A method and device for improving user experience of a network service platform

Abstract

This invention discloses a method and apparatus for improving the user experience of a network service platform, comprising: normalizing and denormalizing processed and predicted data using a reversible instance normalization module; inputting the normalized and denormalized data into a multi-scale time refinement module, extracting multi-scale features and suppressing noise using a pyramid decomposition-fusion strategy and a time attention mechanism; inputting the extracted and noise-suppressed multi-scale features into a patch-level channel enhancement module, dividing the complete sequence into patch sequences, and capturing local contextual relationships and inter-variable interactions within the patch sequences using a history-aware reconstruction strategy and a channel attention mechanism; inputting the patch-enhanced patch sequences into an inverted Transformer encoder to extract global features, dynamically calibrating and filtering the global features obtained by the encoder through a multi-view gating mechanism; and generating the final traffic volume prediction result through a linear prediction layer using the calibrated and filtered results.

Description

Technical Field

[0001] This invention relates to the fields of Internet data processing and artificial intelligence technology, and in particular to a method and apparatus for improving the user experience of a network service platform. Background Technology

[0002] With the rapid development of the mobile internet and the digital economy, the user base of various online business platforms (such as e-commerce platforms, online customer service systems, and IT operation and maintenance platforms) is constantly expanding, resulting in a massive number of business orders. Business volume is a key indicator for measuring platform operational pressure and user activity, exhibiting significant volatility, intermittency, and multivariate correlation. For example, business volume often surges during major promotional events, system failures, or holidays. If the platform cannot accurately predict future business volume, it will lead to unreasonable customer service staff scheduling or uneven allocation of server resources. When the predicted value is too low, it will result in insufficient seats and excessively long user waiting times, seriously affecting user experience and satisfaction. When the predicted value is too high, it will lead to idle staff and wasted computing resources. Therefore, building a high-precision business volume prediction method and device is of great significance for achieving refined operations, optimizing resource scheduling, and reducing operating costs.

[0003] Business volume data is essentially multivariate time series data. Early prediction methods mainly relied on statistical models, such as the Autoregressive Integrated Moving Average (ARIMA) model. These models assume that the data is linear and stationary, making it difficult to capture the complex nonlinear dynamic changes in business data. Subsequently, machine learning-based methods such as Support Vector Regression (SVR) and Gradient Boosting Tree (GBDT) were widely used. Although they solved the nonlinearity problem to some extent, their performance was limited when dealing with long-term time dependencies.

[0004] In recent years, deep learning technology has made significant progress in the field of time series prediction. While models based on Recurrent Neural Networks (RNNs) and their variant, Long Short-Term Memory (LSTM), can capture temporal dependencies, they face problems such as vanishing gradients and low parallel computation efficiency. With the introduction of the Transformer architecture, models based on self-attention mechanisms have gradually become mainstream due to their powerful global modeling capabilities. In particular, the Inverted Transformer architecture, by embedding the entire time series of the same variable as a token, effectively captures the global correlation between variables and performs excellently in multivariate prediction tasks.

[0005] However, existing Transformer-based prediction models still have the following limitations when applied to predicting the business volume of complex online business platforms: First, existing inverted architecture models often focus too much on the overall trend of the sequence, while ignoring fine-grained changes within local time windows and instantaneous interactions between variables. For example, a certain type of business surge may be triggered by a specific event within a short period of time, and directly mapping a long sequence to a single token can easily lead to the loss of these key local features. Second, the original business data usually contains multi-scale noise (e.g., high-frequency random fluctuations and low-frequency trend drift). Most existing models lack multi-scale refinement and denoising mechanisms for the original signal, and directly mapping noisy data to a high-dimensional space can easily lead to model overfitting and reduce the robustness of the prediction. Third, the high-dimensional features output by the encoder often contain redundant information and potential noise, and it is difficult to effectively filter these interferences by directly predicting through linear layers, thus limiting the final prediction accuracy and reducing the user experience. Summary of the Invention

[0006] This invention provides a method and apparatus for improving the user experience of a network service platform. The invention effectively extracts local and global features of multivariate time series data, suppresses noise interference through multi-scale refinement and multi-view gating mechanisms, effectively improves the accuracy and reliability of traffic volume prediction, achieves refined operation, and allows for timely adjustment of scheduling strategies. This avoids wasted manpower or response delays caused by prediction errors, thereby enhancing the user experience. See the description below for details:

[0007] Firstly, a method for improving the user experience of a network service platform, the method comprising:

[0008] The constructed multivariate business volume dataset is input into the MIGformer model for processing and prediction; the MIGformer model includes: a reversible instance normalization module, a multi-scale time refinement module, a patch-level channel enhancement module, an inverted Transformer encoder, and a multi-view gating mechanism.

[0009] The reversible instance normalization module is used to normalize and denormalize the processed and predicted data; the normalized and denormalized data are then input into the multi-scale temporal refinement module, where a pyramid decomposition-fusion strategy and a temporal attention mechanism are used to extract multi-scale features and suppress noise.

[0010] The extracted and noise-suppressed multi-scale features are input into the patch-level channel enhancement module, which divides the complete sequence into patch sequences and uses a history-aware reconstruction strategy and channel attention mechanism to capture the local contextual relationships and the interaction relationships between variables within the patch sequences.

[0011] The patched sequence, enhanced by the patching, is input into the inverted Transformer encoder to extract global features. The global features obtained by the encoder are dynamically calibrated and filtered through a multi-view gating mechanism. The calibrated and filtered results are then passed through a linear prediction layer to generate the final traffic volume prediction result.

[0012] The process of extracting multi-scale features and suppressing noise using a pyramid decomposition-fusion strategy and a temporal attention mechanism is as follows:

[0013] A temporal pattern pyramid is generated using average pooling layers with different kernel sizes; a bottom-up recursive fusion strategy is employed, adjusting the dimensions through linear layers and then summing the residuals; and a temporal attention mechanism is used for refinement.

[0014]

[0015] in, The input features to be refined; The output features are those refined over time. It is a sigmoid activation function; Represents one-dimensional convolution; This represents the average pooling layer.

[0016] The process of segmenting the complete sequence into patch sequences and using a history-aware reconstruction strategy and channel attention mechanism to capture the local contextual relationships and inter-variable interactions within the patch sequences is as follows:

[0017] The current patch and the features obtained by aggregating historical information are overlaid and fused to obtain the initial patch features that include historical context. :

[0018]

[0019] in, This represents the patch characteristics at the current moment; The patch features are those of the merged patch. This represents the patch feature from the previous moment.

[0020] Channel descriptors are extracted by compressing the time dimension through an adaptive average pooling layer, then channel weights are generated through an activation network containing two fully connected layers, and finally the channel weights are multiplied element-wise with the initial patch features.

[0021]

[0022] in, These are patch features after channel attention recalibration; For adaptive average pooling layer; This is a multilayer perceptron network used to generate weights within the attention mechanism; It is a sigmoid activation function.

[0023] Specifically, the enhanced patch sequence is input into the inverted Transformer encoder to extract global features. The encoder then embeds entire sequences of the same variable into a single token, completing the dimensionality inversion transformation.

[0024]

[0025] in, This is the transpose of the enhanced patch sequence output in the previous step. This is the embedded token. Subsequently, a self-attention mechanism is used to calculate the attention weights between different variables, extracting global correlations across variables and outputting high-dimensional latent features. :

[0026]

[0027] in, Includes self-attention computation and feedforward neural networks; This is a high-dimensional latent feature for extracting cross-variable global correlations.

[0028] The inverted encoder outputs high-dimensional latent features of cross-variable global correlation. Then, a multi-view gating mechanism is used to extract and concatenate features from four perspectives:

[0029] (1) Global semantic perspective Linear: ;

[0030] (2) Conv from the perspective of local structure: ;

[0031] (3) Statistical perspective Max: ;

[0032] (4) Statistical perspective (Avg): ;

[0033] Use the four After concatenation, a gated mask is generated through a convolutional layer. :

[0034] G =

[0035] Finally, the features are calibrated:

[0036]

[0037] in, These are high-dimensional latent features after gating and masking calibration and filtering.

[0038] A second aspect is an apparatus for improving the user experience of a network service platform, the apparatus comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the apparatus to perform the method described in any one of the first aspects.

[0039] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0040] The beneficial effects of the technical solution provided by this invention are:

[0041] 1. This invention proposes the MIGformer (a multivariate time series prediction model based on multi-scale interaction and gating mechanism) architecture, which innovatively integrates the multi-scale time refinement module (MTR), the patch-level channel enhancement module (PCE), and the multi-view gating mechanism (MVG). Compared with the traditional Transformer model, this invention not only retains the ability to model global dependencies of long series, but also significantly enhances the ability to capture local time changes and cross-variable interactions, effectively solving the problems of non-stationarity and multi-scale noise in online business platform data.

[0042] 2. The multi-scale temporal refinement module (MTR) of the present invention adopts a "decomposition-fusion-denoising" strategy, which can extract implicit multi-scale temporal patterns from the original business data; in conjunction with the time attention mechanism, the module can effectively filter out high-frequency random noise (e.g., occasional jitter) in the business data, while retaining key trend information, providing high-quality input features for the model.

[0043] 3. The Patch-Level Channel Enhancement Module (PCE) of this invention addresses the defect of inverted architecture being prone to losing local details by introducing a history-aware reconstruction and channel attention mechanism. This enables the model to keenly capture local mutations in business volume within a short time window and the instantaneous coupling relationship between work orders of different business lines (for example, a surge in a certain type of consultation business leads to an increase in complaint business), thereby improving the model's responsiveness to sudden events.

[0044] 4. The multi-view gating mechanism (MVG) of the present invention, as a post-feature filter, can evaluate and calibrate the high-dimensional output of the encoder from multiple dimensions such as statistics, semantics and structure; by adaptively suppressing redundant features, the mechanism ensures that the features finally used for prediction have high discriminative power.

[0045] 5. This invention uses mean squared error (MSE) and mean absolute error (MAE) as evaluation metrics to assess the effectiveness of the neural network framework. Experiments show that this invention has good generalization performance, high prediction accuracy, and small error.

[0046] 6. Based on predictive data, this invention can realize refined scheduling of customer service personnel and dynamic expansion and contraction of computing resources, thereby reducing operating costs and avoiding a decline in user experience due to insufficient resources, which has significant practical value. Attached Figure Description

[0047] Figure 1 A flowchart illustrating a method for improving the user experience of a network service platform;

[0048] Figure 2 A schematic diagram of the business volume prediction model MIGformer;

[0049] Figure 3 This is a structural diagram of the MIGformer;

[0050] Figure 4 This is a schematic diagram of an inverted Transformer.

[0051] Figure 5 This is a schematic diagram of the structure of the time attention mechanism;

[0052] Figure 6 This is a schematic diagram of the channel attention mechanism. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0054] Example 1

[0055] This invention provides a method for improving the user experience of a network service platform, the method comprising the following steps:

[0056] 101: Collect and analyze historical data on different daily business volumes processed by the online business platform, perform data preprocessing, and establish a multivariate business volume dataset of the online business platform with time-series relationships;

[0057] 102: Input the constructed multivariate business volume dataset into the MIGformer model for processing and prediction;

[0058] The MIGformer model is the overall prediction model, consisting of a reversible instance normalization module, a multi-scale temporal refinement module (MTR-Module), a patch-level channel enhancement module (PCE-Module), an inverted Transformer encoder, and a multi-view gating mechanism (MVG-Mechanism).

[0059] 103: Use the reversible instance normalization module set when inputting data into the model and outputting prediction results to normalize and denormalize the processed and predicted data, thereby reducing the impact of distribution offset on prediction.

[0060] 104: The normalized and denormalized data are input into the multi-scale temporal refinement module (MTR-Module), and multi-scale features are extracted and noise is suppressed by using a pyramid decomposition-fusion strategy and a temporal attention mechanism;

[0061] 105: The extracted and noise-suppressed multi-scale features are input into the Patch-Level Channel Enhancement Module (PCE-Module), which divides the complete sequence into patch sequences. The historical perception reconstruction strategy and channel attention mechanism are used to capture the local contextual relationships and the interaction relationships between variables within the patch sequences.

[0062] 106: Input the patched enhanced sequence into the inverted Transformer encoder to extract global features, and then dynamically calibrate and filter the global features obtained by the encoder through the multi-view gating mechanism (MVG-Mechanism);

[0063] 107: The calibrated and filtered results are passed through a linear prediction layer to generate the final traffic volume prediction results, which will be used to guide personnel in dynamic scheduling and task allocation.

[0064] In summary, the embodiments of the present invention suppress noise interference through multi-scale refinement and multi-view gating mechanisms, effectively improve the accuracy and reliability of business volume prediction, achieve refined operation, adjust scheduling strategies in a timely manner, avoid manpower waste or response delays caused by prediction deviations, and improve user experience.

[0065] Example 2

[0066] The following section provides a detailed explanation of the scheme in Example 1, using specific examples and calculation formulas. See the description below for details:

[0067] 201: Obtain historical data on the daily volume of different work orders processed on the online business platform over the past three years, and analyze and preprocess the historical data to establish a business volume dataset of the network business platform with time-series relationships.

[0068] The collected historical data on different daily transaction volumes includes: daily completed transaction volume, daily uncompleted transaction volume, daily growth rate of completed transaction volume, and daily growth rate of uncompleted transaction volume. The collected data is saved in CSV file format. Python is used to preprocess outliers and missing values ​​in the raw data to avoid missing values ​​affecting the prediction results.

[0069] Linear interpolation is used to impute missing values ​​in the dataset. The calculation formula is as follows:

[0070]

[0071] In the formula, The value for the missing part. for The previously known values, for The known values ​​for a, b, and c are natural numbers greater than or equal to 1.

[0072] After completing the preprocessing of the business volume data of the online business platform, a business volume dataset of the online business platform with time-series relationships is established for predictive modeling and training.

[0073] 202: Use the reversible instance normalization module set when inputting data into the model and when outputting prediction results to normalize and denormalize the processed and predicted data;

[0074] Reversible instance normalization is a normalization method specifically designed for time series data. Its core idea is to alleviate the distribution skew problem in time series by dynamically adjusting the statistical properties (mean and variance) of the data. At the same time, after model processing, the original data distribution can be restored through reversible operations, thereby improving the predictive performance of the model.

[0075] Specifically, the mean and standard deviation of the input time series data are calculated for each instance (i.e., each sliding window or time series sample), and then standardized.

[0076]

[0077] in, and These are the mean and standard deviation of the current instance, respectively. The original input data, This is data that has undergone normalization. This step transforms the data into a standard distribution with a mean of 0 and a variance of 1, reducing distributional differences between different time windows.

[0078] After the model outputs the prediction results, the mean and standard deviation saved during the normalization stage are used to restore the predicted values ​​to the original scale:

[0079]

[0080] in, The normalized prediction results output by the model. This is the final predicted value after inverse normalization.

[0081] This not only effectively solves the problem of distribution offset, but also makes the prediction results closer to the true distribution.

[0082] 203: Input the normalized and denormalized data into the multi-scale temporal refinement module (MTR-Module), and use the MTR-Module to extract multi-scale features and perform denoising processing;

[0083] This multi-scale temporal refinement module employs a "decomposition-fusion" strategy. First, it generates a temporal pattern pyramid using average pooling layers with different kernel sizes.

[0084]

[0085] in, For the first Layer-scale characteristics; It is a one-dimensional average pooling layer.

[0086] Next, a bottom-up recursive fusion strategy is adopted, adjusting the dimensions through a linear layer before summing the residuals:

[0087] in, Features are the result of fusing multi-scale information.

[0088] Finally, it is refined through the time-attention mechanism:

[0089]

[0090] in, The input features to be refined; The output features are those refined over time. It is a sigmoid activation function; Represents one-dimensional convolution; This represents the average pooling layer.

[0091] The temporal attention mechanism compresses channels through global average pooling and then learns temporal step weights through a convolutional network to weight features in order to suppress noise.

[0092] 204: The extracted and noise-suppressed multi-scale features are input into the Patch-Level Channel Enhancement Module (PCE-Module) to enhance local interactions. This module (code class name DDI) segments the refined sequence into non-overlapping patches. Using a history-aware strategy, the previous... Historical patches are aggregated through a linear layer:

[0093]

[0094] in, Features derived from aggregating historical information; The activation function for the Gaussian error linear unit; For the first A historical patch.

[0095] The current patch and the features obtained by aggregating historical information are overlaid and fused to obtain the initial patch features that include historical context. :

[0096]

[0097] in, This represents the patch characteristics at the current moment; The patch features are those of the merged patch. This represents the patch feature from the previous moment.

[0098] Next, the merged patch features The input channel attention mechanism performs feature recalibration. This mechanism first compresses the time dimension through an adaptive average pooling layer to extract channel descriptors, then generates channel weights through an activation network containing two fully connected layers, and finally multiplies the channel weights element-wise with the initial patch features.

[0099]

[0100] in, These are patch features after channel attention recalibration; For adaptive average pooling layer; This is a multilayer perceptron network used to generate weights within the attention mechanism; It is a sigmoid activation function.

[0101] Subsequently, the recalibrated patch features are input into another independent multilayer perceptron for deep mixing of channel features:

[0102]

[0103] in, The output features are those obtained after feature mixing. This is a multilayer perceptron used for feature depth mixing.

[0104] Finally, the blended features are scaled by a scaling factor and residually connected with the initial patch features to output the enhanced patch, as shown in the formula:

[0105]

[0106] In the formula, The final output is the enhanced patch sequence. These are the weighting coefficients. After the aforementioned historical perception reconstruction and channel attention recalibration, a series of enhancement patch sequences arranged in chronological order are obtained. ,in This represents the total number of patches.

[0107] Finally, through the patch restoration operation, all the enhanced patches are sequentially concatenated along the time dimension to restore the enhanced time series features with a complete time length, so that they can be directly used as input to the inverted Transformer encoder.

[0108] 205: Generate prediction results using an inverted Transformer encoder and a multi-view gating mechanism (MVG-Mechanism).

[0109] First, the enhanced patch sequence is input into the inverted Transformer encoder. The encoder embeds the entire sequence of the same variable into a single token, completing the dimensional inversion transformation:

[0110]

[0111] in, This is the transpose of the enhanced patch sequence output in the previous step. This is the embedded token. Subsequently, a self-attention mechanism is used to calculate the attention weights between different variable lexical units, extracting global correlations across variables and outputting high-dimensional latent features. :

[0112]

[0113] in, Includes self-attention computation and feedforward neural networks; This is a high-dimensional latent feature for extracting cross-variable global correlations.

[0114] The inverted encoder outputs high-dimensional latent features of cross-variable global correlation. Then, a multi-view gating mechanism is used. This mechanism extracts and concatenates features from four perspectives:

[0115] (1) Global semantic perspective (Linear): ;

[0116] (2) Local structural perspective (Conv): ;

[0117] (3) Statistical perspective (Max): ;

[0118] (4) Statistical perspective (Avg): .

[0119] Use the four After concatenation, a gated mask is generated through a convolutional layer. :

[0120] G =

[0121] Finally, the features are calibrated:

[0122]

[0123] in, "High-dimensional latent features after gating, masking, and filtering;" " indicates element-wise multiplication.

[0124] 206: The calibrated and filtered results are passed through a linear prediction layer to generate the final traffic volume prediction result. The calibration features output by the multi-view gating mechanism are then used. A linear layer is used to perform dimensionality mapping, transforming it into sequence features of the target prediction time length. The calculation formula is:

[0125]

[0126] in, Represents the calculation of the linear prediction layer; The weight matrix of the linear layer; For bias terms; This is the normalized prediction result output by the linear layer of the model.

[0127] Subsequently, the mean from the data input phase was used. and standard deviation For the mapped features Perform a reversible instance denormalization operation to restore it to the true data distribution scale:

[0128]

[0129] Obtain the final business volume forecast results .

[0130] In summary, the traffic volume prediction results generated by the embodiments of the present invention will be directly output to the management system of the network business platform to predict future traffic peaks and valleys in advance. This will guide operators in dynamically scheduling customer service agents and allocating server computing resources, effectively avoiding network congestion and excessive user waiting, and improving the user experience of the network business platform.

[0131] Example 3

[0132] A business volume prediction device for an online business platform includes a processor and a memory. The memory stores program instructions, and the processor calls the program instructions stored in the memory to cause the device to execute the following method steps in Embodiment 1:

[0133] The constructed multivariate business volume dataset is input into the MIGformer model for processing and prediction; the MIGformer model includes: a reversible instance normalization module, a multi-scale time refinement module, a patch-level channel enhancement module, an inverted Transformer encoder, and a multi-view gating mechanism.

[0134] The reversible instance normalization module is used to normalize and denormalize the processed and predicted data; the normalized and denormalized data are then input into the multi-scale temporal refinement module, where a pyramid decomposition-fusion strategy and a temporal attention mechanism are used to extract multi-scale features and suppress noise.

[0135] The extracted and noise-suppressed multi-scale features are input into the patch-level channel enhancement module, which divides the complete sequence into patch sequences and uses a history-aware reconstruction strategy and channel attention mechanism to capture the local contextual relationships and the interaction relationships between variables within the patch sequences.

[0136] The patched sequence, enhanced by the patching, is input into the inverted Transformer encoder to extract global features. The global features obtained by the encoder are dynamically calibrated and filtered through a multi-view gating mechanism. The calibrated and filtered results are then passed through a linear prediction layer to generate the final traffic volume prediction result.

[0137] Among them, the pyramid decomposition-fusion strategy and temporal attention mechanism are used to extract multi-scale features and suppress noise as follows:

[0138] A temporal pattern pyramid is generated using average pooling layers with different kernel sizes; a bottom-up recursive fusion strategy is employed, adjusting the dimensions through linear layers and then summing the residuals; and a temporal attention mechanism is used for refinement.

[0139]

[0140] in, The input features to be refined; The output features are those refined over time.

[0141] The process of segmenting the complete sequence into patch sequences and using a history-aware reconstruction strategy and channel attention mechanism to capture the local contextual relationships and inter-variable interactions within the patch sequences is as follows:

[0142] The current patch and the features obtained by aggregating historical information are overlaid and fused to obtain the initial patch features that include historical context. :

[0143]

[0144] in, This represents the patch characteristics at the current moment; The patch features are those of the merged patch. Patch features representing the previous moment

[0145] Channel descriptors are extracted by compressing the time dimension through an adaptive average pooling layer, then channel weights are generated through an activation network containing two fully connected layers, and finally the channel weights are multiplied element-wise with the initial patch features.

[0146]

[0147] in, These are patch features after channel attention recalibration; For adaptive average pooling layer; This is a multilayer perceptron network used to generate weights within the attention mechanism; It is a sigmoid activation function.

[0148] The enhanced patch sequence is then fed into the inverted Transformer encoder. The encoder embeds the entire sequence of the same variable into a single token, completing the dimensional inversion transformation.

[0149]

[0150] in, This is the transpose of the enhanced patch sequence output in the previous step. This is the embedded token. Subsequently, a self-attention mechanism is used to calculate the attention weights between different variables, extracting global correlations across variables and outputting high-dimensional latent features. :

[0151]

[0152] in, Includes self-attention computation and feedforward neural networks; This is a high-dimensional latent feature for extracting cross-variable global correlations.

[0153] The inverted encoder outputs high-dimensional latent features of cross-variable global correlation. Then, a multi-view gating mechanism is used to extract and concatenate features from four perspectives:

[0154] (1) Global semantic perspective Linear: ;

[0155] (2) Conv from the perspective of local structure: ;

[0156] (3) Statistical perspective Max: ;

[0157] (4) Statistical perspective (Avg): ;

[0158] Use the four After concatenation, a gated mask is generated through a convolutional layer. :

[0159] G =

[0160] Finally, the features are calibrated:

[0161]

[0162] in, These are high-dimensional latent features after gating and masking calibration and filtering.

[0163] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0164] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0165] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0166] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0167] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0168] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0169] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0170] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for improving the user experience of a network service platform, characterized in that, The method includes: The constructed multivariate business volume dataset is input into the MIGformer model for processing and prediction; the MIGformer model includes: a reversible instance normalization module, a multi-scale time refinement module, a patch-level channel enhancement module, an inverted Transformer encoder, and a multi-view gating mechanism. The reversible instance normalization module is used to normalize and denormalize the processed and predicted data; the normalized and denormalized data are then input into the multi-scale temporal refinement module, where a pyramid decomposition-fusion strategy and a temporal attention mechanism are used to extract multi-scale features and suppress noise. The extracted and noise-suppressed multi-scale features are input into the patch-level channel enhancement module, which divides the complete sequence into patch sequences and uses a history-aware reconstruction strategy and channel attention mechanism to capture the local contextual relationships and the interaction relationships between variables within the patch sequences. The patched sequence, enhanced by the patching, is input into the inverted Transformer encoder to extract global features. The global features obtained by the encoder are dynamically calibrated and filtered through a multi-view gating mechanism. The calibrated and filtered results are then passed through a linear prediction layer to generate the final traffic volume prediction result.

2. The method for improving the user experience of a network service platform according to claim 1, characterized in that, The method of extracting multi-scale features and suppressing noise using a pyramid decomposition-fusion strategy and a temporal attention mechanism is as follows: A temporal pattern pyramid is generated using average pooling layers with different kernel sizes; a bottom-up recursive fusion strategy is employed, adjusting the dimensions through linear layers and then summing the residuals; and a temporal attention mechanism is used for refinement. ； in, The input features to be refined; The output features are those refined over time. It is a sigmoid activation function; Represents one-dimensional convolution; This represents the average pooling layer.

3. The method for improving the user experience of a network service platform according to claim 1, characterized in that, The process of dividing the complete sequence into patch sequences and using a history-aware reconstruction strategy and channel attention mechanism to capture the local contextual relationships and inter-variable interactions within the patch sequences is as follows: The current patch and the features obtained by aggregating historical information are overlaid and fused to obtain the initial patch features that include historical context. : ； in, This represents the patch characteristics at the current moment; The patch features are those of the merged patch. The patch feature represents the previous moment; Channel descriptors are extracted by compressing the time dimension through an adaptive average pooling layer, then channel weights are generated through an activation network containing two fully connected layers, and finally the channel weights are multiplied element-wise with the initial patch features. ； in, These are patch features after channel attention recalibration; For adaptive average pooling layer; This is a multilayer perceptron network used to generate weights within the attention mechanism; It is a sigmoid activation function.

4. The method for improving the user experience of a network service platform according to claim 1, characterized in that, The process of inputting the patched and enhanced sequence into the inverted Transformer encoder to extract global features is as follows: in, Includes self-attention computation and feedforward neural networks; This represents the input sequence after passing through the inverted encoder embedding layer; To extract high-dimensional latent features of cross-variable global correlation; The inverted encoder outputs high-dimensional latent features of cross-variable global correlation. Then, a multi-view gating mechanism is used to extract and concatenate features from four perspectives: (1) Global semantic perspective Linear: ; (2) Conv from the perspective of local structure: ; (3) Statistical perspective Max: ; (4) Statistical perspective (Avg): ; Use the four After concatenation, a gated mask is generated through a convolutional layer. : G = ； Finally, the features are calibrated: ； in, These are high-dimensional latent features after gating and masking calibration and filtering.

5. An apparatus for improving the user experience of a network service platform, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-4.

Images