Hot spot prediction and content scheduling method for private domain live broadcast peak period based on deep learning

Abstract

The present invention provides a method for predicting hot spots and scheduling content during peak periods of private domain live broadcasts based on deep learning, which relates to the technical field of private domain live broadcasts, including: extracting multi-scale time series features in private domain live broadcast history data using a hybrid memory-enhanced neural network, generating live broadcast peak period prediction results through a spatial pyramid attention module and a progressive feature distillation network, constructing a dynamic heterogeneous graph neural network, combining multimodal data features and hierarchical comparative learning, outputting a live broadcast content scheduling strategy through a multi-agent reinforcement learning framework, constructing a two-way interactive evaluation network, and utilizing a fast adaptation algorithm of meta-reinforcement learning to perform online optimization of the scheduling strategy according to real-time live broadcast data to obtain an optimal content scheduling strategy.

Description

Technical Field

[0001] The present invention relates to the field of private domain live broadcast technology, and in particular to a method for predicting hot spots and scheduling content during peak periods of private domain live broadcast based on deep learning. Background Art

[0002] The rapid development of deep learning technology has brought new opportunities to all walks of life. In the field of live broadcasting, how to accurately predict the peak period of live broadcasting and effectively schedule content is crucial to improving user experience and live broadcasting effects.

[0003] Traditional methods for predicting peak hours of live broadcasts mainly rely on statistical analysis and rule matching, which makes it difficult to capture complex and changeable user behavior patterns and has limited prediction accuracy. In addition, traditional content scheduling methods lack intelligence and personalization, and are unable to dynamically adjust live broadcast content according to real-time conditions, making it difficult to meet the diverse needs of users.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a method for predicting hot spots and scheduling content during peak periods of private domain live broadcasts based on deep learning, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for predicting hot spots and scheduling content during peak periods of private domain live broadcasts based on deep learning, comprising:

[0007] Collect private domain live broadcast history data, build a hybrid memory enhanced neural network including a fast memory pathway and a slow memory pathway, the fast memory pathway performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory pathway captures long-term dependencies through a hierarchical sparse attention structure, input the private domain live broadcast history data into the hybrid memory enhanced neural network, extract multi-scale temporal features through the fast memory pathway and the slow memory pathway, build a spatial pyramid attention module, perform feature reorganization on the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, adjust the feature receptive field in combination with deformable convolution, obtain a hierarchical feature representation, input the hierarchical feature representation into a progressive feature distillation network, perform multi-level knowledge transfer and feature reorganization operations, and generate a live broadcast peak prediction result;

[0008] Collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node association relationships, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through a goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output a live content scheduling strategy;

[0009] Real-time live broadcast data is collected and added to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between predicted effect and actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, an adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy. Collection and adjustment are repeated to obtain the optimal content scheduling strategy.

[0010] In an optional embodiment,

[0011] Collecting private domain live broadcast history data, constructing a hybrid memory enhanced neural network including a fast memory pathway and a slow memory pathway, wherein the fast memory pathway performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory pathway captures long-term dependencies through a hierarchical sparse attention structure, inputting the private domain live broadcast history data into the hybrid memory enhanced neural network, extracting multi-scale temporal features through the fast memory pathway and the slow memory pathway, constructing a spatial pyramid attention module, and reorganizing the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, including:

[0012] The historical data of private domain live broadcast is sampled every minute, and the number of viewers, likes, comments, and reward amount data are collected. The weighted average calculation is performed on the number of viewers, likes, comments, and reward amount data to obtain historical trend characteristics;

[0013] A fast memory path and a slow memory path of a hybrid memory-enhanced neural network are set, a bidirectional gated recurrent unit network is configured in the fast memory path, a candidate state generation unit is set in the bidirectional gated recurrent unit network, and a nonlinear combination of a current input and a historical state is calculated by the candidate state generation unit to generate a candidate hidden state; the historical state is randomly forgotten by a reset gate, and the candidate hidden state is updated by an update gate;

[0014] A converter network is configured in the slow memory path, a multi-head self-attention layer is combined with a feedforward neural network layer to construct a computing block, a query vector, a key vector, and a value vector are projected into multiple feature subspaces in the multi-head self-attention layer, a scaled dot product calculation is performed in each feature subspace to obtain an attention weight, dropout regularization is performed on the attention weight, two layers of perceptrons are connected in series in the feedforward neural network layer, wherein the first layer of perceptrons performs dimensional expansion, and the second layer of perceptrons performs dimensional compression, and a normalization layer is added between the layers;

[0015] Performing an adaptive batch normalization operation on the private domain live broadcast history data and the historical trend features to obtain normalized data, reorganizing the normalized data into a three-dimensional array including a day dimension, a minute dimension, and an indicator dimension, inputting the three-dimensional array into the fast memory path, bidirectionally scanning along the day dimension to obtain local time series features, inputting the three-dimensional array into the slow memory path, adding time series position information through position coding, and obtaining global time series features;

[0016] Configure the first attention branch, the second attention branch and the third attention branch of the spatial pyramid attention module, set a sparse attention mechanism in the first attention branch, generate an adaptive attention mask, perform information aggregation of key temporal positions through Top-k selection, set a dense attention mechanism in the second attention branch, perform feature extraction of local areas through a deformable sliding window, set a hybrid attention mechanism in the third attention branch, cascade channel attention and spatial attention, extract cross-channel and cross-spatial dimension dependencies, align the local temporal features with the global temporal features in the time dimension and input them into the spatial pyramid attention module, and perform weighted fusion on the output features of the three attention branches to obtain multi-scale fusion features.

[0017] In an optional embodiment,

[0018] Combined with deformable convolution to adjust the feature receptive field, a hierarchical feature representation is obtained, and the hierarchical feature representation is input into the progressive feature distillation network to perform multi-level knowledge transfer and feature reorganization operations to generate the live broadcast peak period prediction results including:

[0019] The offset prediction layer, the modulation factor prediction layer and the feature sampling layer of the multi-layer deformable convolutional network are configured, the two-dimensional offset of the sampling position of the pre-acquired multi-scale fusion feature is calculated in the offset prediction layer, the modulation factor of the sampling position is generated in the modulation factor prediction layer, the feature response is adaptively adjusted, the bilinear interpolation sampling is performed on the feature map in combination with the two-dimensional offset and the modulation factor in the feature sampling layer, the multi-scale fusion feature is sequentially input into each layer of the multi-layer deformable convolutional network, cross-layer connections are added between adjacent layers, and a hierarchical feature representation is generated;

[0020] A teacher network and a multi-level student network of a progressive feature distillation network are configured, parameters of the teacher network are initialized according to the hierarchical feature representation, each level of the student network corresponds to a feature layer of different depths, features of the student networks at adjacent levels are reused through dense residual connections, an attention distillation module is introduced into the teacher network, an attention map is used as an intermediate supervision signal, the hierarchical feature representation is input into the progressive feature distillation network, the L2 loss and the attention map loss of the feature maps of the teacher network and the student network are calculated, the KL divergence loss corresponding to the soft label distribution is calculated, a total loss is obtained by weighted combination of multiple loss items, network parameters are optimized based on the total loss, and a distilled feature representation is generated;

[0021] The distilled feature representation is input into a multilayer perceptron, in which the feature dimension is reduced layer by layer, and in the last layer the output dimension is set to the number of discrete time periods in the prediction time window; the output is normalized by the Softmax function to obtain the peak probability of each time period, and the optimal threshold is determined by combining the dynamic probability threshold selection algorithm, and the time period above the optimal threshold is marked as the predicted peak period.

[0022] In an optional embodiment,

[0023] Collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node associations, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through the goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output the live content scheduling strategy, including:

[0024] Perform feature extraction on each frame in the video stream data through a feature extraction network, the feature extraction network includes a feature pyramid network and an attention enhancement module, extract multi-scale visual features through the feature pyramid network, perform feature enhancement in the channel dimension and the spatial dimension through the attention enhancement module, perform short-time Fourier transform on the audio data in combination with the multi-head self-attention network to obtain a spectrogram, input the spectrogram into a deep network to extract audio features, use a bidirectional encoding network to perform sequence modeling on the barrage text and the comment text, and map the word vector sequence into a text feature vector; input the visual features, the audio features, and the text feature vector into a temporal encoding network, perform temporal encoding to obtain a multimodal feature tensor;

[0025] Construct a dynamic heterogeneous graph neural network, take anchor nodes, audience nodes, video clip nodes, audio clip nodes, bullet screen nodes, and comment nodes as graph nodes, assign the multimodal feature tensor to the corresponding graph nodes, calculate the similarity between node pairs through a topology generator, perform a normalization transformation on the similarity to obtain edge weights, construct a node adjacency matrix in combination with a sparse matrix representation, and use a feature matrix to store node attributes to obtain a feature representation of the node;

[0026] The dynamic heterogeneous graph neural network is dynamically updated. When new data arrives, the corresponding modal features are extracted and new nodes are created. The feature similarity between the newly added nodes and the existing nodes corresponding to the new data is calculated. Connecting edges are added between node pairs whose similarity is greater than a preset first threshold. The product of the degree centrality of the existing nodes and the time decay factor is calculated. Nodes whose product is less than a preset second threshold are deleted. The exponential sliding average of the edge weights is calculated. The edges whose sliding average is less than a preset third threshold are deleted to obtain an optimized graph structure.

[0027] The feature matrix and the node adjacency matrix are input into a multi-layer graph neural network through a feature interaction engine, wherein the feature matrix is ​​composed of a multimodal feature tensor of nodes, representing the initial features of each node, and a learnable parameter matrix is ​​constructed according to the input feature dimension and the hidden layer dimension. First, the input dimension and output dimension of each layer of the graph neural network are set, and the learnable parameter matrix is ​​initialized using a uniform distribution or a normal distribution. The learnable parameter matrix is ​​used as the weight matrix of each layer of the graph neural network, and the node adjacency matrix is ​​degree-normalized to obtain a symmetric normalized matrix. The normalized matrix is ​​multiplied by the node feature matrix of the current layer, and then multiplied by the learnable parameter matrix. A nonlinear activation function is used to obtain the symmetric normalized matrix. Get the node features of the next layer, apply the hierarchical contrast learning method in each layer, construct positive sample pairs and negative sample pairs, the positive sample pairs are sampled from adjacent nodes, and the negative sample pairs are sampled from non-adjacent nodes at different semantic levels, calculate the feature contrast loss, and perform weighted average of the contrast losses at different levels to obtain the total loss. Input the node representation output by the graph neural network into the hierarchical strategy optimization network, set the interaction quality improvement sub-goal, content form optimization sub-goal, and user experience optimization sub-goal in the goal decomposition module, calculate the sub-goal weights, and set the resource scheduling sub-network in the execution planning module, wherein the resource scheduling sub-network includes a multi-layer encoder, and the output of the encoder is mapped to obtain the scheduling action space;

[0028] The nodes in the dynamic heterogeneous graph neural network are mapped to multiple agents. Each agent maintains a local state representation and a policy network. The policy network consists of an action generation network and a state evaluation network. The action generation network outputs the action probability distribution, and the state evaluation network evaluates the state value function. The agents exchange information through a competitive cooperation mechanism, optimize strategies based on non-cooperative games, and calculate rewards based on the joint actions of the agents. The reward function comprehensively considers multiple indicators such as interactivity, content quality, and user feedback to obtain a comprehensive evaluation score.

[0029] Each agent generates a probability distribution of live content scheduling strategy based on its strategy network, and performs parameterized sampling on the probability distribution to obtain discrete actions. The discrete actions of all agents are combined into a live content scheduling strategy, including the host interaction instruction sequence, the audience activity push sequence, and the platform resource allocation sequence. The structure and node characteristics of the dynamic heterogeneous graph neural network are updated according to the feedback after executing the scheduling strategy.

[0030] In an optional embodiment,

[0031] The agents exchange information through a competitive cooperation mechanism, optimize strategies based on non-cooperative games, and calculate rewards based on the joint actions of the agents, including:

[0032] An agent is deployed at each node of a dynamic heterogeneous graph, wherein the nodes of the dynamic heterogeneous graph include video clip nodes, audio clip nodes, and bullet-screen nodes. The state space of the agent is composed of the attribute feature vector of the corresponding node and the attribute feature vector of the adjacent node. The state vector of the video clip node includes the visual features of the video frame, the acoustic features of the audio clip, and the text features of the surrounding bullet-screen nodes. The action space of the agent includes picture clarity adjustment, interactive question insertion, and related video push.

[0033] The agent broadcasts its own action probability distribution to neighboring nodes at each decision time step, receives the action probability distribution of neighboring nodes, and uses the received action probability distribution as an additional input signal to update the corresponding strategy network, wherein the strategy network includes an action generation network and a value evaluation network;

[0034] Strategy optimization is performed based on a non-cooperative game framework. The objective functions of different types of nodes in the non-cooperative game include the audience participation of the anchor node, the viewing time of the video node, and the number of interactions of the bullet comment node. The agent takes its own actions to obtain immediate rewards based on the observed state and the actions of other agents, and receives trajectory data of the interaction between other agents and the environment to estimate the joint action value function;

[0035] Iteratively update the joint action value function, calculate the value function of the current agent state and the joint action, and perform weighted average of the difference between the value function and the sum of the instant reward and the discounted value of the state value function at the next moment to obtain the updated value function, wherein the instant reward is obtained according to the environment, and the weight of the weighted average is the learning rate;

[0036] The policy function is optimized based on the policy gradient algorithm, and the Monte Carlo estimation method is used to calculate the policy gradient of each agent's objective function. The product of the policy gradient and the learning rate is added to the parameters of the agent's policy function to obtain the updated policy function parameters. The agent generates a scheduling strategy based on the optimized policy function, and performs picture clarity adjustment, interactive question insertion, and related video push according to the scheduling strategy.

[0037] In an optional embodiment,

[0038] Collect real-time live broadcast data and add it to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between the predicted effect and the actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, the adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy. Repeated collection and adjustment to obtain the optimal content scheduling strategy includes:

[0039] Adding the collected real-time video stream, real-time audio stream, and real-time text stream data to a pre-built two-way interactive evaluation network, wherein the two-way interactive evaluation network includes a state encoder and an effect evaluator, wherein the state encoder includes a video feature extraction module, an audio feature extraction module, and a text feature extraction module, and the effect evaluator includes a causal graph construction module, a structural equation model, and a multi-layer perceptron network;

[0040] Extracting frame-level visual features from a real-time video stream through a convolutional neural network in the video feature extraction module, extracting audio spectrogram features from a real-time audio stream through a convolutional neural network in the audio feature extraction module, extracting semantic features from a real-time text stream through a recurrent neural network in the text feature extraction module, inputting the frame-level visual features, the audio spectrogram features, and the semantic features into an attention network to generate a feature weight matrix, and performing weighted fusion on the three features based on the feature weight matrix to obtain a multimodal representation vector;

[0041] The causal graph construction module constructs a live broadcast effect causal graph based on historical data, the live broadcast effect causal graph includes a scheduling strategy node, a live broadcast status node, an external factor node and an effect indicator node, the edges between the nodes represent causal relationships, the structural equation model defines the functional relationship between the nodes, the multimodal representation vector is spliced ​​with the preset scheduling strategy vector to form an input vector, causal reasoning is performed based on the causal graph and the structural equation model, the policy intervention effect is obtained, the input vector and the policy intervention effect are input into the multi-layer perceptron network, the multi-layer perceptron network calculates the predicted number of viewers, the predicted number of likes, and the predicted reward amount, counts the real-time number of viewers, the real-time number of likes, and the real-time reward amount, and calculates the weighted average deviation between the predicted value and the actual value to obtain the real-time performance evaluation index;

[0042] Calculate the deviation value between the real-time performance evaluation index and the stored live broadcast peak period prediction result, compare the deviation value with a preset deviation threshold, and trigger strategy adjustment when the deviation value is greater than the deviation threshold;

[0043] The live content scheduling task is decomposed into content recommendation subtask, anchor interaction subtask, and audience activity subtask. Each subtask is configured with an independent policy network and value function network. The policy network generates an action probability distribution based on the current state, and the value function network evaluates the value of the state-action pair. A gradient meta-learner is trained for each subtask. The gradient meta-learner adopts a model-independent meta-learning algorithm. In the inner loop, the policy gradient method is used to update the network parameters based on the latest state data, action data, and reward data. In the outer loop, the meta-parameters are updated by minimizing the meta-objective loss across tasks. The action probability distributions output by the policy network corresponding to the three updated subtasks are normalized, and the action probability distribution of the global scheduling strategy is obtained by fusion through a soft voting mechanism. The live content scheduling action is sampled and selected based on the action probability distribution, and the optimal content scheduling strategy is obtained by combination.

[0044] While executing the scheduling action, new live broadcast data is continuously collected, and the newly collected data is repeatedly input into the two-way interactive evaluation network for evaluation and optimization until the deviation value between the real-time performance evaluation index and the live broadcast peak period prediction result is less than the deviation threshold.

[0045] In an optional embodiment,

[0046] The gradient meta-learner adopts a model-independent meta-learning algorithm, uses a policy gradient method in the inner loop to update network parameters based on the latest state data, action data, and reward data, and minimizes the cross-task meta-objective loss in the outer loop to update the meta-parameters, including:

[0047] Construct a meta-learner network structure, set a policy network to generate action probability distribution, configure a meta-optimizer to adjust policy network parameters, implement the policy network as a deep neural network structure, and initialize meta-parameters shared by the policy network and the meta-optimizer;

[0048] For the pre-obtained subtasks, collect the historical status data, action data, reward data, and effect data of each type of subtask to construct a training task, and obtain new status data, action data, reward data, and effect data from the real-time live broadcast scene to construct a test task;

[0049] Sampling training tasks to obtain state transition data, copying meta-parameters as initial values ​​of task-specific parameters, starting the policy network based on the task-specific parameters, inputting the current state to generate an action probability distribution and sampling actions, executing the actions to obtain immediate rewards and next states, calculating the state value functions of the immediate rewards and next states, using the Bellman equation to calculate the temporal difference error to obtain the action value function, multiplying the action value function with the logarithm of the action probability output by the policy network and taking the negative value to obtain the policy loss, calculating the gradient of the policy loss with respect to the task-specific parameters, updating the task-specific parameters along the gradient direction, and repeatedly executing the policy network forward propagation and gradient update until the inner loop training is completed;

[0050] Calculate the cumulative reward difference of the task-specific parameters before and after each training task is updated on the corresponding task as a fitness metric, construct a meta-objective loss function by weighted average of the fitness metrics of all tasks, calculate the second-order derivative of the meta-objective loss function with respect to the meta-parameters, and drive the meta-optimizer to update the meta-parameters based on the second-order derivatives to minimize the meta-objective loss function;

[0051] Sample the test task, copy the trained meta-parameters as the initial values ​​of the task-specific parameters, and execute the inner loop update to obtain the policy network that adapts to the current task.

[0052] A second aspect of an embodiment of the present invention provides a private domain live broadcast peak period hotspot prediction and content scheduling system based on deep learning, including:

[0053] The first unit is used to collect private domain live broadcast history data, build a hybrid memory enhanced neural network including a fast memory path and a slow memory path, wherein the fast memory path performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory path captures long-term dependency through a hierarchical sparse attention structure, input the private domain live broadcast history data into the hybrid memory enhanced neural network, extract multi-scale temporal features through the fast memory path and the slow memory path, build a spatial pyramid attention module, perform feature recombination on the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, adjust the feature receptive field in combination with deformable convolution, obtain a hierarchical feature representation, input the hierarchical feature representation into a progressive feature distillation network, perform multi-level knowledge transfer and feature recombination operations, and generate a live broadcast peak prediction result;

[0054] The second unit is used to collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node association relationships, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through a goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output the live content scheduling strategy;

[0055] The third unit is used to collect real-time live broadcast data and add it to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between predicted effect and actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, an adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy, and the collection and adjustment are repeated to obtain the optimal content scheduling strategy.

[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] A fourth aspect of the embodiments of the present invention is:

[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 fast and slow memory pathways and a spatial pyramid attention module, the short-term and long-term dependencies in live broadcast data can be more effectively captured, thereby more accurately predicting the peak period of live broadcast, providing a reliable basis for subsequent content scheduling, and utilizing multimodal data, dynamic heterogeneous graph neural networks and hierarchical comparative learning to perform more refined scheduling of live broadcast content, and optimize it in the strategy space through a multi-agent reinforcement learning framework, thereby improving the attractiveness of the live broadcast content and user participation, using a two-way interactive evaluation network to perform real-time evaluation of the live broadcast effect, and using a meta-reinforcement learning algorithm to perform online optimization of the content scheduling strategy, so that the live broadcast content can be dynamically adjusted according to actual conditions to ensure that the live broadcast effect reaches the best state. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of a method for predicting hot spots and scheduling content during peak periods of private domain live broadcasts based on deep learning according to an embodiment of the present invention;

[0065] Figure 2 This is a structural diagram of a private domain live broadcast peak period hotspot prediction and content scheduling system based on deep learning 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 Schematic diagram of the process of the method for predicting hot spots and scheduling content during peak hours of private domain live broadcast based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] S1. Collect private domain live broadcast history data, build a hybrid memory enhanced neural network including a fast memory pathway and a slow memory pathway, wherein the fast memory pathway performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory pathway captures long-term dependencies through a hierarchical sparse attention structure, input the private domain live broadcast history data into the hybrid memory enhanced neural network, extract multi-scale temporal features through the fast memory pathway and the slow memory pathway, build a spatial pyramid attention module, perform feature reorganization on the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, adjust the feature receptive field in combination with deformable convolution, obtain a hierarchical feature representation, input the hierarchical feature representation into a progressive feature distillation network, perform multi-level knowledge transfer and feature reorganization operations, and generate a live broadcast peak prediction result;

[0070] In an optional embodiment,

[0071] Collecting private domain live broadcast history data, constructing a hybrid memory enhanced neural network including a fast memory pathway and a slow memory pathway, wherein the fast memory pathway performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory pathway captures long-term dependencies through a hierarchical sparse attention structure, inputting the private domain live broadcast history data into the hybrid memory enhanced neural network, extracting multi-scale temporal features through the fast memory pathway and the slow memory pathway, constructing a spatial pyramid attention module, and reorganizing the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, including:

[0072] The historical data of private domain live broadcast is sampled every minute, and the number of viewers, likes, comments, and reward amount data are collected. The weighted average calculation is performed on the number of viewers, likes, comments, and reward amount data to obtain historical trend characteristics;

[0073] A fast memory path and a slow memory path of a hybrid memory-enhanced neural network are set, a bidirectional gated recurrent unit network is configured in the fast memory path, a candidate state generation unit is set in the bidirectional gated recurrent unit network, and a nonlinear combination of a current input and a historical state is calculated by the candidate state generation unit to generate a candidate hidden state; the historical state is randomly forgotten by a reset gate, and the candidate hidden state is updated by an update gate;

[0074] A converter network is configured in the slow memory path, a multi-head self-attention layer is combined with a feedforward neural network layer to construct a computing block, a query vector, a key vector, and a value vector are projected into multiple feature subspaces in the multi-head self-attention layer, a scaled dot product calculation is performed in each feature subspace to obtain an attention weight, dropout regularization is performed on the attention weight, two layers of perceptrons are connected in series in the feedforward neural network layer, wherein the first layer of perceptrons performs dimensional expansion, and the second layer of perceptrons performs dimensional compression, and a normalization layer is added between the layers;

[0075] Performing an adaptive batch normalization operation on the private domain live broadcast history data and the historical trend features to obtain normalized data, reorganizing the normalized data into a three-dimensional array including a day dimension, a minute dimension, and an indicator dimension, inputting the three-dimensional array into the fast memory path, bidirectionally scanning along the day dimension to obtain local time series features, inputting the three-dimensional array into the slow memory path, adding time series position information through position coding, and obtaining global time series features;

[0076] Configure the first attention branch, the second attention branch and the third attention branch of the spatial pyramid attention module, set a sparse attention mechanism in the first attention branch, generate an adaptive attention mask, perform information aggregation of key temporal positions through Top-k selection, set a dense attention mechanism in the second attention branch, perform feature extraction of local areas through a deformable sliding window, set a hybrid attention mechanism in the third attention branch, cascade channel attention and spatial attention, extract cross-channel and cross-spatial dimension dependencies, align the local temporal features with the global temporal features in the time dimension and input them into the spatial pyramid attention module, and perform weighted fusion on the output features of the three attention branches to obtain multi-scale fusion features.

[0077] In an optional embodiment,

[0078] Get historical data from private live broadcast platforms, including the number of viewers, likes, comments, and reward amounts per minute. For example, within 10 minutes of a live broadcast, 100, 120, 150, 180, 200, 220, 250, 280, 300, and 320 people watched the live broadcast, and the number of likes was 10, 12, 15, 18, 20, 22, 25, 28, 30, and 32, respectively. The number of comments was 5, 6, 7, 8, 9, 10, 11, 12, 13, and 14, and the reward amounts were 10, 12, 15, 18, 20, 22, 25, 28, 30, and 32 yuan, respectively.

[0079] Perform weighted average calculation to obtain historical trend characteristics. For example, set the weight of the number of viewers to 0.5, the weight of the number of likes to 0.2, the weight of the number of comments to 0.2, and the weight of the amount of reward to 0.1. Then the historical trend characteristic value at the 10th minute is (320×0.5+32×0.2+14×0.2+32×0.1)=172.4.

[0080] Build a hybrid memory-enhanced neural network, including a fast memory path and a slow memory path. The fast memory path uses a bidirectional gated recurrent unit network to control the flow and forgetting of information through reset gates and update gates. For example, the reset gate determines how much previous memory to keep, and the update gate determines how much information of the current input needs to be added to the memory.

[0081] The slow memory path uses a transformer network consisting of a multi-head self-attention layer and a feed-forward neural network layer. The multi-head self-attention layer projects the input data into multiple feature subspaces and calculates attention weights in each subspace to capture long-term dependencies. The feed-forward neural network layer consists of two layers of perceptrons and inter-layer normalization layers for nonlinear transformation of features.

[0082] Adaptive batch normalization is performed on the collected private domain live broadcast historical data and the calculated historical trend features. For example, the number of viewers per minute, the number of likes, the number of comments, the amount of reward, and the historical trend feature values ​​are all normalized so that their value range is within a reasonable range, such as between 0 and 1.

[0083] The normalized data is reorganized into a three-dimensional array, including the day dimension, minute dimension, and indicator dimension. For example, the live broadcast data for 7 days, the number of viewers per minute, the number of likes, the number of comments, the amount of reward, and the historical trend characteristics are organized into a 7x1440x5 three-dimensional array (assuming 24 hours of live broadcast per day).

[0084] The three-dimensional array is input into the fast memory pathway and the slow memory pathway. The fast memory pathway scans bidirectionally along the day dimension to extract local temporal features. The slow memory pathway adds temporal position information through position encoding to extract global temporal features.

[0085] A spatial pyramid attention module is constructed, which contains three parallel attention branches: the first branch uses a sparse attention mechanism, the second branch uses a dense attention mechanism, and the third branch uses a hybrid attention mechanism. The local temporal features and the global temporal features are aligned in the time dimension and input into the spatial pyramid attention module. The output features of the three attention branches are weightedly fused to obtain multi-scale fusion features for predicting the live broadcast effect.

[0086] In this embodiment, by hybridizing the memory pathway and the spatial pyramid attention module, multi-scale temporal features can be extracted more effectively, thereby improving prediction accuracy. The application of techniques such as adaptive batch normalization and dropout regularization enhances the robustness of the model, making it more resistant to noise and abnormal data. Through the different attention branches of the spatial pyramid attention module, the impact of different time scales and spatial positions on the live broadcast effect can be analyzed, thereby improving the interpretability of the model.

[0087] In an optional embodiment,

[0088] Combined with deformable convolution to adjust the feature receptive field, a hierarchical feature representation is obtained, and the hierarchical feature representation is input into the progressive feature distillation network to perform multi-level knowledge transfer and feature reorganization operations to generate the live broadcast peak period prediction results including:

[0089] The offset prediction layer, the modulation factor prediction layer and the feature sampling layer of the multi-layer deformable convolutional network are configured, the two-dimensional offset of the sampling position of the pre-acquired multi-scale fusion feature is calculated in the offset prediction layer, the modulation factor of the sampling position is generated in the modulation factor prediction layer, the feature response is adaptively adjusted, the bilinear interpolation sampling is performed on the feature map in combination with the two-dimensional offset and the modulation factor in the feature sampling layer, the multi-scale fusion feature is sequentially input into each layer of the multi-layer deformable convolutional network, cross-layer connections are added between adjacent layers, and a hierarchical feature representation is generated;

[0090] A teacher network and a multi-level student network of a progressive feature distillation network are configured, parameters of the teacher network are initialized according to the hierarchical feature representation, each level of the student network corresponds to a feature layer of different depths, features of the student networks at adjacent levels are reused through dense residual connections, an attention distillation module is introduced into the teacher network, an attention map is used as an intermediate supervision signal, the hierarchical feature representation is input into the progressive feature distillation network, the L2 loss and the attention map loss of the feature maps of the teacher network and the student network are calculated, the KL divergence loss corresponding to the soft label distribution is calculated, a total loss is obtained by weighted combination of multiple loss items, network parameters are optimized based on the total loss, and a distilled feature representation is generated;

[0091] The distilled feature representation is input into a multilayer perceptron, in which the feature dimension is reduced layer by layer, and in the last layer the output dimension is set to the number of discrete time periods in the prediction time window; the output is normalized by the Softmax function to obtain the peak probability of each time period, and the optimal threshold is determined by combining the dynamic probability threshold selection algorithm, and the time period above the optimal threshold is marked as the predicted peak period.

[0092] In an optional embodiment,

[0093] Prepare training data and collect historical live broadcast data, including multi-dimensional data such as the number of viewers, number of bullet comments, number of likes, number of shares, and corresponding timestamps. Divide according to the predetermined time window, for example, 15 minutes as a time window, and count the various indicator data in each time window. According to the peak number of viewers in each time window in the historical data, set a threshold, mark the time window above the threshold as the peak period, and mark the time window below the threshold as the off-peak period. In this way, construct a training data set, which contains multi-dimensional feature data and corresponding labels (peak or off-peak).

[0094] A multi-layer deformable convolutional network is constructed, which includes multiple deformable convolutional layers. Each layer consists of an offset prediction layer, a modulation factor prediction layer, and a feature sampling layer. First, the multi-scale fusion feature is input into the first deformable convolutional layer. In the offset prediction layer, the network calculates the two-dimensional offset of each sampling point based on the input multi-scale fusion feature. The offset represents the offset position of the sampling point on the feature map. In the modulation factor prediction layer, the network generates a modulation factor for each sampling position to adjust the intensity of the feature response. The modulation factor ranges from 0 to 1, and the larger the value, the higher the importance of the sampling point. In the feature sampling layer, the network performs bilinear interpolation sampling on the feature map based on the calculated two-dimensional offset and modulation factor. Specifically, the network finds a new sampling position based on the offset and performs weighted summation of the sampling values ​​based on the modulation factor. The output of the first deformable convolutional layer is input into the next deformable convolutional layer, and the above operation is repeated. Between adjacent deformable convolutional layers, cross-layer connections are added to fuse the output of the previous layer with the input of the current layer to generate a hierarchical feature representation. For example, assuming the input feature map size is 256x256, the offset output by the offset prediction layer is (0.5, -0.2), and the modulation factor is 0.8, the actual sampling position is (256×0.5, 256×(-0.2)) = (128, -51.2). Since the offset can be a decimal, bilinear interpolation is needed to calculate the new sampling value.

[0095] Construct a progressive feature distillation network, which consists of a teacher network and multiple student networks. Use the previously generated hierarchical feature representation to initialize the parameters of the teacher network. Each student network corresponds to a feature layer of different depth. For example, the first student network corresponds to shallow features, the second student network corresponds to intermediate features, and so on. Adjacent levels of student networks reuse features through dense residual connections, and the output of the previous student network is used as part of the input of the next student network. Introduce the attention distillation module in the teacher network, and use the attention map generated by the teacher network as an intermediate supervision signal to guide the learning of the student network. Input the hierarchical feature representation into the progressive feature distillation network, and calculate the sum of squares of the differences between the feature maps of the teacher network and each student network, as well as the sum of squares of the differences in the attention maps. At the same time, calculate the KL divergence between the soft label distribution output by the teacher network and the soft label distribution output by the student network. These loss terms are weighted and combined to obtain the total loss. For example, the feature map loss weight is 0.5, the attention map loss weight is 0.3, and the soft label loss weight is 0.2. Based on the total loss, use the back propagation algorithm to optimize the network parameters and generate the distilled feature representation.

[0096] The distilled feature representation is input into a multi-layer perceptron. In the multi-layer perceptron, the feature dimension is reduced layer by layer. For example, the first layer reduces the feature dimension to 256, the second layer reduces it to 128, and so on. In the last layer, the output dimension is set to the number of discrete time periods in the prediction time window, for example, if a day is divided into 96 15-minute time windows, the output dimension is 96. The output is normalized using the Softmax function to obtain the peak probability of each time window. Combined with the dynamic probability threshold selection algorithm, for example, according to the distribution of peak probability statistics of historical data, the optimal threshold is adaptively determined. The time window above the optimal threshold is marked as the predicted peak period. For example, assuming the prediction result is [0.1, 0.2, 0.8, 0.3, ..., 0.9], the optimal threshold is 0.7, then the predicted peak period is the third and last time window.

[0097] In this embodiment, the deformable convolution can adaptively adjust the receptive field to capture the dynamic change pattern in the live broadcast data, the progressive feature distillation can effectively transfer knowledge and improve the generalization ability of the model, thereby improving the accuracy of peak period prediction, the multi-scale feature fusion can better cope with the diversity and complexity of the live broadcast data, the attention distillation module can guide the model to focus on important feature information and enhance the robustness of the model, and the progressive feature distillation network can reduce the computational complexity of the model while ensuring the prediction performance and improve the prediction efficiency.

[0098] S2. Collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node association relationships, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through the goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output the live content scheduling strategy;

[0099] In an optional embodiment,

[0100] Collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node associations, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through the goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output the live content scheduling strategy, including:

[0101] Perform feature extraction on each frame in the video stream data through a feature extraction network, the feature extraction network includes a feature pyramid network and an attention enhancement module, extract multi-scale visual features through the feature pyramid network, perform feature enhancement in the channel dimension and the spatial dimension through the attention enhancement module, perform short-time Fourier transform on the audio data in combination with the multi-head self-attention network to obtain a spectrogram, input the spectrogram into a deep network to extract audio features, use a bidirectional encoding network to perform sequence modeling on the barrage text and the comment text, and map the word vector sequence into a text feature vector; input the visual features, the audio features, and the text feature vector into a temporal encoding network, perform temporal encoding to obtain a multimodal feature tensor;

[0102] Construct a dynamic heterogeneous graph neural network, take anchor nodes, audience nodes, video clip nodes, audio clip nodes, bullet screen nodes, and comment nodes as graph nodes, assign the multimodal feature tensor to the corresponding graph nodes, calculate the similarity between node pairs through a topology generator, perform a normalization transformation on the similarity to obtain edge weights, construct a node adjacency matrix in combination with a sparse matrix representation, and use a feature matrix to store node attributes to obtain a feature representation of the node;

[0103] The dynamic heterogeneous graph neural network is dynamically updated. When new data arrives, the corresponding modal features are extracted and new nodes are created. The feature similarity between the newly added nodes and the existing nodes corresponding to the new data is calculated. Connecting edges are added between node pairs whose similarity is greater than a preset first threshold. The product of the degree centrality of the existing nodes and the time decay factor is calculated. Nodes whose product is less than a preset second threshold are deleted. The exponential sliding average of the edge weights is calculated. The edges whose sliding average is less than a preset third threshold are deleted to obtain an optimized graph structure.

[0104] The feature matrix and the node adjacency matrix are input into a multi-layer graph neural network through a feature interaction engine, wherein the feature matrix is ​​composed of a multimodal feature tensor of nodes, representing the initial features of each node, and a learnable parameter matrix is ​​constructed according to the input feature dimension and the hidden layer dimension. First, the input dimension and output dimension of each layer of the graph neural network are set, and the learnable parameter matrix is ​​initialized using a uniform distribution or a normal distribution. The learnable parameter matrix is ​​used as the weight matrix of each layer of the graph neural network, and the node adjacency matrix is ​​degree-normalized to obtain a symmetric normalized matrix. The normalized matrix is ​​multiplied by the node feature matrix of the current layer, and then multiplied by the learnable parameter matrix. A nonlinear activation function is used to obtain the symmetric normalized matrix. Get the node features of the next layer, apply the hierarchical contrast learning method in each layer, construct positive sample pairs and negative sample pairs, the positive sample pairs are sampled from adjacent nodes, and the negative sample pairs are sampled from non-adjacent nodes at different semantic levels, calculate the feature contrast loss, and perform weighted average of the contrast losses at different levels to obtain the total loss. Input the node representation output by the graph neural network into the hierarchical strategy optimization network, set the interaction quality improvement sub-goal, content form optimization sub-goal, and user experience optimization sub-goal in the goal decomposition module, calculate the sub-goal weights, and set the resource scheduling sub-network in the execution planning module, wherein the resource scheduling sub-network includes a multi-layer encoder, and the output of the encoder is mapped to obtain the scheduling action space;

[0105] The nodes in the dynamic heterogeneous graph neural network are mapped to multiple agents. Each agent maintains a local state representation and a policy network. The policy network consists of an action generation network and a state evaluation network. The action generation network outputs the action probability distribution, and the state evaluation network evaluates the state value function. The agents exchange information through a competitive cooperation mechanism, optimize strategies based on non-cooperative games, and calculate rewards based on the joint actions of the agents. The reward function comprehensively considers multiple indicators such as interactivity, content quality, and user feedback to obtain a comprehensive evaluation score.

[0106] Each agent generates a probability distribution of live content scheduling strategy based on its strategy network, and performs parameterized sampling on the probability distribution to obtain discrete actions. The discrete actions of all agents are combined into a live content scheduling strategy, including the host interaction instruction sequence, the audience activity push sequence, and the platform resource allocation sequence. The structure and node characteristics of the dynamic heterogeneous graph neural network are updated according to the feedback after executing the scheduling strategy.

[0107] In an optional embodiment,

[0108] Collect multimodal data of live broadcast, including video stream, audio stream, bullet text and comment text. For video stream, use feature pyramid network to extract multi-scale visual features. For example, for a 1080p video frame, you can extract feature maps of different scales such as 256x144, 512x288, 1024x576, etc. to capture visual information of different granularity, and then use attention enhancement module to enhance features in channel dimension and spatial dimension. For example, through attention mechanism such as SE-Net, enhance the features of important channels and regions. For audio stream, perform short-time Fourier transform to obtain spectrogram, and then input it into deep network (such as CNN) to extract audio features. For example, convert audio clips into Mel spectrogram, and then input it into deep network to extract feature vector. For bullet and comment text, use bidirectional encoding network (such as BERT) for sequence modeling, and map word vector sequence to text feature vector. For example, after each bullet text is segmented, each word is converted into word vector, and then input into bidirectional encoding network model to obtain feature representation of text. The extracted visual features, audio features, and text features are input into the temporal coding network, and temporal coding is performed to obtain a multimodal feature tensor. For example, the features of 5 consecutive seconds of video, audio, and text are concatenated and input into the temporal coding network to obtain a feature tensor that integrates temporal information.

[0109] Construct a dynamic heterogeneous graph neural network. Take the anchor, audience, video clips, audio clips, bullet comments and other elements as graph nodes, and assign the extracted multimodal feature tensors to the corresponding graph nodes. For example, the feature of a video clip node is the multimodal feature tensor of the clip. Calculate the similarity between node pairs through the topology generator. For example, calculate the cosine similarity of the visual features between two video clip nodes. Perform a normalization transformation on the similarity to obtain the edge weight. For example, normalize the similarity value to between 0 and 1. Construct a node adjacency matrix in combination with sparse matrix representation. For example, if the similarity between two nodes is greater than 0.8, the value of the corresponding position in the adjacency matrix is ​​set to 1, otherwise it is set to 0. Use the feature matrix to store node attributes to obtain the feature representation of the node.

[0110] Dynamically update the graph neural network structure. When new data arrives, extract the corresponding modal features and create new nodes. Calculate the feature similarity between the newly added nodes and the existing nodes. For example, calculate the similarity between the new video clip and the existing video clip. Add connecting edges between node pairs whose similarity is greater than the preset first threshold (for example, 0.7). Calculate the product of the degree centrality of existing nodes and the time decay factor. For example, if a node has many connecting edges and the time is relatively long, the product will be relatively small. Delete nodes whose product is less than the preset second threshold (for example, 0.1). Calculate the exponential sliding average of edge weights. For example, record the historical weight of each edge and calculate its weighted average. Delete edges whose sliding average is less than the preset third threshold (for example, 0.2) to obtain the optimized graph structure.

[0111] The feature matrix and node adjacency matrix are input into the multi-layer graph neural network through the feature interaction engine. The feature matrix is ​​composed of the multimodal feature tensors of the nodes, representing the initial features of each node. The learnable parameter matrix is ​​constructed according to the input feature dimension and the hidden layer dimension. For example, if the input feature dimension is 1024 and the hidden layer dimension is 512, the dimension of the parameter matrix is ​​1024x512. Initialize the parameter matrix. Use the parameter matrix as the weight matrix of each layer of the graph neural network. Degree normalize the node adjacency matrix. Multiply the normalized matrix with the node feature matrix of the current layer, and then multiply it with the learnable parameter matrix, and obtain the node features of the next layer through a nonlinear activation function (such as ReLU). Apply a hierarchical contrastive learning method at each layer. Sample positive sample pairs from adjacent nodes and negative sample pairs from non-adjacent nodes at different semantic levels. Calculate the feature contrast loss. Weighted average the contrast losses at different levels to get the total loss.

[0112] The node representation output by the graph neural network is input into the hierarchical strategy optimization network. In the goal decomposition module, sub-goals such as interaction quality improvement, content form optimization, and user experience optimization are set, and the sub-goal weights are calculated. For example, according to the actual situation of the current live broadcast, the weight of interaction quality improvement can be set to 0.5, the weight of content form optimization can be set to 0.3, and the weight of user experience optimization can be set to 0.2. In the execution planning module, a resource scheduling subnetwork is set, which includes a multi-layer encoder. The output of the encoder is mapped to obtain the scheduling action space.

[0113] Map the nodes in the graph neural network to multiple agents. Each agent maintains a local state representation and a policy network. The policy network consists of an action generation network and a state evaluation network. The action generation network outputs the action probability distribution, and the state evaluation network evaluates the state value function. Agents exchange information through a competitive cooperation mechanism and optimize strategies based on non-cooperative games. Rewards are calculated based on the joint actions of the agents. The reward function takes into account multiple indicators such as interactivity, content quality, and user feedback to obtain a comprehensive evaluation score. For example, if the interactivity increases, the content quality improves, and the user feedback is positive, the reward score will be higher.

[0114] Each agent generates a probability distribution of the live content scheduling strategy based on its policy network, and performs parameterized sampling to obtain discrete actions. The discrete actions of all agents are combined into a live content scheduling strategy, including the host interaction instruction sequence, the audience activity push sequence, and the platform resource allocation sequence. For example, instruction sequences such as "the host asks questions to the audience", "pushes coupons to active viewers", and "recommends the live broadcast to the homepage" can be generated according to the strategy. The structure and node features of the graph neural network are updated according to the feedback after executing the scheduling strategy.

[0115] In this embodiment, through multimodal data analysis and multi-agent reinforcement learning, user interests and behavior patterns can be captured more accurately, so as to formulate more effective interaction strategies and improve the quality of live interaction. By optimizing the network through dynamic heterogeneous graph neural networks and hierarchical strategies, the live content arrangement can be dynamically adjusted according to the real-time status of the live broadcast and user feedback, making the live content more attractive and interesting. Through personalized content recommendations and interaction methods, the needs of different users can be met, and the user experience and satisfaction can be improved.

[0116] In an optional embodiment,

[0117] The agents exchange information through a competitive cooperation mechanism, optimize strategies based on non-cooperative games, and calculate rewards based on the joint actions of the agents, including:

[0118] An agent is deployed at each node of a dynamic heterogeneous graph, wherein the nodes of the dynamic heterogeneous graph include video clip nodes, audio clip nodes, and bullet-screen nodes. The state space of the agent is composed of the attribute feature vector of the corresponding node and the attribute feature vector of the adjacent node. The state vector of the video clip node includes the visual features of the video frame, the acoustic features of the audio clip, and the text features of the surrounding bullet-screen nodes. The action space of the agent includes picture clarity adjustment, interactive question insertion, and related video push.

[0119] The agent broadcasts its own action probability distribution to neighboring nodes at each decision time step, receives the action probability distribution of neighboring nodes, and uses the received action probability distribution as an additional input signal to update the corresponding strategy network, wherein the strategy network includes an action generation network and a value evaluation network;

[0120] Strategy optimization is performed based on a non-cooperative game framework. The objective functions of different types of nodes in the non-cooperative game include the audience participation of the anchor node, the viewing time of the video node, and the number of interactions of the bullet comment node. The agent takes its own actions to obtain immediate rewards based on the observed state and the actions of other agents, and receives trajectory data of the interaction between other agents and the environment to estimate the joint action value function;

[0121] Iteratively update the joint action value function, calculate the value function of the current agent state and the joint action, and perform weighted average of the difference between the value function and the sum of the instant reward and the discounted value of the state value function at the next moment to obtain the updated value function, wherein the instant reward is obtained according to the environment, and the weight of the weighted average is the learning rate;

[0122] The policy function is optimized based on the policy gradient algorithm, and the Monte Carlo estimation method is used to calculate the policy gradient of each agent's objective function. The product of the policy gradient and the learning rate is added to the parameters of the agent's policy function to obtain the updated policy function parameters. The agent generates a scheduling strategy based on the optimized policy function, and performs picture clarity adjustment, interactive question insertion, and related video push according to the scheduling strategy.

[0123] In an optional embodiment,

[0124] Construct a dynamic heterogeneous graph. The node types of the graph include video clip nodes, audio clip nodes, and bullet comment nodes. An agent is deployed at each node. The state vector of the video clip node is composed of the visual features of the video frame (e.g., brightness, color histogram, motion vector, etc.), the acoustic features of the audio clip (e.g., volume, pitch, MFCC coefficient, etc.), and the text features of the surrounding bullet comment nodes (e.g., sentiment tendency, keywords, word embedding vectors, etc.). The state vectors of the audio clip node and the bullet comment node are constructed in a similar way, each containing its own attribute features and the attribute features of the adjacent nodes. The action space of the agent includes picture clarity adjustment (e.g., improving clarity, reducing clarity), interactive question insertion (e.g., inserting questions related to the video content), and related video push (e.g., pushing other videos related to the current video topic).

[0125] Agents exchange information and update strategies. At each decision time step, each agent broadcasts its own action probability distribution to neighboring nodes and receives action probability distributions from neighboring nodes. These received action probability distributions will be used as additional input signals to update the agent's own policy network. The policy network consists of an action generation network and a value evaluation network. The action generation network outputs the agent's own action probability distribution based on the agent's state and the action probability distribution of neighboring nodes. The value evaluation network evaluates the long-term value of taking a specific action in the current state.

[0126] Strategy optimization is performed based on a non-cooperative game framework. Different types of nodes have different objective functions. For example, the objective function of the anchor node is to maximize audience engagement (e.g., the number of likes, comments, gifts, etc.), the objective function of the video node is to maximize viewing time, and the objective function of the barrage node is to maximize the number of interactions (e.g., the number of replies, likes, etc.). Each agent takes its own actions based on the observed state and the actions of other agents, and receives immediate rewards. At the same time, the agent also receives trajectory data of other agents interacting with the environment to estimate the joint action value function. For example, a video clip node can observe the user's behavior after watching the clip (e.g., whether to continue watching, whether to send barrages, etc.), as well as the states and actions of other related nodes (e.g., audio clip nodes, barrage nodes), to estimate the joint action value function.

[0127] Iterate and update the joint action value function. Specifically, calculate the value function of the current agent state and the joint action, and perform weighted average of the difference between the value function and the sum of the discounted value of the immediate reward and the value function of the next moment state to obtain the updated value function. The weight of the weighted average is the learning rate, for example 0.01. Assuming that the value function of the current state is 10, the immediate reward is 2, the value function of the next moment state is 12, the discount factor is 0.9, and the learning rate is 0.01, the updated value function is 10+0.01×(2+0.9×12-10)=10.128.

[0128] The policy function is optimized based on the policy gradient algorithm, and the Monte Carlo estimation method is used to calculate the policy gradient of each agent's objective function. The product of the policy gradient and the learning rate is added to the parameters of the agent's policy function to obtain the updated policy function parameters. For example, if the policy gradient is 0.5 and the learning rate is 0.01, the parameters of the policy function will increase by 0.005. The agent generates a scheduling strategy based on the optimized policy function, and performs operations such as adjusting the picture clarity, inserting interactive questions, and pushing related videos according to the scheduling strategy. For example, if the policy function indicates that the picture clarity should be improved, the agent will perform operations to improve the picture clarity.

[0129] In this embodiment, by intelligently adjusting the picture clarity, inserting interactive questions and pushing related videos, the user's viewing experience can be improved and user stickiness can be increased. Personalized recommendations can be made based on user behavior and preferences to improve content distribution efficiency and push more suitable content to more interested users. By increasing audience engagement, viewing time and number of interactions, platform operations can be optimized and the overall benefits of the platform can be improved.

[0130] S3. Collect real-time live broadcast data and add it to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between predicted effect and actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, an adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy. Collection and adjustment are repeated to obtain the optimal content scheduling strategy.

[0131] In an optional embodiment,

[0132] Collect real-time live broadcast data and add it to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between the predicted effect and the actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, the adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy. Repeated collection and adjustment to obtain the optimal content scheduling strategy includes:

[0133] Adding the collected real-time video stream, real-time audio stream, and real-time text stream data to a pre-built two-way interactive evaluation network, wherein the two-way interactive evaluation network includes a state encoder and an effect evaluator, wherein the state encoder includes a video feature extraction module, an audio feature extraction module, and a text feature extraction module, and the effect evaluator includes a causal graph construction module, a structural equation model, and a multi-layer perceptron network;

[0134] Extracting frame-level visual features from a real-time video stream through a convolutional neural network in the video feature extraction module, extracting audio spectrogram features from a real-time audio stream through a convolutional neural network in the audio feature extraction module, extracting semantic features from a real-time text stream through a recurrent neural network in the text feature extraction module, inputting the frame-level visual features, the audio spectrogram features, and the semantic features into an attention network to generate a feature weight matrix, and performing weighted fusion on the three features based on the feature weight matrix to obtain a multimodal representation vector;

[0135] The causal graph construction module constructs a live broadcast effect causal graph based on historical data, the live broadcast effect causal graph includes a scheduling strategy node, a live broadcast status node, an external factor node and an effect indicator node, the edges between the nodes represent causal relationships, the structural equation model defines the functional relationship between the nodes, the multimodal representation vector is spliced ​​with the preset scheduling strategy vector to form an input vector, causal reasoning is performed based on the causal graph and the structural equation model, the policy intervention effect is obtained, the input vector and the policy intervention effect are input into the multi-layer perceptron network, the multi-layer perceptron network calculates the predicted number of viewers, the predicted number of likes, and the predicted reward amount, counts the real-time number of viewers, the real-time number of likes, and the real-time reward amount, and calculates the weighted average deviation between the predicted value and the actual value to obtain the real-time performance evaluation index;

[0136] Calculate the deviation value between the real-time performance evaluation index and the stored live broadcast peak period prediction result, compare the deviation value with a preset deviation threshold, and trigger strategy adjustment when the deviation value is greater than the deviation threshold;

[0137] The live content scheduling task is decomposed into content recommendation subtask, anchor interaction subtask, and audience activity subtask. Each subtask is configured with an independent policy network and value function network. The policy network generates an action probability distribution based on the current state, and the value function network evaluates the value of the state-action pair. A gradient meta-learner is trained for each subtask. The gradient meta-learner adopts a model-independent meta-learning algorithm. In the inner loop, the policy gradient method is used to update the network parameters based on the latest state data, action data, and reward data. In the outer loop, the meta-parameters are updated by minimizing the meta-objective loss across tasks. The action probability distributions output by the policy network corresponding to the three updated subtasks are normalized, and the action probability distribution of the global scheduling strategy is obtained by fusion through a soft voting mechanism. The live content scheduling action is sampled and selected based on the action probability distribution, and the optimal content scheduling strategy is obtained by combination.

[0138] While executing the scheduling action, new live broadcast data is continuously collected, and the newly collected data is repeatedly input into the two-way interactive evaluation network for evaluation and optimization until the deviation value between the real-time performance evaluation index and the live broadcast peak period prediction result is less than the deviation threshold.

[0139] In an optional embodiment,

[0140] Build a two-way interactive evaluation network, which consists of two parts: a state encoder and an effect evaluator. The state encoder is responsible for extracting features from multi-source data. It uses a convolutional neural network to extract visual features of each frame from the real-time video stream, such as color, texture, edge and other information. Similarly, a convolutional neural network is used to extract audio spectrogram features from the real-time audio stream, such as pitch, loudness, rhythm and other information. A recurrent neural network is used to extract semantic features from real-time text streams (such as bullet screens and comments), such as user emotions, topic keywords and other information, and the extracted visual features, audio features and semantic features are input into an attention network. The attention network generates a feature weight matrix based on the importance of different features, and performs weighted fusion of the three features based on the weight matrix to obtain a multimodal representation vector that comprehensively reflects the current live broadcast status.

[0141] The effect evaluator is responsible for evaluating the effect of the current live content scheduling strategy and constructing a live effect causal graph based on historical live data. The causal graph contains scheduling strategy nodes (such as game sessions, talent shows, and interactions with the audience), live status nodes (such as the number of online users, the number of likes, and the number of comments), external factor nodes (such as competitors' live content and holiday effects), and effect indicator nodes (such as the number of viewers, the number of likes, and the amount of rewards). The edges between nodes represent causal relationships. For example, the scheduling strategy node "game session" will affect the live status node "number of online users", and the external factor node "holiday effect" will also affect the "number of online users". The structural equation model is used to define the functional relationship between the nodes and quantify the degree of influence of the causal relationship. The multimodal representation vector output by the state encoder is concatenated with the preset scheduling strategy vector to form an input vector. Based on the constructed causal graph and structural equation model, causal reasoning is performed to obtain the effect of the strategy intervention, that is, to predict the impact of the strategy on the effect indicator. The input vector and the strategy intervention effect are input into a multi-layer perceptron network, which will output the predicted number of viewers, the predicted number of likes, and the predicted amount of rewards. At the same time, the real-time number of viewers, the number of likes, and the amount of rewards are counted. By calculating the weighted average deviation between the predicted value and the actual value, the real-time performance evaluation index is obtained.

[0142] Compare the real-time performance evaluation index with the deviation value of the pre-stored live peak period prediction result. The pre-stored live peak period prediction result is the live peak period effect index predicted based on historical data. If the deviation value between the real-time performance evaluation index and the live peak period prediction result is greater than the pre-set deviation threshold, the policy adjustment mechanism is triggered.

[0143] The policy adjustment mechanism is based on the fast adaptation algorithm of meta-reinforcement learning, which decomposes the live content scheduling task into content recommendation subtasks, anchor interaction subtasks, and audience activity subtasks. Each subtask is configured with an independent policy network and value function network. The policy network generates an action probability distribution based on the current state. For example, in the content recommendation subtask, the policy network outputs the probability of recommending different content. The value function network evaluates the value of the state-action pair, that is, predicts the long-term benefits that can be obtained after taking a certain action, and trains a gradient meta-learner for each subtask. The gradient meta-learner uses a model-independent meta-learning algorithm. In the inner loop, the policy gradient method is used to update the parameters of the policy network and the value function network based on the latest state data, action data, and reward data (such as the number of viewers, the number of likes, and the amount of rewards). In the outer loop, the meta-objective loss across tasks is minimized, and the meta-parameters are updated so that the model can quickly adapt to new tasks. The action probability distributions output by the updated policy network corresponding to the three subtasks are normalized. The policies of the three subtasks are fused through the soft voting mechanism to obtain the action probability distribution of the global scheduling strategy. Based on the probability distribution sampling, the live content scheduling action is selected and combined to obtain the optimal content scheduling strategy. For example, a certain game content can be recommended according to the global strategy, and the anchor can interact with it accordingly, and relevant audience activities can be carried out at the same time.

[0144] While executing the scheduling action, new live broadcast data is continuously collected. The newly collected data is repeatedly input into the two-way interactive evaluation network for evaluation and optimization until the deviation value between the real-time performance evaluation index and the live broadcast peak prediction result is less than the preset deviation threshold. For example, in a certain period of time, the real-time number of viewers is 1000, the real-time number of likes is 5000, and the real-time reward amount is 1000 yuan. The predicted number of viewers is 1200, the predicted number of likes is 6000, and the predicted reward amount is 1200 yuan. Assuming that the weights are all 0.33, the deviation value of the real-time performance evaluation index is (|1000-1200|+|5000-6000|+|1000-1200|) / 3×0.33=133.33. Assuming that the live broadcast peak prediction results are 1500, 10000, 2000, and the deviation threshold is 200, the current deviation value is less than the threshold, and there is no need to adjust the strategy.

[0145] In this embodiment, through cross-modal information fusion and causal reasoning, the factors affecting the live broadcast effect are considered more comprehensively, thereby improving the accuracy of the live broadcast effect prediction. The meta-reinforcement learning algorithm can quickly adjust the strategy according to real-time data, so that the live broadcast content scheduling plan can adapt to the ever-changing live broadcast environment and user preferences. By dynamically optimizing the live broadcast content scheduling strategy, it can better meet user needs, improve the attractiveness of the live broadcast content and user participation, and ultimately improve the overall effect of the live broadcast.

[0146] In an optional embodiment,

[0147] The gradient meta-learner adopts a model-independent meta-learning algorithm, uses a policy gradient method in the inner loop to update network parameters based on the latest state data, action data, and reward data, and minimizes the cross-task meta-objective loss in the outer loop to update the meta-parameters, including:

[0148] Construct a meta-learner network structure, set a policy network to generate action probability distribution, configure a meta-optimizer to adjust policy network parameters, implement the policy network as a deep neural network structure, and initialize meta-parameters shared by the policy network and the meta-optimizer;

[0149] For the pre-obtained subtasks, collect the historical status data, action data, reward data, and effect data of each type of subtask to construct a training task, and obtain new status data, action data, reward data, and effect data from the real-time live broadcast scene to construct a test task;

[0150] Sampling training tasks to obtain state transition data, copying meta-parameters as initial values ​​of task-specific parameters, starting the policy network based on the task-specific parameters, inputting the current state to generate an action probability distribution and sampling actions, executing the actions to obtain immediate rewards and next states, calculating the state value functions of the immediate rewards and next states, using the Bellman equation to calculate the temporal difference error to obtain the action value function, multiplying the action value function with the logarithm of the action probability output by the policy network and taking the negative value to obtain the policy loss, calculating the gradient of the policy loss with respect to the task-specific parameters, updating the task-specific parameters along the gradient direction, and repeatedly executing the policy network forward propagation and gradient update until the inner loop training is completed;

[0151] Calculate the cumulative reward difference of the task-specific parameters before and after each training task is updated on the corresponding task as a fitness metric, construct a meta-objective loss function by weighted average of the fitness metrics of all tasks, calculate the second-order derivative of the meta-objective loss function with respect to the meta-parameters, and drive the meta-optimizer to update the meta-parameters based on the second-order derivatives to minimize the meta-objective loss function;

[0152] Sample the test task, copy the trained meta-parameters as the initial values ​​of the task-specific parameters, and execute the inner loop update to obtain the policy network that adapts to the current task.

[0153] In an optional embodiment,

[0154] Construct a meta-learner network structure. The policy network uses a deep neural network structure to output action probability distribution. The meta-optimizer is used to adjust the policy network parameters. The policy network and meta-optimizer share a set of initial meta-parameters, such as the network weights and biases.

[0155] Prepare training and test tasks. For a specific type of decision-making problem, such as game control or robot navigation, collect historical data of multiple subtasks, including state, action, reward, and effect data, and build a set of training tasks. At the same time, obtain new state, action, reward, and effect data from the real-time live scene to build a set of test tasks. For example, in a game scene, the state can be the pixels of the game screen, the action can be the direction of controlling the character's movement, the reward can be the game score, and the effect can be the change in the character's health. Suppose we collect historical data of 100 subtasks as a training set, and each subtask contains 1,000 state transition data.

[0156] Perform meta-learning training, randomly sample a task from the training task set, and copy the current meta parameters as the task-specific initial parameters. Initialize the policy network with task-specific parameters. Input the current state into the policy network, obtain the action probability distribution, and sample an action according to the distribution. After executing the action, the environment returns the immediate reward and the next state. Use a value network to evaluate the value of the current state and the next state, and use the Bellman equation to calculate the temporal difference error to obtain the value of the action. Multiply the value of the action by the logarithm of the corresponding action probability output by the policy network, and take the negative value to obtain the policy loss. Calculate the gradient of the policy loss with respect to the task-specific parameters, and update the task-specific parameters along the gradient direction. Repeat the forward propagation and parameter update process of the policy network, for example, 10 iterations, to complete the inner loop training. Assume that the learning rate is set to 0.001.

[0157] After completing the inner loop update of a training task, calculate the cumulative reward difference of the task-specific parameters before and after the task update on the task as the fitness measure of the task. Take a weighted average of the fitness measures of all sampled training tasks to construct a meta-objective loss function. Calculate the second-order derivative of the meta-objective loss function with respect to the meta-parameters, for example, using the Hessian vector product method. Drive a meta-optimizer, such as the Adam optimizer, to update the meta-parameters based on the calculated second-order derivatives to minimize the meta-objective loss function.

[0158] To test, randomly sample a task from the test task set and copy the trained meta-parameters as the initial parameters specific to the task. Perform the same inner loop update process as in the training phase to obtain a policy network that adapts to the current test task.

[0159] In this embodiment, the model can quickly adapt to new tasks and learn effective strategies with only a small amount of data. For example, when faced with a new game level, the model can quickly learn a strategy to pass the level. Through the meta-learning process, the model can learn cross-task knowledge, thereby improving the generalization performance on unseen tasks, and can effectively use a small amount of samples for learning, reducing the cost of data collection and training.

[0160] Figure 2 Schematic diagram of the structure of a system for predicting hot spots and scheduling content during peak hours of private domain live broadcast based on deep learning according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0161] The first unit is used to collect private domain live broadcast history data, build a hybrid memory enhanced neural network including a fast memory path and a slow memory path, wherein the fast memory path performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory path captures long-term dependency through a hierarchical sparse attention structure, input the private domain live broadcast history data into the hybrid memory enhanced neural network, extract multi-scale temporal features through the fast memory path and the slow memory path, build a spatial pyramid attention module, perform feature recombination on the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, adjust the feature receptive field in combination with deformable convolution, obtain a hierarchical feature representation, input the hierarchical feature representation into a progressive feature distillation network, perform multi-level knowledge transfer and feature recombination operations, and generate a live broadcast peak prediction result;

[0162] The second unit is used to collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node association relationships, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through a goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output the live content scheduling strategy;

[0163] The third unit is used to collect real-time live broadcast data and add it to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between predicted effect and actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, an adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy, and the collection and adjustment are repeated to obtain the optimal content scheduling strategy.

[0164] According to a third aspect of the embodiments of the present invention,

[0165] An electronic device is provided, comprising:

[0166] processor;

[0167] a memory for storing processor-executable instructions;

[0168] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0169] A fourth aspect of the embodiments of the present invention is:

[0170] 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.

[0171] 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.

[0172] 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 predicting hot spots and scheduling content during peak periods of private domain live broadcasts based on deep learning, characterized in that: include: Collect private domain live broadcast history data, build a hybrid memory enhanced neural network including a fast memory pathway and a slow memory pathway, the fast memory pathway performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory pathway captures long-term dependencies through a hierarchical sparse attention structure, input the private domain live broadcast history data into the hybrid memory enhanced neural network, extract multi-scale temporal features through the fast memory pathway and the slow memory pathway, build a spatial pyramid attention module, perform feature reorganization on the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, adjust the feature receptive field in combination with deformable convolution, obtain a hierarchical feature representation, input the hierarchical feature representation into a progressive feature distillation network, perform multi-level knowledge transfer and feature reorganization operations, and generate a live broadcast peak prediction result; Collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node association relationships, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through a goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output a live content scheduling strategy; Real-time live broadcast data is collected and added to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between predicted effect and actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, an adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy. Collection and adjustment are repeated to obtain the optimal content scheduling strategy.

2. The method according to claim 1, characterized in that Collecting private domain live broadcast history data, constructing a hybrid memory enhanced neural network including a fast memory pathway and a slow memory pathway, wherein the fast memory pathway performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory pathway captures long-term dependencies through a hierarchical sparse attention structure, inputting the private domain live broadcast history data into the hybrid memory enhanced neural network, extracting multi-scale temporal features through the fast memory pathway and the slow memory pathway, constructing a spatial pyramid attention module, and reorganizing the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, including: The historical data of private domain live broadcast is sampled every minute, and the number of viewers, likes, comments, and reward amount data are collected. The weighted average calculation is performed on the number of viewers, likes, comments, and reward amount data to obtain historical trend characteristics; A fast memory path and a slow memory path of a hybrid memory-enhanced neural network are set, a bidirectional gated recurrent unit network is configured in the fast memory path, a candidate state generation unit is set in the bidirectional gated recurrent unit network, and a nonlinear combination of a current input and a historical state is calculated by the candidate state generation unit to generate a candidate hidden state; the historical state is randomly forgotten by a reset gate, and the candidate hidden state is updated by an update gate; A converter network is configured in the slow memory path, a multi-head self-attention layer is combined with a feedforward neural network layer to construct a computing block, a query vector, a key vector, and a value vector are projected into multiple feature subspaces in the multi-head self-attention layer, a scaled dot product calculation is performed in each feature subspace to obtain an attention weight, dropout regularization is performed on the attention weight, two layers of perceptrons are connected in series in the feedforward neural network layer, wherein the first layer of perceptrons performs dimensional expansion, and the second layer of perceptrons performs dimensional compression, and a normalization layer is added between the layers; Performing an adaptive batch normalization operation on the private domain live broadcast history data and the historical trend features to obtain normalized data, reorganizing the normalized data into a three-dimensional array including a day dimension, a minute dimension, and an indicator dimension, inputting the three-dimensional array into the fast memory path, bidirectionally scanning along the day dimension to obtain local time series features, inputting the three-dimensional array into the slow memory path, adding time series position information through position coding, and obtaining global time series features; Configure the first attention branch, the second attention branch and the third attention branch of the spatial pyramid attention module, set a sparse attention mechanism in the first attention branch, generate an adaptive attention mask, perform information aggregation of key temporal positions through Top-k selection, set a dense attention mechanism in the second attention branch, perform feature extraction of local areas through a deformable sliding window, set a hybrid attention mechanism in the third attention branch, cascade channel attention and spatial attention, extract cross-channel and cross-spatial dimension dependencies, align the local temporal features with the global temporal features in the time dimension and input them into the spatial pyramid attention module, and perform weighted fusion on the output features of the three attention branches to obtain multi-scale fusion features.

3. The method according to claim 1, characterized in that: Combined with deformable convolution to adjust the feature receptive field, a hierarchical feature representation is obtained, and the hierarchical feature representation is input into the progressive feature distillation network to perform multi-level knowledge transfer and feature reorganization operations to generate the live broadcast peak period prediction results including: The offset prediction layer, the modulation factor prediction layer and the feature sampling layer of the multi-layer deformable convolutional network are configured, the two-dimensional offset of the sampling position of the pre-acquired multi-scale fusion feature is calculated in the offset prediction layer, the modulation factor of the sampling position is generated in the modulation factor prediction layer, the feature response is adaptively adjusted, the bilinear interpolation sampling is performed on the feature map in combination with the two-dimensional offset and the modulation factor in the feature sampling layer, the multi-scale fusion feature is sequentially input into each layer of the multi-layer deformable convolutional network, cross-layer connections are added between adjacent layers, and a hierarchical feature representation is generated; A teacher network and a multi-level student network of a progressive feature distillation network are configured, parameters of the teacher network are initialized according to the hierarchical feature representation, each level of the student network corresponds to a feature layer of different depths, features of the student networks at adjacent levels are reused through dense residual connections, an attention distillation module is introduced into the teacher network, an attention map is used as an intermediate supervision signal, the hierarchical feature representation is input into the progressive feature distillation network, the L2 loss and the attention map loss of the feature maps of the teacher network and the student network are calculated, the KL divergence loss corresponding to the soft label distribution is calculated, a total loss is obtained by weighted combination of multiple loss items, network parameters are optimized based on the total loss, and a distilled feature representation is generated; The distilled feature representation is input into a multilayer perceptron, in which the feature dimension is reduced layer by layer, and in the last layer the output dimension is set to the number of discrete time periods in the prediction time window; the output is normalized by the Softmax function to obtain the peak probability of each time period, and the optimal threshold is determined by combining the dynamic probability threshold selection algorithm, and the time period above the optimal threshold is marked as the predicted peak period.

4. The method according to claim 1, characterized in that: Collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node associations, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through the goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output the live content scheduling strategy, including: Perform feature extraction on each frame in the video stream data through a feature extraction network, the feature extraction network includes a feature pyramid network and an attention enhancement module, extract multi-scale visual features through the feature pyramid network, perform feature enhancement in the channel dimension and the spatial dimension through the attention enhancement module, perform short-time Fourier transform on the audio data in combination with the multi-head self-attention network to obtain a spectrogram, input the spectrogram into a deep network to extract audio features, use a bidirectional encoding network to perform sequence modeling on the barrage text and the comment text, and map the word vector sequence into a text feature vector; input the visual features, the audio features, and the text feature vector into a temporal encoding network, perform temporal encoding to obtain a multimodal feature tensor; Construct a dynamic heterogeneous graph neural network, take anchor nodes, audience nodes, video clip nodes, audio clip nodes, bullet screen nodes, and comment nodes as graph nodes, assign the multimodal feature tensor to the corresponding graph nodes, calculate the similarity between node pairs through a topology generator, perform a normalization transformation on the similarity to obtain edge weights, construct a node adjacency matrix in combination with a sparse matrix representation, and use a feature matrix to store node attributes to obtain a feature representation of the node; The dynamic heterogeneous graph neural network is dynamically updated. When new data arrives, the corresponding modal features are extracted and new nodes are created. The feature similarity between the newly added nodes and the existing nodes corresponding to the new data is calculated. Connecting edges are added between node pairs whose similarity is greater than a preset first threshold. The product of the degree centrality of the existing nodes and the time decay factor is calculated. Nodes whose product is less than a preset second threshold are deleted. The exponential sliding average of the edge weights is calculated. The edges whose sliding average is less than a preset third threshold are deleted to obtain an optimized graph structure. The feature matrix and the node adjacency matrix are input into a multi-layer graph neural network through a feature interaction engine, wherein the feature matrix is ​​composed of a multimodal feature tensor of nodes, representing the initial features of each node, and a learnable parameter matrix is ​​constructed according to the input feature dimension and the hidden layer dimension. First, the input dimension and output dimension of each layer of the graph neural network are set, and the learnable parameter matrix is ​​initialized using a uniform distribution or a normal distribution. The learnable parameter matrix is ​​used as the weight matrix of each layer of the graph neural network, and the node adjacency matrix is ​​degree-normalized to obtain a symmetric normalized matrix. The normalized matrix is ​​multiplied by the node feature matrix of the current layer, and then multiplied by the learnable parameter matrix. A nonlinear activation function is used to obtain the symmetric normalized matrix. Get the node features of the next layer, apply the hierarchical contrast learning method in each layer, construct positive sample pairs and negative sample pairs, the positive sample pairs are sampled from adjacent nodes, and the negative sample pairs are sampled from non-adjacent nodes at different semantic levels, calculate the feature contrast loss, and perform weighted average of the contrast losses at different levels to obtain the total loss. Input the node representation output by the graph neural network into the hierarchical strategy optimization network, set the interaction quality improvement sub-goal, content form optimization sub-goal, and user experience optimization sub-goal in the goal decomposition module, calculate the sub-goal weights, and set the resource scheduling sub-network in the execution planning module, wherein the resource scheduling sub-network includes a multi-layer encoder, and the output of the encoder is mapped to obtain the scheduling action space; The nodes in the dynamic heterogeneous graph neural network are mapped to multiple agents. Each agent maintains a local state representation and a policy network. The policy network consists of an action generation network and a state evaluation network. The action generation network outputs the action probability distribution, and the state evaluation network evaluates the state value function. The agents exchange information through a competitive cooperation mechanism, optimize strategies based on non-cooperative games, and calculate rewards based on the joint actions of the agents. The reward function comprehensively considers multiple indicators such as interactivity, content quality, and user feedback to obtain a comprehensive evaluation score. Each agent generates a probability distribution of live content scheduling strategy based on its strategy network, and performs parameterized sampling on the probability distribution to obtain discrete actions. The discrete actions of all agents are combined into a live content scheduling strategy, including the host interaction instruction sequence, the audience activity push sequence, and the platform resource allocation sequence. The structure and node characteristics of the dynamic heterogeneous graph neural network are updated according to the feedback after executing the scheduling strategy.

5. The method according to claim 4, characterized in that The agents exchange information through a competitive cooperation mechanism, optimize strategies based on non-cooperative games, and calculate rewards based on the joint actions of the agents, including: An agent is deployed at each node of a dynamic heterogeneous graph, wherein the nodes of the dynamic heterogeneous graph include video clip nodes, audio clip nodes, and bullet-screen nodes. The state space of the agent is composed of the attribute feature vector of the corresponding node and the attribute feature vector of the adjacent node. The state vector of the video clip node includes the visual features of the video frame, the acoustic features of the audio clip, and the text features of the surrounding bullet-screen nodes. The action space of the agent includes picture clarity adjustment, interactive question insertion, and related video push. The agent broadcasts its own action probability distribution to neighboring nodes at each decision time step, receives the action probability distribution of neighboring nodes, and uses the received action probability distribution as an additional input signal to update the corresponding strategy network, wherein the strategy network includes an action generation network and a value evaluation network; Strategy optimization is performed based on a non-cooperative game framework. The objective functions of different types of nodes in the non-cooperative game include the audience participation of the anchor node, the viewing time of the video node, and the number of interactions of the bullet comment node. The agent takes its own actions to obtain immediate rewards based on the observed state and the actions of other agents, and receives trajectory data of the interaction between other agents and the environment to estimate the joint action value function; Iteratively update the joint action value function, calculate the value function of the current agent state and the joint action, and perform weighted average of the difference between the value function and the sum of the instant reward and the discounted value of the state value function at the next moment to obtain the updated value function, wherein the instant reward is obtained according to the environment, and the weight of the weighted average is the learning rate; The policy function is optimized based on the policy gradient algorithm, and the Monte Carlo estimation method is used to calculate the policy gradient of each agent's objective function. The product of the policy gradient and the learning rate is added to the parameters of the agent's policy function to obtain the updated policy function parameters. The agent generates a scheduling strategy based on the optimized policy function, and performs picture clarity adjustment, interactive question insertion, and related video push according to the scheduling strategy.

6. The method according to claim 1, characterized in that Collect real-time live broadcast data and add it to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between the predicted effect and the actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, the adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy. Repeated collection and adjustment to obtain the optimal content scheduling strategy includes: Adding the collected real-time video stream, real-time audio stream, and real-time text stream data to a pre-built two-way interactive evaluation network, wherein the two-way interactive evaluation network includes a state encoder and an effect evaluator, wherein the state encoder includes a video feature extraction module, an audio feature extraction module, and a text feature extraction module, and the effect evaluator includes a causal graph construction module, a structural equation model, and a multi-layer perceptron network; Extracting frame-level visual features from a real-time video stream through a convolutional neural network in the video feature extraction module, extracting audio spectrogram features from a real-time audio stream through a convolutional neural network in the audio feature extraction module, extracting semantic features from a real-time text stream through a recurrent neural network in the text feature extraction module, inputting the frame-level visual features, the audio spectrogram features, and the semantic features into an attention network to generate a feature weight matrix, and performing weighted fusion on the three features based on the feature weight matrix to obtain a multimodal representation vector; The causal graph construction module constructs a live broadcast effect causal graph based on historical data, the live broadcast effect causal graph includes a scheduling strategy node, a live broadcast status node, an external factor node and an effect indicator node, the edges between the nodes represent causal relationships, the structural equation model defines the functional relationship between the nodes, the multimodal representation vector is spliced ​​with the preset scheduling strategy vector to form an input vector, causal reasoning is performed based on the causal graph and the structural equation model, the policy intervention effect is obtained, the input vector and the policy intervention effect are input into the multi-layer perceptron network, the multi-layer perceptron network calculates the predicted number of viewers, the predicted number of likes, and the predicted reward amount, counts the real-time number of viewers, the real-time number of likes, and the real-time reward amount, and calculates the weighted average deviation between the predicted value and the actual value to obtain the real-time performance evaluation index; Calculate the deviation value between the real-time performance evaluation index and the stored live broadcast peak period prediction result, compare the deviation value with a preset deviation threshold, and trigger strategy adjustment when the deviation value is greater than the deviation threshold; The live content scheduling task is decomposed into content recommendation subtask, anchor interaction subtask, and audience activity subtask. Each subtask is configured with an independent policy network and value function network. The policy network generates an action probability distribution based on the current state, and the value function network evaluates the value of the state-action pair. A gradient meta-learner is trained for each subtask. The gradient meta-learner adopts a model-independent meta-learning algorithm. In the inner loop, the policy gradient method is used to update the network parameters based on the latest state data, action data, and reward data. In the outer loop, the meta-parameters are updated by minimizing the meta-objective loss across tasks. The action probability distributions output by the policy network corresponding to the three updated subtasks are normalized, and the action probability distribution of the global scheduling strategy is obtained by fusion through a soft voting mechanism. The live content scheduling action is sampled and selected based on the action probability distribution, and the optimal content scheduling strategy is obtained by combination. While executing the scheduling action, new live broadcast data is continuously collected, and the newly collected data is repeatedly input into the two-way interactive evaluation network for evaluation and optimization until the deviation value between the real-time performance evaluation index and the live broadcast peak period prediction result is less than the deviation threshold.

7. The method according to claim 6, characterized in that The gradient meta-learner adopts a model-independent meta-learning algorithm, uses a policy gradient method in the inner loop to update network parameters based on the latest state data, action data, and reward data, and minimizes the cross-task meta-objective loss in the outer loop to update the meta-parameters, including: Construct a meta-learner network structure, set a policy network to generate action probability distribution, configure a meta-optimizer to adjust policy network parameters, implement the policy network as a deep neural network structure, and initialize meta-parameters shared by the policy network and the meta-optimizer; For the pre-obtained subtasks, collect the historical status data, action data, reward data, and effect data of each type of subtask to construct a training task, and obtain new status data, action data, reward data, and effect data from the real-time live broadcast scene to construct a test task; Sampling training tasks to obtain state transition data, copying meta-parameters as initial values ​​of task-specific parameters, starting the policy network based on the task-specific parameters, inputting the current state to generate an action probability distribution and sampling actions, executing the actions to obtain immediate rewards and next states, calculating the state value functions of the immediate rewards and next states, using the Bellman equation to calculate the temporal difference error to obtain the action value function, multiplying the action value function with the logarithm of the action probability output by the policy network and taking the negative value to obtain the policy loss, calculating the gradient of the policy loss with respect to the task-specific parameters, updating the task-specific parameters along the gradient direction, and repeatedly executing the policy network forward propagation and gradient update until the inner loop training is completed; Calculate the cumulative reward difference of the task-specific parameters before and after each training task is updated on the corresponding task as a fitness metric, construct a meta-objective loss function by weighted average of the fitness metrics of all tasks, calculate the second-order derivative of the meta-objective loss function with respect to the meta-parameters, and drive the meta-optimizer to update the meta-parameters based on the second-order derivatives to minimize the meta-objective loss function; Sample the test task, copy the trained meta-parameters as the initial values ​​of the task-specific parameters, and execute the inner loop update to obtain the policy network that adapts to the current task.

8. A deep learning-based private domain live broadcast peak hotspot prediction and content scheduling system, 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 private domain live broadcast history data, build a hybrid memory enhanced neural network including a fast memory path and a slow memory path, wherein the fast memory path performs short-term dependency modeling through a dynamic gating mechanism, and the slow memory path captures long-term dependency through a hierarchical sparse attention structure, input the private domain live broadcast history data into the hybrid memory enhanced neural network, extract multi-scale temporal features through the fast memory path and the slow memory path, build a spatial pyramid attention module, perform feature recombination on the multi-scale temporal features according to the parallel attention branches in the spatial pyramid attention module, adjust the feature receptive field in combination with deformable convolution, obtain a hierarchical feature representation, input the hierarchical feature representation into a progressive feature distillation network, perform multi-level knowledge transfer and feature recombination operations, and generate a live broadcast peak prediction result; The second unit is used to collect multimodal data related to live broadcast and perform feature extraction and feature encoding, construct a multimodal feature tensor and add it to a pre-set dynamic heterogeneous graph neural network, combine the topology structure generator to build node association relationships, perform cross-modal feature fusion through the feature interaction engine, apply a hierarchical contrast learning training method to the dynamic heterogeneous graph neural network, construct sample pairs at different semantic levels and perform feature comparison, input the feature comparison results into a hierarchical strategy optimization network, divide the global goal into multiple sub-goals through a goal decomposition module, plan the sub-goal realization path according to the execution planning module, input the hierarchical strategy optimization network into a non-cooperative game multi-agent reinforcement learning framework, optimize the strategy space through multi-agent competition and collaboration, and output the live content scheduling strategy; The third unit is used to collect real-time live broadcast data and add it to a pre-built two-way interactive evaluation network including a state encoder and an effect evaluator. The state encoder extracts multi-source information features according to a cross-modal fusion method. The effect evaluator establishes a mapping relationship between predicted effect and actual effect based on a causal reasoning mechanism to obtain a real-time performance evaluation index. If the deviation value between the real-time performance evaluation index and the live peak prediction result is greater than a preset deviation threshold, an adjustment mechanism is triggered. According to a fast adaptation algorithm based on meta-reinforcement learning, the live content scheduling strategy is optimized online through task decomposition and gradient meta-learning. The live content scheduling plan is dynamically adjusted based on the optimized strategy, and the collection and adjustment are repeated to obtain the optimal content scheduling strategy.

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