Cross-platform unmanned live broadcast resource scheduling method and system based on agent cooperation

Abstract

The application provides a cross-platform unmanned live broadcast resource scheduling method and system based on agent cooperation, and belongs to the technical field of live broadcast resource scheduling. The method disassembles multi-source heterogeneous data from multiple live broadcast platforms according to a unified element dimension and constructs a multi-agent heterogeneous graph, defines five types of nodes, namely live room nodes, virtual human agent nodes, content segment nodes, commodity nodes and strategy rule nodes, and five types of edges, namely scheduling edges, content supply edges, conversion edges, cooperative sharing edges and constraint edges, in the graph. Dynamic analysis is performed on the heterogeneous graph in time and space sequence, the coupling risk cooperation strength of multiple nodes is calculated through phase space reconstruction and covariance matrix eigenvalue decomposition, and a critical state early warning signal is generated. The technical scheme breaks through the heterogeneous barrier of cross-platform data, realizes global cooperative scheduling of virtual human resources, can early warn the risk of illegal diffusion and resource cascade collapse, improves conversion income and reduces the probability of live broadcast interruption.

Description

Technical Field

[0001] This invention relates to the field of live streaming resource scheduling technology, and more specifically, to a cross-platform unmanned live streaming resource scheduling method and system based on intelligent agent collaboration. Background Technology

[0002] With the widespread application of unmanned live streaming technology in e-commerce, cross-platform, multi-room live streaming and automated virtual human broadcasting have become important means to reduce labor costs and achieve 24 / 7 operation. However, the data formats, scheduling interfaces, and compliance rules of different live streaming platforms are heterogeneous, requiring virtual human resources to be dynamically allocated among hundreds of live streaming rooms. Most existing scheduling solutions are based on static scheduling or rule engines within a single platform, making it difficult to perceive the global resource status in real time during traffic surges and rapid content iterations, and even more difficult to provide early warnings of system-level risks such as the spread of violations and cascading resource crashes. This passive response mode results in low utilization of virtual humans, limited conversion revenue, and susceptibility to large-scale live streaming interruptions due to localized failures. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, the present invention provides a cross-platform unmanned live streaming resource scheduling method based on intelligent agent collaboration, the method comprising: Multi-source heterogeneous data from multiple live streaming platforms are decomposed into a structured dataset, extracted and classified according to a unified element dimension, and multi-type association relationships between data are pre-established. Based on the structured association dataset, a multi-agent heterogeneous graph is constructed, wherein node types and edge types are defined, and the multi-agent heterogeneous graph containing specific nodes and specific edges is instantiated according to the entities, attributes and associations in the structured association dataset. The node types include live streaming room nodes, virtual human intelligent agent nodes, content segment nodes, product nodes, and strategy rule nodes; the edge types include scheduling edges, content supply edges, conversion edges, collaborative sharing edges, and constraint edges. On the multi-agent heterogeneous graph, perform spatiotemporal sequential dynamic analysis, calculate the dynamic activation value of each node, and perform conflict detection and critical state early warning. Based on the results of the dynamic analysis, adaptive shaping and topology reconstruction are implemented, including applying personalized attention masks to the edges with content transitive attributes in the multi-agent heterogeneous graph, and performing pruning, growth and reconnection operations on the multi-agent heterogeneous graph. The system collects feedback data and adjusts the edge weights of the multi-agent heterogeneous graph, the node trust factor of the virtual human agent nodes, and the triggering conditions for topology reconstruction based on the feedback signals, thus forming a closed-loop evolution.

[0004] Furthermore, the present invention also provides a cross-platform unmanned live streaming resource scheduling system based on intelligent agent collaboration, which is used to store and execute the above-mentioned methods.

[0005] Based on the above aspects, this technical solution has the following beneficial effects.

[0006] First, by structurally decomposing multi-source heterogeneous data from multiple live streaming platforms according to a unified element dimension and constructing a multi-agent heterogeneous graph, the heterogeneous barriers of data formats and scheduling interfaces between platforms are eliminated. The five types of nodes—live streaming rooms, virtual humans, content fragments, products, and strategy rules—and their relationships are integrated into a unified graph structure, enabling global perception and collaborative scheduling of cross-platform resource status, and increasing the utilization rate of virtual human resources from 62% to 89%.

[0007] Second, by performing spatiotemporal sequential dynamic analysis on a multi-agent heterogeneous graph, and calculating the collaborative strength of multi-node coupling risks based on phase space reconstruction and covariance matrix eigenvalue decomposition, critical instability warning signals can be generated in advance before nonlinear abrupt changes occur in the state of the live broadcast room and virtual human. Simultaneously, violation phase transition warning and resource collapse warning mechanisms are constructed to address the collaborative diffusion tendency of violations and the cascading vulnerability of resource networks, respectively, transforming passive response into proactive prevention, reducing the number of live broadcast room interruptions caused by violations by 70%.

[0008] Third, through adaptive shaping and topology reconstruction operations, including content delivery edge weight adjustment based on personalized attention masks, and three types of topology operations such as pruning, growth, and reconnection, the system can autonomously expand the scheduling network during traffic surges and when new platforms are accessed. When risks occur, it can proactively isolate high-risk nodes and allocate dedicated resource pools for high-value live streaming rooms. The 95th percentile of scheduling response latency is controlled within 2 seconds.

[0009] Fourth, a closed-loop evolutionary mechanism is formed through differentiated credit attribution and adaptive threshold tuning. The multi-branch backtracking tree attribution method accurately allocates the net revenue of a single live broadcast to each participating node along the scheduling path, guiding the update of value estimation; the threshold tuning method based on group kernel density estimation and Bayesian classification continuously optimizes each critical threshold with operational experience, reducing the false alarm rate by 28%. The global conversion revenue is improved by approximately 35% compared to the static scheduling scheme. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the hierarchical mechanism for critical state early warning in this invention; Figure 3 A schematic diagram illustrating the system provided in this embodiment. Detailed Implementation

[0011] like Figures 1-3 As shown.

[0012] This technical solution provides a cross-platform unmanned live streaming resource scheduling method and system based on intelligent agent collaboration. The technical details are explained in detail below using a specific application scenario.

[0013] This embodiment illustrates the application of a cross-platform unmanned live streaming resource scheduling system in a large-scale promotional event. The scenario involves an organization operating over 200 live streaming rooms across five different platforms. Each platform provides differentiated data interfaces and rules, requiring the scheduling of approximately 500 virtual human agents to automatically complete tasks such as product explanations and fan interaction. In the context of peak traffic, rapid content updates, and increasingly stringent platform compliance, traditional static scheduling and single-platform scheduling methods are highly susceptible to problems such as competition for virtual human resources, the spread of compliance risks, and interruptions of high-value live streaming rooms due to malfunctions.

[0014] The method in this embodiment runs on a set of central scheduling servers. The servers obtain real-time status of the live stream, user behavior, product information, and other content through the API interfaces provided by each platform, and simultaneously issue instructions to each platform for virtual human on / off streaming, content switching, and other functions. Internally, the servers are divided into a data preprocessing module, a heterogeneous graph construction module, a dynamic analysis module, an adaptive shaping and topology reconstruction module, and a feedback adjustment module. Data exchange between modules occurs through shared memory and message queues. The following description details the specific steps.

[0015] Step 1: Structured decomposition and association construction of multi-source heterogeneous data; This step corresponds to the data preprocessing module. First, raw data is retrieved from the API interfaces of five platforms, including real-time online viewers in the live stream, bullet comment lists, like and share action logs, product showcase clicks and transaction records, virtual human audio and text interaction records, and platform violation notices. The formats of this data vary; for example, platform A returns a JSON structure using `room_id` to identify the live stream room, while platform B uses `broadcast_id`. Product identifiers are sometimes numeric IDs and sometimes strings containing letters. In addition, the data also contains semi-structured log text and unstructured live stream screenshot tags.

[0016] To eliminate heterogeneity, a unified set of element dimensions is defined, comprising seven dimensions: entity identifier, entity type, timestamp, platform code, event type, event payload, and session identifier. The entity type enumeration includes six categories: live room, virtual person, content fragment, product, policy rule, and platform user. The event type enumeration includes 38 types, such as live start, live end, content exposure, product click, transaction, violation warning, and virtual person switching. All raw data is parsed, and fields are extracted according to these seven dimensions and mapped to a standardized format. For example, the `room_id` field from platform A and the `broadcast_id` field from platform B are both mapped to the entity identifier in the unified dimension, with the entity type marked as "live room." Timestamps are uniformly converted to Unix millisecond timestamps, and platform codes use pre-assigned integer numbers from 1 to 5.

[0017] Based on standardization, multiple types of association relationships are pre-built to form a structured association dataset. Four types of association relationships are defined: attribution relationships (e.g., content fragments belong to the live stream room, virtual characters belong to the scheduling pool); triggering relationships (e.g., a bullet comment triggers the exposure of a product link); conversion relationships (e.g., a product click ultimately results in a transaction within 15 minutes); and similarity relationships (e.g., the cosine similarity of the voice style embedding vectors of two virtual characters is greater than 0.8). These association relationships are stored together with entity attributes in a relational database. Each association table contains fields such as source entity, target entity, relationship type, relationship weight, and timestamp. Simultaneously, to support graph computation, an additional mirror image of the adjacency list structure is maintained in memory, with each record accompanied by an association direction identifier.

[0018] Step 2: Construct a multi-agent heterogeneous graph A multi-agent heterogeneous graph is a directed attribute graph containing multiple node types and edge types, used to uniformly describe the complex interaction relationships between live streaming rooms, virtual human agents, content, products, and policy rules. This step instantiates specific nodes and edges based on the aforementioned structured association dataset.

[0019] The diagram defines five node types: live stream node, virtual human agent node, content segment node, product node, and strategy rule node. The live stream node includes the following attributes: platform identifier (an integer from 1 to 5); audience profile vector (a 64-dimensional floating-point array representing the label distribution obtained by clustering historical user behavior on the platform); real-time traffic status (a triple containing the current number of online users, the number of new users in the last 5 minutes, and the interaction rate, where the interaction rate is the ratio of bullet comments to online users); conversion status (a two-dimensional vector consisting of the number of completed orders and the transaction amount in the last 30 minutes); and compliance status (a discrete value, where 0 indicates normal, 1 indicates warning, 2 indicates traffic restriction, and 3 indicates ban). The virtual human agent node includes the following attributes: style embedding, a 128-dimensional vector extracted from pre-trained voiceprint and language style models; current load, the number of live streaming tasks the virtual human is simultaneously undertaking; fatigue level, a cumulative value calculated according to a preset fatigue curve, with the formula F = 1 - exp(-0.02 * t), where F is the fatigue level of the virtual human agent node, a continuous value between 0 and 1; t is the current continuous working time in hours; compliance history, a list of timestamps of violations within the last 90 days; and a node trust factor, the initial value of which is jointly determined based on the matching degree of historical compliance records and the compliance tags of the supplied content, calculated using the following formula: T0 = ​​0.5 + 0.5 * exp(-number of violations / 10); T0: The initial value of the node trust factor.

[0020] M: The total number of historical violations by this virtual human agent node.

[0021] If a serious violation exists, it will be multiplied by an additional 0.5. Content fragment nodes contain attributes including: type (enumerated values ​​1 to 4, representing product explanation, fan interaction, talent showcase, and promotional announcements respectively); compliance tag (a continuous value from 0 to 1, output by the content review model); and a time-lapse function, using an exponential decay form f(Δt) = exp(-λΔt), where Δt is the number of hours since the content was created, and λ defaults to 0.05. Product nodes contain attributes including: commission rate (a percentage value); inventory status (a Boolean value, 1 indicating inventory); and conversion window attribute (represented by the remaining valid seconds). Strategy rule nodes encapsulate a set of constraint predicates, each predicate being a logical expression. For example, a compliance status of 3 in the live stream would prohibit scheduling, or a virtual human fatigue level greater than 0.8 would restrict new tasks.

[0022] There are five edge types. **Dispatch Edge:** A dispatch edge points from a livestream node to a virtual agent node, indicating that a virtual agent is assigned to a livestream. This edge carries a fit weight, calculated as the cosine similarity between the style embedding vector and the livestream audience profile vector, linearly scaled to the 0-1 range. **Content Supply Edge:** A content fragment edge points from a content segment node to a livestream node, indicating that a content segment is delivered to that livestream. This edge weight is the content fit, initially set to 1. **Conversion Edge:** A conversion edge points from a content fragment node to a product node, carrying a historical conversion rate weight, which is the ratio of product clicks to impressions after the content exposure. **Collaboration and Sharing Edge:** An undirected edge connecting different virtual agent nodes, carrying a style distance weight, calculated as 1 minus the style embedding cosine similarity. Smaller distances indicate higher substitutability. All virtual agent nodes and collaboration and sharing edges constitute a collaboration and sharing subgraph. A constraint edge points from any node to a policy rule node, or from a policy rule node to another node, representing a certain constraint relationship. When the scheduling path matches the constraint predicate set, the total path cost increases by a preset high cost. In this embodiment, the high cost is fixed at 100,000.

[0023] The instantiation process is as follows: traverse the structured association dataset, create corresponding nodes according to entity type and populate attributes; create content supply edges based on affiliation, create conversion edges based on conversion, and create collaborative sharing edges based on similarity; create scheduling edges based on virtual human scheduling records and calculate matching weights; create constraint edges based on violation rules. The final result is a multi-agent heterogeneous graph containing over 200 live stream nodes, 500 virtual human agent nodes, thousands of content fragment nodes, and product nodes, with approximately 50,000 edges.

[0024] Step 3: Spatiotemporal sequential dynamic analysis and critical state early warning The dynamic analysis module performs a state update and analysis on the multi-agent heterogeneous graph once every decision cycle, which is set to 10 seconds in this embodiment. First, the dynamic activation value of each node is calculated.

[0025] The activation value A_room of the live stream node is obtained by weighted summation of the number of online users, interaction rate, compliance risk index, and lifecycle factor in its real-time traffic status. The compliance risk index is mapped by compliance status: status 0 maps to 0, status 1 to 0.3, status 2 to 0.7, and status 3 to 1. The lifecycle factor is determined based on the duration of the live stream from its initial broadcast to the present, combined with traffic trends, using a segmented mapping function: Cold start phase, traffic increases but the base is low, factor is 0.6; growth phase, factor is 1.0; stable phase, factor is 0.8; decline phase, traffic continues to decrease, factor is 0.3. The phase determination is based on the slope and magnitude of the number of online users in the last 3 hours. The weighting coefficients are as follows: normalized online user value weight 0.2, interaction rate weight 0.3, compliance risk index weight -0.4, lifecycle factor weight 0.1, calculated using the following formula: A_room = 0.2 * N_norm + 0.3 * I_rate - 0.4 * R_comp + 0.1 * L_fac; A_room: The dynamic activation value of the live stream node.

[0026] N_norm: The normalized value of the current number of online users divided by the historical peak number of online users and truncated to 1.

[0027] I_rate: The decimal representation of the interaction rate.

[0028] R_comp: Compliance risk index.

[0029] L_fac: Life cycle factor.

[0030] Experiments show that when the compliance risk weight is 0.4, it can effectively reduce the activation value of non-compliant live streaming rooms quickly. When it is below 0.3, the risk suppression is insufficient, and when it is above 0.5, it may lead to conservative scheduling and loss of conversion opportunities.

[0031] The activation value A_agent of the virtual human agent node is calculated as a weighted sum of current load, fatigue level, and recent violation records. Recent violation records are the number of violations within the last 7 days, normalized, with the following weighting: load 0.5, fatigue level 0.3, and violation record weight 0.2. The formula is: A_agent = 0.5 * L + 0.3 * F + 0.2 * V .

[0032] A_agent: The dynamic activation value of the virtual human agent node.

[0033] L: Current load value.

[0034] F: Fatigue level.

[0035] V_norm: The normalized value of the number of violations in the past 7 days. The normalization method is to divide the number of violations by 5 and truncate it to 1.

[0036] The activation value A_content of a content fragment node is calculated based on the current value of the time decay function f(Δt), the sum of the weights of all content supply edges pointing to that node, and the frequency of the node's references. The formula is as follows: A_content = f(Δt) * ( Σw_supply + ln(1 + ref_cnt) ); A_content: The dynamic activation value of the content fragment node.

[0037] f(Δt): The value of the time decay function at the current Δt, f(Δt) = exp(-λ * Δt), where λ is 0.05 by default.

[0038] Σw_supply: The sum of the weights of all content supply edges pointing to this node.

[0039] ref_cnt: Frequency of citation.

[0040] The activation value A_product for a product node is calculated based on the proportion of remaining time in the conversion window and the maximum value of the historical conversion rate weights of all inbound conversion edges. The activation value A_rule for a strategy rule node is equal to the frequency of its triggering within the most recent 5-minute time window multiplied by the preset penalty strength, which defaults to 10.

[0041] Conflict detection involves retrieving the scheduling path before each scheduling decision and determining whether the virtual agent nodes along the path already have scheduling tasks and their load has reached the limit. If the load exceeds the threshold of 3, a conflict is identified. The core of critical state early warning is the prediction of system-level instability. A phase space analysis method based on activation values ​​is employed. The activation values ​​of all live-streaming nodes and virtual agent nodes are combined to form the joint state vector of the current system's high-dimensional phase space. To capture dynamic evolution characteristics, a phase space reconstruction with a preset time delay is applied to this joint state vector. The time delay parameter τ is initially set to 5 periods, and the embedding dimension is determined to be 8 using the spurious nearest neighbor method. This generates a multivariate delay embedding matrix, with each row corresponding to the delay vector at each time step. The covariance matrix of this matrix is ​​then calculated and eigenvalue decomposition is performed to extract the largest eigenvalue λ1 and the second largest eigenvalue λ2. The multi-node coupling risk coordination strength R_c = λ1 / λ2 is then calculated. When R_c monotonically increases within three consecutive time windows and exceeds the preset critical coordination threshold of 3.5, a critical instability early warning signal is generated. Experiments show that when the threshold is set to 3.5, there are early warnings for all 15 cascading crash events in the simulation, while when the threshold is set to 2.8, the false alarm rate increases by 22%.

[0042] Step 4: Adaptive Shaping and Topology Reconstruction After receiving dynamic analysis results and early warning signals, the adaptive shaping and topology reconstruction module performs three types of operations. The first type is to apply personalized attention masks to edges with content delivery attributes. Content supply edges are typical content delivery edges. The mask generation method is as follows: for each content supply edge, the audience profile vector of the connected live room node and the content embedding vector of the content fragment node are multiplied by the dot product, mapped by the Sigmoid function, and then multiplied by the time decay function value of the content fragment node to obtain the mask coefficient. This coefficient is then multiplied element-wise with the weight of the edge. This mechanism enables differentiated amplification of the same content fragment by different live rooms. For example, for live rooms that prefer promotional content, the weight of promotional content supply edges is enhanced.

[0043] The second type of operation is pruning. A pruning priority score is calculated for each livestream node, using the formula: P_prune = (the node's marginal contribution to global conversion revenue divided by the node's associated violation risk exposure value plus 1). The marginal contribution value is obtained by calculating the decrease in overall conversion revenue after removing the node, and the violation risk exposure value is equal to the product of the node's violation count and average severity over the past 30 days. When a livestream node's activation value is below the dynamic decay threshold of 0.2 for five consecutive observation periods, and its lifecycle factor drops below 0.3, if its pruning priority score is below the retention threshold of 0.1, then the node and all its first-level scheduling edges are soft-deleted. During soft deletion, the node's state is marked as dormant, all outgoing edge weights are multiplied by the residual factor of 0.01, and a 300-second recovery observation period is initiated. If the node's lifecycle factor recovers to above 0.4 during the observation period, its weights and activation state are restored.

[0044] The third type of operation includes growth and reconnection. A new platform growth operation is triggered when a new external platform data source is received from the new platform and the structural fusion barrier index exceeds the fusion barrier threshold of 0.6. The structural fusion barrier index is defined as the complement ratio of the number of successful attempts in 10 consecutive attempts when a temporary probe node establishes a content supply edge or collaborative sharing edge with an existing node; that is, barrier index = 1 - number of successful attempts / 10. The growth operation instantiates a new live streaming room node, copies its current scheduling subgraph from a historically high-conversion live streaming room on the same platform as a template subgraph, applies a Gaussian noise random mutation mask with a mean of 0 and a standard deviation of 0.1 to the weight of each content supply edge in the template subgraph, then attaches the mutated subgraph to the new node, and selects the virtual human agent node with the closest style embedding vector distance and current load below the load threshold of 3 from the collaborative sharing subgraph to establish a scheduling edge.

[0045] When a violation phase transition warning signal is received, the scheduling edge switching in the reconnection operation is performed. A risk-reward balance plane is constructed with style distance as the horizontal axis and trust factor as the vertical axis. After mapping the candidate replacement virtual human node set to this plane, Pareto front is used to select nodes that are both style-similar and have high trust to replace the virtual human node pointed to by the original scheduling edge. When a resource collapse warning signal is received, the scheduling edge of the nodes marked as high-value live streaming rooms (i.e., the top 20% of the cumulative transaction volume in the past 7 days) is locked for protection, and pruning and normal reconnection are prohibited during the locking period. At the same time, the scheduling request of the node is preferentially routed to a pre-reserved dedicated virtual human agent resource pool. This resource pool consists of 10 virtual humans with the highest trust and a load that is consistently below 1.5, ensuring stable service.

[0046] Step 5: Feedback-Loop Evolution The feedback adjustment module is responsible for collecting long-term feedback data, including global conversion revenue, total cost of violations, virtual agent switching success rate, and average scheduling response latency, all calculated every 24 hours. Based on the feedback signals, the edge weights and node trust factors in the multi-agent heterogeneous graph are adjusted. For scheduling edges that contribute high conversions and maintain compliance, their initial weights are increased by 5%, and the corresponding virtual agent node trust factors are increased by a step value of 0.05. For scheduling subgraphs that repeatedly trigger violation risks, the weights of all related edges are reduced by 15%, and the trust factors of the corresponding virtual agent nodes are reduced by a reduction factor of 0.9. At the same time, a stricter penalty coefficient is added to the associated policy rule nodes, increasing the coefficient by 20% from the original level.

[0047] Feedback allocation employs a differentiated credit attribution method. The net revenue for a single livestream cycle is defined as the total transaction amount minus returns and penalties. Tracing back along the scheduling and content supply edges to the associated virtual agent nodes and content fragment nodes, the marginal contribution of each participating node is estimated using the Shapley value, and this value is used to update the node attributes. This value estimate guides pruning protection priorities, meaning livestream nodes with higher value estimates are less likely to be pruned; it also guides the growth initialization strategy, prioritizing the reuse of scheduling templates from high-value nodes by new nodes.

[0048] Through the cyclical execution of the above 5 steps, a closed-loop evolution is formed. The system continuously optimizes resource allocation during each major promotion. The utilization rate of virtual humans during peak periods increases from the initial 62% to 89%, the global conversion revenue increases by about 35% compared to the static scheduling scheme, and the number of live broadcast interruptions caused by violations decreases by 70%.

[0049] Following step two above, the basic framework for node and edge types in multi-agent heterogeneous graphs has been given. To enable those skilled in the art to implement it, the following provides a specific definition of the attribute details and derived concepts in the context of a major sales event scheduling scenario.

[0050] The live stream node contains five core attributes. The platform identifier is an integer number from 1 to 5, corresponding to the five live stream platforms accessed. The audience profile vector is a 64-dimensional floating-point array, derived by clustering the behavioral tags of active users in the live stream over the past 30 days; each dimension represents the distribution weight of an interest category. The real-time traffic status is a triplet consisting of the current number of online users, the number of new users in the last 5 minutes, and the interaction rate, defined as the ratio of the total number of bullet comments to the number of online users multiplied by 100%. The conversion status is a two-dimensional vector recording the number of completed transactions and the transaction amount in yuan within the last 30 minutes. The compliance status is a discrete enumeration value: 0 indicates normal, 1 indicates warning, 2 indicates platform-limited traffic, and 3 indicates ban. Furthermore, to support subsequent activation value calculations, two derived variables are introduced: the compliance risk index is linearly mapped from the compliance status, with statuses 0, 1, 2, and 3 corresponding to index values ​​of 0, 0.3, 0.7, and 1.0, respectively; the lifecycle factor is dynamically determined based on the duration and traffic trends of the live stream since its initial broadcast, rather than being permanently stored. The phase determination rules are as follows: when the slope of the number of online users in the last 3 hours is positive and the average is less than 30% of the average of the entire platform during the same period, it is determined to be the cold start phase, with a lifecycle factor of 0.6; when the slope is consistently positive and exceeds the average by 30%, it is the growth phase, with a factor of 1.0; when the absolute value of the slope is less than 0.05 for 6 consecutive periods, it is the stable phase, with a factor of 0.8; when the slope is negative and declines for more than 10 consecutive periods, it is the decline phase, with a factor of 0.3. This segmented mapping allows the system to give higher activation weight to live streaming rooms in the growth phase, in order to seize traffic dividends.

[0051] The virtual human agent node includes style embedding, current load, fatigue level, compliance history, and node trust factor. The style embedding is a 128-dimensional vector extracted through a pre-trained voiceprint and language style model, used to measure the semantic features of the virtual human's appearance, timbre, and speech. Current load is the number of live-streaming tasks the virtual human is simultaneously undertaking, and its value is a non-negative integer. Fatigue level is the cumulative fatigue value, obtained based on the mapping relationship between continuous working hours and a preset fatigue curve, with the formula F = 1 - exp(-0.02 * t), where t is the cumulative number of hours worked continuously since the last rest. When F exceeds 0.9, the virtual human is forced into a dormant state. The compliance history stores the timestamps and violation types of each violation event within the last 90 days. When calculating the activation value, records from the last 7 days are extracted from this list, and the number of violations is recorded as V. The initial value T0 of the node trust factor is jointly determined by the matching degree of historical compliance records and compliance tags of the supplied content fragments. The calculation formula is T0 = 0.5 + 0.5 * exp(-M / 10), where M is the total number of historical violations. If there are serious violations in the history of the virtual human, T0 is additionally multiplied by a decay coefficient of 0.5. Thereafter, the trust factor is dynamically updated with feedback signals. The update mechanism is detailed in step five.

[0052] The attribute set of each content fragment node includes type, compliance tag, time-lapse function, and content embedding vector for mask calculation. The type enumeration value is 1 to 4, representing product explanation, fan interaction, talent showcase, and promotional announcement, respectively. The compliance tag is a continuous value from 0 to 1, output by the platform's content review model; the closer to 1, the higher the compliance level. The time-lapse function uses an exponential decay form f(Δt) = exp(-λ * Δt), where λ defaults to 0.05, and Δt is the number of hours since the content was created. In addition, each content fragment node implicitly maintains a reference frequency attribute, which is equal to the total number of content supply edges pointing to that node in the current graph. This value is automatically updated by the heterogeneous graph construction module with each topology change, requiring no additional storage.

[0053] The product node includes commission rate, inventory status, and conversion window attributes. The commission rate is a percentage value, such as 15% meaning a 15% commission. The inventory status is a Boolean value, where 1 indicates sufficient stock and 0 indicates out of stock. The conversion window attribute is represented by the remaining valid seconds. For limited-time promotional products, this value decreases over time, and the product node enters a dormant state after reaching zero.

[0054] Each policy rule node encapsulates a set of constraint predicates, each predicate being a parsable logical expression. For example, a live stream's compliance status of 3 would prohibit scheduling, while a virtual human's fatigue level greater than 0.8 and current load greater than 2 would restrict the addition of new tasks. Simultaneously, the policy rule node internally stores a penalty intensity parameter, with a default value of 10. When a scheduling path triggers a constraint predicate, the total cost of that path is increased by the product of the penalty intensity and the number of matches. This penalty intensity is also used as a weighting factor when calculating its own activation value.

[0055] The specific definitions of edge types are defined concurrently. The scheduling edge points from the live stream node to the virtual agent node, carrying a fit weight. This weight is calculated by performing a cosine similarity operation between the virtual agent's style embedding vector and the live stream's audience profile vector, with the result linearly scaled to a range of 0 to 1. A larger value indicates a better fit between the virtual agent and the live stream's audience preferences. The content supply edge points from the content fragment node to the live stream node, indicating that content is being delivered to that live stream. This edge carries a content fit weight, initially set to 1.0, which can be dynamically adjusted based on the content's performance in the live stream. The conversion edge points from the content fragment node to the product node, carrying a historical conversion rate weight. This weight is the ratio of product clicks resulting from content exposure to the number of content exposures. If the number of exposures is 0, the weight is reset to 0.

[0056] A collaborative shared edge connects two virtual human agent nodes. This edge is undirected and carries a style distance weight, defined as 1 minus the cosine similarity of their style embeddings; therefore, a smaller distance indicates higher substitutability. The graph structure consisting of all virtual human agent nodes and all collaborative shared edges connecting them is defined as a collaborative shared subgraph. This subgraph contains at least one connected component, and the number and size of these connected components directly reflect the topological connectivity and redundancy of the virtual human resource network. Constraint edges point from any node to a policy rule node, or vice versa, indicating that the node is constrained by the policy rule. When constructing a scheduling path, if a node on the path matches the constraint predicate set with an edge, the total path cost is increased by a preset high cost. In this embodiment, this high cost is fixed at 100,000, ensuring that such paths are automatically avoided during optimization.

[0057] After completing the above-mentioned refined graph structure instantiation, the dynamic analysis module in step three performs spatiotemporal sequential dynamic activation value calculation and critical state early warning in each decision cycle, which is set to 10 seconds in this example.

[0058] First, the dynamic activation value of each node is calculated. The activation value A_room of the live stream node is obtained by weighted summation of the normalized value of online users in real-time traffic status, interaction rate, compliance risk index, and lifecycle factor. The weight coefficients are determined by grid search optimization, and the formula is A_room = 0.2 * N_norm + 0.3 * I_rate - 0.4 * R_comp + 0.1 * L_fac. Where N_norm is the normalized value obtained by dividing the current online users by the historical peak online users and truncating to 1, I_rate is the decimal representation of the interaction rate, R_comp is the compliance risk index, and L_fac is the lifecycle factor. Experiments show that when the compliance risk coefficient is 0.4, the activation value of non-compliant live streams is effectively suppressed. If it is below 0.3, the risk isolation effect is insufficient, and if it is above 0.5, some slightly non-compliant live streams are overweighted and miss conversions.

[0059] The activation value A_agent of the virtual human agent node is calculated by weighted summation of current load, fatigue level, and recent violation records, using the formula: A_agent = 0.5 * L + 0.3 * F + 0.2 * V_norm. L is the current load value, F is the fatigue level, and V_norm is the normalized value of the number of violations in the past 7 days, normalized by dividing the number of violations by a preset maximum acceptable number of 5 and truncating it to 1. When A_agent exceeds 0.8, the virtual human will no longer accept new scheduling tasks.

[0060] The activation value A_content of a content fragment node is jointly calculated based on the current value of the node's time decay function, the sum of the weights of all incoming content supply edges, and the frequency of citations. The formula is A_content = f(Δt) * (Σw_supply + ln(1 + ref_cnt) ), where f(Δt) is the value of the time decay function at the current Δt, Σw_supply is the sum of the weights of all content supply edges pointing to this node, and ref_cnt is the frequency of citations. The natural logarithm is used to mitigate the marginal effect of frequency growth.

[0061] The activation value A_product of the product node is calculated jointly based on the remaining time proportion of the conversion window attributes and the maximum historical conversion rate weight of the inbound conversion edge. The formula is as follows: A_product = ratio_remain * max_hist_cvr, A_product: The dynamic activation value of the product node.

[0062] ratio_remain: The ratio of the remaining valid seconds in the conversion window to the total valid seconds.

[0063] max_hist_cvr: The maximum value of the historical conversion rate weights of all incoming edges that are transformed.

[0064] This design allows products that are about to expire and products with high conversion rates to receive higher activation values.

[0065] The activation value A_rule of a policy rule node is equal to the frequency with which the constraint predicate set encapsulated by that node is triggered by conflict detection within the most recent 5-minute time window, multiplied by the penalty intensity stored in that node. When a certain type of violation frequently triggers constraints, the activation value of the corresponding policy rule node increases, thereby increasing the cost weight of the relevant path in the entire graph.

[0066] The specific implementation of conflict detection is as follows: before each scheduling decision is generated, the candidate scheduling paths are traversed, and it is checked whether the current load of the virtual human intelligent agent nodes traversed by the path has reached its upper limit of 3. If it has, it is marked as a conflict and the path is excluded.

[0067] The core of critical state early warning is the prediction of overall system instability. The activation values ​​of all live-stream nodes and virtual agent nodes at the current moment are concatenated into a joint state vector, thereby constructing a high-dimensional phase space of the current system with the operating state as the dimension. To capture the dynamic evolution characteristics of the system, phase space reconstruction is applied to this joint state vector. The optimal embedding dimension and optimal time delay required for reconstruction are determined by the following method: using the false nearest neighbor method, candidate embedding matrices are constructed incrementally starting from embedding dimension 2. The distance between each state point and its nearest neighbor is calculated. When the proportion of state points whose distance decreases abruptly with increasing dimension is lower than the false neighbor proportion threshold of 5%, the current dimension is determined as the optimal embedding dimension. In the typical operating data of this embodiment, this value is stable at 8.

[0068] The mutual information method is used to calculate the mutual information value between the joint state vector and its translation vector after shifting according to different candidate time delays. A decay curve of the mutual information value as a function of delay is plotted. The candidate time delay at which the curve first drops to its initial value 1 / e is selected as the optimal time delay, calculated to be 5 periods. Based on the optimal embedding dimension of 8 and the optimal time delay of 5 periods, a multivariate delay embedding matrix is ​​generated. Next, the covariance matrix of this delay embedding matrix is ​​calculated, and eigenvalue decomposition is performed to extract the largest eigenvalue λ1 and the second largest eigenvalue λ2. The ratio of these two eigenvalues ​​is defined as the multi-node coupling risk coordination strength R_c = λ1 / λ2, which reflects the degree of synchronous amplification of the fluctuations in the live broadcast room state and the virtual human state. When R_c shows a monotonically increasing trend within three consecutive time windows, and the final value exceeds the preset critical coordination threshold of 3.5, the system generates a critical instability warning signal. This signal indicates that a nonlinear state change is about to occur between the live broadcast room and the virtual human agent, such as a simultaneous drop in traffic in many live broadcast rooms and collective overload of virtual humans. Simulation experiments show that when the threshold is 3.5, warnings are successfully issued at least 2 windows in advance for all 15 simulated cascading events, while the false alarm rate increases by 22% when the threshold is 2.8.

[0069] The critical state early warning also covers the integration monitoring of new platform ecosystems. When a new external platform data source is detected in the multi-agent heterogeneous graph, such as an institution joining a sixth live streaming platform, and this data source cannot establish an effective connection with existing live streaming room nodes through content supply edges or collaborative sharing edges, the system generates a temporary probe node in the graph. The probe node, represented by a sample live streaming room of the new platform, attempts to establish content supply edges and collaborative sharing edges with 10 randomly selected nodes of different types in the existing graph in sequence, recording the number of successful attempts. The structural integration barrier index is defined as 1 minus the ratio of the number of successful attempts to the total number of attempts. If this index exceeds the preset integration barrier threshold of 0.6 for five consecutive observation windows, each lasting one minute, a new platform growth trigger signal is generated. This signal will be transmitted to the adaptive shaping and topology reconstruction module in step four, driving independent network building and growth operations for the new platform.

[0070] Through the explicit definition of node attributes, joint calculation of activation values, and multi-level critical state early warning mechanism, the system can perceive resource bottlenecks and structural risks in real time in complex cross-platform environments, providing accurate quantitative basis for subsequent adaptive topology adjustments.

[0071] Within the overall framework of critical state early warning, there are also several in-depth analysis mechanisms targeting specific risk patterns, which are respectively aimed at the tendency of collaborative diffusion of violations and the cascading vulnerability of virtual human resource networks.

[0072] First, the evolution trajectory of violation events is monitored. For each virtual human agent node, the trigger penalty records of all its associated constraint edges are sorted in ascending order of timestamps to construct a violation event stream. Each record in the event stream contains the trigger time and the corresponding penalty intensity. The event interval sequence, i.e., the time difference between two adjacent violation events, in seconds, is extracted from this event stream. Based on this, the local coefficient of variation (LCV) is calculated. The LCV is defined as: the mean of the event interval sequence divided by the standard deviation within a sliding window, i.e., Cv_local = μ_inter / σ_inter; Cv_local: local coefficient of variation; μ_inter: mean of the event interval sequence within the sliding window; σ_inter: standard deviation of the event interval sequence within the sliding window. When the standard deviation is zero, Cv_local is set to a preset large number. It is important to note that the LCV here is expressed as the mean divided by the standard deviation, which is the reciprocal of the common coefficient of variation in statistics. Its purpose is to decrease, rather than increase, this value when violation events tend to become more regular, thus aligning with the meaning of "downward crossing" in the subsequent threshold judgment logic.

[0073] The sliding window length adaptively adjusts based on the event density of the violation event stream. Event density is defined as the number of violation events occurring per unit time; in this example, it's the number of times a violation event occurs within one hour. When the event density increases, the window length shrinks, specifically by fixing the number of target events within the window to 8. When the event density decreases, the window length expands to include 8 events. The upper and lower limits of the window length are set to 20 and 4 events, respectively. Within each sliding window, the local coefficient of variation (LCV) is calculated using the mean and standard deviation of the event interval sequence within that window. The LCVs of each window are then concatenated in chronological order to generate the evolution trajectory of the LCVs over time.

[0074] To accurately determine the evolution trend, an exponentially weighted temporal smoothing filter is applied to the evolution trajectory, with a smoothing coefficient α of 0.3. The smoothed trajectory sequence is denoted as S(t), and its first derivative is calculated as a continuous sequence of signs. When this sequence exhibits alternating positive and negative signs within a preset convergence determination period (15 consecutive periods in this embodiment), and the range between the number of positive and negative signs is lower than a preset sign fluctuation threshold of 2, the evolution trajectory is determined to show a trend from disordered fluctuation to quasi-periodic convergence. If the moving average of the local coefficient of variation further falls below the preset ordering threshold of 0.8, the violation event flow of the virtual human agent node is determined to have entered a critical pre-colonial diffusion state. The ordering threshold of 0.8 is set based on the fact that in the retrospective analysis of eight group violation events in historical data, the moving average of the local coefficient of variation had fallen below 0.8 before all events occurred, while this value is usually higher than 1.2 during normal operation.

[0075] After detecting a node entering a critical pre-colonial diffusion state, the virtual agent node and all live-streaming nodes directly connected to it via scheduling edges are extracted to form a local violation subgraph. Two topological indices are calculated on this subgraph: clustering coefficient and degree-degree correlation. The clustering coefficient is defined as the ratio of the actual number of triangles to the possible number of triangles in the subgraph, reflecting whether the violating nodes tend to form close groups. The degree-degree correlation is obtained by calculating the Pearson correlation coefficient between the degrees of the nodes at both ends of each edge in the subgraph. A positive value indicates that high-degree nodes tend to connect to other high-degree nodes, while a negative value indicates that high-degree nodes tend to connect to low-degree nodes. When the clustering coefficient continuously increases within five consecutive observation windows, and the degree-degree correlation flips from negative to positive, it indicates that the topological propagation conditions for the violation have matured. The system generates a violation phase transition warning signal, which characterizes the transformation of the violation from an isolated event to a collective collaborative diffusion.

[0076] In assessing the vulnerability of virtual human resource networks, the algebraic connectivity of the collaboratively shared subgraph is used as the core indicator. First, a weight matrix of collaboratively shared edges is constructed, where each element represents the style distance weight between two corresponding virtual human agent nodes, with diagonal elements set to zero. The largest connected component with the most nodes in the collaboratively shared subgraph is extracted, and its Laplacian matrix is ​​decomposed using eigenvalue decomposition. The Laplacian matrix is ​​defined as the degree matrix minus the weight matrix; the degree matrix is ​​a diagonal matrix, and its diagonal elements are the sum of the weights of all edges for each node. The second smallest eigenvalue of this Laplacian matrix, namely the algebraic connectivity, is extracted, denoted as λ². A higher algebraic connectivity indicates better network connectivity and stronger resistance to node failures.

[0077] Next, the identification of vulnerable hub nodes is performed. Each virtual agent node in the collaborative shared subgraph is traversed, and the removal of a single node is simulated sequentially. The algebraic connectivity of the largest connected component after removal is recalculated, and the decrease relative to before removal is recorded. The decrease is defined as the difference in algebraic connectivity before and after removal divided by the algebraic connectivity before removal. Nodes with a decrease exceeding a preset vulnerability threshold of 0.3 are marked as vulnerable hub nodes; the failure of these nodes will significantly impact the global connectivity of the network.

[0078] Simultaneously, a two-dimensional state plane is constructed, consisting of the current load and fatigue level of the virtual human agent nodes, with load on the horizontal axis and fatigue level on the vertical axis. All nodes in the collaborative sharing subgraph are mapped to this plane, resulting in a scattered distribution. The Mahalanobis distance of these points relative to the entire distribution is calculated. Mahalanobis distance is defined as the distance between a point and the sample mean vector, weighted by the inverse of the covariance matrix. The mean of the Mahalanobis distances of all nodes is taken as an indicator of the degree to which the group deviates from its normal state. A resource collapse warning signal is generated when three conditions are simultaneously met: algebraic connectivity is below the connectivity warning threshold of 0.05, there is at least one marked vulnerable hub node, and the mean Mahalanobis distance shows a monotonically increasing trend within three consecutive observation windows. This signal indicates that a cascading collapse of the virtual human resource network is imminent. Experiments show that the algebraic connectivity threshold of 0.05 can issue warnings two windows in advance for 10 resource overload events in simulation tests, and the monotonically increasing mean Mahalanobis distance effectively filters out false alarms caused by occasional anomalies in individual nodes.

[0079] Based on the above dynamic analysis results, the adaptive shaping and topology reconstruction module performs three types of operations: personalized attention mask application, pruning operation, and growth and reconnection operation.

[0080] Personalized attention masks are applied to edges with content delivery attributes, including at least content supply edges in this embodiment. The mask generation process is as follows: For each content supply edge, the audience profile vector of the connected livestream node and the content embedding vector of the content fragment node are obtained. The content embedding vector is a pre-defined 128-dimensional feature vector of the content fragment node, extracted by embedding text descriptions and audio features of the content. The audience profile vector and the content embedding vector are multiplied by a dot product, and the result is mapped to the 0-1 range using a Sigmoid function. This result is then multiplied by the current value of the time-decrease function of the content fragment node to obtain the mask coefficient. This mask coefficient is then applied to the current weight of the content supply edge in an element-wise multiplication manner. This mechanism allows the same content fragment to receive differentiated information delivery intensity in livestreams with different audience profiles. For example, promotional content receives enhanced weight in livestreams with audiences who prefer discounts, while it is suppressed in livestreams with audiences who prefer knowledge.

[0081] Pruning operations are performed on inefficient or high-risk live streaming nodes. First, a pruning priority score is calculated for each live streaming node, using the following formula: P_prune = Δ_revenue / (R_exposure + ε); P_prune: Pruning priority score.

[0082] Δ_revenue: The marginal contribution of this node to the global conversion revenue.

[0083] R_exposure: The exposure value of the violation risk associated with this node.

[0084] ε: Zero constant, taken as 0.01.

[0085] Δ_revenue represents the marginal contribution of this node to the global conversion revenue. It is estimated by the difference between the predicted total transaction volume over the global 30-minute period before and after the removal of the subgraph rooted at this node. The prediction model uses a linear regression model, with features including the node's online users, interaction rate, and current conversion status. R_exposure represents the violation risk exposure value associated with this node, equal to the sum of the penalty intensity of all violation events associated with this node in the past 30 days. ε is set to 0.01 to prevent division by zero. When the activation value of a live streaming node is below the dynamic decay threshold of 0.2 for five consecutive observation periods, and the lifecycle factor continues to decline and falls into the low-activity range (i.e., the factor value is less than or equal to 0.3), its pruning priority score is further determined. If this score is below the preset retention threshold of 0.1, soft deletion is performed on the node and all first-level scheduling edges it connects to. The specific operation of soft deletion is as follows: the node state is marked as dormant, all outgoing edge weights are multiplied by the residual factor of 0.01, and the dormant node no longer participates in scheduling calculations, but its data is completely retained. At the same time, a 300-second recovery observation period is initiated. If the lifecycle factor of the node rises to above 0.4 during the observation period, the residual weight will be restored to 50% of the original value, and the node status will be restored to active. If the node has not recovered by the end of the observation period, the node and its related edges will be permanently removed.

[0086] The growth operation is triggered upon receiving a new platform growth trigger signal or detecting an abnormal traffic spike on the platform. An abnormal traffic spike is defined as a platform's total online users increasing by more than 200% of its historical average within 5 minutes. During growth, a new livestream node is first instantiated, its attributes initialized to the target platform's average. Then, nodes ranking in the top 5% of total transaction volume over the past 7 days from historically high-conversion livestreams on the same platform are selected, and their current scheduling subgraph is extracted as a template subgraph. The template subgraph includes the livestream node, its connected scheduling edges, content supply edges, and the content fragment nodes and virtual agent nodes connected to these edges. A Gaussian noise-based random mutation mask is applied to the weights of each content supply edge in the template subgraph. The noise has a mean of 0 and a standard deviation of 0.1, meaning the mutated weight equals the original weight multiplied by 1 + ε, where ε follows N(0, 0.1²). After generating the differentiated scheduling subgraph, it is attached to the new livestream node. Simultaneously, select the virtual human agent node with the closest style embedding vector distance, the smallest cosine distance, and a current load below the load threshold of 3 from the collaborative sharing subgraph, establish a new scheduling edge, and initialize the edge weight to the mean of the corresponding scheduling edge weight in the template subgraph.

[0087] The reconnection operation includes two scenarios. Upon receiving a violation phase transition warning signal, a risk-reward balance plane is constructed. The horizontal axis represents the style embedding vector distance between the candidate virtual human node and the original virtual human node, and the vertical axis represents the node trust factor of the candidate node. The set of replacement virtual human agent nodes—that is, all nodes in the collaborative shared subgraph except the current node—is mapped to this plane. Then, the Pareto front algorithm is used to select nodes located on the non-dominated front, i.e., nodes that are not strictly dominated by other candidate nodes in both style distance and trust factors. One node is randomly selected from these nodes to perform a scheduling edge switch, and the switch process seamlessly continues after the current live stream segment is completed. Upon receiving a resource collapse warning signal, locking protection is implemented on scheduling edges marked as high-value live stream nodes. High-value live stream nodes are defined as those in the top 20% of cumulative transaction volume over the past 7 days. Locking protection is achieved by marking the state of the scheduling edge as locked. In the locked state, the edge weight cannot be modified by pruning operations, nor can it be switched by ordinary reconnection operations; it can only be changed upon receiving a release signal or manual intervention. Meanwhile, the scheduling request of the high-value live streaming room node is preferentially allocated to a pre-reserved dedicated virtual human intelligent agent resource pool. This resource pool consists of 10 virtual human nodes with a trust factor that is consistently higher than 0.9 and a historical average load that is lower than 1.5. It is independently divided from the ordinary scheduling pool and is only open to high-value live streaming rooms during resource collapse warnings. After the warning is lifted, sharing is restored.

[0088] Following the adaptive shaping and topology reconstruction operations in step four, the system enters the closed-loop evolution phase, namely step five, feedback acquisition and parameter adjustment. This phase also includes phase space reconstruction parameter optimization methods, serving as a refined supplement to the joint state vector processing flow in the dynamic analysis module.

[0089] In step five, feedback data collection and adaptive adjustment, the feedback adjustment module performs a core adjustment once after each complete live streaming cycle. In this embodiment, a complete live streaming cycle is defined as 24 hours, consistent with the daily statistical rhythm of the promotional event. The collected feedback data includes four categories: long-cycle conversion revenue, which refers to the total transaction amount generated by all live streaming rooms through virtual human sales within this cycle, minus return amount and platform fines, in yuan; violation cost, which refers to the sum of direct losses and indirect opportunity losses caused by platform penalties such as traffic restrictions, bans, and fines due to compliance violations within this cycle. Indirect opportunity losses are estimated based on the average conversion revenue of the affected live streaming rooms under normal conditions; virtual human switching success rate, which refers to the proportion of instructions where, after the system issues a virtual human replacement instruction, the new virtual human successfully takes over the live stream within 5 seconds without any interruption of the screen, sound lag, or product link failure; and scheduling response latency, which refers to the time interval between the output of the scheduling decision by the dynamic analysis module and the actual start of the corresponding virtual human executing the new task, using the 95th percentile of all scheduling events within this cycle as the representative value.

[0090] Based on the feedback data mentioned above, the system performs targeted adjustments to the edge weights and node trust factors in the multi-agent heterogeneous graph. For scheduling edges that contribute high conversion rates and maintain compliance, their weights are increased. The specific criteria are: the net revenue generated by the livestream room and virtual human combination associated with the scheduling edge during the current period ranks in the top 30% of all scheduling edges, and the virtual human node has not triggered any violations during the current period. For scheduling edges that meet these conditions, their fit weight is increased by 5% from its current value, with an upper limit of 2.0 to prevent infinite weight growth leading to value overflow. The node trust factor of the corresponding virtual human agent node is increased in a preset increment step, with the step value set at 0.05. Experiments show that when the step value is 0.05, the trust factor can recover from the initial minimum value of about 0.5 to close to 1.0 within 30 cycles, which matches the natural recovery cycle of compliance records. If the step value is 0.1, the recovery is too fast, which may cause the short-term compliance appearance to cover up potential risks. If the step value is 0.02, the recovery is too slow, which affects the speed of increasing the scheduling priority of compliant virtual humans.

[0091] For scheduling subgraphs that repeatedly trigger violation risks, a punitive decay is applied. A scheduling subgraph is a directed path rooted at a live-streaming room node, extending through scheduling edges to virtual agent nodes, then through content supply edges to content fragment nodes, and finally through conversion edges to product nodes. When any node in this subgraph triggers a constraint predicate and violates the rules more than three times within the current period, the weights of all edges in the subgraph are reduced by a preset decay factor of 0.9, meaning the current weight of each related edge is multiplied by 0.9. Simultaneously, the node trust factor of the corresponding virtual agent node is also reduced by the decay factor, i.e., the current trust factor is multiplied by 0.9. The decay factor of 0.9 considers the memory effect of violations; after five consecutive violations, the trust factor will drop to approximately 59% of its original value, creating a significant scheduling inhibition effect, while the impact of a single, occasional violation is relatively mild. Furthermore, a stricter penalty coefficient is added to the corresponding policy rule nodes. The penalty intensity parameter stored internally in the policy rule node is increased by 20% from its original value. That is, if the original penalty intensity is 10, it becomes 12 after the addition; if it was originally 100, it becomes 120 after the addition. This addition operation has an upper limit, which is no more than 5 times the initial penalty intensity, to avoid the penalty coefficient growing indefinitely and causing the entire scheduling optimization to degenerate into a pure constraint avoidance problem.

[0092] The allocation of feedback signals employs a differentiated credit attribution method, the core of which is to trace back the net revenue of a single live-streaming cycle along the scheduling edge and content supply edge to the associated virtual agent nodes and content fragment nodes. This embodiment uses a Shapley value-based attribution algorithm. Specifically, for a transaction event within a live-streaming cycle, all participating nodes along the complete path from the live-streaming room node to the product node are extracted, including the live-streaming room node, virtual agent nodes, content fragment nodes, and product nodes, forming a player set in a cooperative game. The net revenue of this transaction is used as the total revenue, and the average marginal contribution of each player across all possible player alliance sequences is calculated, i.e., the Shapley value. When the number of nodes on the path is large, leading to excessive Shapley value calculation, an approximate alternative based on attention weights is adopted: the attention score calculated in step five when generating personalized attention masks is normalized and used as the contribution weight of each node. This weight is multiplied by the total revenue to obtain an approximate value estimate for each node. The value estimates of each node are accumulated and stored in the node attributes to guide subsequent pruning protection priorities and growth initialization strategies. During the pruning operation, the marginal contribution value of live streaming nodes with higher value estimates increases accordingly, thereby raising their pruning priority score and making them less likely to be pruned. During the growth operation, newly instantiated live streaming nodes prioritize reusing the scheduling template subgraph of the top 10% of high-value live streaming nodes in terms of value estimates.

[0093] Step Six: Dynamic Optimization of Phase Space Reconstruction Parameters This step describes the optimal parameter determination method for the phase space reconstruction process performed by the state analysis module, which is closely linked to the critical state early warning mechanism in step three. Step three mentioned earlier involves applying phase space reconstruction to the joint state vector to obtain a multivariate delay embedding matrix; here, the process of determining the optimal embedding dimension and optimal time delay is elaborated in detail.

[0094] The joint state vector is formed by concatenating the activation values ​​of all live-stream nodes and all virtual human agent nodes at the current moment. Its dimension is equal to the total number of nodes of the two types, which is approximately 700 dimensions in a typical configuration of this embodiment. The purpose of reconstructing the phase space of this joint state vector is to recover the dynamic attractor structure of the system in a high-dimensional space, thereby more accurately determining whether the system is approaching a critical state.

[0095] The optimal embedding dimension is determined using the false nearest neighbor method. Starting from a low embedding dimension (in this embodiment, the initial dimension is set to 2), candidate embedding matrices are constructed incrementally. At dimension d, each row of the candidate embedding matrix is ​​a delay vector, whose components are the values ​​of the joint state vector at times t, t-τ, t-2τ up to t-(d-1)τ, where τ is the currently used time delay, initially preset to 5 periods. The Euclidean distance between each state point and its nearest neighbor in the embedding space is calculated for each embedding dimension. When the embedding dimension increases from d to d+1, the new nearest neighbors of each state point are recalculated. If the nearest neighbor distance of a state point decreases abruptly after the dimension increases (i.e., the distance ratio R = new distance / original distance is greater than the preset false neighbor determination threshold of 10), then the state point is determined to be a false nearest neighbor. The proportion of false nearest neighbors to all state points is statistically analyzed. When this proportion decreases to below the preset false neighbor proportion threshold of 5% as the dimension continues to increase, the current embedding dimension is determined as the optimal embedding dimension. In the historical operating data of this embodiment, the optimal embedding dimension has remained stable at 8.

[0096] The optimal time delay is determined using the mutual information method. The mutual information value between the joint state vector X(t) and its translation vector X(t+τ_c) after shifting according to different candidate time delays is calculated. The mutual information value measures the degree of nonlinear dependence between two variables, and its calculation formula is as follows: I(τ_c) = Σ_{i,j} P(x_i, y_j) * log[ P(x_i, y_j) / ( P(x_i) * P(y_j) ) ] ; Where I(τ_c): the mutual information value corresponding to the candidate time delay τ_c, X represents the value of the joint state vector at time t, denoted as X(t); Y represents the value of the joint state vector at time t+τ_c, i.e., the shifted vector, denoted as X(t+τ_c). τ_c is the candidate time delay, in units of decision cycles, which in this embodiment iterates from 1 to 30.

[0097] x_i and y_j are the discretization intervals in the value spaces of X and Y, respectively. Since the joint state vector is a high-dimensional vector of approximately 700 dimensions, directly calculating the joint probability distribution would be computationally extremely expensive. Therefore, in practice, a dimension reduction projection method is used: the top three components with the highest activation values ​​in the joint state vector are selected, and the value ranges of these components are divided into 10 equally spaced intervals to form a three-dimensional grid. Each grid cell is a discretization interval. x_i corresponds to the i-th grid cell of X(t) in this three-dimensional grid, and y_j corresponds to the j-th grid cell of X(t+τ_c) in this three-dimensional grid.

[0098] P(x_i, y_j) is the joint probability that X(t) falls into grid cell x_i and X(t+τ_c) falls into grid cell y_j. Its estimate is the number of sample points satisfying this condition divided by the total number of sample points. P(x_i) is the marginal probability that X(t) falls into grid cell x_i. Its estimate is the number of sample points falling into x_i divided by the total number of sample points. P(y_j) is the marginal probability that X(t+τ_c) falls into grid cell y_j, estimated in the same way. Sample points are derived from snapshots of all joint state vectors within the current sliding time window. The window length is the data from the most recent 500 decision periods to ensure statistical stability.

[0099] In this embodiment, the base of the logarithmic function is taken as the natural constant e, and the unit of mutual information is knight. The summation is performed on all grid cell pairs (i, j). When the number of samples in a certain grid cell is zero, the joint probability P(x_i, y_j) of that cell is 0. According to the convention 0 * log(0) = 0, this term does not contribute to the summation.

[0100] By using the discretization estimation method described above, the mutual information calculation of high-dimensional continuous variables is transformed into frequency statistics and weighted summation on a finite grid, which balances computational feasibility and estimation accuracy.

[0101] The probability distribution is obtained by projecting the components of the joint state vector and plotting a two-dimensional histogram. Mutual information values ​​are calculated for each candidate time delay from 1 to 30 periods, and a decay curve of the mutual information value as a function of the candidate time delay is plotted. The candidate time delay at which the mutual information value first decreases to its initial value (τ_c = 0) and the predetermined decay ratio 1 / e is selected as the optimal time delay. If the curve shows a clear first local minimum point during its descent (i.e., the mutual information value at this point is lower than the values ​​of the two adjacent points), the candidate time delay corresponding to this point is preferentially selected as the optimal time delay. In the typical operating data of this embodiment, the optimal time delay is 5 periods, consistent with the initial predetermined value, indicating that the setting of the predetermined value is statistically reasonable.

[0102] Based on the optimal embedding dimension of 8 and the optimal time delay of 5 periods determined by the above method, phase space reconstruction is performed on the joint state vector to generate a multivariate delay embedding matrix. Each row of this matrix corresponds to a reconstructed state vector. The number of columns in the matrix is ​​equal to the optimal embedding dimension multiplied by the original dimension of the joint state vector, and the number of rows is equal to the original time series length minus the product of the optimal embedding dimension and the optimal time delay. Subsequent calculations of the covariance matrix, eigenvalue decomposition, and the collaborative strength calculation of multi-node coupling risk are all based on this delay embedding matrix, as detailed in step three.

[0103] It should be noted that the computational overhead of the spurious nearest neighbor method and the mutual information method is directly related to the dimension of the joint state vector. In actual system operation, optimal parameter estimation is not re-executed every decision cycle. Instead, parameter re-estimation is triggered every 300 decision cycles (approximately 50 minutes), or when a significant change in the system topology is detected, such as the addition or removal of more than 10% of nodes, to balance computational efficiency and analytical accuracy. The optimal embedding dimension and optimal time delay determined most recently are used between two re-estimations. Experimental data shows that during cross-platform promotional periods, the system topology changes frequently but the dynamic characteristics remain relatively stable. The 50-minute re-estimation interval can keep the additional consumption of computational resources within 5% while ensuring early warning sensitivity.

[0104] In the critical state early warning mechanism, the accuracy of the analysis of the violation event flow directly affects the timeliness of the early warning for the spread of group violations. This embodiment further refines the calculation of the local variation coefficient and the method for determining its evolution trajectory, which will be described in detail below in the context of a cross-platform promotional event.

[0105] For a single virtual human agent node, the violation event stream is already sorted in ascending order of timestamps. The event interval sequence is extracted from this event stream, denoted as I = {Δt1, Δt2, …, Δt}. n}, where Δtᵢ = tᵢ − tᵢ₋1, in seconds. To capture the micro-changes in the rhythm of violation events, a sliding window analysis is applied to the event interval sequence. The window length of the sliding window is not a fixed value, but is adaptively adjusted according to the event density of the violation event stream. The event density ρ is defined as the number of violation events occurring per unit time, which is 3600 seconds in this example. The window adjustment rule is as follows: the number of events expected to be included in each sliding window is set to a constant value K, which is 8 in this embodiment. When the event density increases, that is, when violation events tend to be more frequent, maintaining the original window length will result in a longer time span across the window, containing too many events. At this time, the window length is shrunk so that the window contains exactly the intervals of the most recent 8 violation events. When the event density decreases, the window length is expanded to include 8 intervals. The upper and lower limits of the window length are set to 20 and 4 times, respectively, to avoid window drift and loss of control in extremely sparse or dense cases. This adaptive windowing mechanism ensures that local statistics are always calculated based on the same number of event samples under different violation frequencies, enhancing the comparability of local variation coefficients over time.

[0106] Within each sliding window, the mean μ_inter and standard deviation σ_inter of the event interval sequence within that window are calculated to obtain the local coefficient of variation Cv_local = μ_inter / σ_inter for that window. When the standard deviation is zero, meaning all intervals within the window are equal, it indicates that the violation events are completely periodic. In this case, Cv_local is set to 0.01 to represent extremely high regularity. The local coefficients of variation corresponding to each sliding window are concatenated in chronological order according to the window center time to generate the evolution trajectory of the local coefficients of variation over time.

[0107] To suppress the interference of high-frequency noise in the original trajectory on trend determination, an exponentially weighted temporal smoothing filter is used to smooth the evolved trajectory. The smoothing filter is a first-order low-pass filter, and its recursive formula is as follows: S(t) = α * Cv_local(t) + (1-α) * S(t-1); S(t): The smoothed trajectory value.

[0108] Cv_local(t): The local variation coefficient of the current window.

[0109] α: Smoothing coefficient, set to 0.3.

[0110] S(t-1): The smoothed trajectory value of the previous window.

[0111] The choice of a smoothing coefficient of 0.3 balances the response speed to abrupt signals and the ability to suppress random fluctuations. Experiments show that when α is 0.3, the smoothed trajectory can reflect the trend reversal within about 3 window periods after the change in the violation rhythm, while attenuating the pseudo-periodic fluctuation amplitude to less than 40% of the original trajectory.

[0112] The first derivative of the smoothed trajectory is approximately calculated using backward difference, i.e., S'(t) = S(t) - S(t-1). The sign of the first derivative is continuously taken: positive values ​​are denoted as +1, negative values ​​as -1, and zero values ​​retain the previous non-zero sign. This yields a symbol sequence. When this symbol sequence exhibits an alternating positive and negative oscillation pattern within a preset convergence determination period (15 consecutive sliding windows in this embodiment), and the absolute value of the difference between the number of positive and negative signs (i.e., the difference between the number of positive and negative signs) is lower than a preset symbol fluctuation threshold of 2, the evolution trajectory is determined to show a trend of convergence from disordered fluctuations to quasi-periodicity. The core of this determination logic is that when violations tend to become regularized, the directionality of the rate of change of the local coefficient of variation disappears before it crosses the ordering threshold, instead exhibiting a high-frequency oscillation around a stable mean, reflected in the symbol sequence as alternating positive and negative signs with balanced numbers. By combining this trend judgment with the threshold of the local coefficient of variation moving average, the critical pre-diffusion state can be identified in advance, providing earlier triggering conditions for subsequent illegal phase transition warnings.

[0113] Connect the above process with the aforementioned critical collaborative diffusion precursor determination: when the evolution trajectory is determined to show a quasi-periodic convergence trend, and the moving average of the local variation coefficient further falls below the ordering threshold of 0.8, it is confirmed that the violation event flow of the virtual human intelligent agent node has entered the critical collaborative diffusion precursor state. Subsequently, the local violation subgraph is extracted to calculate the clustering coefficient and degree correlation, and the violation phase transition early warning signal is generated.

[0114] The core of adaptive tuning of topology reconstruction triggering conditions is to enable each critical threshold in the system to be dynamically corrected based on the performance feedback of the actual reconstruction operation, thereby reducing false alarms and false negatives and gradually approaching the optimal early warning sensitivity.

[0115] First, all critical thresholds in the system are uniformly constructed into an adjustable threshold vector Θ = [θ_coop, θ_order, θ_conn, θ_decay, θ_retain], where θ_coop is the critical cooperation threshold with an initial value of 3.5; θ_order is the ordering threshold with an initial value of 0.8; θ_conn is the connectivity warning threshold with an initial value of 0.05; θ_decay is the dynamic decay threshold with an initial value of 0.2; and θ_retain is the retention threshold with an initial value of 0.1. Each component is processed independently in subsequent adjustments; the vector form is only used for unified recording and management.

[0116] Threshold adjustments are performed at the end of each preset tuning cycle. In this embodiment, the tuning cycle is set to 24 hours, synchronized with the complete live broadcast cycle of the feedback adjustment module. All topology reconstruction operation execution records triggered by each critical state warning signal are collected within this cycle. These records include: pruning operations triggered by critical instability warnings, growth operations triggered by new platform growth warnings, reconnection scheduling edge switching operations triggered by violation phase transition warnings, and lockout protection and dedicated pool allocation operations triggered by resource collapse warnings. Each record includes the operation type, trigger signal source, corresponding actual state indicator value at the time of triggering, and snapshots of global conversion gains and violation costs before and after the operation execution. The change in global conversion gains is defined as the net gain within 30 minutes after the operation execution minus the net gain within 30 minutes before execution; the change in violation costs is defined similarly.

[0117] Based on the above records, a reconstruction efficiency evaluation matrix is ​​constructed. The matrix represents each topology reconstruction operation, and its columns include the operation type identifier, trigger signal, actual state indicator value before execution, revenue increment, violation cost increment, and the number of scheduling interruptions caused by the operation. The number of scheduling interruptions is defined as follows: for pruning operations, it is the number of ongoing live streams that are forced to stop due to soft deletion of live stream nodes; for reconnection switching operations, it is the number of live stream screen switches that occur when virtual humans are replaced due to scheduling edge switching; for growth operations, the number of scheduling interruptions is set to 0; for lock protection and pool allocation operations, since these operations themselves do not cause interruptions, the number is also set to 0. Through this definition, the number of interruptions directly reflects the instantaneous disturbance of the online service caused by the reconstruction operation.

[0118] Calculate a performance score (Score) for each operation record in the matrix. The formula for calculating the performance score is: Score = (ΔR − ΔC) / (N_interrupt + 1); Score: Performance rating of topology remodeling operations.

[0119] ΔR: Global conversion revenue increment.

[0120] ΔC: Incremental cost of violations.

[0121] N_interrupt: Number of scheduled interrupts. The denominator is incremented by 1 to avoid division by zero.

[0122] This score represents the net benefit gained from each refactoring operation at the cost of a unit of scheduling interruption. A higher performance score indicates that the operation, given the actual state indicators and threshold values ​​used, produces good benefits and causes minimal disturbance to the system.

[0123] All operations within this period are sorted from highest to lowest performance score. Operations in the top 50% are classified as high-efficiency refactoring operations, and those in the bottom 50% are classified as low-efficiency refactoring operations. For each category, a threshold adjustment analysis is performed.

[0124] For efficient reconstruction operations, the actual state index value of the corresponding critical state warning signal before its execution is extracted. For example, if the trigger for an efficient pruning operation originates from a critical instability warning, the actual state index value is the multi-node coupling risk coordination strength R_c at that time; if it is a growth operation, it corresponds to the structural fusion barrier index; if it is a reconnection switch, it corresponds to the aggregation coefficient or degree correlation correction during the violation phase transition judgment, but for unified scalar processing, this embodiment uses the aggregation coefficient of the local violation subgraph at the judgment as the actual value; if it is a resource collapse operation, it corresponds to the algebraic connectivity. These actual values ​​are compared with the corresponding thresholds in the adjustable threshold vector to calculate the deviation direction and magnitude. The deviation magnitude is defined as |actual value - threshold| / threshold.

[0125] If the actual state index value of the efficient reconstruction operation is on the lenient side of the corresponding threshold—that is, for cases where the threshold must be exceeded to trigger, the actual value is further away from the trigger condition (e.g., the critical coordination threshold needs to exceed 3.5 to trigger, but the actual value is 4.2, thus it is on the lenient side); and for cases where the threshold needs to be lowered to trigger, the actual value is further away from the trigger condition (e.g., the ordering threshold needs to be lower than 0.8 to trigger, but the actual value is 0.6, also on the lenient side), and the deviation exceeds the preset hysteresis tolerance range, it indicates that the existing threshold is too sensitive and triggers the reconstruction operation too early. Although the operation is classified as efficient, the system is still far from a dangerous state when it is executed. The threshold should be adjusted towards the stricter direction to postpone the triggering of similar warnings and reduce unnecessary operational overhead. The hysteresis tolerance range is set to a deviation within 15%, meaning that adjustment is only initiated if the actual value deviates from the threshold by more than 15%. The adjustment is made by one step in the stricter direction, with the step size set to 10% of the current threshold. For example, the critical coordination threshold of 3.5 increases by 10% to 3.85, and the ordering threshold of 0.8 decreases by 10% to 0.72.

[0126] For inefficient reconstruction operations, extract their actual state index values. If the actual value is on the strict side of the corresponding threshold, meaning it has exceeded the triggering condition but the operation itself is inefficient (e.g., the critical coordination threshold is 3.5, and the actual value is 3.2, meaning it wasn't triggered but the operation was falsely triggered), it's important to note that inefficient operations are also triggered by warning signals. Therefore, the actual value must have already exceeded the threshold. "On the strict side" means the actual value has just exceeded the threshold, very close to it, and the operation is inefficient, indicating that this trigger might be a false alarm or premature triggering. If the deviation exceeds the preset false alarm tolerance range (also set to within 15%), the corresponding threshold is adjusted to a more lenient direction by one step, also 10% of the current threshold, so that similar situations are less likely to trigger in the future. For example, the critical coordination threshold is lowered to 3.15, and the ordering threshold is raised to 0.88.

[0127] It is important to note that the adjustment of each threshold is independent and is performed separately only during analysis based on the corresponding trigger signals. An adjustment log is recorded after each adjustment, and preset safety boundary limits are applied to each component of the threshold vector. For example, the lower limit of the critical collaborative threshold is 2.0 and the upper limit is 5.0, and the lower limit of the ordering threshold is 0.3 and the upper limit is 1.5, etc., to prevent the thresholds from drifting into unreasonable ranges.

[0128] Through the aforementioned threshold adaptive tuning mechanism based on performance feedback, the system gradually learns the optimal warning trigger point during multiple consecutive major promotional events. Experimental data shows that after approximately seven tuning cycles (7 days), the fluctuations of various thresholds tend to stabilize. The critical coordination threshold converges to around 3.8, the ordering threshold converges to around 0.75, the connectivity warning threshold converges to 0.04, the dynamic decay threshold converges to 0.18, and the retention threshold converges to 0.12. Compared to the initial fixed thresholds, the adaptively tuned thresholds reduce the false alarm rate by 28%, the false negative rate by 17%, and improve the average performance score of the reconstruction operation by 35%, indicating that the system effectively reduces unnecessary scheduling interruptions caused by false triggers while maintaining high conversion gains.

[0129] In the feedback loop evolution of step five, differentiated credit attribution is used to backtrack the net revenue of a single live broadcast cycle to each participating node along the scheduling path to update its value estimate. This embodiment provides an attribution weight propagation method based on a multi-branch backtracking tree as a specific implementation of credit attribution. Its calculation process does not require traversing all node permutations, resulting in higher computational efficiency and a transparent and interpretable attribution path.

[0130] First, construct a multi-branch backtracking tree structure. After a complete live stream cycle, select any completed live stream node as the backtracking starting point. Using this backtracking starting point as the root node, perform reverse backtracking along the edges in the graph. Specifically, starting from the root node, backtrack along all content supply edges pointing to the root node to each content fragment node; this is the first backtracking branch. Starting from the root node, backtrack along all scheduling edges pointing to the root node to each virtual agent node; this is the second backtracking branch. Further, starting from the backtracked content fragment nodes, backtrack along all conversion edges pointing to these content fragment nodes to each product node; this is the third-level backtracking branch. If other directional relationships exist in the graph, they can continue to expand in the reverse edge direction. In this way, all nodes and edges on the reverse path form a multi-branch tree rooted at the backtracking starting point, and the depth of the tree depends on the maximum number of levels that the reverse backtracking can reach. In this embodiment, the tree node types include live streaming room nodes, content segment nodes, product nodes, and virtual human intelligent agent nodes, and the direction of the edges is opposite to the original scheduling and supply direction.

[0131] Next, on the constructed multi-branch backtracking tree structure, attribution weights are propagated according to a preset layer-by-layer decay factor. The attribution weight of the root node is initially set to 1, a unit value, representing that all the basic weight contributed by the backtracking starting point within this live stream cycle is attributed to this live stream. Then, starting from the root node, the attribution weights are propagated to the next layer of nodes along each edge. Each time an edge is traversed, the weight is multiplied by a decay factor γ, where γ is a preset constant ranging from 0 to 1; in this embodiment, γ is set to 0.6. This means that the first-order adjacent nodes directly connected to the root node—namely, content fragment nodes and virtual human agent nodes—receive an attribution weight equal to the root node weight multiplied by γ, i.e., w1 = 1 * γ = 0.6. The second-order adjacent nodes—namely, the product nodes backtracked from the first-order content fragment nodes along the transformation edge—receive an attribution weight equal to the first-order node weight multiplied by γ, i.e., w2 = w1 * γ = 0.36. Similarly, the attribution weight obtained by the nth adjacent node is γ^n. The attribution weight decays exponentially with the number of layers. Weight propagation continues along a path until the attribution weight calculated by a node of a certain order falls below a preset propagation cutoff threshold, at which point propagation on that path immediately stops. The propagation cutoff threshold is set to 0.01, stopping when γ^n < 0.01. In this embodiment, when γ is 0.6, the weight drops to 0.028 when n≥7 and below 0.01 when n≥9, thus automatically controlling the propagation depth.

[0132] Because the same node may appear at different positions in a multi-branch backtracking tree via multiple different paths—for example, a content segment node might be directly connected to the root node via a content supply edge, or it might be indirectly backtracked via a longer path—this node will receive attribution weights from multiple paths. Therefore, it is necessary to sum all the attribution weights obtained by the same node through different paths in the multi-branch backtracking tree to obtain the node's comprehensive attribution weight. This comprehensive attribution weight serves as the node's attribution allocation weight for the net revenue of this live streaming period. The attribution allocation weight is a scalar used to measure the relative contribution of the node to the final conversion revenue.

[0133] After calculating the comprehensive attribution weight, the node's value estimate is weighted and updated using this weight. The node's value estimate is stored in the node attributes and used to guide pruning protection priorities and growth initialization strategies. The update method uses an exponential moving average, meaning that the attribution weight calculated in this operation is fused with the node's historical value estimate using a preset smoothing coefficient. Let the node's current historical value estimate be V_old, the comprehensive attribution weight calculated in this operation be w_sum, and the smoothing coefficient be β. Then, the updated value estimate V_new is calculated using the following formula: V_new = β * w_sum + (1 - β) * V_old; The smoothing coefficient β is set to 0.2. When β is small, the inertia of historical value estimation is greater, which can smooth the fluctuations of single attribution and make the value estimation gradually approach the long-term average contribution. Experiments show that when β is 0.2, the value estimation of a node can be basically stable after about 10 effective attribution updates, while when β is 0.1, the update is too slow, and when β is 0.4, it is easily disturbed by single abnormal fluctuations.

[0134] To avoid the value estimate of the same node being updated repeatedly in a short period of time, a minimum update interval can be set. That is, the value estimate of the same node is only updated once in a single tuning cycle. If the node appears again in the multi-branch backtracking tree corresponding to other backtracking starting points in the same cycle, its newly calculated comprehensive attribution weight can be added to the existing weight and then an exponential moving average is performed with the historical value estimate at the end of the current cycle to reduce computational overhead.

[0135] Using the multi-branch backtracking tree attribution method described above, the system accurately quantifies the contributions of participating nodes after each live broadcast cycle. The value estimate is continuously updated and used to guide subsequent topology reconstruction operations. The layer-by-layer decay mechanism of this method intuitively reflects the objective law that contributions decrease as the propagation path lengthens. The propagation truncation threshold avoids the over-amplification of weak contributions from deep nodes, while the exponential moving average ensures the stability and reliability of the value estimate. Compared with the Shapley value method, the computational complexity of this method is only proportional to the size of the backtracking tree, exhibiting a significant computational efficiency advantage in multi-agent heterogeneous graphs with a large number of nodes.

[0136] During the threshold adaptive tuning process, each topology reconstruction operation needs to be divided into efficient and inefficient reconstruction operations based on its performance score, so as to analyze its relationship with the trigger threshold separately. This embodiment further provides an adaptive partitioning method based on group statistics and Bayesian classification, replacing the fixed ratio partitioning, so that the determination of efficient and inefficient operations is more in line with the distribution characteristics of each type of operation.

[0137] First, all topology reconstruction operations recorded during this tuning cycle are grouped by operation type into three groups: pruning operations, growth operations, and reconnection operations. The reconnection operations group includes scheduling edge switching triggered by illegal phase transitions and lock protection and pool allocation triggered by resource crashes; both belong to the reconnection category in terms of operation characteristics. Subsequent statistical and classification steps are performed independently within each group.

[0138] For the performance score data within each group, extract its first quartile (Q1) and third quartile (Q3). The first quartile is the value at the 25th percentile after the performance score of the group is sorted in ascending order, and the third quartile is the value at the 75th percentile. Calculate the interquartile range (IQR) = Q3 - Q1. Within this group, the upper whisker is determined as the third quartile plus the interquartile range of a preset whisker multiple, i.e., upper whisker = Q3 + k * IQR. In this embodiment, the preset whisker multiple k is 1.5, consistent with the outlier determination threshold of the standard box plot. Topology reconstruction operations with performance scores exceeding this upper whisker are judged as outlier high-efficiency operations, and these operations are directly marked as high-efficiency reconstruction operations. Outlier high-efficiency operations usually represent an operation that achieved a net benefit far exceeding the average level of similar operations within the current period, and their success is worth referring to in subsequent threshold adjustments.

[0139] Simultaneously, the median of the performance scores for this group is identified. Topology reconstruction operations with performance scores below the group median are extracted and, together with the aforementioned outlier efficient operations, form the labeled sample set for this group. In the labeled sample set, outlier efficient operations are assigned positive class labels, while the remaining operations are tentatively considered inefficient operations and assigned negative class labels. This construction method ensures that the labeled sample set contains two classes of samples with clear distributional differences: efficient samples at the tail end of the distribution and inefficient samples in the lower half, which is beneficial for subsequent learning of classification boundaries.

[0140] Next, a one-dimensional kernel density estimation model is constructed for each group of operations, using the performance score as input. For positive class samples in the labeled sample set, i.e., outlier efficient operations, a kernel density estimation function f_high(x) is independently fitted; for negative class samples, i.e., tentatively inefficient operations, a kernel density estimation function f_low(x) is independently fitted. The kernel density estimation uses a Gaussian kernel function, and the bandwidth is automatically selected through cross-validation. Its calculation formula is as follows: f(x) = (1 / (n*h)) * Σ K((x - x_i) / h) Where n is the number of samples in the corresponding category, h is the bandwidth, K is the standard normal density function, and x_i is the performance score of each sample. Kernel density estimation can smoothly approximate the true probability density distribution, avoiding pre-defined assumptions about the shape of the performance score distribution.

[0141] After obtaining the conditional probability density functions for both classes, a Bayesian optimal classification boundary is constructed. It is assumed that the prior probabilities of the positive and negative classes are estimated from the class proportions in the labeled sample set, i.e., P(high) = N_high / (N_high + N_low), P(low) = N_low / (N_high + N_low), where N_high is the number of outlier efficient operations, and N_low is the number of tentatively inefficient operations. According to the minimum error rate Bayesian decision rule, the classification boundary is the point where the posterior probabilities are equal, i.e., satisfying: P(high|x) = P(low|x); By Bayes' theorem, it is equivalent to: f_high(x) * P(high) = f_low(x) * P(low); The efficiency score x obtained by solving this equation is the efficiency threshold θ_score. When the efficiency score of an operation is greater than θ_score, its posterior probability is biased towards the efficient category, and it is judged as an efficient reconstruction operation; when it is less than θ_score, it is judged as an inefficient reconstruction operation. If there are multiple solutions to the equation within the effective interval of the efficiency score, the solution located between the mean of outlier efficient operations and the provisional mean of inefficient operations is selected as the threshold. If there is no solution within the interval, the median is used as the default threshold.

[0142] Through the above steps, an independent performance threshold is determined for each group of operations. The pruning, growth, and reconnection operation groups each have their own thresholds that reflect their specific score distribution characteristics, rather than using a uniform fixed ratio. In the next tuning cycle, the updated performance thresholds for each group are directly used to classify newly generated topology reconstruction operations as efficient or inefficient, without needing to re-estimate kernel density and perform Bayesian boundary fitting, thus reducing computational overhead. The entire partitioning process is repeated only every 5 tuning cycles or when a significant drift in the performance score distribution is detected—that is, when the difference between the mean performance scores of two groups exceeds twice the historical standard deviation—to update the performance thresholds.

[0143] After adopting the above adaptive partitioning method, the determination of efficient and inefficient reconstruction operations no longer relies on a preset fixed ratio, but is based on the statistical characteristics of the efficiency scores of various operations and the Bayesian optimal classification criterion, making the partitioning results more robust. For operation types with generally high efficiency scores, such as growth operations which often bring significant benefits during peak traffic periods, their partitioning threshold naturally shifts upward, avoiding misclassifying normal high-yield operations as outliers. For operation types with large fluctuations in efficiency scores, such as pruning operations which occasionally recover losses by prematurely terminating declining live streams, the outlier efficient operations identified above are more meaningful. Experimental data shows that compared to fixed-ratio partitioning, this method improves the accuracy of threshold adjustment direction by about 18%, reducing the situation where thresholds are adjusted in the wrong direction due to misclassification.

[0144] This embodiment provides a cross-platform unmanned live streaming resource scheduling system based on intelligent agent collaboration. This system stores and executes the methods described in any of the preceding embodiments. The system is deployed on a central scheduling cluster consisting of multiple servers, which are interconnected via 10 Gigabit Ethernet and establish long-lived connections with the open API interfaces of five external live streaming platforms for data retrieval and instruction issuance. Internally, the system is functionally divided into a data preprocessing module, a heterogeneous graph construction module, a dynamic analysis module, an adaptive shaping and topology reconstruction module, and a feedback adjustment module. Modules exchange data through shared memory and message queues based on a publish-subscribe pattern.

[0145] The data preprocessing module is responsible for the structured decomposition and association construction of multi-source heterogeneous data. This module consists of a data adapter subunit, a standardization engine subunit, and an association mining subunit. The data adapter subunit maintains an independent set of data retrieval processes for each external platform. Internally, each process encapsulates the corresponding platform's API authentication protocol, pagination logic, and rate limiting and retry strategies. The process writes the retrieved raw JSON messages, log text, and real-time video stream metadata into the module's internal raw data buffer. The standardization engine subunit reads data line by line from the raw data buffer. According to seven preset unified element dimensions—entity identifier, entity type, timestamp, platform encoding, event type, event payload, and session identifier—it maps and extracts the fields of each data entry, uniformly converting heterogeneous fields into an internal standard format. The converted data is then written to the standard data bus. The association mining subunit subscribes to the standard data bus and performs rule engine matching on the incoming standardized event stream. The rule engine has built-in matching logic for four types of association rules: attribution, triggering, transformation, and similarity. The matching results are output to the structured association data storage area in the form of a triple: source entity, target entity, and relationship type, along with relationship weight and timestamp. The storage area uses a relational database for persistent storage on disk, and also maintains a mirror image in memory using an adjacency list structure. The adjacency list uses the entity identifier as the key and the outgoing and incoming edge lists as values, providing a high-speed access interface for the downstream heterogeneous graph construction module.

[0146] The heterogeneous graph construction module is used to construct multi-agent heterogeneous graphs based on structured association datasets. This module contains a graph pattern definition subunit and an instantiation engine subunit. The graph pattern definition subunit maintains metadata about node and edge types. Node types include five types: live streaming nodes, virtual human agent nodes, content fragment nodes, product nodes, and strategy rule nodes. Edge types include five types: scheduling edges, content supply edges, conversion edges, collaborative sharing edges, and constraint edges. The attribute schema for each node and edge type is managed in configuration files, supporting dynamic expansion at runtime. The instantiation engine subunit periodically reads incremental data from the structured association data storage area. For newly added entities, it calls the corresponding node constructor, fills in the attribute fields, and inserts them into the graph database. The specific definitions of node attribute fields are consistent with the aforementioned method implementation and will not be repeated here. For newly added relationships, the instantiation engine calls the corresponding edge constructor, calculates the edge weight, and establishes a connection in the graph database as a directed or unrelated edge. After instantiation, the multi-agent heterogeneous graph is fully stored in the in-memory graph database to support millisecond-level traversal queries, and snapshots are periodically persisted to the distributed file system.

[0147] The dynamic analysis module is used to perform spatiotemporally sequential dynamic analysis and critical state early warning on multi-agent heterogeneous graphs. This module includes an activation value calculation subunit, a conflict detection subunit, and a critical state early warning subunit. The activation value calculation subunit, within each decision cycle (10 seconds), traverses all nodes in the graph and calculates the dynamic activation value according to the respective calculation formulas for livestream nodes, virtual agent nodes, content fragment nodes, product nodes, and policy rule nodes. The weight coefficients, mapping functions, and threshold parameters involved in the calculation formulas are defined as configurable constants in the module initialization file. Before each scheduling decision is generated, the conflict detection subunit traverses candidate scheduling paths and checks whether the current load of the virtual agent nodes associated with the path has reached a preset limit. If it has, the path is marked as a conflict path and excluded from the candidate set.

[0148] The critical state early warning subunit incorporates a phase space reconstruction engine, a covariance analysis engine, and a new platform fusion monitoring engine. The phase space reconstruction engine performs embedding dimension estimation based on spurious nearest neighbors and time delay estimation based on mutual information on the joint state vector, generating a multivariate delayed embedding matrix. The covariance analysis engine calculates the covariance matrix on the delayed embedding matrix and performs eigenvalue decomposition to obtain the multi-node coupling risk coordination strength. When the strength monotonically increases for three consecutive windows and exceeds the critical coordination threshold, a critical instability early warning signal is output. The new platform fusion monitoring engine monitors the structural fusion barrier index of external data sources. When the index exceeds the threshold, a new platform growth trigger signal is output. Furthermore, the critical state early warning subunit also includes a violation event flow analysis engine and a resource network vulnerability analysis engine, used to generate violation phase transition early warning signals and resource collapse early warning signals, respectively. Their analysis logic has been detailed in the aforementioned method embodiments.

[0149] The adaptive shaping and topology reconstruction module is used to perform adaptive shaping and topology reconstruction based on the results of dynamic analysis. This module includes attention masking sub-units, pruning sub-units, growth sub-units, and reconnection sub-units. After receiving the node activation values ​​output by the dynamic analysis module, the attention masking sub-unit calculates a personalized attention mask for each edge in the graph that has content transitivity attributes. The mask is jointly generated by the audience profile vectors, content embedding vectors, and content timeliness decay function values ​​of the nodes at both ends of the edge. After generation, it is applied to the edge weights in an element-wise multiplication manner, and the updated edge weights are written back to the graph database.

[0150] The pruning subunit maintains a pruning priority score calculation thread. For livestream nodes whose activation values ​​are continuously below the dynamic decay threshold and whose lifecycle factors fall into the low-activity range, the ratio of their marginal contribution value to their violation risk exposure value is calculated as the pruning priority score. Nodes below the retention threshold are marked as dormant, their edge weights are multiplied by the residual factor, and a recovery timer is started. Nodes that fail to recover within the timeout period are completely removed from the graph database by the pruning subunit. The growth subunit listens for new platform growth trigger signals and abnormal traffic peak events. When an event occurs, it selects a high-conversion livestream subgraph template for the corresponding platform from the template library, calls the mutation engine to apply Gaussian noise to the edge weights, instantiates a new node, and attaches it to the subgraph. At the same time, it selects low-load style neighbor virtual humans from the collaborative shared subgraph to establish scheduling edges. The reconnection subunit listens for violation phase transition warning signals and resource collapse warning signals. The former triggers the scheduling edge switching process, selecting Pareto front candidate nodes on the risk-reward balance plane for replacement; the latter triggers the locking protection of high-value livestream scheduling edges and the allocation of dedicated resource pools.

[0151] The feedback adjustment module collects feedback data and adjusts graph weights, node trust factors, and topology reconstruction triggering conditions based on feedback signals, forming a closed-loop evolution. This module includes a feedback collection subunit, a weight adjustment subunit, and a threshold optimization subunit. At the end of each 24-hour live streaming cycle, the feedback collection subunit retrieves transaction reports and violation records from various platforms, combines them with internal system logs to calculate virtual human switching success rates and scheduling response latency, and summarizes them into a feedback data packet. The weight adjustment subunit parses the feedback data packet, identifies scheduling edges that contribute high conversion rates and are compliant, and enhances their weights. It identifies scheduling subgraphs that repeatedly trigger violation risks, reduces their edge weights, lowers the trust factor of the corresponding virtual human nodes, and adds penalty coefficients to policy rule nodes. The threshold optimization subunit maintains an adjustable threshold vector and the safety boundaries of its components. At the end of each optimization cycle, it calculates the efficiency score of each operation based on the reconstruction efficiency evaluation matrix. After grouping by operation type, it uses a Bayesian classification method based on kernel density estimation to determine the high-efficiency and low-efficiency boundary thresholds for each group. Based on this, it classifies operations and performs targeted adjustments to each threshold.

[0152] The five modules described above are deployed as independent processes on the central scheduling server. These processes read and write graph data through a unified shared memory area and transmit control signals and warning notifications via message queues. Within each module, sub-units execute concurrently using multi-threading, fully utilizing the server's multi-core CPU resources to ensure that the entire graph's node activation values ​​are updated and a critical state scan is completed within each 10-second decision cycle. The system supports hot updates during continuous operation; node type schemas, threshold parameters, and decay coefficients can all be modified online via configuration files and periodically hot-loaded by each module without requiring system downtime or restart. Through this system architecture, this embodiment achieves the engineering deployment of a cross-platform unmanned live streaming resource scheduling method. During peak periods, it can stably manage over 200 live streaming rooms and 500 virtual human agents, completing weight updates for approximately 50,000 edges within a single decision cycle, with the 95th percentile of scheduling response latency controlled within 2 seconds.

[0153] As can be seen from the above embodiments, this technical solution abstracts the resource scheduling problem of cross-platform unmanned live streaming into a dynamic analysis and adaptive topology reconstruction problem of multi-agent heterogeneous graphs. Within each decision cycle, the system completes a full-process analysis from multi-source heterogeneous data standardization, graph structure instantiation, node activation value calculation to critical state early warning, and triggers three types of topology reconstruction operations—pruning, growth, and reconnection—based on the early warning signal. The effects of the reconstruction operations are fed back to the system in the form of increased conversion revenue and changes in violation costs. On the one hand, the revenue contribution is accurately allocated to each participating node through a layer-by-layer decay attribution mechanism of a multi-branch backtracking tree, updating its value estimate and trust factor. On the other hand, each critical threshold is adaptively optimized through a Bayesian classification method based on grouped kernel density estimation. This closed-loop mechanism of analysis, decision-making, execution, feedback, and optimization enables the scheduling system to continuously and autonomously evolve in environments with traffic surges and frequent changes in platform rules. Experimental data shows that in a cross-platform promotional scenario managing more than 200 live streaming rooms and 500 virtual human agents, the utilization rate of virtual human resources increased from the initial 62% to 89%, the global conversion revenue increased by about 35% compared to the static scheduling scheme, the number of live streaming room interruptions caused by violations decreased by 70%, the 95th percentile of scheduling response latency was controlled within 2 seconds, and the false alarm rate of critical state warnings decreased by 28% after adaptive optimization.

Claims

1. A cross-platform unmanned live streaming resource scheduling method based on intelligent agent collaboration, characterized in that, The method includes: Multi-source heterogeneous data from multiple live streaming platforms are decomposed into a structured dataset, extracted and classified according to a unified element dimension, and multi-type association relationships between data are pre-established. Based on the structured association dataset, a multi-agent heterogeneous graph is constructed, wherein node types and edge types are defined, and the multi-agent heterogeneous graph containing specific nodes and specific edges is instantiated according to the entities, attributes and associations in the structured association dataset. The node types include live streaming room nodes, virtual human intelligent agent nodes, content segment nodes, product nodes, and strategy rule nodes; the edge types include scheduling edges, content supply edges, conversion edges, collaborative sharing edges, and constraint edges. On the multi-agent heterogeneous graph, perform spatiotemporal sequential dynamic analysis, calculate the dynamic activation value of each node, and perform conflict detection and critical state early warning. Based on the results of the dynamic analysis, adaptive shaping and topology reconstruction are implemented, including applying personalized attention masks to the edges with content transitive attributes in the multi-agent heterogeneous graph, and performing pruning, growth and reconnection operations on the multi-agent heterogeneous graph. The system collects feedback data and adjusts the edge weights of the multi-agent heterogeneous graph, the node trust factor of the virtual human agent nodes, and the triggering conditions for topology reconstruction based on the feedback signals, thus forming a closed-loop evolution.

2. The method according to claim 1, characterized in that, The construction of the multi-agent heterogeneous graph includes: The live streaming room node includes platform identifier, audience profile vector, real-time traffic status, conversion status, and compliance status. The virtual human agent node includes style embedding, current load, fatigue level, compliance history, and node trust factor; The content fragment node includes a type, compliance tag, and time-limit decay function; The product node includes commission rate, inventory status, and conversion window attributes; The strategy rule node encapsulates a set of constraint predicates; The scheduling edge represents the scheduling relationship between the live broadcast room node and the virtual human intelligent agent node, and carries a matching weight; The content supply edge represents the supply relationship between the live room node and the content segment node; The conversion edge represents the conversion relationship between the content fragment node and the product node, and carries a historical conversion rate weight; The collaborative shared edge represents the substitutability between the virtual human agent nodes and carries style distance weights; the graph structure consisting of all the virtual human agent nodes and all the collaborative shared edges connecting the virtual human agent nodes is defined as a collaborative shared subgraph, which contains at least one connected component to characterize the topological connectivity and redundancy of the virtual human resource network. The constraint edge represents the constraint relationship between the node and the policy rule node, and is used to make the scheduling path matching the constraint predicate set generate a preset high cost; The initial value of the node trust factor is determined jointly based on the historical compliance records of the virtual human intelligent agent node and the degree of matching of the compliance label, and is dynamically updated according to the differentiated credit attribution results of the feedback signal.

3. The method according to claim 2, characterized in that, The process of performing spatiotemporally sequential dynamic analysis on the multi-agent heterogeneous graph, calculating the dynamic activation value of each node, and performing conflict detection and critical state early warning includes: The activation value of the live streaming node is calculated by weighted summation based on its real-time traffic status, interaction rate, compliance risk index and life cycle factor, wherein the life cycle factor is determined according to a preset mapping function of the live streaming node from cold start, growth, stability to decline. The activation value of the virtual human intelligent agent node is calculated based on a weighted sum of its current load, accumulated fatigue value, and recent violation records, wherein the accumulated fatigue value is determined according to the mapping relationship between the node's continuous working time and a preset fatigue curve. Based on the live streaming room node, the virtual human intelligent agent node, and the scheduling edge connecting the two, a high-dimensional phase space of the current system with the running state as the dimension is constructed; In the high-dimensional phase space of the current system, a joint state vector composed of the activation values ​​of the live broadcast room node and the activation values ​​of the virtual human agent node is calculated, and a phase space reconstruction with a preset time delay is applied to the joint state vector to obtain a multivariate delay embedding matrix. Calculate the covariance matrix of the multivariate delayed embedding matrix, and perform eigenvalue decomposition on the covariance matrix to extract its largest and second largest eigenvalues; Calculate the ratio of the largest eigenvalue to the second largest eigenvalue to obtain the multi-node coupling risk coordination strength; When the multi-node coupling risk collaboration strength shows a monotonically increasing trend within multiple consecutive time windows and exceeds a preset critical collaboration threshold, a critical instability warning signal is generated to characterize the impending nonlinear state change between the live broadcast room and the virtual human agent. The activation value of the content fragment node is calculated jointly based on the current value of the node's time decay function, the weight of the content supply edge, and the frequency of the node's references. The activation value of the product node is calculated jointly based on the remaining time proportion of the conversion window attribute of the node and the historical conversion rate weight of the conversion edge; The activation value of the policy rule node is calculated by weighting and summing the frequency of the conflict detection and the corresponding penalty intensity of the set of constraint predicates encapsulated by the node within a preset time window. The critical state early warning also includes generating a temporary probe node in the graph when a new external platform data source is detected in the multi-agent heterogeneous graph and the data source cannot establish an effective connection with the existing live broadcast room node through the content supply edge or the collaborative sharing edge. The structural fusion barrier index is calculated by the success rate of the connection attempt between the temporary probe node and the existing node. When the structural fusion barrier index exceeds the preset fusion barrier threshold within the continuous observation window, a new platform growth trigger signal is generated to characterize the need for the new platform ecosystem to build an independent network.

4. The method according to claim 3, characterized in that, The critical state early warning also includes: The penalty records of all the constraint edges associated with the virtual human intelligent agent node are used to construct a violation event stream in time sequence; Extract the event interval sequence from the violation event stream, calculate the local variation coefficient of the event interval sequence, and plot the evolution trajectory of the local variation coefficient over time; When the evolutionary trajectory shows a trend of convergence from disordered fluctuations to quasi-periodicity, and the moving average of the local variation coefficients falls below a preset ordering threshold, the violation event flow is determined to have entered a critical pre-diffusion state. Extract the local violation subgraph formed by the virtual human agent node in the critical pre-diffusion state and the live broadcast room node it is connected to, and calculate the clustering coefficient and degree correlation of the local violation subgraph. When the clustering coefficient continues to rise within a continuous observation window and the degree correlation flips from a negative value to a positive value, a violation phase transition early warning signal is generated to characterize the transformation of violation behavior from an isolated event to a collective coordinated diffusion. Construct the weight matrix of the collaborative shared edges, set the diagonal elements to zero, extract the largest connected component with the most nodes in the collaborative shared subgraph, calculate the second smallest eigenvalue of the Laplacian matrix of the largest connected component, and obtain the algebraic connectivity of the collaborative shared subgraph. Traverse each virtual human agent node in the collaborative sharing subgraph, remove individual nodes in turn, and calculate the decrease in algebraic connectivity after removal. Mark nodes whose decrease exceeds a preset vulnerability threshold as hub vulnerable nodes. Construct a two-dimensional state plane consisting of the current load and fatigue of the virtual human intelligent agent nodes, map each node in the collaborative sharing subgraph to the two-dimensional state plane, and calculate the mean Mahalanobis distance of the node distribution; When the algebraic connectivity is lower than the connectivity warning threshold, there are marked vulnerable nodes in the hub, and the mean Mahalanobis distance shows a monotonically increasing trend within a continuous observation window, a resource collapse warning signal is generated to characterize the impending cascading collapse of the virtual human resource network.

5. The method according to claim 4, characterized in that, The adaptive shaping and topology reconstruction based on the results of the dynamic analysis includes: The personalized attention mask is generated jointly by the audience profile vector of the live room node, the content embedding vector of the content segment node, and the time decay function of the content segment node. The mask is applied to the weight of the edge with content delivery attribute in an element-wise multiplication manner. The edge with content delivery attribute includes at least the content supply edge. The pruning operation includes calculating a pruning priority score for each live streaming node in the multi-agent heterogeneous graph. The pruning priority score is determined by the ratio of the node's marginal contribution to global conversion revenue to the node's associated violation risk exposure value. When the activation value of a live streaming node is lower than the dynamic decay threshold for a consecutive preset number of observation periods, and its life cycle factor continues to decline and falls into a preset low-activity range, if the node's pruning priority score is lower than a preset retention threshold, then the node and all its connected first-level scheduling edges are soft-deleted, the re-wakeable residual weights are retained, and a recovery observation period is set. The growth operation includes, upon receiving the new platform growth trigger signal or detecting an abnormal peak in platform traffic, instantiating a new live streaming room node, extracting the current scheduling subgraph from the high-conversion live streaming room node on the same platform as a template subgraph, applying a random mutation mask based on Gaussian noise to each content supply edge in the template subgraph, generating a differentiated scheduling subgraph and attaching it to the new live streaming room node, and simultaneously selecting the virtual human agent node with the closest style embedding vector distance and current load below the load threshold from the collaborative sharing subgraph to establish a scheduling edge; The reconnection operation includes, upon receiving the violation phase transition warning signal, constructing a risk-reward balance plane with style embedding vector distance as the horizontal axis and node trust factor as the vertical axis, mapping the set of replacement virtual human agent nodes onto this plane, and selecting candidate nodes located on the Pareto front to perform scheduling edge switching; upon receiving the resource collapse warning signal, performing locking protection on the scheduling edge marked as a high-value live streaming room node, and prioritizing the allocation of the scheduling request of the high-value live streaming room node to a pre-reserved dedicated virtual human agent resource pool.

6. The method according to claim 5, characterized in that, The collected feedback data is adjusted based on the feedback signal, including: The collected feedback data includes long-term conversion benefits, cost of violations, success rate of virtual human switching, and scheduling response latency; Based on the feedback data, the initial weight of the scheduling edge that contributes high conversion and maintains compliance is increased, and the node trust factor of the corresponding virtual human intelligent agent node is increased in a preset incremental step. For scheduling subgraphs that repeatedly trigger violation risks, the weights of all relevant edges are attenuated, and the node trust factor of the corresponding virtual human intelligent agent node is reduced by a preset attenuation factor. At the same time, a stricter penalty coefficient is added to the corresponding policy rule node. The feedback signal allocation adopts differentiated credit attribution, which traces the net revenue of a single live broadcast cycle back to the associated virtual human agent node and content fragment node along the scheduling edge and the content supply edge, and updates the value estimate of each node using Shapley value or attention weight. The value estimate is used to guide the pruning protection priority and growth initialization strategy.

7. The method according to claim 3, characterized in that, The phase space reconstruction by applying a preset time delay to the joint state vector to obtain a multivariate delay embedding matrix includes: The optimal embedding dimension of the multivariate delayed embedding matrix is ​​determined by using the false nearest neighbor method. Specifically, candidate embedding matrices are constructed by progressively increasing the dimensions of the joint state vector starting from the low embedding dimension. The distance between each state point and its nearest neighbor in the embedding space under each candidate embedding dimension is calculated. When the proportion of state points whose distance decreases abruptly with the increase of dimension is lower than the preset false neighbor proportion threshold, the current embedding dimension is determined as the optimal embedding dimension. The optimal time delay of the multivariate delay embedding matrix is ​​determined using the mutual information method. Specifically, the mutual information value between the joint state vector and its translation vector after being translated according to different candidate time delays is calculated. The decay curve of the mutual information value as the candidate time delay changes is plotted. The candidate time delay when the decay curve first drops to a preset decay ratio of its initial value is selected, or the candidate time delay corresponding to the first local minimum point of the decay curve is selected as the optimal time delay. Based on the optimal embedding dimension and the optimal time delay, phase space reconstruction is performed on the joint state vector to generate the multivariate delay embedding matrix.

8. The method according to claim 4, characterized in that, The step of extracting the event interval sequence from the violation event stream, calculating the local coefficient of variation of the event interval sequence, and plotting the evolution trajectory of the local coefficient of variation over time includes: A sliding window analysis is applied to the event interval sequence. The window length of the sliding window is adaptively adjusted according to the event density of the violation event stream. When the event density increases, the window length is shrunk, and when the event density decreases, the window length is expanded, so that the number of events contained in each sliding window remains constant. Within each sliding window, the ratio of the mean to the standard deviation of the event interval sequence is calculated to obtain the local coefficient of variation corresponding to that window; The local variation coefficients of each sliding window are connected in chronological order to generate the evolution trajectory of the local variation coefficients over time. The evolution trajectory is smoothed using an exponentially weighted temporal smoothing filter, and a continuous symbol sequence of the first derivative of the smoothed trajectory is calculated. When the continuous symbol sequence alternates between positive and negative symbols within a preset convergence determination period and the range of positive and negative symbols is lower than a preset symbol fluctuation threshold, it is determined that the evolution trajectory shows a trend of convergence from disordered fluctuation to quasi-periodicity.

9. The method according to claim 6, characterized in that, The collection of feedback data, and the adjustment of the triggering conditions for topology reconstruction based on the feedback signals, include: The critical collaboration threshold, the ordering threshold, the connectivity early warning threshold, the dynamic decay threshold, and the retention threshold are uniformly constructed into an adjustable threshold vector; At the end of each preset tuning cycle, the execution records of the pruning operation, the growth operation and the reconnection operation triggered by each critical state warning signal during the cycle are collected, as well as the changes in the global transformation revenue and the changes in the violation cost of the multi-agent heterogeneous graph after execution, and a reconstruction efficiency evaluation matrix is ​​constructed. Based on the aforementioned reconstruction performance evaluation matrix, the performance score for each topology reconstruction operation is calculated. The performance score is the global conversion revenue increment after the operation minus the violation cost increment, and the difference is divided by the number of scheduling interruptions caused by the operation. Each topology reconstruction operation is divided into high-efficiency reconstruction operations and low-efficiency reconstruction operations according to its performance score. The actual state index values ​​corresponding to each critical state warning signal before the execution of the high-efficiency reconstruction operation are extracted and compared with the corresponding thresholds in the adjustable threshold vector to calculate the deviation direction and deviation magnitude. If the actual state index value of the efficient reconstruction operation is located on the lenient side of the corresponding threshold and the deviation exceeds the preset hysteresis tolerance range, then the corresponding threshold is adjusted by one step towards the strict direction. If the actual state index value of the inefficient reconstruction operation is on the strict side of the corresponding threshold and the deviation exceeds the preset false alarm tolerance range, then the corresponding threshold is adjusted by one step towards the lenient side.

10. A cross-platform unmanned live streaming resource scheduling system based on intelligent agent collaboration, the system being used to store and execute the method described in any one of claims 1-9, characterized in that, include: The data preprocessing module is used to structurally decompose multi-source heterogeneous data from multiple live streaming platforms, extract and classify the data according to a unified element dimension, and pre-build multiple types of relationships between the data to form a structured relational dataset. The heterogeneous graph construction module is used to construct a multi-agent heterogeneous graph based on the structured association dataset. It defines node types and edge types, and instantiates the multi-agent heterogeneous graph containing specific nodes and edges based on the entities, attributes, and relationships in the structured association dataset. The node types include live streaming room nodes, virtual human agent nodes, content fragment nodes, product nodes, and strategy rule nodes. The edge types include scheduling edges, content supply edges, conversion edges, collaborative sharing edges, and constraint edges. The dynamic analysis module is used to perform spatiotemporal sequential dynamic analysis on the multi-agent heterogeneous graph, calculate the dynamic activation value of each node, and perform conflict detection and critical state early warning. The adaptive shaping and topology reconstruction module is used to implement adaptive shaping and topology reconstruction based on the results of the dynamic analysis, including applying personalized attention masks to the edges with content transitive attributes in the multi-agent heterogeneous graph, and performing pruning, growth and reconnection operations on the multi-agent heterogeneous graph. The feedback adjustment module is used to collect feedback data and adjust the edge weights of the multi-agent heterogeneous graph, the node trust factor of the virtual human agent nodes, and the triggering conditions of the topology reconstruction based on the feedback signals, thereby forming a closed-loop evolution.

Images