A short video traffic multi-dimensional prediction method based on TSGTST

Abstract

The application discloses a short video traffic multidimensional prediction method based on TSGTST, aiming at solving the problems of existing models, such as dependence on single behavior data, insufficient long-short term time sequence feature extraction, and low prediction accuracy of play quantity, best-seller probability, recommendation weight and user stay duration. The method collects and pre-processes short video basic behavior data and multi-source emotion data, and constructs an emotion index; through RevIN normalization, adaptive time enhancement, sliding block and channel independent coding, feature enhancement and time sequence structure processing are realized; a mid-term cross-modal fusion layer is designed to deeply fuse behavior and emotion features, and combined with mixed global-local attention, the long-short term change rule of traffic is captured; time decay gated feedforward neural network is used to filter and weight features, and finally multidimensional traffic prediction is realized. The application effectively improves the prediction accuracy and robustness, and can provide support for platform recommendation algorithm optimization and content operation decision.

Description

Technical Field

[0001] This invention relates to the fields of computer technology, short video algorithm optimization and time series prediction, and in particular to a multi-dimensional prediction method for short video traffic based on TSGTST. Background Technology

[0002] With the rapid development of the short video industry, short video platforms are placing increasingly higher demands on the accuracy of traffic prediction for content recommendation and operational decisions. Accurate traffic prediction helps platforms optimize their recommendation algorithms and helps creators refine their content creation strategies. Currently, most short video traffic prediction methods rely on basic behavioral data such as completion rate and likes, neglecting the impact of emotional factors such as comments, bullet screen sentiment, and content sentiment tags on traffic. However, emotional dissemination is one of the core factors that enables short videos to become viral hits and gain high recommendation weight from platforms. Meanwhile, existing prediction models mostly employ traditional time series models or basic Transformer models, which have the following drawbacks: First, traditional time series models struggle to capture the long-term trends and short-term fluctuations of short video traffic, and are insufficient in modeling the temporal correlation between posting time and account level. Second, basic Transformer models have high attention computational complexity, making them inefficient when processing massive amounts of short video time series data, and their indiscriminate feature processing easily introduces redundant information. Third, multi-source feature fusion often uses pre- or post-fusion methods, which can easily cause feature interference and fail to achieve deep collaborative interaction between basic behavioral features and emotional features. Fourth, the prediction dimensions are singular, mostly only predicting play counts, lacking joint prediction of core indicators such as the probability of viral hits, recommendation weights, and user dwell time, making it difficult to meet the actual needs of platforms and creators. While the Transformer model, as an efficient long-sequence time series prediction model, is widely used in fields such as financial time series prediction, directly applying it to short video traffic prediction scenarios cannot adapt to the emotional feature fusion and multi-dimensional prediction needs of short video data. Targeted improvements are needed to enhance the accuracy and efficiency of multi-dimensional traffic prediction. Summary of the Invention

[0003] This invention addresses the characteristics of short video traffic, such as rapid temporal decay, strong coupling between emotion and behavior, and separation between short-term fluctuations and long-term trends. It provides a multi-dimensional prediction method for short video traffic based on TSGTST, achieving cross-modal deep fusion of basic behavioral features and emotional features. This method accurately captures the long-term and short-term temporal features of short video traffic and simultaneously completes multi-dimensional predictions of play count, probability of viral hits, recommendation weight, and user dwell time, thereby improving prediction accuracy and robustness.

[0004] A multi-dimensional prediction method for short video traffic based on TSGTST, the steps of which are as follows: Step 1: Data Acquisition and Preprocessing Basic behavioral data and multi-source sentiment data of short videos are collected. The two types of data are cleaned, standardized and quantified respectively to construct a short video sentiment index and obtain a dual-input feature sequence. The basic behavioral data includes completion rate, number of likes, number of reposts, video duration, posting time, and account level. The multi-source sentiment data includes comment sentiment, bullet screen sentiment, and content sentiment tags.

[0005] 1.1 Data Collection: Sample data was collected through the open interfaces of short video platforms, spanning from January 1, 2023 to January 1, 2025, covering multiple short video categories such as lifestyle, comedy, technology, and emotion, totaling 1 million valid short video data points. Basic behavioral data included completion rate, likes, shares, video duration (seconds), posting time, and account level; multi-source emotional data included comment sentiment (crawled from short video comment sections), bullet screen sentiment (collected from bullet screen content during playback), and content sentiment tags (labeled by the platform as touching, funny, angry, depressed, etc.).

[0006] 1.2 Data Cleaning: Remove missing and outlier values ​​from basic behavioral data, eliminate meaningless content and pure emoji content from comments / bullet screens; standardize content sentiment tags and unify the tag classification system.

[0007] 1.3 Feature Quantification and Standardization: Z-score standardization is applied to completion rate, number of likes, number of shares, and video duration to eliminate differences in units; 1.4 Sentiment Data Classification and Index Construction: A pre-trained BERT sentiment classification model was used to classify comment sentiment and bullet screen sentiment into positive (… ),neutral( ),negative( The content is categorized into three types; emotional tags are assigned quantitative values ​​based on emotional intensity (touched / funny is positive, +1; angry / depressed is negative, -1; no obvious emotion is neutral, 0). This is done using a formula.

[0008] Calculating the short video sentiment index , ∈[-100,100], A value >0 indicates an overall positive sentiment. The larger the absolute value, the more pronounced the emotional tendency.

[0009] 1.5 Feature Sequence Construction: Based on time steps, construct basic behavioral feature sequences and emotional feature sequences as dual inputs to the model.

[0010] Step 2: Feature Enhancement Processing

[0011] The dual-channel input feature sequences obtained in the first step are subjected to RevIN channel independent reversible normalization to eliminate dimensional differences. The statistical features of the feature sequences are extracted through the adaptive temporal enhancement module, and the enhancement strategy is dynamically generated and the feature enhancement is completed to improve the model's adaptability to the temporal fluctuations of short video traffic.

[0012] 2.1 RevIN Channel-Independent Reversible Normalization: Channel-independent reversible normalization is performed on the dual-channel input feature sequences to eliminate the dimensional differences between different channels. The calculation formula is as follows:

[0013] in Let be the mean of the c-th channel of the b-th sample. Standard deviation To minimize the value, the denominator should not be zero; during the prediction phase, the original dimensions are restored through inverse normalization.

[0014] 2.2 Adaptive Time Augmentation: Extract the mean, variance, maximum, and minimum values ​​of each channel for each sample to construct a four-dimensional statistical feature space.

[0015] Enhancement weights are generated using two independent MLPs. and enhanced parameters : The weights for enabling three enhancement methods—noise injection, time warp, and frequency perturbation—are controlled. Adjusting the enhancement intensity weakens the enhancement for short videos with stable traffic and strengthens the enhancement for short videos with large traffic fluctuations, thereby improving the robustness of the model.

[0016] Step 3: Sequence Blocking and Channel-Independent Encoding

[0017] Short video traffic features are one-dimensional continuous time-series data. Directly inputting them into the attention module leads to a quadratic increase in computational complexity. Furthermore, single-dimensional time-series data lacks structured modeling and is prone to losing channel-independent features and positional information. This step transforms the one-dimensional time series into a fixed-length patch two-dimensional structure, significantly reducing the complexity of subsequent attention calculations. At the same time, channel-independent encoding preserves the independence of each feature channel, and positional encoding supplements the temporal positional information, achieving structured and low-complexity modeling of time-series features and laying the foundation for cross-modal fusion.

[0018] 3.1 Sequence Blocking: The enhanced feature sequences are padded with zeros to fill in the sequence edges and prevent information loss. A sliding block strategy is adopted, with a patch length P=16 and a step size S=8, transforming the original one-dimensional temporal features into a two-dimensional patch structure. Where B is the batch size and C is the number of feature channels. For the number of patches, The length of a single patch.

[0019] 3.2 Channel-independent encoding: For each channel's patch subsequence, perform independent linear projection, mapping the dimensional vector to the model dimension (set to 128), using the following formula:

[0020] in For projection weights, For bias; a learnable zero-initialized positional code (PE) is introduced, positional information is overlaid, and a Dropout operation is performed: ,in The features are flattened by batch and channel dimensions. The flattened dimensions are: .

[0021] Step 4: Cross-modal feature fusion

[0022] Basic behavioral features and emotional features are two core heteromodal factors influencing short video traffic. Simple feature concatenation easily leads to feature interference between modalities and fails to uncover the inherent synergistic interaction between the two types of features. Pre- or post-fusion methods can also easily result in the loss of feature information. This step adopts a mid-stage fusion strategy, completing cross-modal fusion after feature encoding and before attention calculation. Dimension alignment ensures feature interactivity, and a gating network adaptively learns the interaction weights of bimodal features. Then, the original features and interaction features are concatenated to achieve deep fusion, which strengthens the modeling of the association between behavior and emotion while preserving the independent information of the original features, and outputs a feature format adapted to the attention module.

[0023] 4.1 Dimension Flattening and Alignment: The batch channel dimensions of the dual-channel encoded features are split and recombined to restore the batch dimensions while preserving channel independence, resulting in...

[0024] in The number of emotion feature channels, Number of channels for basic behavioral characteristics; 4.2 Gated Interaction Fusion: Interaction weights are learned through a single-hidden-layer gating network, using the following formula:

[0025] in It is a single-hidden-layer gated network. For the gated weight matrix, For element-wise product; 4.3 Feature Enhancement and Fusion: The original encoded features and interaction features are concatenated to obtain the final fused features.

[0026] Dimensions ; 4.4 Dimensional Reorganization: The fused features are flattened again to obtain:

[0027] Adapt to the input format of subsequent attention modules.

[0028] Step 5: Hybrid Global-Local Attention Feature Extraction

[0029] Short video traffic exhibits clear long-term trends and short-term fluctuations. Traditional single attention mechanisms either struggle to capture local short-term fluctuations or are inefficient in global trend modeling. This step designs a multi-path attention head, using dynamic local masks to limit the computational scope of local attention and accurately capture the details of short-term traffic fluctuations. Unconstrained global attention is used to achieve global modeling of long-term trends. Finally, concatenation and linear projection are used to achieve complementary enhancement of the two types of features, allowing the model to simultaneously and accurately uncover the long- and short-term temporal evolution patterns of short video traffic. The attention heads (total number of heads set to 12) are divided into local attention heads (8) and global attention heads (4), respectively capturing short-term fluctuations and long-term trends of short video traffic. 5.1 Local Attention Calculation: The attention range is limited by a dynamic local mask. The mask window W=4, allowing each patch to only pay attention to its four adjacent patches before and after it. The mask formula is as follows:

[0030] Calculate feature responses based on masks using scaled dot product attention: Let h be the query, key, and value vector of the h-th local attention head. For attention head dimension; 5.2 Global Attention Calculation: Without mask constraints, each patch can interact with all other patches in terms of attention. The calculation formula is as follows:

[0031] 5.3 Attention Fusion: Concatenating the outputs of all local and global attention heads

[0032] The final attention fusion features are obtained by mapping back to model dimension 128 through a linear projection layer.

[0033] Step 6: Time-decay gated feedforward neural network processing

[0034] Attention fusion features contain some low-value redundant information that decays over time, and traditional feedforward neural networks treat features indiscriminately, easily retaining invalid information that interferes with subsequent predictions. Furthermore, deep network training is prone to the vanishing gradient problem. This step, as the core of feature optimization, introduces a time decay factor to adapt to the temporal decay characteristics of short video traffic, and combines it with Sigmoid gating to achieve accurate selection and weighting of high-value features. Simultaneously, through the paradigm of "layer normalization-transformation-residual connection," gradient vanishing is effectively avoided, improving the propagation efficiency of features in the network, and ultimately outputting redundancy-free, high-value fused features.

[0035] 6.1 Time-decay-gated FFN calculation: Input attention fusion features into a time-decay-gated feedforward neural network (FFN) Introducing a time decay factor

[0036] in This is a learnable attenuation coefficient (initial value set to 0.1). The time difference is calculated using the following formula:

[0037] in Sigmoid gate, To achieve element-wise multiplication, dual filtering and weighting of features are implemented; 6.2 Residual Connectivity and Layer Normalization: In Layer normalization is introduced before and after processing. The input and output are added together through a residual connection to construct a "normalization-transformation-residual" paradigm, which avoids gradient vanishing, improves feature propagation efficiency, and finally outputs the features. .

[0038] Step 7: Multi-dimensional prediction output

[0039] After multiple rounds of feature processing, fusion, and optimization, the features have been mapped to a high-dimensional feature space. A transformation from this feature space to the prediction space is needed. Furthermore, the previous normalization process caused the features to lose their original dimensions, making it impossible to directly output prediction results interpretable for actual business needs. This step transforms the high-dimensional structured features into one-dimensional vectors through feature flattening. Independent linear heads are used to decode multiple prediction targets separately, adapting to different prediction needs such as play counts, probabilities, and weights. Then, RevIN denormalization restores the original dimensions of the features, ultimately outputting multi-dimensional traffic metrics that fit the actual business of short video platforms, realizing the implementation from feature modeling to business prediction.

[0040] 7.1 Feature Flattening: The fused feature Z is flattened into a two-dimensional vector using the Flatten operation.

[0041] 7.2 Independent Decoding: Independent linear head decoding is performed on each fusion channel, and the flattened feature vectors are mapped to the target prediction window length respectively, realizing the transformation from feature space to prediction space; 7.3 Denormalization: The RevIN denormalization operation restores the original dimensions, outputting four prediction results: Short video views: The predicted cumulative views over the next 7 days; Viral probability: The predicted views are compared with the platform's viral threshold (dynamically set by category, e.g., 1 million views for comedy short videos), combined with a comprehensive score of sentiment index and recommendation weight, outputting a viral probability value of 0-1; Recommendation weight: The predicted weight of the platform's recommendation algorithm for short videos, ranging from 0-10, with higher values ​​indicating stronger recommendation; User dwell time: The predicted average dwell time (in seconds) for users watching short videos.

[0042] Step 8: Model Training and Optimization

[0043] Using preprocessed short video data as training samples, MSE, MAE, and RSE were used as evaluation metrics. The Adam optimizer and dynamic learning rate were configured, and an early stopping mechanism was introduced to avoid overfitting, thus completing the optimization of model parameters.

[0044] 8.1 Dataset partitioning: The 1 million preprocessed samples were divided into a training set (700,000 samples), a validation set (100,000 samples), and a test set (200,000 samples) in a ratio of 7:1:2. Stratified sampling was used to ensure that the distribution of short video categories, account levels, and posting times was consistent across the datasets. 8.2 Training parameter configuration: The number of model channels is set to 12, the batch size is set to 32, the learning rate is set to 0.001; the optimizer is Adam, the first moment decay coefficient β1=0.9, the second moment decay coefficient β2=0.999, and the numerical stability constant ε=1e-8. 8.3 Evaluation Metrics: MSE (mean squared error), MAE (mean absolute error), and RSE (relative squared error) are used as model evaluation metrics. The loss function is MSE loss.

[0045] in For the number of channels, To predict the number of steps, For predicted values, The actual value; 8.4 Early Stopping Mechanism: During the training process, after each epoch, the model loss value is evaluated on the validation set. If the loss value on the validation set does not decrease for 10 consecutive epochs, the early stopping mechanism is triggered, training is stopped and the current optimal model parameters are saved. 8.5 Model Optimization: Based on the evaluation results of the test set, fine-tune parameters such as the model's patch length, attention mask window, and decay coefficient to further improve the model's prediction accuracy.

[0046] The beneficial effects of this invention are as follows: 1. High prediction accuracy: The model achieves deep integration of basic behavioral data and sentiment data through cross-modal fusion, accurately captures long-term and short-term temporal features through a hybrid global-local attention module, and reduces redundant information interference through time decay gating FFN. Experiments have verified that the MSE, MAE, and RSE of the model of this invention are all superior to traditional models such as DLinear, Transformer, and Autoformer. 2. Inter-module collaboration: The innovation of this invention lies in the collaboration between modules. RevIN+ adaptive enhancement ensures that emotion and behavioral features are co-distributed input. Mid-term cross-modal fusion completes deep coupling of behavior and emotion before attention. Hybrid global-local attention captures both short-term fluctuations and long-term trends. Time decay gating TD-GFFN adapts to the rapid decay characteristics of short video traffic. 3. Strong model robustness: RevIN normalization and adaptive time enhancement modules effectively eliminate data distribution bias and noise interference, and the early stopping mechanism avoids model overfitting. The model can adapt to short video traffic prediction scenarios of different categories, account levels and release times. 4. High computational efficiency: By dividing the sequence into blocks, the computational complexity of attention is reduced from O(N²) to O(N). When processing massive amounts of short video time-series data, it balances computational efficiency and feature capture capabilities, meeting the platform's real-time prediction needs. 5. High adaptability: Parameters such as attenuation coefficient, viral threshold, and attention mask window in the model can be dynamically adjusted according to the category characteristics and traffic patterns of short video platforms, which can quickly adapt to the application needs of different short video platforms. Attached Figure Description

[0047] Figure 1 This is a framework diagram of a multi-dimensional prediction method for short video traffic based on TSGTST. Figure 2 Diagram of the TD-GFFN Transformer encoder; Figure 3 This is a block diagram of a Time Decay Gated Feedforward Neural Network (TD-GFFN). Detailed Implementation

[0048] This embodiment provides a multi-dimensional prediction method for short video traffic based on TSGTST. The specific steps are as follows: S1. Data Acquisition and Preprocessing 1.1 Data Collection: Sample data was collected through the open interfaces of short video platforms, spanning from January 1, 2023 to January 1, 2025, covering multiple short video categories such as lifestyle, comedy, technology, and emotion, totaling 1 million valid short video data points. Basic behavioral data included completion rate, likes, shares, video duration (seconds), posting time, and account level; multi-source emotional data included comment sentiment (crawled from short video comment sections), bullet screen sentiment (collected from bullet screen content during playback), and content sentiment tags (labeled by the platform as touching, funny, angry, depressed, etc.).

[0049] 1.2 Data Cleaning: Remove missing and outlier values ​​from basic behavioral data, eliminate meaningless content and pure emoji content from comments / bullet screens; standardize content sentiment tags and unify the tag classification system.

[0050] 1.3 Feature Quantification and Standardization: Z-score standardization is applied to completion rate, number of likes, number of shares, and video duration to eliminate differences in units; 1.4 Sentiment Data Classification and Index Construction: A pre-trained BERT sentiment classification model was used to classify comment sentiment and bullet screen sentiment into "positive" (…). ),neutral( ),negative( The content is categorized into three types: emotional, emotional, and negative. Emotional tags are quantified and assigned values ​​based on emotional intensity (e.g., moved / funny = positive, +1; angry / suppressed = negative, -1; no obvious emotion = neutral, 0). This is achieved using a formula...

[0051] Calculating the short video sentiment index , ∈[-100,100], A value >0 indicates an overall positive sentiment. The larger the absolute value, the more pronounced the emotional tendency.

[0052] 1.5 Feature Sequence Construction: Based on time steps, construct basic behavioral feature sequences and emotional feature sequences as dual inputs to the model.

[0053] S2, Feature Enhancement Processing

[0054] 2.1 RevIN Channel-Independent Reversible Normalization: Channel-independent reversible normalization is performed on the dual-channel input feature sequences to eliminate the dimensional differences between different channels. The calculation formula is as follows:

[0055] in Let be the mean of the c-th channel of the b-th sample. Standard deviation To minimize the value, the denominator should not be zero; during the prediction phase, the original dimensions are restored through inverse normalization.

[0056] 2.2 Adaptive Time Augmentation: Extract the mean, variance, maximum, and minimum values ​​of each channel for each sample to construct a four-dimensional statistical feature space.

[0057] Enhancement weights are generated using two independent MLPs. and enhanced parameters : The weights for enabling three enhancement methods—noise injection, time warp, and frequency perturbation—are controlled. Adjusting the enhancement intensity weakens the enhancement for short videos with stable traffic and strengthens the enhancement for short videos with large traffic fluctuations, thereby improving the robustness of the model.

[0058] S3, Sequence Blocking and Channel Independent Encoding

[0059] 3.1 Sequence Blocking: The enhanced feature sequences are padded with zeros to fill in the sequence edges and prevent information loss. A sliding block strategy is adopted, with a patch length P=16 and a step size S=8, transforming the original one-dimensional temporal features into a two-dimensional patch structure. Where B is the batch size and C is the number of feature channels. For the number of patches, The length of a single patch.

[0060] 3.2 Channel-independent encoding: For each channel's patch subsequence, perform independent linear projection, mapping the dimensional vector to the model dimension (set to 128), using the following formula:

[0061] in For projection weights, For bias; a learnable zero-initialized positional code (PE) is introduced, positional information is overlaid, and a Dropout operation is performed: ,in The features are flattened by batch and channel dimensions. The flattened dimensions are: .

[0062] S4, Cross-modal feature fusion

[0063] 4.1 Dimension Flattening and Alignment: The batch channel dimensions of the dual-channel encoded features are split and recombined to restore the batch dimensions while preserving channel independence, resulting in...

[0064] in The number of emotion feature channels, Number of channels for basic behavioral characteristics; 4.2 Gated Interaction Fusion: Interaction weights are learned through a single-hidden-layer gating network, using the following formula:

[0065] in It is a single-hidden-layer gated network. For the gated weight matrix, For element-wise product; 4.3 Feature Enhancement and Fusion: The original encoded features and interaction features are concatenated to obtain the final fused features.

[0066] Dimensions ; 4.4 Dimensional Reorganization: The fused features are flattened again to obtain:

[0067] Adapt to the input format of subsequent attention modules.

[0068] S5. Hybrid Global-Local Attention Feature Extraction

[0069] The attention heads (total number of heads set to 12) are divided into local attention heads (8) and global attention heads (4) to capture short-term fluctuations and long-term trends in short video traffic, respectively: 5.1 Local Attention Calculation: The attention range is limited by a dynamic local mask. The mask window W=4, allowing each patch to only pay attention to its four adjacent patches before and after it. The mask formula is as follows:

[0070] Calculate feature responses based on masks using scaled dot product attention: Let h be the query, key, and value vector of the h-th local attention head. For attention head dimension; 5.2 Global Attention Calculation: Without mask constraints, each patch can interact with all other patches in terms of attention. The calculation formula is as follows:

[0071] 5.3 Attention Fusion: Concatenating the outputs of all local and global attention heads

[0072] The final attention fusion features are obtained by mapping back to model dimension 128 through a linear projection layer.

[0073] S6, Time-decay gated feedforward neural network processing

[0074] 6.1 Time-decay-gated FFN calculation: Input attention fusion features into a time-decay-gated feedforward neural network (FFN) The model introduces a time decay factor and assigns differentiated weights to features at different time steps based on the temporal variation of short video traffic. This weakens the influence of outdated historical features, strengthens the expression of recent effective temporal features, and enhances the model's ability to perceive dynamic changes in short video traffic. By combining Sigmoid gating, the model achieves adaptive feature selection and weighting, highlighting key features that are highly correlated with traffic, suppressing redundant noise interference, and further enhancing the model's temporal feature extraction capability and prediction stability.

[0075]

[0076] in This is a learnable attenuation coefficient (initial value set to 0.1). The time difference is calculated using the following formula:

[0077] in Sigmoid gate, To achieve element-wise multiplication, dual filtering and weighting of features are implemented; 6.2 Residual Connectivity and Layer Normalization: In Layer normalization is introduced before and after processing. The input and output are added together through a residual connection to construct a "normalization-transformation-residual" paradigm, which avoids gradient vanishing, improves feature propagation efficiency, and finally outputs the features. .

[0078] S7, Multi-dimensional Prediction Output

[0079] 7.1 Feature Flattening: The fused feature Z is flattened into a two-dimensional vector using the Flatten operation.

[0080] 7.2 Independent Decoding: Independent linear head decoding is performed on each fusion channel, and the flattened feature vectors are mapped to the target prediction window length respectively, realizing the transformation from feature space to prediction space; 7.3 Denormalization: The RevIN denormalization operation restores the original dimensions, outputting four prediction results: Short video views: The predicted cumulative views over the next 7 days; Viral probability: The predicted views are compared with the platform's viral threshold (dynamically set by category, e.g., 1 million views for comedy short videos), combined with a comprehensive score of sentiment index and recommendation weight, outputting a viral probability value of 0-1; Recommendation weight: The predicted weight of the platform's recommendation algorithm for short videos, ranging from 0-10, with higher values ​​indicating stronger recommendation; User dwell time: The predicted average dwell time (in seconds) for users watching short videos.

[0081] S8, Model Training and Optimization

[0082] 8.1 Dataset partitioning: The 1 million preprocessed samples were divided into a training set (700,000 samples), a validation set (100,000 samples), and a test set (200,000 samples) in a ratio of 7:1:2. Stratified sampling was used to ensure that the distribution of short video categories, account levels, and posting times was consistent across the datasets. 8.2 Training parameter configuration: The number of model channels is set to 12, the batch size is set to 32, the learning rate is set to 0.001; the optimizer is Adam, the first moment decay coefficient β1=0.9, the second moment decay coefficient β2=0.999, and the numerical stability constant ε=1e-8. 8.3 Evaluation Metrics: MSE (mean squared error), MAE (mean absolute error), and RSE (relative squared error) are used as model evaluation metrics. The loss function is MSE loss.

[0083] in For the number of channels, To predict the number of steps, For predicted values, The actual value; 8.4 Early Stopping Mechanism: During the training process, after each epoch, the model loss value is evaluated on the validation set. If the loss value on the validation set does not decrease for 10 consecutive epochs, the early stopping mechanism is triggered, training is stopped and the current optimal model parameters are saved. 8.5 Model Optimization: Based on the evaluation results of the test set, fine-tune parameters such as the model's patch length, attention mask window, and decay coefficient to further improve the model's prediction accuracy.

[0084] Experimental verification

[0085] To verify the effectiveness of the model in this invention, four classic models in the field of short video traffic prediction (DLinear, Linear, Transformer, and Autoformer) were selected as benchmarks for comparison. They were trained and tested under the same experimental environment and parameter settings. The model performance was evaluated using three core metrics: MSE, MAE, and RSE. The experimental results are shown in Table 1.

[0086] Table 1 Model Comparison Table

[0087] Experimental results show that the TSGTST model of this invention outperforms the comparative models in all three core metrics: MSE is reduced by 28.0% compared to Transformer, MAE by 18.3% compared to Autoformer, and RSE by 6.3% compared to Linear. Compared to the better-performing DLinear model, the model of this invention is comparable in MAE, but superior in MSE and RSE, demonstrating stronger control over extreme errors. Furthermore, this invention features a dedicated algorithm improvement for specific scenarios. Through a unique combination of mid-term fusion, hybrid attention, and time decay gating, it solves the long-standing technical challenges in short video traffic prediction, such as the difficulty in fusing sentiment and behavior, the difficulty in balancing long-term and short-term features, and the difficulty in modeling time decay, thus meeting the practical needs of real-time prediction and operational decision-making for short video platforms.

Claims

1. A multi-dimensional prediction method for short video traffic based on TSGTST, characterized in that, In response to the rapid temporal decay of short video traffic and the strong coupling between behavioral and emotional characteristics, the following steps are included: Step 1: Data Collection and Preprocessing: Collect basic behavioral data of short videos and multi-source sentiment data such as comments, bullet comments, and content sentiment tags. Clean and standardize the basic behavioral data, quantify and statistically calculate the multi-source sentiment data, construct the short video sentiment index, and obtain the dual-input feature sequence. Step 2, Feature Enhancement Processing: Perform RevIN channel independent reversible normalization on the dual-channel input feature sequences obtained in Step 1 to eliminate dimensional differences; extract statistical features of the feature sequences through the adaptive time enhancement module, dynamically generate enhancement strategies, and complete feature enhancement; Step 3, Sequence Blocking and Independent Channel Encoding: The enhanced feature sequence from Step 2 is padded with zeros and subjected to sliding block processing to obtain fixed-length Patch subsequences; each channel's Patch subsequence is subjected to independent linear projection and learnable zero-initialized position encoding, and the encoded features are obtained after flattening the batch and channel dimensions; Step 4, Cross-modal feature fusion: The basic behavior encoding features and emotion encoding features are flattened and aligned in dimensions through the mid-term cross-modal fusion layer. The interaction weights of the dual-path features are learned by using a gating network. The original encoding features and interaction features are concatenated to obtain the final fused features, which are adapted to the input format of the subsequent attention module. Step 5, Hybrid Global-Local Attention: Divide the attention head into local attention head and global attention head. The local attention head captures the short-term fluctuation features of short video traffic through dynamic masking, while the global attention head captures long-term trend features without constraints. The outputs of all attention heads are stitched together and linearly projected to achieve feature fusion. Step 6: Time-decay gated feedforward neural network: Input the attention fusion features into the time-decay gated feedforward neural network. The features are dual-selected and weighted through the time decay factor and Sigmoid gating. The gradient vanishing is avoided by combining residual connections and layer normalization. Step 7, Prediction Output: After flattening the processed fused features, decode them through independent linear heads, and output the prediction results of short video views, probability of becoming a hit, recommendation weight, and user dwell time after RevIN inverse normalization; Step 8: Model Training and Optimization: Using the preprocessed short video data as training samples, MSE, MAE, and RSE are used as evaluation metrics. The Adam optimizer and dynamic learning rate are configured, and an early stopping mechanism is introduced to avoid overfitting, thus completing the optimization of model parameters.

2. The method for multi-dimensional prediction of short video traffic based on TSGTST according to claim 1, characterized in that, In step S1, the basic behavioral data includes completion rate, number of likes, number of shares, video duration, posting time, and account level, while the multi-source sentiment data includes comment sentiment, bullet screen sentiment, and content sentiment tags.

3. The method for multi-dimensional prediction of short video traffic based on TSGTST according to claim 1, characterized in that, In step 1, a short video sentiment index is constructed. Specifically, the sentiment of comments and bullet screen comments are categorized as "positive, neutral, and negative." Content sentiment tags are quantified and assigned values ​​based on whether they evoke emotion, humor, anger, or depression. The percentage of each emotion is then calculated using the formula: Calculate the sentiment index, where For the quantity of positive emotions, The number of negative emotions, The emotion index represents the number of neutral emotions, with a value range of [-100, 100]. The positive or negative value and the magnitude of the value reflect the overall emotional tendency and intensity of the short video.

4. The method for multi-dimensional prediction of short video traffic based on TSGTST according to claim 1, characterized in that, In step 2, the adaptive time enhancement module specifically involves: extracting the mean, variance, maximum, and minimum values ​​of each channel for each sample to construct a four-dimensional statistical feature space, and generating enhancement weights using an MLP. and enhanced parameters , The weights for enabling three enhancement methods—noise injection, time warp, and frequency perturbation—are controlled. Adjust the intensity of each enhancement method and dynamically adjust the enhancement strategy based on the stationarity of the short video feature sequence.

5. The method for multi-dimensional prediction of short video traffic based on TSGTST according to claim 1, characterized in that, In step 5, a hybrid global-local attention module is used. Local attention calculation restricts the attention range through a dynamic local mask, allowing each patch to focus only on its W adjacent patches before and after it. The mask formula is as follows: Global attention has no mask constraints, and each patch interacts with all other patches in terms of attention. The outputs of the local attention head and the global attention head are concatenated and linearly projected back to the model dimension, achieving complementary enhancement of short-term fluctuations and long-term trend features.

6. The method for multi-dimensional prediction of short video traffic based on TSGTST according to claim 1, characterized in that, In step 6, the time-decay-gated feedforward neural network introduces a time decay factor. Based on the temporal variation patterns of short video traffic, it assigns differentiated weights to features at different time steps, weakening the influence of outdated historical features and strengthening the expression of recent effective temporal features, thereby improving the model's ability to perceive dynamic changes in short video traffic. Combined with a gating mechanism, it achieves adaptive feature selection and weighting, highlighting key features highly correlated with traffic, suppressing redundant noise interference, and further enhancing the model's temporal feature extraction capability and prediction stability. The formula for the time decay factor is as follows: in The attenuation coefficient is... The time difference is calculated using the following formula: in Sigmoid gate, Element-wise product , , For projection weights, For bias; attenuation coefficient These are learnable parameters that are dynamically adjusted based on the short video's release time and traffic fluctuation characteristics.

7. The method for multi-dimensional prediction of short video traffic based on TSGTST according to claim 1, characterized in that, In step 8, model training and optimization specifically involves dividing the dataset into training, validation, and test sets in a 7:1:2 ratio, and using stratified sampling to ensure that the distribution of short video features in each dataset is consistent. The initial learning rate is set to 0.001, the first moment decay coefficient β1 of the Adam optimizer is set to 0.9, the second moment decay coefficient β2 is set to 0.999, and the numerical stability constant ε is set to 1e-8. If the validation set loss value does not decrease for 10 consecutive rounds, the early stopping mechanism is triggered to save the optimal model parameters.

8. A multi-dimensional prediction method for short video traffic based on TSGTST as described in any one of claims 1-7, characterized in that, Built specifically for multi-dimensional prediction of short video traffic, including: data acquisition module, data preprocessing module, feature enhancement module, sequence encoding module, cross-modal fusion module, attention feature extraction module, gating feature processing module, multi-dimensional prediction module, and model training optimization module; The data acquisition module is used to collect basic behavioral data and multi-source sentiment data from short videos; The data preprocessing module is used to clean, standardize, and quantify the collected data, construct a short video sentiment index, and output a dual-input feature sequence. The feature enhancement module includes a RevIN normalization unit and an adaptive temporal enhancement unit, which are used to normalize and dynamically enhance the feature sequence. The sequence encoding module includes a sequence segmentation unit and a channel-independent encoding unit, used to convert the feature sequence into a fixed-length encoded feature; The cross-modal fusion module is used to achieve deep fusion of basic behavioral coding features and emotion coding features, and output fused features; The attention feature extraction module is a hybrid global-local attention module used to capture the long and short-term temporal features of short video traffic; The gated feature processing module is a time-decay gated feedforward neural network, which combines residual connections and layer normalization units for feature selection, weighting, and propagation. The multi-dimensional prediction module is used to decode and denormalize the fused features, and output the prediction results of play count, hit probability, recommendation weight, and user dwell time. The model training and optimization module is used to set model training parameters, evaluation metrics, and optimization mechanisms to complete the learning and optimization of model parameters.

9. The prediction method according to claim 8, characterized in that, The data preprocessing module includes a posting time period feature quantification unit and an account level feature mapping unit. The posting time period is divided into four intervals—morning, noon, evening, and early morning—according to the peak traffic hours of the short video platform and quantified and assigned values. The account level is mapped to continuous numerical features according to the platform level system.

10. The prediction method according to claim 7, characterized in that, The multi-dimensional prediction module includes a viral probability determination unit. By comparing the predicted play count with the platform's viral threshold and combining it with a comprehensive score of sentiment index and recommendation weight, it outputs the probability value of the short video becoming a viral hit. The threshold can be dynamically adjusted according to different categories and account levels on the platform.

Images