 of the user traffic prediction method according to an embodiment of this application is shown schematically.

[0027] Figure 10 This schematic diagram illustrates a structural block diagram of a user traffic prediction device according to an embodiment of the present application;

[0028] as well as

[0029] Figure 11 A block diagram of an electronic device suitable for implementing a user traffic prediction method according to an embodiment of this application is shown schematically. Detailed Implementation

[0030] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0031] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0032] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0033] When using expressions such as "at least one of A, B, and C," the meaning should generally be interpreted according to the understanding of someone skilled in the art. For example, "a system having at least one of A, B, and C" should include, but is not limited to, systems having A alone, having B alone, having C alone, having A and B, having A and C, having B and C, and / or having A, B, and C. Similarly, when using expressions such as "at least one of A, B, or C," the meaning should generally be interpreted according to the understanding of someone skilled in the art. For example, "a system having at least one of A, B, or C" should include, but is not limited to, systems having A alone, having B alone, having C alone, having A and B, having A and C, having B and C, and / or having A, B, and C.

[0034] The user traffic prediction method, apparatus, device, medium, and program product of this application can be used in the field of artificial intelligence technology, and can also be used in any field other than artificial intelligence. The application fields of the user traffic prediction method, apparatus, device, medium, and program product of this application are not limited.

[0035] As the physical carriers of banking operations, bank branches play a crucial role in meeting customers' financial needs, providing diversified financial services, and serving the real economy and people's livelihoods. The service quality and brand image of bank branches, especially customer experience and satisfaction, have become important factors for banks to maintain their core competitiveness in a fiercely competitive market. However, in actual operation, bank branches generally face resource misallocation problems caused by huge fluctuations in customer traffic. On the one hand, peak traffic periods lead to long waiting times for customers; on the other hand, low traffic periods may result in redundant service personnel. This mismatch between service supply and customer demand not only leads to business losses, but more seriously, the long waiting times severely damage customer experience and even pose a risk of customer churn. Therefore, in order to predict relevant short-term traffic data in advance, an accurate traffic prediction model is urgently needed to provide accurate decision-making basis for branches to "plan the appropriate arrangement of service personnel," thereby optimizing resource allocation, minimizing business losses caused by excessive traffic, and ultimately ensuring the customer experience and satisfaction of every customer.

[0036] Currently, existing technical solutions for addressing this problem have significant limitations. One mainstream approach to pedestrian flow prediction relies heavily on image recognition technology. A typical process involves acquiring surveillance video images from bank branches within a preset time period, then processing these images frame by frame, grouping consecutive images for enhanced imaging, and finally analyzing the images to determine the pedestrian flow for the current time period. This approach has proven to have serious flaws: First, it is overly complex and redundant; the entire process of image recognition followed by image processing is cumbersome and redundant, resulting in significant technical costs and processing overhead. Second, the prediction method is too simplistic; it is essentially a form of real-time or short-term statistical analysis based on images rather than true prediction, making it difficult to construct an effective prediction model and thus unable to effectively address the problem of long customer wait times. In the field of machine learning prediction, although there are some time series prediction techniques, they also face their own bottlenecks. For example, although convolutional neural network models can extract data feature values ​​more accurately and filter out interference factors in the data to the greatest extent, they are not good at handling long-term sequence dependencies in time series. While long short-term memory network models have a certain ability to mine information from long-term and long-distance time series data, they may also experience instability and gradient vanishing when faced with excessively long sequences, leading to a decrease in prediction accuracy.

[0037] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, application, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0038] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0039] Figure 1 The illustration shows an application scenario diagram of the user traffic prediction method, apparatus, device, medium, and program product according to embodiments of this application.

[0040] like Figure 1 As shown, application scenario 100 according to this embodiment may include a user traffic prediction method, apparatus, device, medium, and program product. Network 104 serves as a medium for providing a communication link between a first terminal device 101, a second terminal device 102, a third terminal device 103, and a server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0041] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as financial service applications, web browser applications, search applications, email clients, social media platform software, etc. (for example only).

[0042] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0043] Server 105 can be a server that provides various services, such as a backend management server that supports information browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0044] It should be noted that, in the first aspect, the user traffic prediction method provided in this application embodiment can be executed by server 105; in the second aspect, the user traffic prediction method provided in this application embodiment can be executed by first terminal device 101, second terminal device 102, or third terminal device 103. Correspondingly, in the third aspect, the user traffic prediction device provided in this application embodiment can be located in server 105; in the fourth aspect, the user traffic prediction device provided in this application embodiment can be located in first terminal device 101, second terminal device 102, or third terminal device 103. The user traffic prediction method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with first terminal device 101, second terminal device 102, third terminal device 103, and / or server 105. Correspondingly, the user traffic prediction device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with first terminal device 101, second terminal device 102, third terminal device 103, and / or server 105.

[0045] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0046] The following will be based on Figure 1 The described scene, through Figures 2-9 The user traffic prediction method of the application embodiment is described in detail.

[0047] Figure 2 A flowchart illustrating a user traffic prediction method according to an embodiment of this application is shown.

[0048] like Figure 2 As shown, this application provides a user traffic prediction method, including operations S100 to S500.

[0049] Specifically, in operation S100, the first user traffic time series data in three-dimensional tensor form is acquired. In some specific embodiments, the first user traffic time series data in three-dimensional tensor form is a structured dataset constructed for subsequent deep learning model processing. This three-dimensional tensor form has three dimensions, which respectively represent the number of samples, the time step contained in each sample, and the number of features observed at each time step. The reason for acquiring this three-dimensional tensor form of data is that the original user traffic data (e.g., the hourly traffic sequence of bank branches) usually exists in the form of a one-dimensional time series. This one-dimensional data lacks a clear correspondence between input and output, making it unsuitable for direct use in supervised learning. By converting the one-dimensional time series data into a three-dimensional tensor, the original prediction problem can be reconstructed into a supervised learning problem, which can be trained to learn the mapping from sequence to value, that is, to predict the value of a future point in time immediately adjacent to the input sequence using an input sequence containing multiple historical time steps. This three-dimensional tensor form provides a standard data input format for subsequent local feature extraction and long-term time-series dependency modeling.

[0050] like Figure 3 As shown, in some exemplary embodiments, the acquisition of the first user traffic time series data in three-dimensional tensor form includes: operations S110 to S120.

[0051] Specifically, in operation S110, the second user traffic time series data after scaling range transformation is obtained. In some specific embodiments, the second user traffic time series data after scaling range transformation refers to one-dimensional time series data that has undergone data preprocessing and whose values ​​have been uniformly mapped to a preset, fixed numerical range. The numerical range of the original user traffic data may fluctuate greatly, with significant differences in scale. If such unprocessed data is directly input into a deep learning model for training, data of different scales will dominate the gradient descent process, leading to unstable model weight updates and difficulty in converging or extremely slow convergence during training. Therefore, the processing logic of operation S110 is to perform a normalization process, i.e., scaling range transformation, on the original time series data before performing the sliding window transformation. Operation S110 scales all data points proportionally to a uniform numerical range according to their relative position in the entire dataset, eliminating the influence of scale and ensuring that all input data contribute to model training on the same scale, thereby accelerating convergence and improving the stability of subsequent model training and the final prediction accuracy.

[0052] like Figure 4 As shown, in some exemplary embodiments, the step of obtaining the second user traffic time series data after proportional range conversion includes: operations S111 to S113.

[0053] Specifically, in operation S111, raw user traffic time series data is acquired. In some specific embodiments, the raw user traffic time series data refers to the unprocessed initial dataset directly obtained from the business system or data warehouse. Raw user traffic time series data is typically organized chronologically, recording user traffic statistics observed over consecutive time units (e.g., hourly granularity) (e.g., customer traffic at bank branches). This raw user traffic time series data is the starting point for all subsequent data processing steps; however, it may contain various noises, such as "missing or outliers" due to statistical interruptions or equipment malfunctions, and therefore cannot be directly used for model training. It must first be cleaned and transformed.

[0054] Specifically, in operation S112, missing value imputation processing is performed based on the original user traffic time series data to obtain complete user traffic time series data. In some specific embodiments, performing missing value imputation processing is a step to ensure the continuity and integrity of the time series. During execution, the acquired original user traffic time series data is first scanned for integrity to detect whether there are any data points that are empty or marked as invalid time steps. When a missing data point is detected, the imputation logic is initiated; one specific imputation implementation method is to use the "values ​​before and after" of the missing point for imputation. For example, a linear imputation method can be used, that is, based on the most recent valid data point before and the most recent valid data point after the missing point, a straight line is established between these two values, and the corresponding value of the missing time point on the straight line is calculated as the imputation value; in other embodiments, more complex imputation can also be used, such as using the previous valid value for forward filling or using the next valid value for backward filling, or imputation based on the average value of a longer time window. Operation S112 obtains complete user traffic time series data for data points, which avoids the problem of the model being unable to correctly learn time dependencies due to data fragmentation for subsequent modeling that relies on fixed time steps and continuous sequences.

[0055] Specifically, in operation S113, based on the complete user traffic time series data, a normalization process based on a preset numerical range is performed to obtain the second user traffic time series data. In some specific embodiments, performing the normalization process based on the preset numerical range is a data scaling operation performed on the interpolated complete user traffic time series data: the normalization process first traverses the complete user traffic time series data to determine the global maximum and global minimum values ​​in the entire dataset; then it traverses each data point in the data again and calculates the difference between the data point and the global minimum value; next, it calculates the difference between the global maximum and the global minimum value (i.e., the total range of the data); finally, it divides the first difference (the difference between the data point and the minimum value) by the second difference (the total range) to obtain the new normalized value. Through the normalization process, all original data values, regardless of their absolute size, are proportionally mapped to the "preset numerical range".

[0056] Operations S111 to S113 constitute a complete data preprocessing stage. Imputation ensures data integrity, and normalization eliminates the influence of data dimensions, jointly outputting regular and uniformly scaled second user traffic time series data. This provides a stable data foundation for subsequent feature engineering and model training, improving the model's prediction accuracy.

[0057] Specifically, in operation S120, based on the second user traffic time series data, a sliding window-based time series transformation process is performed to obtain the first user traffic time series data in three-dimensional tensor form. In some specific embodiments, the sliding window-based time series transformation process is the core step in realizing the transformation from one-dimensional time series to three-dimensional supervised learning samples; wherein, the second user traffic output by operation S110 is still a continuous one-dimensional sequence, lacking the pairing structure of input features and output labels required for model training. The processing logic of operation S120 is to define a sliding window with a "preset time step" (i.e., window size, such as 64 time steps), and make the sliding window slide along the time axis of the second user traffic time series data from beginning to end. At each position of the slide, the continuous sequence data contained in the sliding window is extracted as input features, and the data of the next time point immediately adjacent to the sliding window is extracted as labels. In this way, the one-dimensional sequence is transformed into a large number of supervised input-output sample pairs.

[0058] Operations S110 and S120 ensured data scale uniformity through scaling range transformation, constructed the sample form required for supervised learning through sliding window transformation, and obtained the first user traffic time series data in three-dimensional tensor form, providing morphologically correct input data for the input layer of subsequent models.

[0059] like Figure 5 As shown, in some exemplary embodiments, the step of performing a sliding window-based time series transformation process based on the second user traffic time series data to obtain the first user traffic time series data in three-dimensional tensor form includes: operations S121 to S122.

[0060] Specifically, in operation S121, a sliding window sampling process is performed along the time axis of the second user traffic time series data with a preset time step as the window size, obtaining multiple input-output sample pairs. The input of each input-output sample pair is a sequence of data for consecutive time steps, and the output of each input-output sample pair is data for a single time step immediately following the consecutive time steps. In some specific embodiments, the preset time step is first determined as the window size; for example, the time step can be set to 64, meaning the model will use data from the past 64 consecutive time units as the basis for prediction. At the start of the sampling process, the sliding window covers the first 64 data points of the second user traffic time series data, using this entire 64-point sequence as the input of the first input-output sample pair; simultaneously, the 65th data point immediately adjacent to this sequence is used as the output of the sample pair. Subsequently, the sliding window slides forward one step along the time axis, covering the sequence data from the 2nd to the 65th data point, using this as the input of the second sample pair, and the 66th data point as the output. Repeat this sliding and sampling process until the sliding window reaches the end of the sequence, meaning the last available input sequence has been extracted. Through operation S121, a set containing multiple input-output sample pairs is finally obtained.

[0061] Specifically, in operation S122, based on the input sequence data within the multiple input-output sample pairs, a combination process is performed to generate the first user traffic time series data in a three-dimensional tensor form. The three dimensions of the three-dimensional tensor form represent the number of samples, the time step, and the number of features, respectively. In some specific embodiments, the combination process involves regularizing and stacking the multiple input-output sample pairs obtained in S121 to form a standard three-dimensional tensor structure that meets the model input requirements. Specifically, the combination process extracts the input sequence data from all sample pairs and stacks these sequences along the new dimension of the number of samples. The three dimensions of the finally generated first user traffic time series data in a three-dimensional tensor form have the following specific meanings: "representing the number of samples" refers to the total number of usable input-output sample pairs generated in operation S121; "representing the time step" refers to the fixed length of each input sample sequence; and "representing the number of features" refers to the number of indicators observed at each time step. In this embodiment, only the user traffic indicator is observed. Therefore, operations S121 and S122 together transform one-dimensional time series data into a three-dimensional tensor of the form (N, 64, 1), thus transforming the time series prediction problem into a supervised learning problem that can be processed by deep learning models.

[0062] Specifically, in operation S200, based on the first user traffic time series data, local feature extraction and dimensionality reduction processing are performed to obtain dimensionality-reduced feature sequence data. In some specific embodiments, although the first user traffic time series data has been constructed into a three-dimensional tensor form, it is still raw, unabstracted sequence information in the time dimension. If this raw sequence is directly input into a long-term time series modeling network, it may be difficult to effectively capture key patterns in a short period of time, and the long time step will also bring a huge computational burden. Therefore, the purpose of performing local feature extraction and dimensionality reduction processing is to first automatically learn and extract local time patterns in the data through a dedicated feature extraction network, such as short-term features like "rise / fall" or "stable / surge" in traffic over several hours. Local feature extraction processing can maximize the filtering of interfering factors in the data and extract valuable key features. Dimensionality reduction processing is used to compress the dimensions of the data, especially the time dimension, in order to retain usable feature information, reducing the computational complexity of the subsequent time series modeling network while retaining the most significant local features. The processing logic of the S200 operation is to sequentially feed the three-dimensional tensor data into a series structure that includes convolution processing, nonlinear transformation processing, and dimensionality reduction processing, and finally output a dimensionality-reduced feature sequence data that is richer in the feature dimension but more concise in the time dimension, laying the data foundation for subsequent long-term time-series dependency modeling.

[0063] like Figure 6As shown, in some exemplary embodiments, the step of performing local feature extraction and dimensionality reduction processing based on the first user traffic time series data to obtain dimensionality-reduced feature sequence data includes: operations S210 to S230.

[0064] Specifically, in operation S210, one-dimensional convolution processing is performed based on the first user traffic time series data to obtain local feature data representing local time patterns. In some specific embodiments, performing one-dimensional convolution processing is a method specifically for feature extraction from time series data, using a preset number of filters to perform operation S210. Each filter contains a one-dimensional convolution kernel with a first preset size, enabling it to cover multiple consecutive time steps. During one-dimensional convolution processing, the one-dimensional convolution kernel slides along the time step dimension of the first user traffic time series data, performing convolution operations at each sliding position to capture specific time patterns within the local window. Since multiple different filters are configured, each filter focuses on recognizing a different local time pattern after training. To ensure that the time dimension of the output sequence is consistent with the input sequence (e.g., if the input has 64 time steps, the output also has 64 time steps), one-dimensional convolution processing can also employ padding at both ends of the input sequence. In one specific embodiment, the local feature data output by the one-dimensional convolution process is transformed from (N, 64, 1) to (N, 64, 64), where the 64 feature dimensions represent the 64 different local temporal patterns extracted.

[0065] Specifically, in operation S220, a nonlinear transformation is performed on the local feature data to obtain nonlinear feature data. In some specific embodiments, the nonlinear transformation is performed by the activation function layer and is a necessary supplement to the convolution process in operation S210. Since the convolution operation in operation S210 is essentially a linear weighted sum, without a nonlinear transformation, even stacking multiple convolution layers can only achieve a linear mapping and cannot learn the complex nonlinear patterns in the data. Therefore, the nonlinear transformation performs a nonlinear mapping once for each data point in the local feature data output by operation S210. In one specific embodiment, the value of the data point is first determined; if the value is greater than or equal to a preset threshold, the value is kept unchanged; if the value is less than the preset threshold, the value is reset. This processing method can effectively alleviate the gradient vanishing problem, making the model more likely to respond quickly to the data. After the nonlinear transformation, the shape of the obtained nonlinear feature data is consistent with the local feature data, but the data distribution of the nonlinear feature data has been nonlinearized.

[0066] Specifically, in operation S230, based on the nonlinear feature data, a dimensionality reduction process based on max pooling is performed to obtain the dimensionality-reduced feature sequence data that retains the temporal dimension. In some specific embodiments, performing the dimensionality reduction process based on max pooling is a process of compressing and downsampling the "nonlinear feature data" output by operation S220; its purpose is to reduce the feature dimension of the data while retaining usable feature information. The dimensionality reduction process is implemented through a pooling window with a second preset size. The pooling window slides along the time dimension of the data at a preset step size. At each sliding position, the dimensionality reduction process based on max pooling examines all feature values ​​within multiple time steps covered by the pooling window and retains only the maximum value as the output of the pooling window, discarding the rest of the data. In this way, the length of the data in the time dimension is effectively shortened, while the feature dimension remains unchanged. The final obtained dimensionality-reduced feature sequence data is a more compact sequence that retains the temporal dimension so that it can be received by subsequent long-term temporal dependency modeling processing.

[0067] Specifically, in operation S300, based on the dimensionality-reduced feature sequence data, multi-layer long-term time-series dependency modeling processing is performed to obtain deep time-series feature sequence data. In some specific embodiments, although the dimensionality-reduced feature sequence data already contains highly condensed local patterns, the temporal dependencies between these patterns, especially information from long-term, long-distance time-series data, have not yet been fully explored. For example, operation S200 may identify local features of the morning peak and the midday trough, but it cannot understand how the occurrence of the morning peak affects the midday trough, i.e., it cannot capture the long-distance dependencies between features. To address this, the S300 operation employs a multi-layered structure, connecting two temporal modeling layers. The processing logic of the S300 is as follows: the dimensionality-reduced feature sequence data is input into the first-layer temporal modeling network, which traverses the sequence, learns, and outputs a sequence representing the surface temporal features. This new sequence is then used as input to the second-layer temporal modeling network, allowing it to further learn and abstract deeper temporal features based on the surface features. This stacked, multi-layered processing approach enables the model to construct more complex and hierarchical representations of temporal dependencies, thereby comprehensively characterizing temporal data, solving the problem that a single model struggles to grasp long-distance dependencies between features, and obtaining deep temporal feature sequence data.

[0068] like Figure 7 As shown, in some exemplary embodiments, the step of performing multi-layer long temporal dependency modeling processing based on the dimensionality-reduced feature sequence data to obtain deep temporal feature sequence data includes: operations S310 to S320.

[0069] Specifically, in operation S310, based on the reduced-dimensionality feature sequence data, a first long-term dependency modeling process is performed to obtain a first hidden state sequence data representing the surface-level temporal dependencies. In some specific embodiments, the first long-term dependency modeling process is implemented through a first temporal modeling unit, such as a Long Short-Term Memory (LSTM) network unit with 128 internal units. The first long-term dependency modeling process receives the reduced-dimensionality feature sequence data, and the first temporal modeling unit processes the feature vector of each time step sequentially along the temporal dimension of the sequence. When processing each time step, the first temporal modeling unit uses its internal gating mechanism to update its internal long-term memory (unit state) and short-term memory (hidden state) by combining the input features of the current time step and the hidden state of the previous time step. In order to pass information to the next layer, the first long-term dependency modeling process is specially configured to return a sequence, that is, the first long-term dependency modeling process outputs the hidden states of all time steps, not just the last one. The first hidden state sequence data output by the first long-term dependency modeling process encodes the surface-level temporal dependencies in the input feature sequence.

[0070] Specifically, in operation S320, based on the first hidden state sequence data, a second long temporal dependency modeling process is performed to obtain the deep temporal feature sequence data representing deep temporal dependencies. In some specific embodiments, the second long temporal dependency modeling process is implemented through a second temporal modeling unit (e.g., also having 128 internal units). This second long temporal dependency modeling process receives the first hidden state sequence data as its input; that is, the second temporal modeling unit no longer processes the original local features, but processes the sequence of features that have already undergone the first layer of temporal abstraction. This stacking method enables the second temporal modeling unit to learn deeper temporal dependencies that are more complex and abstract than those of the first temporal modeling unit. To cooperate with the subsequent feature weighting processing (i.e., attention mechanism), the second long temporal dependency modeling process is also configured to return a sequence. If the second long temporal dependency modeling process does not return a complete sequence, the subsequent feature weighting processing will not be able to be executed due to the lack of a weightable sequence. Therefore, the deep temporal feature sequence data finally output by the second long temporal dependency modeling process is a deep feature sequence with the same time dimension as the input sequence.

[0071] Specifically, in operation S400, feature weighting processing is performed based on the deep temporal feature sequence data to obtain context vector data. In some specific embodiments, feature weighting processing is performed to address the information bottleneck problem in long sequence prediction. Deep temporal feature sequence data is a sequence containing information from multiple time steps. When predicting the future, not all information from historical time steps is equally important. For example, when predicting pedestrian traffic in the next hour, the weight of data from the most recent hours may be much higher than that from the more distant hours; if this information is not differentiated, the prediction effect will be limited. Therefore, a soft attention mechanism is introduced as a feature weighting processing method. The processing logic of operation S400 is that instead of treating each time step in the sequence equally, a weight is calculated for each time step in the sequence output by operation S300 by probability allocation; based on these weights, the sequence is weighted and summed, folding the entire sequence into a single context vector data. This context vector data condenses the information from all time steps and intelligently assigns different weights to different information, thereby enabling focus on the historical information most critical to the prediction.

[0072] like Figure 8 As shown, in some exemplary embodiments, the step of performing feature weighting processing based on the deep temporal feature sequence data to obtain context vector data includes: operations S410 to S420.

[0073] Specifically, in operation S410, probabilistic weight allocation processing is performed on the feature vectors at different time steps in the deep temporal feature sequence data to obtain the weight corresponding to each time step. In some specific embodiments, the probabilistic weight allocation processing is the core of feature weighting. The probabilistic weight allocation processing first calculates the attention score of the feature vectors at different time steps in the deep temporal feature sequence data; there are various ways to calculate this attention score, for example, each feature vector can be mapped to a scalar score through a trainable weight matrix, or the similarity between each vector in the sequence and a query vector can be calculated. After obtaining the original scores of all time steps, in order to give these original scores a probabilistic meaning, the probabilistic weight allocation processing transforms these original scores based on probability normalization; the probability normalization processing amplifies the items with high scores, suppresses the items with low scores, and ensures that all weight values ​​after transformation are between 0 and 1, and their sum is 1. Finally, the probabilistic weight allocation processing generates a corresponding weight for each time step in the sequence, which quantitatively represents the contribution of that time step to the final prediction.

[0074] Specifically, in operation S420, based on the weights, a weighted summation process is performed on the feature vectors of all time steps in the deep temporal feature sequence data to fold the time dimension and obtain the context vector data representing the weighted information. In some specific embodiments, the feature vectors of all time steps in the deep temporal feature sequence data are traversed; for the i-th time step, the weighted summation process obtains the feature vector corresponding to the i-th time step and the weight corresponding to the i-th time step; then, a scalar multiplication is performed on all time steps to obtain 32 weighted feature vectors; these 32 weighted feature vectors are vector-summed to obtain the vector, which is the context vector data. The context vector data no longer has a time dimension. The context vector data condenses all the key information marked with high weights in the entire sequence and is used to represent the weighted comprehensive information.

[0075] Specifically, in operation S500, based on the context vector data, regression-based prediction processing is performed to obtain user traffic prediction data. In some specific embodiments, the context vector data represents the final understanding of the historical sequence, but it is not a predicted value; to predict a single, continuous value (i.e., the traffic flow at the next time point), regression-based prediction processing is performed. The processing logic of operation S500 is to use a regression layer to take the high-dimensional context vector data as input and transform and map the context vector data to a single output dimension. The regression-based prediction processing converges the complex features extracted by all previous processing (local feature extraction, multi-layer temporal modeling, feature weighting) to generate quantified user traffic prediction data.

[0076] like Figure 9 As shown, in some exemplary embodiments, the step of performing regression-based prediction processing based on the context vector data to obtain user traffic prediction data includes: operations S510 to S520.

[0077] Specifically, in operation S510, the context vector data is input to a preset regression layer, which is configured with a mapping structure having a single output node. In some specific embodiments, the preset regression layer is a specific network layer used to perform the final prediction. The core configuration of the regression layer lies in the output dimension of its mapping structure, which is configured to have a single output node. The "single output node" configuration is based on the prediction target being the predicted value of pedestrian traffic at the next specified time period. During processing, each element in the context vector data is connected to a trainable weight parameter within the regression layer. The single output node mapping structure ensures that all computational power, from the initial local feature extraction to the final feature weighting, works together to generate a unique output value.

[0078] Specifically, in operation S520, based on the regression layer, a weighted summation and linear mapping process is performed on the context vector data to obtain the user traffic prediction data. In some specific embodiments, the regression layer performs two key operations after receiving the context vector data. First, the regression layer performs a weighted summation process: multiplying each input signal by its corresponding internal weight (learned during model training), and then summing all these products. Second, the regression layer performs a linear mapping process on the result of the weighted summation to finally obtain the user traffic prediction data.

[0079] In some exemplary embodiments, the user traffic prediction method further includes operation S600, specifically, in operation S600, based on the user traffic prediction data, performing a reverse normalization transformation based on a preset numerical range to obtain actual user traffic prediction data. In some specific embodiments, the user traffic prediction data is a standardized value within a preset numerical range; the user traffic prediction data is necessary for model training, but lacks direct guiding significance for business decisions. Therefore, performing a reverse normalization transformation restores the predicted value to its original physical dimensions (e.g., pedestrian flow). Through this reverse normalization transformation, the standardized user traffic prediction data is converted back into actual user traffic prediction data with practical business significance, realizing the final application of the model prediction results.

[0080] In this embodiment, the operations S100~S500, S110, S120, S111, S112, S113, S121, S122, S210, S220, S230, S310, S320, S410, S420, S510, S520, and S600 described above can also be executed on the system side, and will not be repeated here.

[0081] Based on the above-described user traffic prediction method, this application also provides a user traffic prediction device, which will be described below in conjunction with... Figure 10 The device is described in detail.

[0082] Figure 10 A schematic block diagram of a user traffic prediction device according to an embodiment of this application is shown.

[0083] like Figure 10 As shown, the user traffic prediction device 200 in this embodiment includes a first acquisition module 210, a first processing module 220, a second processing module 230, a third processing module 240, and a fourth processing module 250.

[0084] The first acquisition module 210 is configured to acquire first user traffic time series data in three-dimensional tensor form. In one embodiment, the first acquisition module 210 can be used to execute the operation S100 described above, as well as operations S110, S120, S111, S112, S113, S121, and S122, which will not be described again here.

[0085] The first processing module 220 is configured to perform local feature extraction and dimensionality reduction processing based on the first user traffic time series data to obtain dimensionality-reduced feature sequence data. In some embodiments, the first processing module 220 can be used to execute the operation S200, as well as operations S210, S220, and S230 described above, which will not be repeated here.

[0086] The second processing module 230 is configured to perform multi-layer long temporal dependency modeling based on the dimensionality-reduced feature sequence data to obtain deep temporal feature sequence data. In one embodiment, the second processing module 230 can be used to execute the operation S300, as well as operations S310 and S320 described above, which will not be repeated here.

[0087] The third processing module 240 is configured to perform feature weighting processing based on the deep temporal feature sequence data to obtain context vector data. In one embodiment, the third processing module 240 can be used to execute the operation S400, as well as operations S410 and S420 described above, which will not be repeated here.

[0088] The fourth processing module 250 is configured to perform regression-based prediction processing based on the context vector data to obtain user traffic prediction data. In one embodiment, the fourth processing module 250 can be used to execute the operation S500, as well as operations S510 and S520 described above, which will not be repeated here.

[0089] According to embodiments of this application, any multiple modules among the first acquisition module 210, first processing module 220, second processing module 230, third processing module 240, and fourth processing module 250 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the first acquisition module 210, first processing module 220, second processing module 230, third processing module 240, and fourth processing module 250 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the first acquisition module 210, the first processing module 220, the second processing module 230, the third processing module 240, and the fourth processing module 250 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0090] Figure 11 A block diagram of an electronic device suitable for implementing a user traffic prediction method according to an embodiment of this application is shown schematically.

[0091] like Figure 10 As shown, an electronic device 900 according to an embodiment of this application includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage portion 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include onboard memory for caching purposes. The processor 901 may include a single processing module or multiple processing modules for performing different actions of the method flow according to an embodiment of this application.

[0092] RAM 903 stores various programs and data required for the operation of electronic device 900. Processor 901, ROM 902, and RAM 903 are interconnected via bus 904. Processor 901 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 902 and / or RAM 903. It should be noted that programs may also be stored in one or more memories other than ROM 902 and RAM 903. Processor 901 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in one or more memories.

[0093] According to embodiments of this application, the electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to a bus 904. The electronic device 900 may also include one or more of the following components connected to the input / output (I / O) interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 910 as needed so that computer programs read from it can be installed into the storage section 908 as needed.

[0094] This application also provides a computer-readable storage medium, which may be included in the apparatus described in the above embodiments; or it may exist independently and not assembled into the apparatus. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0095] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 902 and / or RAM 903 and / or one or more memories other than ROM 902 and RAM 903 described above.

[0096] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to enable the computer system to implement the user traffic prediction method provided in the embodiments of this application.

[0097] When the computer program is executed by the processor 901, it performs the functions defined in the apparatus of the embodiments of this application. According to the embodiments of this application, the apparatus and the like described above can be implemented by computer program modules.

[0098] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 909, and / or installed from a removable medium 911. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0099] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 909, and / or installed from the removable medium 911. When the computer program is executed by the processor 901, it performs the functions defined in the system of the embodiments of this application. According to embodiments of this application, the apparatuses described above can be implemented by computer program modules.

[0100] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0101] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0102] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0103] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A user traffic prediction method characterized by, The method comprises: obtaining a first user traffic time series data in a three-dimensional tensor form; based on the first user traffic time series data, performing local feature extraction and dimension reduction processing to obtain dimension reduction feature sequence data; based on the dimension reduction feature sequence data, performing multi-layer long time series dependence modeling processing to obtain deep time series feature sequence data; based on the deep time series feature sequence data, performing feature weighting processing to obtain context vector data; based on the context vector data, performing regression-based prediction processing to obtain user traffic prediction data.

2. The method of claim 1, wherein, The method comprises: obtaining second user traffic time series data converted based on a scale range; based on the second user traffic time series data, performing sliding window-based time series conversion processing to obtain the first user traffic time series data in a three-dimensional tensor form.

3. The method of claim 2, wherein, The method comprises: obtaining original user traffic time series data; based on the original user traffic time series data, performing missing value imputation processing to obtain complete user traffic time series data; based on the complete user traffic time series data, performing normalization processing based on a preset numerical interval to obtain the second user traffic time series data.

4. The method of claim 2, wherein, The method comprises: taking a preset time step as a window size, performing sliding window sampling processing along a time axis of the second user traffic time series data to obtain a plurality of input-output sample pairs, wherein the input of each input-output sample pair is a sequence data of consecutive time steps, and the output of each input-output sample pair is a single time step data immediately after the consecutive time steps; based on the input sequence data in the plurality of input-output sample pairs, performing combination processing to generate the first user traffic time series data in a three-dimensional tensor form, wherein the three dimensions of the three-dimensional tensor form represent sample number, time step and feature number respectively.

5. The method of claim 1, wherein, The method comprises: based on the first user traffic time series data, performing one-dimensional convolution processing to obtain local feature data representing local time patterns; performing nonlinear transformation processing on the local feature data to obtain nonlinear feature data; based on the nonlinear feature data, performing dimension reduction processing based on maximum pooling to obtain the dimension reduction feature sequence data retaining time series dimensions.

6. The method of claim 1, wherein, The method comprises: based on the dimension reduction feature sequence data, performing first long time series dependence modeling processing to obtain first hidden state sequence data representing surface time series dependence relationships; based on the first hidden state sequence data, performing second long time series dependence modeling processing to obtain second hidden state sequence data representing deep time series dependence relationships; According to the first hidden state sequence data, a second long time sequence dependence modeling process is performed to obtain the deep time sequence feature sequence data representing deep time sequence dependence.

7. The method of claim 1, wherein, The feature weighting process is performed based on the deep time sequence feature sequence data to obtain context vector data, specifically including: The probability weight distribution process is performed on the feature vectors at different time steps in the deep time sequence feature sequence data to obtain the weight corresponding to each time step; Based on the weight, the weighted sum process is performed on the feature vectors at all time steps in the deep time sequence feature sequence data to fold the time dimension and obtain the context vector data representing the weighted information.

8. The method of claim 3, wherein, The regression-based prediction process is performed based on the context vector data to obtain user traffic prediction data, specifically including: The context vector data is input into a preset regression layer, wherein the regression layer is configured to have a mapping structure with a single output node; Based on the regression layer, the weighted sum and linear mapping process is performed on the context vector data to obtain the user traffic prediction data.

9. The method of claim 8, wherein, The method further includes: Based on the user traffic prediction data, the inverse normalization transformation process based on a preset numerical interval is performed to obtain actual user traffic prediction data.

10. A user traffic prediction device characterized by comprising: The device includes: The first acquisition module is configured to perform acquisition of first user traffic time sequence data in three-dimensional tensor form; The first processing module is configured to perform local feature extraction and dimension reduction processing based on the first user traffic time sequence data to obtain reduced dimension feature sequence data; The second processing module is configured to perform multi-layer long time sequence dependence modeling processing based on the reduced dimension feature sequence data to obtain deep time sequence feature sequence data; The third processing module is configured to perform feature weighting processing based on the deep time sequence feature sequence data to obtain context vector data; The fourth processing module is configured to perform regression-based prediction processing based on the context vector data to obtain user traffic prediction data.

11. An electronic device, comprising: one or more processors; a memory for storing one or more computer programs, characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1-9.

12. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer program or instructions are executed by the processor to implement the steps of the method according to any one of claims 1-9.

13. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions are executed by the processor to implement the steps of the method according to any one of claims 1-9. The computer program or instructions are executed by the processor to implement the steps of the method according to any one of claims 1-9.

Images