Hotspot prediction and adaptive routing method for private live broadcast rooms based on swarm intelligence

Abstract

The present invention provides a hotspot prediction and adaptive routing method for a private domain live broadcast room based on swarm intelligence, which relates to the field of data processing technology, including: building a dynamic interest community through a dynamic graph neural network, extracting group interaction pattern features using a multi-layer graph attention network, generating temporal correlation features in combination with an improved bidirectional gated recurrent unit, extracting high-order semantic features, performing hotspot prediction through a multi-task learning framework, building a multi-level network topology model according to the prediction results, and performing path exploration in combination with an improved heuristic search algorithm to obtain an optimal transmission path, performing traffic distribution through a weighted polling algorithm, and dynamically adjusting the data transmission rate in combination with a distributed load balancing controller and a congestion control algorithm based on deep learning to achieve adaptive routing push of live broadcast content.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a hotspot prediction and adaptive routing method for a private domain live broadcast room based on swarm intelligence. Background Art

[0002] With the rapid development and popularization of Internet technology, live streaming has become an emerging e-commerce model. The rise of private live streaming rooms has provided merchants with more accurate user reach and more efficient conversion paths. However, how to accurately predict hot topics in private live streaming rooms and efficiently distribute content and manage traffic has become a major challenge facing private live streaming operations.

[0003] Existing hotspot prediction methods mainly rely on simple statistical analysis or rule-based expert systems, which are difficult to capture the dynamic changes and complex relationships of user behavior, resulting in low prediction accuracy. At the same time, traditional content distribution methods usually use static routing or simple load balancing strategies, which cannot adapt to the fluctuations in traffic in private domain live broadcast rooms and the diversity of user interests, and are prone to network congestion and reduced user experience. In addition, the existing private domain live broadcast management system lacks in-depth analysis and understanding of user group behavior, making it difficult to achieve refined traffic control and personalized content recommendations. Summary of the invention

[0004] The embodiment of the present invention provides a private domain live broadcast room hotspot prediction and adaptive routing method based on swarm intelligence, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of an embodiment of the present invention provides a method for hotspot prediction and adaptive routing in a private domain live broadcast room based on swarm intelligence, comprising:

[0006] According to the preset sliding time window, data is collected from the private domain live broadcast room to obtain user behavior data, and feature extraction is performed through a time series convolutional network to obtain time series features and add them to the gated recurrent unit. The update gate and the reset gate are combined for information flow, and the long-term dependency features are calculated. Based on the correlation of different positions, feature fusion is performed through a feature fusion network and the fused features are upgraded to a high-dimensional semantic space in combination with a feedforward neural network to obtain a time series feature matrix. A dynamic graph neural network is constructed based on the time series feature matrix, and the multi-dimensional similarity between user nodes is calculated through a dynamic edge weight update mechanism. The user nodes are dynamically connected based on the multi-dimensional similarity and a preset similarity threshold to obtain a dynamic interest community.

[0007] The dynamic interest community is added to a pre-set multi-layer graph attention network, and the group interaction pattern features are extracted through the spatial attention layer and the channel attention layer and added to the improved bidirectional gated recurrent unit to obtain long-term and short-term temporal dependencies and generate temporal correlation features. The temporal correlation features are added to the deep residual network and high-order semantic features are extracted through residual learning and feature recalibration. A multi-task learning framework is constructed based on the high-order semantic features. Multi-task prediction is performed through the multi-task learning framework to obtain hotspot prediction results and a multi-level network topology model is constructed based on the hotspot prediction results. Edge weights are set according to network transmission constraints, and path exploration is performed in combination with an improved heuristic search algorithm to obtain multiple optimal transmission paths.

[0008] Assign a weight to each optimal transmission path and build a set of candidate paths. Perform traffic distribution on each path in the candidate path set through a weighted polling algorithm. Combined with a distributed load balancing controller based on deep learning, update the weight of each path and detect the network smoothness. If there is local congestion, send a congestion indication signal, establish a content classification table and deploy predetermined high-heat content to the neighboring nodes of the user node. Collect content access data and update the content classification table according to a preset content distribution controller based on an online learning algorithm. Obtain real-time network status and dynamically adjust the data transmission rate through a congestion control algorithm. Combined with a token bucket algorithm to limit the network speed, update the optimal transmission path based on the adjusted network status to obtain a live broadcast push path.

[0009] In an optional embodiment,

[0010] According to the preset sliding time window, data is collected from the private domain live broadcast room to obtain user behavior data and extract features through the time series convolutional network to obtain time series features and add them to the gated recurrent unit. The update gate and the reset gate are combined for information flow, and the long-term dependency features are calculated. Based on the correlation of different positions, feature fusion is performed through the feature fusion network and the fused features are upgraded to a high-dimensional semantic space in combination with the feedforward neural network to obtain a time series feature matrix. A dynamic graph neural network is constructed based on the time series feature matrix and the multi-dimensional similarity between user nodes is calculated through a dynamic edge weight update mechanism. The user nodes are dynamically connected based on the multi-dimensional similarity and the preset similarity threshold to obtain dynamic interest communities including:

[0011] According to the pre-set sliding time window, user behavior data of the private domain live broadcast room is collected, and the user behavior data includes user viewing time, number of interactions, number of likes, and comment content;

[0012] The user behavior data is input into a temporal convolutional network for feature extraction. The temporal convolutional network extracts temporal features from the user behavior data through multiple one-dimensional convolutional layers and pooling layers to obtain temporal features that reflect changes in user behavior over time;

[0013] The time series feature is input into a gated recurrent unit, and the gated recurrent unit controls the information flow through an update gate and a reset gate. The update gate controls the retention degree of the hidden state information at the previous moment, and the reset gate controls the neglect degree of the hidden state information at the previous moment, so as to obtain a long-term dependent feature reflecting the user's long-term interest tendency, and the long-term dependent feature is input into a feature fusion network. The feature fusion network adopts an attention mechanism, and adaptively allocates weights according to the correlation between features at different moments, so as to obtain a fused feature;

[0014] The fused features are processed by a feedforward neural network, wherein the feedforward neural network includes multiple fully connected layers and nonlinear activation functions, and the features are mapped to a high-dimensional semantic space to obtain a time series feature matrix;

[0015] Construct a dynamic graph neural network based on the time series feature matrix, set each user as a node in the graph, calculate the multidimensional similarity between user nodes, the multidimensional similarity includes cosine similarity and Pearson correlation coefficient, and according to a preset similarity threshold, establish a dynamic edge between user nodes whose multidimensional similarity is higher than the similarity threshold, and determine the weight of the dynamic edge according to the multidimensional similarity between user nodes;

[0016] Based on the changes in the user behavior data, the multidimensional similarity between user nodes is continuously updated. When the updated multidimensional similarity is lower than the similarity threshold, the corresponding dynamic edge is deleted. When the updated multidimensional similarity is higher than the similarity threshold, a new dynamic edge is established. A dynamic interest community is constructed through the subgraph formed by the dynamic edge connection.

[0017] In an optional embodiment,

[0018] The weight of the dynamic edge is determined according to the multi-dimensional similarity between user nodes as shown in the following formula:

[0019] ;

[0020] in, EW represents the weight value of the dynamic edge, w k ( t ) indicates the k The similarity index at time t Adaptive weights, similarity k For the k The value of the similarity index,p k Indicates k The power term of similarity indexes is used to control the nonlinear combination relationship of different indexes. t k Indicates k The time decay factor of the similarity index is λ represents the time attenuation coefficient, I ij Indicates user i and users j The interaction weight between tt j Indicates user j The behavior timestamp, α Represents the weight coefficient of multi-view similarity, SV m Indicates m User from the perspective i and users j The similarity.

[0021] In an optional embodiment,

[0022] The dynamic interest community is added to a pre-set multi-layer graph attention network, and the group interaction pattern features are extracted through the spatial attention layer and the channel attention layer and added to the improved bidirectional gated recurrent unit to obtain long-term and short-term temporal dependencies and generate temporal correlation features. The temporal correlation features are added to the deep residual network and high-order semantic features are extracted through residual learning and feature recalibration. A multi-task learning framework is constructed based on the high-order semantic features. Multi-task prediction is performed through the multi-task learning framework to obtain hotspot prediction results and a multi-level network topology model is constructed based on the hotspot prediction results. The edge weights are set according to the network transmission constraints, and the improved heuristic search algorithm is combined to perform path exploration to obtain multiple optimal transmission paths including:

[0023] The dynamic interest community is input into a multi-layer graph attention network composed of a spatial attention layer and a channel attention layer. The spatial attention layer performs weighted aggregation on the node features, and obtains the aggregate node representation by calculating the attention weights between nodes and weighted summing the node features. The channel attention layer obtains the channel-level feature descriptor through global average pooling, and generates the channel importance weight through two fully connected layers and a sigmoid activation function. The channel importance weight is multiplied element by element with the original feature map to obtain a recalibrated feature map, and the group interaction pattern features are extracted;

[0024] The group interaction pattern feature is input into an improved bidirectional gated recurrent unit including a forget gate, an input gate, a time memory gate, a unit state and an output gate, the forgetting degree of historical state information in the group interaction pattern feature is controlled by the forget gate, the current information retention degree is controlled by the input gate, the dependency between time steps is captured by the time memory gate, the long-term and short-term temporal dependency is obtained by combining the dependency between time steps and the output information flow is controlled by the output gate to obtain a forward hidden state and a backward hidden state, and the forward hidden state and the backward hidden state are concatenated to obtain a temporal correlation feature;

[0025] The temporal correlation features are input into a deep residual network comprising a plurality of residual blocks, each of which comprises two convolutional layers and an identity mapping connection, the original features are transmitted through the identity mapping connection, the residual mapping relationship is learned through the convolutional layer, a feature recalibration mechanism is introduced between the residual blocks, the features are adaptively scaled and offset, and high-order semantic features are obtained;

[0026] The high-order semantic features are input into a multi-task learning framework including a shared feature extractor and multiple task predictors, wherein the shared feature extractor extracts a common feature representation, and the multiple task predictors perform community evolution trend prediction, community influence prediction, and community topic popularity prediction based on the common feature representation to obtain a hotspot prediction result;

[0027] A multi-level network topology model is constructed according to the hotspot prediction result, different level nodes are set in the multi-level network topology model to represent different community granularities, and inter-level connections are set to represent the association relationship between communities. Weights are assigned to the edges in the multi-level network topology model according to bandwidth restrictions and latency requirements to obtain a weighted network topology model;

[0028] The weighted network topology model is input into an improved heuristic search algorithm including a multi-objective optimization mechanism and an adaptive strategy adjustment mechanism. The path bandwidth, transmission delay, and node load are balanced through the multi-objective optimization mechanism. The search strategy is adjusted according to the search feedback information through the adaptive strategy adjustment mechanism to obtain multiple optimal transmission paths that meet the network transmission constraints.

[0029] In an optional embodiment,

[0030] The multiple task predictors perform community evolution trend prediction, community influence prediction, and community topic popularity prediction based on the general feature representation, and obtain hotspot prediction results including:

[0031] The universal feature representations output by the shared feature extractor are respectively input into a pre-set community evolution trend predictor, and the long-range temporal dependency features are extracted through the gated time convolution network in the temporal feature branch of the community evolution trend predictor. The community topological structure features are extracted through the graph isomorphism network in the structural feature branch of the community evolution trend predictor. The long-range temporal dependency features and the community topological structure features are input into an adaptive feature fusion module, and the correlation weights between feature channels and the importance weights of feature space positions are calculated. The correlation weights are multiplied by the importance weights to obtain fusion weights, and the fusion weights are weightedly combined with the original features to obtain fusion features. The fusion features are input into a multi-layer perceptron to predict the community evolution trend, and the community evolution trend prediction results are obtained.

[0032] A heterogeneous graph structure is constructed based on a preset community influence predictor. Community members are used as nodes, different types of member interactions are used as edges, and the general feature representation is used as the initial node feature. The attention weights on different types of edges are calculated through a multi-head attention layer. The neighbor node features are weighted aggregated to obtain a member-level representation. The member-level representation is input into a differential graph pooling layer to generate a group-level representation. The group-level representation is propagated through a graph diffusion convolutional network. The propagated features are input into a recursive neural network to predict the community influence evolution sequence, and the community influence prediction result is obtained.

[0033] Based on a preset community topic heat predictor, extract topic text from the common features, preprocess the topic text, construct a historical topic sequence in combination with a time window and extract sentiment words, establish a topic-sentiment word mapping table, encode the topic text through a pre-trained language model to obtain semantic features, map the sentiment words into sentiment vectors according to the topic-sentiment word mapping table, input the sentiment vectors into a deep sentiment analysis network to extract sentiment features, input the historical topic sequence into a causal convolutional network to extract temporal features, calculate the cross-modal attention scores between the semantic features, the sentiment features and the temporal features, perform weighted fusion according to the attention scores, input the fused features into a conditional variational autoencoder to obtain latent variables, reparameterize and sample the latent variables, and input the sampling results into a decoder to generate a diversified topic heat prediction distribution;

[0034] Calculate the correlation matrix and task difficulty vector between each prediction task, score the importance of each prediction task based on the correlation matrix and the difficulty vector, assign weights and learning rates to different prediction tasks according to the importance scores, calculate the angle between the gradient directions of each task, and perform a gradient orthogonalization operation when the angle is less than a preset threshold;

[0035] A task adaptability evaluation function is constructed, and task loss and prediction accuracy are used as reward signals. The policy gradient algorithm is used to optimize the task scheduling strategy, and the task weight distribution scheme is updated. The output results of the predictor are exponentially weighted averaged to obtain the integrated prediction results. The Bayesian model is used to average the uncertainty interval of the prediction results to obtain the hotspot prediction results.

[0036] In an optional embodiment,

[0037] Assign a weight to each optimal transmission path and build a candidate path set, distribute traffic to each path in the candidate path set through a weighted polling algorithm, update the weight of each path in combination with a distributed load balancing controller based on deep learning and detect the smoothness of the network. If there is local congestion, send a congestion indication signal, establish a content classification table and deploy predetermined high-heat content to the neighboring nodes of the user node, collect content access data and update the content classification table according to a pre-set content distribution controller based on an online learning algorithm, obtain real-time network status and dynamically adjust the data transmission rate through a congestion control algorithm, combine the token bucket algorithm to limit the network speed, update the optimal transmission path based on the adjusted network status, and obtain the live push path, including:

[0038] Receiving multiple optimal transmission paths acquired in advance, collecting performance indicator data of the optimal transmission paths, the performance indicator data including bandwidth, delay and packet loss rate, constructing a judgment matrix, calculating the maximum eigenvalue and path eigenvector of the judgment matrix, normalizing the path eigenvector to obtain a path weight coefficient, and constructing a candidate path set based on the optimal transmission paths and the corresponding path weight coefficients;

[0039] Input the candidate path set into the weighted polling scheduler, initialize the accumulated weight variable to zero, initialize the last selected path variable to empty, traverse each path in the candidate path set, add the accumulated weight variable to the path weight coefficient corresponding to the current traversed path, and when the accumulated weight variable is greater than a preset weight threshold, select the current traversed path as the data forwarding path, subtract the preset weight threshold from the accumulated weight variable, and record the current traversed path as the last selected path variable;

[0040] Deploy multiple load balancing controller nodes in the network, input the path state features in the candidate path set into the load balancing controller nodes, the path state features include bandwidth utilization, delay jitter and packet loss rate, output the path weight adjustment value through a convolutional neural network model, the convolutional neural network model includes a feature extraction layer and a fully connected layer, the feature extraction layer uses multiple convolution kernels to extract the spatiotemporal correlation of the path state features, the fully connected layer maps the extracted features to weight adjustment values, and adopts a federated learning method to collaboratively train the convolutional neural network model among multiple controller nodes;

[0041] Constructing a content classification table, prioritizing the content based on the historical access popularity and timeliness of the content, wherein the content classification table includes content identification, access frequency, most recent access time and priority level, monitoring path bandwidth utilization and delay jitter, generating a congestion indication signal when the bandwidth utilization exceeds a first preset threshold or the delay jitter exceeds a second preset threshold, and upon receiving the congestion indication signal, deploying a preset number of contents with the highest priority in the content classification table to neighboring nodes of the user node on the congested path;

[0042] Collecting content request information of each node, the content request information including request time, request content identifier and request status, inputting the content request information into a multi-armed bandit algorithm, performing online learning on content access data, calculating access probability and confidence interval of each content, estimating real-time content popularity according to the access probability and the confidence interval, and dynamically adjusting the priority of the content in the content classification table based on the real-time content popularity;

[0043] Performing status evaluation on each path in the candidate path set, calculating the path delay change rate, when the path delay change rate exceeds a third preset threshold, reducing the transmission rate of the corresponding path according to a preset ratio, maintaining a token bucket at the source node to limit the flow, determining the token bucket size and the token generation rate according to the transmission rate, and caching the data to be sent in a local buffer when the number of tokens in the token bucket is insufficient;

[0044] The path state characteristics and the transmission rate are input into a path evaluation model, the path evaluation model calculates the path comprehensive score by weighted summation, updates the weight coefficients of the paths in the candidate path set based on the path comprehensive score, writes the updated weight coefficients into the candidate path set and recalculates the optimal transmission path to obtain the live broadcast push path.

[0045] In an optional embodiment,

[0046] Maintaining a token bucket at the source node to limit traffic, determining a token bucket size and a token generation rate according to the transmission rate, and caching the data to be sent to a local buffer when the number of tokens in the token bucket is insufficient includes:

[0047] Determine the token bucket size according to the preset maximum transmission rate, the token bucket size is the maximum number of bytes that can be transmitted per unit time corresponding to the maximum transmission rate, set the token generation rate to the maximum transmission rate, initialize the number of tokens in the token bucket to the token bucket size, and record the token update timestamp as the current time;

[0048] When data is sent, obtain the current timestamp, calculate the number of newly generated tokens within the time interval according to the current timestamp and the token update timestamp, the number of newly generated tokens is the product of the token generation rate and the time interval, add the number of newly generated tokens to the number of tokens in the token bucket, the number of tokens in the token bucket does not exceed the token bucket size, and update the token update timestamp to the current timestamp;

[0049] Obtain the data size of the data to be sent, determine whether the number of tokens in the token bucket is greater than or equal to the data size, when the number of tokens in the token bucket is greater than or equal to the data size, subtract the number of tokens equal to the data size from the token bucket and send the data to be sent, when the number of tokens in the token bucket is less than the data size, store the data to be sent in a local buffer;

[0050] Checking whether there is cached data in the local buffer at a preset time interval, when the cached data exists in the local buffer, obtaining the data size of the cached data, judging whether the number of tokens in the token bucket is greater than or equal to the data size of the cached data, and when the number of tokens in the token bucket is greater than or equal to the data size of the cached data, subtracting the number of tokens equal to the data size of the cached data from the token bucket and sending the cached data;

[0051] Monitor the buffer size of the local buffer, and when the buffer size reaches a preset buffer threshold, obtain the priority and waiting time of the cached data in the local buffer, sort the cached data based on the priority and the waiting time, and remove a preset number of cached data at the end of the sorting.

[0052] A second aspect of an embodiment of the present invention provides a hotspot prediction and adaptive routing system for a private domain live broadcast room based on swarm intelligence, comprising:

[0053] The first unit is used to collect data from the private domain live broadcast room according to a preset sliding time window, obtain user behavior data, and extract features through a time series convolutional network, obtain time series features and add them to a gated recurrent unit, combine update gates and reset gates for information flow, calculate long-term dependency features, and based on the correlation of different positions, perform feature fusion through a feature fusion network and combine a feedforward neural network to upgrade the fused features to a high-dimensional semantic space to obtain a time series feature matrix, construct a dynamic graph neural network based on the time series feature matrix, and calculate the multi-dimensional similarity between user nodes through a dynamic edge weight update mechanism, dynamically connect user nodes based on the multi-dimensional similarity and a preset similarity threshold, and obtain a dynamic interest community;

[0054] The second unit is used to add the dynamic interest community to a pre-set multi-layer graph attention network, extract group interaction pattern features through a spatial attention layer and a channel attention layer and add them to an improved bidirectional gated recurrent unit, obtain long-term and short-term temporal dependencies and generate temporal correlation features, add the temporal correlation features to a deep residual network and extract high-order semantic features through residual learning and feature recalibration, construct a multi-task learning framework based on the high-order semantic features, perform multi-task prediction through the multi-task learning framework, obtain hotspot prediction results and construct a multi-level network topology model based on the hotspot prediction results, set edge weights according to network transmission constraints, perform path exploration in combination with an improved heuristic search algorithm, and obtain multiple optimal transmission paths;

[0055] The third unit is used to assign weights to each optimal transmission path and build a set of candidate paths, distribute traffic to each path in the candidate path set through a weighted polling algorithm, update the weight of each path in combination with a distributed load balancing controller based on deep learning and detect the smoothness of the network. If there is local congestion, a congestion indication signal is sent, a content classification table is established, and predetermined high-heat content is deployed to the neighboring nodes of the user node. According to a preset content distribution controller based on an online learning algorithm, content access data is collected and updated. The content classification table is updated, the real-time network status is obtained, and the data transmission rate is dynamically adjusted through a congestion control algorithm. The network speed is limited in combination with a token bucket algorithm, and the optimal transmission path is updated based on the adjusted network status to obtain a live broadcast push path.

[0056] According to a third aspect of the embodiments of the present invention,

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] According to a fourth aspect of the embodiments of the present invention,

[0062] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0063] In the present invention, by combining multiple deep learning models such as temporal convolutional networks, gated recurrent units, dynamic graph neural networks, multi-layer graph attention networks and improved bidirectional gated recurrent units, it is possible to more effectively capture the temporal features, long-term dependencies, group interaction patterns and high-order semantic features in user behavior data, thereby achieving accurate prediction of hot spots in private domain live broadcast rooms, and significantly improving the accuracy of hot spot prediction. A multi-level network topology model is constructed based on the hot spot prediction results, and an improved heuristic search algorithm, a weighted polling algorithm and a distributed load balancing controller based on deep learning are combined to perform path exploration and traffic distribution, thereby achieving adaptive routing and load balancing, effectively avoiding network congestion, improving data transmission efficiency, and ensuring the stability and smoothness of live broadcast push. The congestion control mechanism and content classification strategy can monitor the network status in real time, and dynamically adjust the data transmission rate through the congestion control algorithm, and combine the token bucket algorithm to perform network speed limiting, effectively reducing the network load and improving the speed and efficiency of user access to content. In summary, this embodiment achieves the high efficiency and stability of data transmission in private domain live broadcast rooms, and improves user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of a method for hotspot prediction and adaptive routing in a private domain live broadcast room based on swarm intelligence according to an embodiment of the present invention;

[0065] Figure 2 This is a structural diagram of a private domain live broadcast room hotspot prediction and adaptive routing system based on swarm intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0068] Figure 1 FIG. 1 is a flow chart of a method for hotspot prediction and adaptive routing in a private domain live broadcast room based on swarm intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] S1. Collect data from the private domain live broadcast room according to the preset sliding time window, obtain user behavior data, extract features through the time series convolutional network, obtain time series features and add them to the gated recurrent unit, combine the update gate and the reset gate for information flow, calculate the long-term dependency features, and based on the correlation of different positions, perform feature fusion through the feature fusion network and combine the feedforward neural network to upgrade the fused features to a high-dimensional semantic space to obtain a time series feature matrix. Based on the time series feature matrix, a dynamic graph neural network is constructed and the multi-dimensional similarity between user nodes is calculated through a dynamic edge weight update mechanism. Based on the multi-dimensional similarity and the preset similarity threshold, the user nodes are dynamically connected to obtain a dynamic interest community;

[0070] The information flow refers to the process of information propagation and transmission in a network or system, and is usually used to describe the interaction and transmission of data or signals between different nodes or regions. It is widely used in fields such as communications and social network analysis. The dynamic graph neural network is a neural network model that specifically processes dynamically changing graph structures. It can learn and infer on graph data whose nodes or edges change over time. It is suitable for fields such as time-changing social networks and transportation networks. The dynamic edge weight update mechanism means that in a dynamic graph neural network, as time passes and new information is input, the weight of the edge will be dynamically adjusted according to specific rules to reflect the intensity or importance of the interaction between nodes, helping the model to better capture temporal relationships and changes in the graph. The dynamic interest community refers to a community structure formed by the dynamic changes in user interests and behaviors over time in a social network or information dissemination system, and is usually used to analyze the evolution of individual interests and the evolution of the community.

[0071] In an optional embodiment,

[0072] According to the preset sliding time window, data is collected from the private domain live broadcast room to obtain user behavior data and extract features through the time series convolutional network to obtain time series features and add them to the gated recurrent unit. The update gate and the reset gate are combined for information flow, and the long-term dependency features are calculated. Based on the correlation of different positions, feature fusion is performed through the feature fusion network and the fused features are upgraded to a high-dimensional semantic space in combination with the feedforward neural network to obtain a time series feature matrix. A dynamic graph neural network is constructed based on the time series feature matrix and the multi-dimensional similarity between user nodes is calculated through a dynamic edge weight update mechanism. The user nodes are dynamically connected based on the multi-dimensional similarity and the preset similarity threshold to obtain dynamic interest communities including:

[0073] According to the pre-set sliding time window, user behavior data of the private domain live broadcast room is collected, and the user behavior data includes user viewing time, number of interactions, number of likes, and comment content;

[0074] The user behavior data is input into a temporal convolutional network for feature extraction. The temporal convolutional network extracts temporal features from the user behavior data through multiple one-dimensional convolutional layers and pooling layers to obtain temporal features that reflect changes in user behavior over time;

[0075] The time series feature is input into a gated recurrent unit, and the gated recurrent unit controls the information flow through an update gate and a reset gate. The update gate controls the retention degree of the hidden state information at the previous moment, and the reset gate controls the neglect degree of the hidden state information at the previous moment, so as to obtain a long-term dependent feature reflecting the user's long-term interest tendency, and the long-term dependent feature is input into a feature fusion network. The feature fusion network adopts an attention mechanism, and adaptively allocates weights according to the correlation between features at different moments, so as to obtain a fused feature;

[0076] The fused features are processed by a feedforward neural network, wherein the feedforward neural network includes multiple fully connected layers and nonlinear activation functions, and the features are mapped to a high-dimensional semantic space to obtain a time series feature matrix;

[0077] Construct a dynamic graph neural network based on the time series feature matrix, set each user as a node in the graph, calculate the multidimensional similarity between user nodes, the multidimensional similarity includes cosine similarity and Pearson correlation coefficient, and according to a preset similarity threshold, establish a dynamic edge between user nodes whose multidimensional similarity is higher than the similarity threshold, and determine the weight of the dynamic edge according to the multidimensional similarity between user nodes;

[0078] Based on the changes in the user behavior data, the multidimensional similarity between user nodes is continuously updated. When the updated multidimensional similarity is lower than the similarity threshold, the corresponding dynamic edge is deleted. When the updated multidimensional similarity is higher than the similarity threshold, a new dynamic edge is established. A dynamic interest community is constructed through the subgraph formed by the dynamic edge connection.

[0079] The long-term dependency feature refers to the fact that in time series data, the state at the current moment is affected by the state at a more distant moment. Capturing this long-term dependency is crucial to the prediction accuracy of the model, and is especially widely used in natural language processing and time series analysis. The dimensionality increase process refers to the process of mapping low-dimensional data to high-dimensional space, usually through mathematical transformation or feature expansion to increase the representation ability of the data, and is often used in machine learning algorithms such as support vector machines to better segment data categories.

[0080] Collect user behavior data. Set a sliding time window, such as 30 minutes, and collect live room user data every 5 minutes. The collected data includes user viewing time, number of interactions (such as the number of bullet comments sent), number of likes, and comment content. Suppose user A watched for 3 minutes in the first 5 minutes, sent 2 bullet comments, liked 1 time, and commented "the content is wonderful".

[0081] The collected user behavior data will be input into the time series convolutional network for feature extraction. The network consists of multiple one-dimensional convolutional layers and pooling layers. For example, the first convolutional layer uses 32 convolution kernels, the convolution kernel size is 3, and the step size is 1; the second convolutional layer uses 64 convolution kernels, the convolution kernel size is 3, and the step size is 1; each convolutional layer is followed by a maximum pooling layer, the pooling window size is 2, and the step size is 2. Through convolution and pooling operations, the time series features in the user behavior data can be extracted, such as the changing trend of the user's viewing time, the change of the interaction frequency, etc. Assuming that the viewing time of user A in four consecutive time windows is 3 minutes, 4 minutes, 3 minutes, and 2 minutes respectively, after passing through the time series convolutional network, a feature vector can be extracted, such as [0.2, 0.3, 0.2, 0.1], reflecting the trend that the user's viewing time first increases and then decreases.

[0082] The extracted time series features will be input into the gated recurrent unit (GRU). GRU controls the flow of information through the update gate and the reset gate. The update gate determines how much hidden state information from the previous moment is retained, and the reset gate determines how much hidden state information from the previous moment is ignored. Through GRU, long-term dependencies in user behavior data can be captured, such as a user's continued attention to a specific topic. Assuming that user A frequently comments on content related to "beauty" in the past period of time, GRU can learn user A's long-term interest in beauty.

[0083] The long-term dependency features obtained from the GRU will be input into the feature fusion network. The network uses an attention mechanism to adaptively assign weights based on the correlation between features at different times. For example, if user A has a high viewing time and number of interactions in a certain time period, the features of that time period will be given a higher weight.

[0084] The fused features will be input into the feedforward neural network for dimensionality increase. The network consists of multiple fully connected layers and nonlinear activation functions (such as ReLU). Through the feedforward neural network, the features can be mapped to a high-dimensional semantic space to obtain a temporal feature matrix. For example, a feature vector of user A [0.2, 0.3, 0.2, 0.1] may be mapped to a higher-dimensional vector [0.15, 0.25, 0.20, 0.10, 0.05, 0.25] after passing through the feedforward neural network.

[0085] Construct a dynamic graph neural network based on the obtained time series feature matrix. Set each user as a node in the graph. Calculate the multidimensional similarity between user nodes, including cosine similarity and Pearson correlation coefficient. Preset a similarity threshold, such as 0.8. If the multidimensional similarity between two user nodes is higher than the threshold, a dynamic edge is established between the two nodes, and the weight of the edge is determined by the multidimensional similarity. For example, the multidimensional similarity between user A and user B is 0.85, which is higher than the threshold 0.8, then an edge is established between A and B with a weight of 0.85.

[0086] As user behavior data changes, the multidimensional similarity between user nodes is continuously updated. If the updated multidimensional similarity is lower than the threshold, the corresponding dynamic edge is deleted. If the updated multidimensional similarity is higher than the threshold, a new dynamic edge is established. The subgraph formed by dynamic edge connection is a dynamic interest community. For example, if the similarity between user A and user B decreases to 0.7, which is lower than the threshold of 0.8, the edge between A and B is deleted; if the similarity between user A and user C rises to 0.82, which is higher than the threshold of 0.8, the edge between A and C is established.

[0087] In this embodiment, through multi-dimensional similarity and dynamic graph neural networks, the interest associations between users can be more accurately identified, thereby constructing more accurate interest communities. The connection relationships between user nodes are dynamically updated according to changes in user behavior data, so that the constructed interest communities can reflect changes in user interests in real time. It comprehensively considers multi-dimensional data such as user viewing time, number of interactions, number of likes, and comment content, and can more comprehensively characterize user interest characteristics. In summary, this embodiment realizes the dynamic construction and update of user interest communities, improves the accuracy and real-time performance of user interest analysis, and provides support for personalized recommendations and precision marketing.

[0088] In an optional embodiment,

[0089] The weight of the dynamic edge is determined according to the multi-dimensional similarity between user nodes as shown in the following formula:

[0090] ;

[0091] in, EW represents the weight value of the dynamic edge, w k ( t ) indicates the k The similarity index at time t Adaptive weights, similarity k For the k The value of the similarity index, p k Indicates k The power term of similarity indexes is used to control the nonlinear combination relationship of different indexes. t k Indicates k The time decay factor of the similarity index is λ represents the time attenuation coefficient, I ij Indicates user i and users j The interaction weight between tt j Indicates user j The behavior timestamp, α Represents the weight coefficient of multi-view similarity, SV m Indicates m User from the perspective i and users j The similarity.

[0092] This embodiment can more accurately characterize the relationship strength between users by comprehensively considering multiple similarity indicators and time decay factors, thereby improving the accuracy of applications such as user relationship prediction and recommendation systems. The adaptive weight and time decay mechanism enables the dynamic edge weight to be dynamically adjusted as user behavior changes, thereby better reflecting the dynamic changes of user relationships, showing the impact of different similarity indicators on user relationships and the impact of time decay on user relationships, and improving the interpretability of user relationship analysis. In summary, this embodiment can more accurately characterize the association strength between user nodes and improve the accuracy and robustness of the construction of dynamic interest communities.

[0093] S2. Add the dynamic interest community to a pre-set multi-layer graph attention network, extract group interaction pattern features through the spatial attention layer and the channel attention layer and add them to the improved bidirectional gated recurrent unit to obtain long-term and short-term temporal dependencies and generate temporal correlation features, add the temporal correlation features to the deep residual network and extract high-order semantic features through residual learning and feature recalibration, build a multi-task learning framework based on the high-order semantic features, perform multi-task prediction through the multi-task learning framework, obtain hotspot prediction results and build a multi-level network topology model based on the hotspot prediction results, set edge weights according to network transmission constraints, and perform path exploration in combination with the improved heuristic search algorithm to obtain multiple optimal transmission paths;

[0094] The group interaction pattern feature refers to the collective behavior pattern or law formed by the interaction between individuals in a multi-agent system or social network. It is usually used to reveal the impact of individual behavior on the overall pattern of the group, such as crowd gathering, information dissemination, etc. The bidirectional gated recurrent unit is an improved recurrent neural network structure that combines forward and reverse information flows and can simultaneously consider the past and future information of the sequence. It is particularly suitable for tasks that need to capture long-term dependencies, such as natural language processing. The feature recalibration refers to standardizing or adjusting the features so that they operate within a specific range or scale. Common recalibration methods include normalization and standardization to improve the efficiency and stability of model training. The multi-task learning framework is a machine learning method that simultaneously learns multiple tasks in the same model to share the learned features and improve the performance of the model on each task. It is suitable for multiple related learning tasks. The multi-level network topology model is a model that describes a complex network structure, in which network nodes are organized at multiple levels with different connections and interactions between levels. It is often used to analyze hierarchical structures in complex systems, such as social networks, ecological networks, etc. The heuristic search algorithm is a type of algorithm used to find approximate optimal solutions in the problem solution space. It guides the search process by applying heuristic strategies and is often used to solve NP-hard problems, such as the traveling salesman problem and the eight queens problem.

[0095] In an optional embodiment,

[0096] The dynamic interest community is added to a pre-set multi-layer graph attention network, and the group interaction pattern features are extracted through the spatial attention layer and the channel attention layer and added to the improved bidirectional gated recurrent unit to obtain long-term and short-term temporal dependencies and generate temporal correlation features. The temporal correlation features are added to the deep residual network and high-order semantic features are extracted through residual learning and feature recalibration. A multi-task learning framework is constructed based on the high-order semantic features. Multi-task prediction is performed through the multi-task learning framework to obtain hotspot prediction results and a multi-level network topology model is constructed based on the hotspot prediction results. The edge weights are set according to the network transmission constraints, and the improved heuristic search algorithm is combined to perform path exploration to obtain multiple optimal transmission paths including:

[0097] The dynamic interest community is input into a multi-layer graph attention network composed of a spatial attention layer and a channel attention layer. The spatial attention layer performs weighted aggregation on the node features, and obtains the aggregate node representation by calculating the attention weights between nodes and weighted summing the node features. The channel attention layer obtains the channel-level feature descriptor through global average pooling, and generates the channel importance weight through two fully connected layers and a sigmoid activation function. The channel importance weight is multiplied element by element with the original feature map to obtain a recalibrated feature map, and the group interaction pattern features are extracted;

[0098] The group interaction pattern feature is input into an improved bidirectional gated recurrent unit including a forget gate, an input gate, a time memory gate, a unit state and an output gate, the forgetting degree of historical state information in the group interaction pattern feature is controlled by the forget gate, the current information retention degree is controlled by the input gate, the dependency between time steps is captured by the time memory gate, the long-term and short-term temporal dependency is obtained by combining the dependency between time steps and the output information flow is controlled by the output gate to obtain a forward hidden state and a backward hidden state, and the forward hidden state and the backward hidden state are concatenated to obtain a temporal correlation feature;

[0099] The temporal correlation features are input into a deep residual network comprising a plurality of residual blocks, each of which comprises two convolutional layers and an identity mapping connection, the original features are transmitted through the identity mapping connection, the residual mapping relationship is learned through the convolutional layer, a feature recalibration mechanism is introduced between the residual blocks, the features are adaptively scaled and offset, and high-order semantic features are obtained;

[0100] The high-order semantic features are input into a multi-task learning framework including a shared feature extractor and multiple task predictors, wherein the shared feature extractor extracts a common feature representation, and the multiple task predictors perform community evolution trend prediction, community influence prediction, and community topic popularity prediction based on the common feature representation to obtain a hotspot prediction result;

[0101] A multi-level network topology model is constructed according to the hotspot prediction result, different level nodes are set in the multi-level network topology model to represent different community granularities, and inter-level connections are set to represent the association relationship between communities. Weights are assigned to the edges in the multi-level network topology model according to bandwidth restrictions and latency requirements to obtain a weighted network topology model;

[0102] The weighted network topology model is input into an improved heuristic search algorithm including a multi-objective optimization mechanism and an adaptive strategy adjustment mechanism. The path bandwidth, transmission delay, and node load are balanced through the multi-objective optimization mechanism. The search strategy is adjusted according to the search feedback information through the adaptive strategy adjustment mechanism to obtain multiple optimal transmission paths that meet the network transmission constraints.

[0103] The channel-level feature descriptor refers to the description or encoding of the features of each channel in a convolutional neural network to enhance the model's ability to capture information from different channels in an image or signal, and is usually used to improve the accuracy of feature extraction. The dependency between time steps refers to the fact that when processing time series data, the state of the current time step not only depends on the state of the previous time step, but may also be related to the state of multiple time steps before. It is usually modeled by a recursive neural network or an attention mechanism. The identity mapping connection is a commonly used connection method in a neural network. It avoids the gradient vanishing problem by directly passing the input signal to the output, and helps retain input information in a deep network. It is commonly found in residual networks. The community granularity refers to the scale for defining the size of a community in a social network or community analysis. Different granularities determine the level of detail in community division. Finer granularity can reveal finer community structures. The weight assignment refers to the assignment of numerical weights to each parameter or connection in the network in a machine learning model. The reasonable assignment of weights directly affects the learning effect of the model and the final prediction accuracy.

[0104] A multi-layer graph attention network for dynamic interest communities is constructed. Each dynamic interest community is regarded as a node in the graph. The feature vector of the node can contain information such as the number of community members, activity, and topic tags. For example, the feature vector of a game interest community can be expressed as [10000, 0.8, "role playing", "strategy"], where 10000 represents the number of community members, 0.8 represents the community activity, and "role playing" and "strategy" represent the community topic tags. The spatial attention layer performs weighted aggregation on node features by calculating the attention weights between nodes. For example, the attention weights between two game communities can be obtained by calculating the similarity of topic tags. The channel attention layer obtains channel-level feature descriptors through global average pooling, and then generates channel importance weights through two fully connected layers and sigmoid activation functions, and multiplies the weights by the original feature map element by element to obtain the recalibrated feature map, thereby extracting group interaction pattern features.

[0105] The extracted group interaction pattern features are input into the improved bidirectional gated recurrent unit. In units of time steps, the forget gate controls the degree of forgetting the information of the previous time step, the input gate controls the degree of retention of the information of the current time step, the temporal memory gate captures the dependencies between time steps, and the output gate controls the flow of output information. For example, assuming that the group interaction pattern feature of the current time step is the increase in the heat of community discussions, the forget gate may reduce the attention paid to the previous discussion heat, the input gate will retain the information of the current heat increase, the temporal memory gate will capture the trend of heat change, and the output gate will output the final timing information. The hidden states of the forward and backward gated recurrent units are concatenated to obtain the temporal correlation features containing long-term and short-term temporal dependencies.

[0106] The temporal correlation features are input into the deep residual network, which contains multiple residual blocks. Each residual block contains two convolutional layers and an identity mapping connection. The convolutional layer learns the residual mapping relationship, and the identity mapping connection transmits the original features. A feature recalibration mechanism is introduced between the residual blocks to adaptively scale and offset the features. For example, the input feature of a residual block is [0.2, 0.5, 0.8]. After being processed by the convolutional layer, the residual mapping [0.1, 0.2, 0.3] is obtained. The residual mapping is added to the original feature to obtain [0.3, 0.7, 1.1]. The feature is then recalibrated, for example, the feature value is limited to between 0 and 1, and finally the high-order semantic feature [0.3, 0.7, 1.0] is obtained.

[0107] The extracted high-order semantic features are input into the multi-task learning framework, where the multi-task learning framework includes a shared feature extractor and multiple task predictors. The shared feature extractor extracts common feature representations, and the multiple task predictors predict community evolution trends, community influence, and community topic popularity based on the common feature representations to obtain hotspot prediction results. For example, the common features extracted by the shared feature extractor are the activity and topic relevance of the community, the community evolution trend predictor predicts the growth trend of the community size, the community influence predictor predicts the degree of influence of the community on other communities, and the community topic popularity predictor predicts the popularity of topics in the community in the future.

[0108] A multi-level network topology model is constructed based on the hotspot prediction results. Nodes at different levels represent different community granularities, and connections between levels represent the association between communities. Weights are assigned to edges based on bandwidth limitations and latency requirements. For example, the connection weight between two communities can be determined based on the frequency of information exchange and the amount of data transmission between them.

[0109] The weighted network topology model is input into the improved heuristic search algorithm. The heuristic search algorithm contains a multi-objective optimization mechanism and an adaptive strategy adjustment mechanism to balance path bandwidth, transmission delay, and node load. For example, when searching for a path, the algorithm will give priority to paths with sufficient bandwidth, low delay, and light load. Through the adaptive strategy adjustment mechanism, the search strategy is adjusted according to the search feedback information, and finally multiple optimal transmission paths that meet the network transmission constraints are obtained.

[0110] In this embodiment, through the multi-layer graph attention network and the improved bidirectional gated recurrent unit, the community interaction patterns and temporal dependencies can be captured more effectively, thereby improving the accuracy of hotspot prediction. The multi-task learning framework can perform multiple prediction tasks at the same time, which can significantly improve the prediction efficiency compared to performing each task separately. The improved heuristic search algorithm can find the optimal transmission path that meets the network transmission constraints, thereby optimizing network resource utilization and reducing transmission costs. In summary, this embodiment can comprehensively characterize the dynamic characteristics and evolution laws of the community, realize the intelligent scheduling and optimal allocation of network resources, improve the accuracy of community analysis and network transmission efficiency, and provide reliable technical support for the operational decision-making and resource scheduling of private domain live broadcast rooms.

[0111] In an optional embodiment,

[0112] The multiple task predictors perform community evolution trend prediction, community influence prediction, and community topic popularity prediction based on the general feature representation, and obtain hotspot prediction results including:

[0113] The universal feature representations output by the shared feature extractor are respectively input into a pre-set community evolution trend predictor, and the long-range temporal dependency features are extracted through the gated time convolution network in the temporal feature branch of the community evolution trend predictor. The community topological structure features are extracted through the graph isomorphism network in the structural feature branch of the community evolution trend predictor. The long-range temporal dependency features and the community topological structure features are input into an adaptive feature fusion module, and the correlation weights between feature channels and the importance weights of feature space positions are calculated. The correlation weights are multiplied by the importance weights to obtain fusion weights, and the fusion weights are weightedly combined with the original features to obtain fusion features. The fusion features are input into a multi-layer perceptron to predict the community evolution trend, and the community evolution trend prediction results are obtained.

[0114] A heterogeneous graph structure is constructed based on a preset community influence predictor. Community members are used as nodes, different types of member interactions are used as edges, and the general feature representation is used as the initial node feature. The attention weights on different types of edges are calculated through a multi-head attention layer. The neighbor node features are weighted aggregated to obtain a member-level representation. The member-level representation is input into a differential graph pooling layer to generate a group-level representation. The group-level representation is propagated through a graph diffusion convolutional network. The propagated features are input into a recursive neural network to predict the community influence evolution sequence, and the community influence prediction result is obtained.

[0115] Based on a preset community topic heat predictor, extract topic text from the common features, preprocess the topic text, construct a historical topic sequence in combination with a time window and extract sentiment words, establish a topic-sentiment word mapping table, encode the topic text through a pre-trained language model to obtain semantic features, map the sentiment words into sentiment vectors according to the topic-sentiment word mapping table, input the sentiment vectors into a deep sentiment analysis network to extract sentiment features, input the historical topic sequence into a causal convolutional network to extract temporal features, calculate the cross-modal attention scores between the semantic features, the sentiment features and the temporal features, perform weighted fusion according to the attention scores, input the fused features into a conditional variational autoencoder to obtain latent variables, reparameterize and sample the latent variables, and input the sampling results into a decoder to generate a diversified topic heat prediction distribution;

[0116] Calculate the correlation matrix and task difficulty vector between each prediction task, score the importance of each prediction task based on the correlation matrix and the difficulty vector, assign weights and learning rates to different prediction tasks according to the importance scores, calculate the angle between the gradient directions of each task, and perform a gradient orthogonalization operation when the angle is less than a preset threshold;

[0117] A task adaptability evaluation function is constructed, and task loss and prediction accuracy are used as reward signals. The policy gradient algorithm is used to optimize the task scheduling strategy, and the task weight distribution scheme is updated. The output results of the predictor are exponentially weighted averaged to obtain the integrated prediction results. The Bayesian model is used to average the uncertainty interval of the prediction results to obtain the hotspot prediction results.

[0118] The topic-sentiment word mapping table refers to a table that associates sentiment words with specific topics or themes, and is used in sentiment analysis tasks to determine the sentiment tendency of the text by identifying sentiment words and their related topics in the text. The conditional variational autoencoder is a generative model that generates new data samples by learning the implicit distribution of data and combining conditional information (such as labels or context). It is widely used in fields such as image generation and speech synthesis. The reparameterized sampling is a technology in the variational autoencoder. By converting the sampling process into a differentiable operation, the gradient can be propagated through the network, so that training can be performed through standard optimization methods. The gradient orthogonalization operation refers to adjusting the gradient during the neural network training process so that it has better independence when updating parameters, avoiding gradient explosion or gradient vanishing problems, and improving the stability and convergence speed of network training.

[0119] Obtain historical data of the target community, including community member information, interactive relationships between members, topics discussed in the community, topic publishing time, and topic popularity value, etc. For example, collect data from an online forum over the past year, including forum user IDs, replies and likes between users, forum post content, posting time, and post views and likes, etc.

[0120] Three different predictors are used for prediction. The first one is the community evolution trend predictor, which receives the universal feature representation output by the shared feature extractor and uses two branches to extract temporal features and structural features respectively. The temporal feature branch uses a gated temporal convolutional network to capture the pattern of community changes over time. The structural feature branch uses a graph isomorphism network to learn the topological structure of the relationship between community members. An adaptive feature fusion module performs a weighted combination of the two features based on the correlation between feature channels and the importance of feature spatial positions to obtain fused features. The fused features are input into a multi-layer perceptron to predict the future evolution trend of the community, such as changes in community size and activity.

[0121] The community influence predictor uses community members as nodes and different types of member interactions as edges to construct a heterogeneous graph. The universal feature representation is used as the initial node feature, and the multi-head attention mechanism is used to calculate the attention weights on different types of edges. The neighbor node features are weighted aggregated to obtain the member-level representation. The differential graph pooling layer is used to aggregate the member-level representation into the group-level representation. The graph diffusion convolutional network is used to propagate the features of the group-level representation. The propagated features are input into the recurrent neural network to predict the evolution sequence of community influence, such as the change trend of the overall influence score of the community over time.

[0122] The community topic heat predictor extracts topic text from common features and performs preprocessing, such as removing stop words and punctuation marks, combining time windows to construct historical topic sequences, and extracting sentiment words. A topic-sentiment word mapping table is established, and the topic text is encoded through a pre-trained language model to obtain semantic features, and sentiment words are mapped to sentiment vectors according to the mapping table. The sentiment features are extracted using a deep sentiment analysis network, and the temporal features of the historical topic sequence are extracted using a causal convolutional network. The cross-modal attention scores between semantic features, sentiment features, and temporal features are calculated, and the three features are weighted and fused according to the scores. The fused features are input into the encoder of the conditional variational autoencoder to obtain latent variables. The latent variables are reparameterized and sampled, and the sampling results are input into the decoder to generate a diverse distribution of topic heat predictions, such as predicting the probability distribution of topic heat in the future.

[0123] While performing the three prediction tasks, the correlation matrix and task difficulty vector between each prediction task are calculated, each prediction task is scored for importance, and weights and learning rates are assigned to different tasks based on the scores. In addition, the angle between the gradient directions of each task is calculated. When the angle is less than the preset threshold, the gradient orthogonalization operation is performed to avoid mutual interference between tasks.

[0124] Construct a task adaptability evaluation function, use task loss and prediction accuracy as reward signals, use the policy gradient algorithm to optimize the task scheduling strategy, update the task weight distribution scheme, perform exponential weighted average on the output results of the three predictors, and obtain the integrated prediction result. And use the Bayesian model to average and calculate the uncertainty interval of the prediction result. For example, predicting that the popularity value of a topic in the next week is 10,000, the uncertainty interval is [9,000, 11,000].

[0125] In this embodiment, by integrating the results of multiple prediction tasks and considering the correlation and difficulty between tasks, the accuracy of hotspot prediction can be improved. By analyzing the contribution and uncertainty of different prediction tasks, the mechanism of hotspot formation can be better understood. By dynamically adjusting task weights and learning rates, the model can be better adapted to different communities and topics. In summary, this embodiment realizes multi-dimensional analysis and prediction of community development trends, provides reliable prediction results with uncertainty assessment, and provides comprehensive data support for community operation decisions.

[0126] S3. Assign a weight to each optimal transmission path and build a set of candidate paths, distribute traffic to each path in the candidate path set through a weighted polling algorithm, update the weight of each path in combination with a distributed load balancing controller based on deep learning and detect the network smoothness. If there is local congestion, send a congestion indication signal, establish a content classification table and deploy predetermined high-heat content to the neighboring nodes of the user node, collect content access data and update the content classification table according to a pre-set content distribution controller based on an online learning algorithm, obtain real-time network status and dynamically adjust the data transmission rate through a congestion control algorithm, combine the token bucket algorithm to limit the network speed, update the optimal transmission path based on the adjusted network status, and obtain the live broadcast push path.

[0127] The weighted polling algorithm is a scheduling algorithm commonly used for load balancing. By assigning different weights to each request and determining the allocation of requests based on the weight of the node, it can achieve reasonable allocation of resources and is suitable for handling uneven loads. The distributed load balancing controller is a controller that achieves load balancing through the coordination of multiple nodes. It is usually used in large-scale distributed systems and can dynamically adjust the workload of each node to avoid overloading some nodes. The local congestion refers to the congestion that occurs in a certain part of the network or system (such as switches, routers, nodes, etc.), which may cause some requests to be delayed or lost, and requires local policy The content classification table refers to a table that classifies different types of data or content in the system according to certain standards. It is commonly found in content distribution networks or cache systems and is used to optimize data access and scheduling. The congestion control algorithm refers to an algorithm used to prevent or alleviate network congestion. It ensures that network resources are effectively utilized by adjusting parameters such as data flow, window size, and transmission rate. Common algorithms include TCP congestion control, etc. The token bucket algorithm is an algorithm used to control network traffic. It controls the sending rate of data packets by maintaining a token bucket. The token generation rate represents the allowed transmission rate and is often used for traffic shaping and speed limiting.

[0128] In an optional embodiment,

[0129] Assign a weight to each optimal transmission path and build a candidate path set, distribute traffic to each path in the candidate path set through a weighted polling algorithm, update the weight of each path in combination with a distributed load balancing controller based on deep learning and detect the smoothness of the network. If there is local congestion, send a congestion indication signal, establish a content classification table and deploy predetermined high-heat content to the neighboring nodes of the user node, collect content access data and update the content classification table according to a pre-set content distribution controller based on an online learning algorithm, obtain real-time network status and dynamically adjust the data transmission rate through a congestion control algorithm, combine the token bucket algorithm to limit the network speed, update the optimal transmission path based on the adjusted network status, and obtain the live push path, including:

[0130] Receiving multiple optimal transmission paths acquired in advance, collecting performance indicator data of the optimal transmission paths, the performance indicator data including bandwidth, delay and packet loss rate, constructing a judgment matrix, calculating the maximum eigenvalue and path eigenvector of the judgment matrix, normalizing the path eigenvector to obtain a path weight coefficient, and constructing a candidate path set based on the optimal transmission paths and the corresponding path weight coefficients;

[0131] Input the candidate path set into the weighted polling scheduler, initialize the accumulated weight variable to zero, initialize the last selected path variable to empty, traverse each path in the candidate path set, add the accumulated weight variable to the path weight coefficient corresponding to the current traversed path, and when the accumulated weight variable is greater than a preset weight threshold, select the current traversed path as the data forwarding path, subtract the preset weight threshold from the accumulated weight variable, and record the current traversed path as the last selected path variable;

[0132] Deploy multiple load balancing controller nodes in the network, input the path state features in the candidate path set into the load balancing controller nodes, the path state features include bandwidth utilization, delay jitter and packet loss rate, output the path weight adjustment value through a convolutional neural network model, the convolutional neural network model includes a feature extraction layer and a fully connected layer, the feature extraction layer uses multiple convolution kernels to extract the spatiotemporal correlation of the path state features, the fully connected layer maps the extracted features to weight adjustment values, and adopts a federated learning method to collaboratively train the convolutional neural network model among multiple controller nodes;

[0133] Constructing a content classification table, prioritizing the content based on the historical access popularity and timeliness of the content, wherein the content classification table includes content identification, access frequency, most recent access time and priority level, monitoring path bandwidth utilization and delay jitter, generating a congestion indication signal when the bandwidth utilization exceeds a first preset threshold or the delay jitter exceeds a second preset threshold, and upon receiving the congestion indication signal, deploying a preset number of contents with the highest priority in the content classification table to neighboring nodes of the user node on the congested path;

[0134] Collecting content request information of each node, the content request information including request time, request content identifier and request status, inputting the content request information into a multi-armed bandit algorithm, performing online learning on content access data, calculating access probability and confidence interval of each content, estimating real-time content popularity according to the access probability and the confidence interval, and dynamically adjusting the priority of the content in the content classification table based on the real-time content popularity;

[0135] Performing status evaluation on each path in the candidate path set, calculating the path delay change rate, when the path delay change rate exceeds a third preset threshold, reducing the transmission rate of the corresponding path according to a preset ratio, maintaining a token bucket at the source node to limit the flow, determining the token bucket size and the token generation rate according to the transmission rate, and caching the data to be sent in a local buffer when the number of tokens in the token bucket is insufficient;

[0136] The path state characteristics and the transmission rate are input into a path evaluation model, the path evaluation model calculates the path comprehensive score by weighted summation, updates the weight coefficients of the paths in the candidate path set based on the path comprehensive score, writes the updated weight coefficients into the candidate path set and recalculates the optimal transmission path to obtain the live broadcast push path.

[0137] The judgment matrix is ​​a matrix used in multi-criteria decision analysis. By comparing the relative importance of different factors, the matrix is ​​constructed and consistency is checked. It is usually used in decision models such as the analytic hierarchy process (AHP). The multi-armed bandit algorithm is a reinforcement learning algorithm that simulates multiple choices of a bandit. The goal is to maximize long-term benefits by trying different options. It is often used for the balance problem of exploration and utilization. The source node refers to the starting node for sending data in the network. It usually serves as the starting point of data transmission and is responsible for the generation or sending of data.

[0138] Obtain multiple candidate optimal transmission paths and collect performance indicator data, including bandwidth, delay, and packet loss rate. For example, the bandwidth of path 1 is 100Mbps, the delay is 20ms, and the packet loss rate is 0.1%; the bandwidth of path 2 is 50Mbps, the delay is 30ms, and the packet loss rate is 0.5%; the bandwidth of path 3 is 75Mbps, the delay is 25ms, and the packet loss rate is 0.2%.

[0139] Construct a judgment matrix to compare the performance of the paths. The elements of this matrix represent the relative performance of the two paths. For example, if the performance of path 1 is better than that of path 2, the value of the corresponding element in the matrix is ​​greater than 1. Based on the collected performance indicator data, a 3x3 matrix can be constructed. By calculating the maximum eigenvalue and corresponding eigenvector of the judgment matrix, the weight coefficient of each path can be obtained. After normalizing the eigenvector, for example, the weight coefficient of path 1 is 0.5, the weight coefficient of path 2 is 0.2, and the weight coefficient of path 3 is 0.3. These paths and their corresponding weight coefficients constitute the candidate path set.

[0140] Use a weighted round-robin scheduler to distribute traffic to the paths in the candidate path set. Initialize a cumulative weight variable to zero and record the last selected path. Traverse the candidate path set and add the cumulative weight variable to the weight coefficient of the current path. Assume that the preset weight threshold is 1. When traversing to path 1, the cumulative weight is 0.5, which is less than threshold 1, and path 1 is not selected. Continue to traverse to path 2, the cumulative weight becomes 0.7, which is still less than threshold 1, and path 2 is not selected. Continue to traverse to path 3, the cumulative weight becomes 1, reaching threshold 1, select path 3 as the data forwarding path, and subtract threshold 1 from the cumulative weight to become 0. Record path 3 as the last selected path. Subsequent data packets will continue to select paths according to this method.

[0141] In order to dynamically adjust the path weight, multiple load balancing controller nodes are deployed in the network. The state features of the candidate paths are collected, including bandwidth utilization, delay jitter, and packet loss rate. For example, the bandwidth utilization of path 1 is 70%, the delay jitter is 5ms, and the packet loss rate is 0.2%. These features are input into a convolutional neural network model, which includes a feature extraction layer and a fully connected layer. The feature extraction layer uses multiple convolution kernels to extract the spatiotemporal correlation of the path state features. The fully connected layer maps the extracted features to weight adjustment values. The convolutional neural network model is trained collaboratively between the controller nodes using federated learning. Assuming that the model outputs the weight adjustment value of path 1 as 0.1, the new weight coefficient of path 1 is 0.5+0.1=0.6.

[0142] In order to cope with network congestion, a content classification table is established to prioritize content according to its historical access popularity and timeliness. For example, a live news broadcast has a higher priority than a review program. The content classification table contains content identification, access frequency, last access time, and priority level. When it is detected that the path bandwidth utilization exceeds the first preset threshold (e.g., 80%) or the delay jitter exceeds the second preset threshold (e.g., 10ms), a congestion indication signal is generated. After receiving the congestion indication signal, the preset number of contents with the highest priority (e.g., the first 10) in the content classification table are deployed to the neighboring nodes of the user node on the congested path.

[0143] In order to make the content rating table reflect the real-time content popularity, an online learning algorithm is used. The content request information of each node is collected, including the request time, the requested content identifier, and the request status. This information is input into the multi-armed bandit algorithm to calculate the access probability and confidence interval of each content. The real-time content popularity is estimated based on the access probability and confidence interval, and the priority of the content in the content rating table is dynamically adjusted. For example, if the number of visits to a review program suddenly surges, its priority will be increased accordingly.

[0144] In order to control the data transmission rate, the status of the candidate path is evaluated and the path delay change rate is calculated. When the path delay change rate exceeds a third preset threshold (e.g., 10%), the transmission rate of the corresponding path is reduced according to a preset ratio (e.g., 50%). The token bucket algorithm is used to limit the flow. The token bucket size and token generation rate are determined according to the transmission rate. When the number of tokens in the token bucket is insufficient, the data to be sent is cached in the local buffer.

[0145] The path status characteristics and transmission rate are input into the path evaluation model, and the path comprehensive score is calculated by weighted summation. The weight coefficients of the paths in the candidate path set are updated based on the path comprehensive score, and the optimal transmission path is recalculated.

[0146] In this embodiment, through weighted polling and deep learning, the optimal transmission path can be dynamically selected according to the real-time status of the network, avoiding the impact of single path failure and congestion, and ensuring the smoothness of live broadcast. Through congestion control and content classification mechanism, network resources can be effectively allocated, and the transmission of high-priority content can be prioritized to avoid waste of resources. By deploying high-popularity content to neighbor nodes of user nodes, the distance of content transmission is shortened and the delay of live broadcast is reduced. In summary, this embodiment realizes adaptive optimization of live data transmission, improves the efficiency of network resource utilization, ensures the quality of live broadcast service, and enhances the reliability and stability of the system.

[0147] In an optional embodiment,

[0148] Maintaining a token bucket at the source node to limit traffic, determining a token bucket size and a token generation rate according to the transmission rate, and caching the data to be sent to a local buffer when the number of tokens in the token bucket is insufficient includes:

[0149] Determine the token bucket size according to the preset maximum transmission rate, the token bucket size is the maximum number of bytes that can be transmitted per unit time corresponding to the maximum transmission rate, set the token generation rate to the maximum transmission rate, initialize the number of tokens in the token bucket to the token bucket size, and record the token update timestamp as the current time;

[0150] When data is sent, obtain the current timestamp, calculate the number of newly generated tokens within the time interval according to the current timestamp and the token update timestamp, the number of newly generated tokens is the product of the token generation rate and the time interval, add the number of newly generated tokens to the number of tokens in the token bucket, the number of tokens in the token bucket does not exceed the token bucket size, and update the token update timestamp to the current timestamp;

[0151] Obtain the data size of the data to be sent, determine whether the number of tokens in the token bucket is greater than or equal to the data size, when the number of tokens in the token bucket is greater than or equal to the data size, subtract the number of tokens equal to the data size from the token bucket and send the data to be sent, when the number of tokens in the token bucket is less than the data size, store the data to be sent in a local buffer;

[0152] Checking whether there is cached data in the local buffer at a preset time interval, when the cached data exists in the local buffer, obtaining the data size of the cached data, judging whether the number of tokens in the token bucket is greater than or equal to the data size of the cached data, and when the number of tokens in the token bucket is greater than or equal to the data size of the cached data, subtracting the number of tokens equal to the data size of the cached data from the token bucket and sending the cached data;

[0153] Monitor the buffer size of the local buffer, and when the buffer size reaches a preset buffer threshold, obtain the priority and waiting time of the cached data in the local buffer, sort the cached data based on the priority and the waiting time, and remove a preset number of cached data at the end of the sorting.

[0154] The local buffer is an area in a computer or network system used to temporarily store data. It can alleviate the delay of data processing, especially in the process of data transmission, and is used to store data that has not been processed or sent.

[0155] Set the maximum transmission rate to 10MB / s, which means that a maximum of 10MB of data can be transmitted per second. Based on this, set the token bucket size to 10MB, which is the maximum number of bytes that can be transmitted per unit time. At the same time, set the token generation rate to 10MB / s, and initialize the number of tokens in the token bucket to 10MB. Record the current time as the token update timestamp, for example, the initial timestamp is January 1, 2024 00:00:00.

[0156] When there is data to be sent, first get the current timestamp, for example, the current timestamp is 00:00:01 on January 1, 2024. Calculate the time interval between the current timestamp and the token update timestamp, which is 1 second here. Calculate the number of newly generated tokens based on the token generation rate and time interval, that is, 10MB / s*1s=10MB. Add the newly generated tokens to the token bucket, and the number of tokens in the token bucket is 10MB (initial 10MB) + 10MB=20MB. Since the token bucket size is limited to 10MB, the actual number of tokens in the token bucket is 10MB, and the token update timestamp is updated to the current timestamp, that is, 00:00:01 on January 1, 2024.

[0157] Assume that the size of the data to be sent is 5MB. Since the number of tokens in the token bucket (10MB) is greater than the size of the data to be sent (5MB), 5MB of tokens are subtracted from the token bucket, leaving 5MB, and the 5MB of data is sent immediately.

[0158] Assume that the size of the data to be sent is 15MB. Since the number of tokens in the token bucket (10MB) is less than the size of the data to be sent (15MB), the 15MB of data is stored in the local buffer.

[0159] The system periodically checks the local buffer at a preset time interval, such as 100 milliseconds. Suppose in one check, it is found that there is 15MB of previously cached data in the buffer. At this time, the current timestamp is obtained again, and the number of newly generated tokens is calculated. Assuming that the number of newly generated tokens is 1MB, the number of tokens in the token bucket becomes 6MB (5MB remaining in the previous example + 1MB newly generated). Since the number of tokens in the token bucket (6MB) is still less than the cached data size (15MB), continue to wait. In the next check, assuming that the number of tokens in the token bucket has reached 15MB or more, 15MB of tokens are subtracted from the token bucket, and the 15MB of data is taken out of the buffer and sent.

[0160] Continuously monitor the size of the local buffer. Assume that the buffer size reaches a preset threshold, such as 50MB. At this time, obtain the priority and waiting time of all cached data in the buffer. Assume that data A has a high priority and a waiting time of 1 second; data B has a low priority and a waiting time of 2 seconds; data C has a medium priority and a waiting time of 3 seconds. Sort the data according to priority and waiting time, with the data with a high priority at the front, and the data with a long waiting time at the front if the priorities are the same. The sorting result is: A, C, B. Remove the preset number of cached data at the end of the sorting, for example, remove the last data, that is, remove data B, to ensure that the buffer does not overflow.

[0161] In this embodiment, the data sending rate is accurately controlled by the token bucket algorithm to ensure smooth traffic output, prevent burst traffic from impacting the network, effectively avoid network congestion, and improve network utilization. When the token is insufficient, the data will be cached in the local buffer instead of being discarded directly, ensuring the integrity of the data, and sending the cached data when the token is sufficient to improve the reliability of data transmission. The local buffer mechanism can effectively absorb burst traffic and avoid system crashes due to instantaneous excessive load. At the same time, the buffer clearing mechanism can remove part of the data according to priority and waiting time, further improving system stability and resource utilization efficiency. In summary, this embodiment effectively balances network bandwidth utilization and transmission smoothness, avoids network congestion and data loss, improves the service quality of live transmission, ensures the timely transmission of important data, and enhances the reliability and stability of the system.

[0162] Figure 2 FIG. 1 is a schematic diagram of a hotspot prediction and adaptive routing system for a private domain live broadcast room based on swarm intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0163] The first unit is used to collect data from the private domain live broadcast room according to a preset sliding time window, obtain user behavior data, and extract features through a time series convolutional network, obtain time series features and add them to a gated recurrent unit, combine update gates and reset gates for information flow, calculate long-term dependency features, and based on the correlation of different positions, perform feature fusion through a feature fusion network and combine a feedforward neural network to upgrade the fused features to a high-dimensional semantic space to obtain a time series feature matrix, construct a dynamic graph neural network based on the time series feature matrix, and calculate the multi-dimensional similarity between user nodes through a dynamic edge weight update mechanism, dynamically connect user nodes based on the multi-dimensional similarity and a preset similarity threshold, and obtain a dynamic interest community;

[0164] The second unit is used to add the dynamic interest community to a pre-set multi-layer graph attention network, extract group interaction pattern features through a spatial attention layer and a channel attention layer and add them to an improved bidirectional gated recurrent unit, obtain long-term and short-term temporal dependencies and generate temporal correlation features, add the temporal correlation features to a deep residual network and extract high-order semantic features through residual learning and feature recalibration, construct a multi-task learning framework based on the high-order semantic features, perform multi-task prediction through the multi-task learning framework, obtain hotspot prediction results and construct a multi-level network topology model based on the hotspot prediction results, set edge weights according to network transmission constraints, perform path exploration in combination with an improved heuristic search algorithm, and obtain multiple optimal transmission paths;

[0165] The third unit is used to assign weights to each optimal transmission path and build a set of candidate paths, distribute traffic to each path in the candidate path set through a weighted polling algorithm, update the weight of each path in combination with a distributed load balancing controller based on deep learning and detect the smoothness of the network. If there is local congestion, a congestion indication signal is sent, a content classification table is established, and predetermined high-heat content is deployed to the neighboring nodes of the user node. According to a preset content distribution controller based on an online learning algorithm, content access data is collected and updated. The content classification table is updated, the real-time network status is obtained, and the data transmission rate is dynamically adjusted through a congestion control algorithm. The network speed is limited in combination with a token bucket algorithm, and the optimal transmission path is updated based on the adjusted network status to obtain a live broadcast push path.

[0166] According to a third aspect of the embodiments of the present invention,

[0167] An electronic device is provided, comprising:

[0168] processor;

[0169] a memory for storing processor-executable instructions;

[0170] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0171] According to a fourth aspect of the embodiments of the present invention,

[0172] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0173] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for hotspot prediction and adaptive routing in private live broadcast rooms based on swarm intelligence, characterized in that: include: According to the preset sliding time window, data is collected from the private domain live broadcast room to obtain user behavior data, and feature extraction is performed through a time series convolutional network to obtain time series features and add them to the gated recurrent unit. The update gate and the reset gate are combined for information flow, and the long-term dependency features are calculated. Based on the correlation of different positions, feature fusion is performed through a feature fusion network and the fused features are upgraded to a high-dimensional semantic space in combination with a feedforward neural network to obtain a time series feature matrix. A dynamic graph neural network is constructed based on the time series feature matrix, and the multi-dimensional similarity between user nodes is calculated through a dynamic edge weight update mechanism. The user nodes are dynamically connected based on the multi-dimensional similarity and a preset similarity threshold to obtain a dynamic interest community. The dynamic interest community is added to a pre-set multi-layer graph attention network, and the group interaction pattern features are extracted through the spatial attention layer and the channel attention layer and added to the improved bidirectional gated recurrent unit to obtain long-term and short-term temporal dependencies and generate temporal correlation features. The temporal correlation features are added to the deep residual network and high-order semantic features are extracted through residual learning and feature recalibration. A multi-task learning framework is constructed based on the high-order semantic features. Multi-task prediction is performed through the multi-task learning framework to obtain hotspot prediction results and a multi-level network topology model is constructed based on the hotspot prediction results. Edge weights are set according to network transmission constraints, and path exploration is performed in combination with an improved heuristic search algorithm to obtain multiple optimal transmission paths. Assign a weight to each optimal transmission path and build a set of candidate paths. Perform traffic distribution on each path in the candidate path set through a weighted polling algorithm. Combined with a distributed load balancing controller based on deep learning, update the weight of each path and detect the network smoothness. If there is local congestion, send a congestion indication signal, establish a content classification table and deploy predetermined high-heat content to the neighboring nodes of the user node. Collect content access data and update the content classification table according to a preset content distribution controller based on an online learning algorithm. Obtain real-time network status and dynamically adjust the data transmission rate through a congestion control algorithm. Combined with a token bucket algorithm to limit the network speed, update the optimal transmission path based on the adjusted network status to obtain a live broadcast push path.

2. The method according to claim 1, characterized in that According to the preset sliding time window, data is collected from the private domain live broadcast room to obtain user behavior data and extract features through the time series convolutional network to obtain time series features and add them to the gated recurrent unit. The update gate and reset gate are combined for information flow to calculate the long-term dependency features. Based on the correlation of different positions, feature fusion is performed through the feature fusion network and the fused features are upgraded to a high-dimensional semantic space in combination with the feedforward neural network to obtain a time series feature matrix. A dynamic graph neural network is constructed based on the time series feature matrix and the multi-dimensional similarity between user nodes is calculated through a dynamic edge weight update mechanism. The user nodes are dynamically connected based on the multi-dimensional similarity and the preset similarity threshold to obtain dynamic interest communities including: According to the pre-set sliding time window, user behavior data of the private domain live broadcast room is collected, and the user behavior data includes user viewing time, number of interactions, number of likes, and comment content; The user behavior data is input into a temporal convolutional network for feature extraction. The temporal convolutional network extracts temporal features from the user behavior data through multiple one-dimensional convolutional layers and pooling layers to obtain temporal features that reflect changes in user behavior over time; The time series feature is input into a gated recurrent unit, and the gated recurrent unit controls the information flow through an update gate and a reset gate. The update gate controls the retention degree of the hidden state information at the previous moment, and the reset gate controls the neglect degree of the hidden state information at the previous moment, so as to obtain a long-term dependent feature reflecting the user's long-term interest tendency, and the long-term dependent feature is input into a feature fusion network. The feature fusion network adopts an attention mechanism, and adaptively allocates weights according to the correlation between features at different moments, so as to obtain a fused feature; The fused features are processed by a feedforward neural network, wherein the feedforward neural network includes multiple fully connected layers and nonlinear activation functions, and the features are mapped to a high-dimensional semantic space to obtain a time series feature matrix; Construct a dynamic graph neural network based on the time series feature matrix, set each user as a node in the graph, calculate the multidimensional similarity between user nodes, the multidimensional similarity includes cosine similarity and Pearson correlation coefficient, and according to a preset similarity threshold, establish a dynamic edge between user nodes whose multidimensional similarity is higher than the similarity threshold, and determine the weight of the dynamic edge according to the multidimensional similarity between user nodes; Based on the changes in the user behavior data, the multidimensional similarity between user nodes is continuously updated. When the updated multidimensional similarity is lower than the similarity threshold, the corresponding dynamic edge is deleted. When the updated multidimensional similarity is higher than the similarity threshold, a new dynamic edge is established. A dynamic interest community is constructed through the subgraph formed by the dynamic edge connection.

3. The method according to claim 2, characterized in that The weight of the dynamic edge is determined according to the multi-dimensional similarity between user nodes as shown in the following formula: ； in, EW represents the weight value of the dynamic edge, w k ( t ) indicates the k The similarity index at time t Adaptive weights, similarity k For the k The value of the similarity index, p k Indicates k The power term of similarity indexes is used to control the nonlinear combination relationship of different indexes. t k Indicates k The time decay factor of the similarity index is λ represents the time attenuation coefficient, I ij Indicates user i and users j The interaction weight between tt j Indicates user j The behavior timestamp, α Represents the weight coefficient of multi-view similarity, SV m Indicates m User from the perspective i and users j The similarity.

4. The method according to claim 1, characterized in that The dynamic interest community is added to a pre-set multi-layer graph attention network, and the group interaction pattern features are extracted through the spatial attention layer and the channel attention layer and added to the improved bidirectional gated recurrent unit to obtain long-term and short-term temporal dependencies and generate temporal correlation features. The temporal correlation features are added to the deep residual network and high-order semantic features are extracted through residual learning and feature recalibration. A multi-task learning framework is constructed based on the high-order semantic features. Multi-task prediction is performed through the multi-task learning framework to obtain hotspot prediction results and a multi-level network topology model is constructed based on the hotspot prediction results. The edge weights are set according to the network transmission constraints, and the improved heuristic search algorithm is combined to perform path exploration to obtain multiple optimal transmission paths including: The dynamic interest community is input into a multi-layer graph attention network composed of a spatial attention layer and a channel attention layer. The spatial attention layer performs weighted aggregation on the node features, and obtains the aggregate node representation by calculating the attention weights between nodes and weighted summing the node features. The channel attention layer obtains the channel-level feature descriptor through global average pooling, and generates the channel importance weight through two fully connected layers and a sigmoid activation function. The channel importance weight is multiplied element by element with the original feature map to obtain a recalibrated feature map, and the group interaction pattern features are extracted; The group interaction pattern feature is input into an improved bidirectional gated recurrent unit including a forget gate, an input gate, a time memory gate, a unit state and an output gate, the forgetting degree of historical state information in the group interaction pattern feature is controlled by the forget gate, the current information retention degree is controlled by the input gate, the dependency between time steps is captured by the time memory gate, the long-term and short-term temporal dependency is obtained by combining the dependency between time steps and the output information flow is controlled by the output gate to obtain a forward hidden state and a backward hidden state, and the forward hidden state and the backward hidden state are concatenated to obtain a temporal correlation feature; The temporal correlation features are input into a deep residual network comprising a plurality of residual blocks, each of which comprises two convolutional layers and an identity mapping connection, the original features are transmitted through the identity mapping connection, the residual mapping relationship is learned through the convolutional layer, a feature recalibration mechanism is introduced between the residual blocks, the features are adaptively scaled and offset, and high-order semantic features are obtained; The high-order semantic features are input into a multi-task learning framework including a shared feature extractor and multiple task predictors, wherein the shared feature extractor extracts a common feature representation, and the multiple task predictors perform community evolution trend prediction, community influence prediction, and community topic popularity prediction based on the common feature representation to obtain a hotspot prediction result; A multi-level network topology model is constructed according to the hotspot prediction result, different level nodes are set in the multi-level network topology model to represent different community granularities, and inter-level connections are set to represent the association relationship between communities. Weights are assigned to the edges in the multi-level network topology model according to bandwidth restrictions and latency requirements to obtain a weighted network topology model; The weighted network topology model is input into an improved heuristic search algorithm including a multi-objective optimization mechanism and an adaptive strategy adjustment mechanism. The path bandwidth, transmission delay, and node load are balanced through the multi-objective optimization mechanism. The search strategy is adjusted according to the search feedback information through the adaptive strategy adjustment mechanism to obtain multiple optimal transmission paths that meet the network transmission constraints.

5. The method according to claim 4, characterized in that The multiple task predictors perform community evolution trend prediction, community influence prediction, and community topic popularity prediction based on the general feature representation, and obtain hotspot prediction results including: The universal feature representations output by the shared feature extractor are respectively input into a pre-set community evolution trend predictor, and the long-range temporal dependency features are extracted through the gated time convolution network in the temporal feature branch of the community evolution trend predictor. The community topological structure features are extracted through the graph isomorphism network in the structural feature branch of the community evolution trend predictor. The long-range temporal dependency features and the community topological structure features are input into an adaptive feature fusion module, and the correlation weights between feature channels and the importance weights of feature space positions are calculated. The correlation weights are multiplied by the importance weights to obtain fusion weights, and the fusion weights are weightedly combined with the original features to obtain fusion features. The fusion features are input into a multi-layer perceptron to predict the community evolution trend, and the community evolution trend prediction results are obtained. A heterogeneous graph structure is constructed based on a preset community influence predictor. Community members are used as nodes, different types of member interactions are used as edges, and the general feature representation is used as the initial node feature. The attention weights on different types of edges are calculated through a multi-head attention layer. The neighbor node features are weighted aggregated to obtain a member-level representation. The member-level representation is input into a differential graph pooling layer to generate a group-level representation. The group-level representation is propagated through a graph diffusion convolutional network. The propagated features are input into a recursive neural network to predict the community influence evolution sequence, and the community influence prediction result is obtained. Based on a preset community topic heat predictor, extract topic text from the common features, preprocess the topic text, construct a historical topic sequence in combination with a time window and extract sentiment words, establish a topic-sentiment word mapping table, encode the topic text through a pre-trained language model to obtain semantic features, map the sentiment words into sentiment vectors according to the topic-sentiment word mapping table, input the sentiment vectors into a deep sentiment analysis network to extract sentiment features, input the historical topic sequence into a causal convolutional network to extract temporal features, calculate the cross-modal attention scores between the semantic features, the sentiment features and the temporal features, perform weighted fusion according to the attention scores, input the fused features into a conditional variational autoencoder to obtain latent variables, reparameterize and sample the latent variables, and input the sampling results into a decoder to generate a diversified topic heat prediction distribution; Calculate the correlation matrix and task difficulty vector between each prediction task, score the importance of each prediction task based on the correlation matrix and the difficulty vector, assign weights and learning rates to different prediction tasks according to the importance scores, calculate the angle between the gradient directions of each task, and perform a gradient orthogonalization operation when the angle is less than a preset threshold; A task adaptability evaluation function is constructed, and task loss and prediction accuracy are used as reward signals. The policy gradient algorithm is used to optimize the task scheduling strategy, and the task weight distribution scheme is updated. The output results of the predictor are exponentially weighted averaged to obtain the integrated prediction results. The Bayesian model is used to average the uncertainty interval of the prediction results to obtain the hotspot prediction results.

6. The method according to claim 1, characterized in that Assign a weight to each optimal transmission path and build a candidate path set, distribute traffic to each path in the candidate path set through a weighted polling algorithm, update the weight of each path in combination with a distributed load balancing controller based on deep learning and detect the smoothness of the network. If there is local congestion, send a congestion indication signal, establish a content classification table and deploy predetermined high-heat content to the neighboring nodes of the user node, collect content access data and update the content classification table according to a pre-set content distribution controller based on an online learning algorithm, obtain real-time network status and dynamically adjust the data transmission rate through a congestion control algorithm, combine the token bucket algorithm to limit the network speed, update the optimal transmission path based on the adjusted network status, and obtain the live push path, including: Receiving multiple optimal transmission paths acquired in advance, collecting performance indicator data of the optimal transmission paths, the performance indicator data including bandwidth, delay and packet loss rate, constructing a judgment matrix, calculating the maximum eigenvalue and path eigenvector of the judgment matrix, normalizing the path eigenvector to obtain a path weight coefficient, and constructing a candidate path set based on the optimal transmission paths and the corresponding path weight coefficients; Input the candidate path set into the weighted polling scheduler, initialize the accumulated weight variable to zero, initialize the last selected path variable to empty, traverse each path in the candidate path set, add the accumulated weight variable to the path weight coefficient corresponding to the current traversed path, and when the accumulated weight variable is greater than a preset weight threshold, select the current traversed path as the data forwarding path, subtract the preset weight threshold from the accumulated weight variable, and record the current traversed path as the last selected path variable; Deploy multiple load balancing controller nodes in the network, input the path state features in the candidate path set into the load balancing controller nodes, the path state features include bandwidth utilization, delay jitter and packet loss rate, output the path weight adjustment value through a convolutional neural network model, the convolutional neural network model includes a feature extraction layer and a fully connected layer, the feature extraction layer uses multiple convolution kernels to extract the spatiotemporal correlation of the path state features, the fully connected layer maps the extracted features to weight adjustment values, and adopts a federated learning method to collaboratively train the convolutional neural network model among multiple controller nodes; Constructing a content classification table, prioritizing the content based on the historical access popularity and timeliness of the content, wherein the content classification table includes content identification, access frequency, most recent access time and priority level, monitoring path bandwidth utilization and delay jitter, generating a congestion indication signal when the bandwidth utilization exceeds a first preset threshold or the delay jitter exceeds a second preset threshold, and upon receiving the congestion indication signal, deploying a preset number of contents with the highest priority in the content classification table to neighboring nodes of the user node on the congested path; Collecting content request information of each node, the content request information including request time, request content identifier and request status, inputting the content request information into a multi-armed bandit algorithm, performing online learning on content access data, calculating access probability and confidence interval of each content, estimating real-time content popularity according to the access probability and the confidence interval, and dynamically adjusting the priority of the content in the content classification table based on the real-time content popularity; Performing status evaluation on each path in the candidate path set, calculating the path delay change rate, when the path delay change rate exceeds a third preset threshold, reducing the transmission rate of the corresponding path according to a preset ratio, maintaining a token bucket at the source node to limit the flow, determining the token bucket size and the token generation rate according to the transmission rate, and caching the data to be sent in a local buffer when the number of tokens in the token bucket is insufficient; The path state characteristics and the transmission rate are input into a path evaluation model, the path evaluation model calculates the path comprehensive score by weighted summation, updates the weight coefficients of the paths in the candidate path set based on the path comprehensive score, writes the updated weight coefficients into the candidate path set and recalculates the optimal transmission path to obtain the live broadcast push path.

7. The method according to claim 6, characterized in that Maintaining a token bucket at the source node to limit traffic, determining a token bucket size and a token generation rate according to the transmission rate, and caching the data to be sent to a local buffer when the number of tokens in the token bucket is insufficient includes: Determine the token bucket size according to the preset maximum transmission rate, the token bucket size is the maximum number of bytes that can be transmitted per unit time corresponding to the maximum transmission rate, set the token generation rate to the maximum transmission rate, initialize the number of tokens in the token bucket to the token bucket size, and record the token update timestamp as the current time; When data is sent, obtain the current timestamp, calculate the number of newly generated tokens within the time interval according to the current timestamp and the token update timestamp, the number of newly generated tokens is the product of the token generation rate and the time interval, add the number of newly generated tokens to the number of tokens in the token bucket, the number of tokens in the token bucket does not exceed the token bucket size, and update the token update timestamp to the current timestamp; Obtain the data size of the data to be sent, determine whether the number of tokens in the token bucket is greater than or equal to the data size, when the number of tokens in the token bucket is greater than or equal to the data size, subtract the number of tokens equal to the data size from the token bucket and send the data to be sent, when the number of tokens in the token bucket is less than the data size, store the data to be sent in a local buffer; Checking whether there is cached data in the local buffer at a preset time interval, when the cached data exists in the local buffer, obtaining the data size of the cached data, judging whether the number of tokens in the token bucket is greater than or equal to the data size of the cached data, and when the number of tokens in the token bucket is greater than or equal to the data size of the cached data, subtracting the number of tokens equal to the data size of the cached data from the token bucket and sending the cached data; Monitor the buffer size of the local buffer, and when the buffer size reaches a preset buffer threshold, obtain the priority and waiting time of the cached data in the local buffer, sort the cached data based on the priority and the waiting time, and remove a preset number of cached data at the end of the sorting.

8. A hotspot prediction and adaptive routing system for private live broadcast rooms based on swarm intelligence, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect data from the private domain live broadcast room according to a preset sliding time window, obtain user behavior data, and extract features through a time series convolutional network, obtain time series features and add them to a gated recurrent unit, combine update gates and reset gates for information flow, calculate long-term dependency features, and based on the correlation of different positions, perform feature fusion through a feature fusion network and combine a feedforward neural network to upgrade the fused features to a high-dimensional semantic space to obtain a time series feature matrix, construct a dynamic graph neural network based on the time series feature matrix, and calculate the multi-dimensional similarity between user nodes through a dynamic edge weight update mechanism, dynamically connect user nodes based on the multi-dimensional similarity and a preset similarity threshold, and obtain a dynamic interest community; The second unit is used to add the dynamic interest community to a pre-set multi-layer graph attention network, extract group interaction pattern features through a spatial attention layer and a channel attention layer and add them to an improved bidirectional gated recurrent unit, obtain long-term and short-term temporal dependencies and generate temporal correlation features, add the temporal correlation features to a deep residual network and extract high-order semantic features through residual learning and feature recalibration, construct a multi-task learning framework based on the high-order semantic features, perform multi-task prediction through the multi-task learning framework, obtain hotspot prediction results and construct a multi-level network topology model based on the hotspot prediction results, set edge weights according to network transmission constraints, perform path exploration in combination with an improved heuristic search algorithm, and obtain multiple optimal transmission paths; The third unit is used to assign weights to each optimal transmission path and build a set of candidate paths, distribute traffic to each path in the candidate path set through a weighted polling algorithm, update the weight of each path in combination with a distributed load balancing controller based on deep learning and detect the smoothness of the network. If there is local congestion, a congestion indication signal is sent, a content classification table is established, and predetermined high-heat content is deployed to the neighboring nodes of the user node. According to a preset content distribution controller based on an online learning algorithm, content access data is collected and updated. The content classification table is updated, the real-time network status is obtained, and the data transmission rate is dynamically adjusted through a congestion control algorithm. The network speed is limited in combination with a token bucket algorithm, and the optimal transmission path is updated based on the adjusted network status to obtain a live broadcast push path.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images