A method for predicting user influence for rapid evaluation of social network information propagation effect

Abstract

The application discloses a kind of social network information propagation effect fast evaluation's prediction user influence method, method includes the following steps: step A, training data and constructing with offline annotation;Step B, seed set-graph joint coding;Step C, comparative learning;Step D, influence value regression and sorting consistency joint training;Step F, agent model auxiliary evolution's intelligent optimization solution.The application significantly reduces the time overhead and the power consumption of single influence evaluation, so that the influence maximization method can be stably landed in large-scale network, real-time or quasi-real-time business system;The application constructs influence similar sample, significantly different sample and marginal gain sorting sample, and applies constraint in representation space and output space simultaneously, so that the agent model can still maintain stable and reliable relative ordering relationship between candidate seed set in the presence of certain numerical estimation error, significantly improve the overall consistency.

Description

Technical Field

[0001] This invention specifically relates to a method for predicting user influence to rapidly assess the effectiveness of information dissemination on social networks. Background Technology

[0002] Influence Maximization (IM) is one of the core problems in social network analysis and combinatorial optimization. Its goal is, given a social network graph and a diffusion model (e.g., Independent Cascade Model (IC) or Linear Threshold Model (LT), to select a set S of seed nodes with a budget of k from the network such that the expected number of activated nodes (influence sphere) σ(S) after diffusion is maximized. IM is widely used in scenarios such as viral marketing, public opinion dissemination, cold start of recommendation systems, and public health intervention resource allocation.

[0003] Influence estimation (IM) has been proven to be an NP-hard problem. Classical solutions typically employ a "greedy algorithm + influence estimation" framework: at each step, an attempt is made to add a candidate node to the current seed set, and the marginal gain is obtained through influence estimation, then the node with the largest gain is selected. The reason why this type of method is difficult to scale is that the key bottleneck lies in influence estimation (IE): under stochastic diffusion models such as IC / LT, accurately calculating σ(S) is #P-hard, and in engineering, it usually relies on a large number of Monte Carlo (MC) simulations to approximate the expectation, resulting in huge computational costs and making it difficult to meet the timeliness requirements of large-scale networks or online decision-making. The subsequent emergence of RIS (Reverse Influence Sampling) / RR set techniques has reduced the complexity to some extent, but in ultra-large-scale graphs, dynamic graphs, or intelligent optimization scenarios that require frequent repeated evaluation of a large number of candidate seed sets, problems such as memory consumption, engineering complexity, and response latency may still be encountered.

[0004] In the past three years, research on deep learning to solve IM has shown the following mainstream directions: (1) End-to-end deep graph representation learning and combinatorial optimization coupling: By learning graph representations and combining the seed selection process with optimization strategies, the dependence on traditional IE is reduced, and the inference speed and generalization ability are improved. Typical works propose deep graph representation learning and matching optimization mechanisms to solve IM.

[0005] (2) Learning-based influence estimator (IE) with GNN as its core: a model is trained to directly output influence estimates from input (graph, seed set), replacing MC as the IE subroutine, and further integrated into greedy / ranking / reinforcement learning strategies. Some works propose GNN influence estimators and combine them with CELF, learnable ranking or Q-learning to form scalable IM schemes, emphasizing inductive generalization to large graphs after training on small graphs.

[0006] (3) Online / Interactive IM: When the diffusion network topology is unknown or incomplete, IM is placed in a multi-round delivery scenario, and the reward is estimated and the exploration-exploitation is balanced by bandit and neural network, and seeds are selected online.

[0007] (4) Enhanced diffusion dynamics modeling: Improve the separability and expressive power of complex diffusion patterns by using novel graph neural networks (such as sheaf neuraldiffusion), and design optimization strategies that take into account the effects of overlap to reduce the combinatorial search space.

[0008] (5) Variant objectives for real social scenarios: such as maximizing the influence of multiple entities / balanced systems, using deep reinforcement learning frameworks and introducing auxiliary signals such as behavioral sequences to reduce sampling estimation overhead.

[0009] (6) Dynamic network / real-time candidate prediction: Characterize the changes in dynamic graph structure and predict potential high-impact nodes to improve real-time performance and applicability.

[0010] While the aforementioned research has promoted the development of learning-based IM, there are still key technical challenges that urgently need to be addressed from the perspective of "patent implementation and engineering reuse":

[0011] A. Insufficient "generalization and transferability" of surrogate training: Many learning methods depend on the size of the training graph, graph distribution, diffusion probability distribution, or specific dataset. Once the target network deviates from the training distribution, the influence estimation error will be significantly amplified, leading to a decrease in the quality of seed selection.

[0012] B. Insufficient constraints on "ranking consistency / marginal gain consistency": IM solutions focus more on the relative merits (ranking) and marginal gains between different candidate seed sets, rather than just absolute regression error. Regression models trained solely on MSE / MAE may be numerically close but incorrectly ranked, leading to greedy or evolutionary searches entering local optima.

[0013] C. Insufficient support for the "high-frequency evaluation-intelligent optimization" closed loop: In intelligent optimization such as genetics, particle swarm optimization, differential evolution, and multi-objective evolution, each generation generates a large number of candidate seed sets, requiring massive IE calls. Existing work often focuses on the "seed selection strategy itself," lacking a general influence value proxy model design for "high-throughput candidate evaluation," so that it can serve as a pluggable component for various optimizations.

[0014] D. The contradiction between data annotation cost and training stability: High-quality labels still need to be generated offline by MC / RIS. If there is a lack of effective self-supervised / contrastive learning mechanisms, the model needs more labeled samples to learn stable representations, resulting in high overall training costs.

[0015] Therefore, there is an urgent need for a technical solution centered on an "influence value assessment proxy model": learning a representation that is sensitive to diffusion dynamics, robust to distribution shifts, and friendly to ranking consistency with limited annotation costs in the offline stage; and completing the rapid estimation of the influence value of the candidate seed set with one or a small amount of forward inference in the online stage, which can be embedded into existing intelligent optimization frameworks to significantly reduce the dependence on MC and improve the solution efficiency and effect. Summary of the Invention

[0016] The purpose of this invention is to provide a method for predicting user influence to quickly assess the effectiveness of information dissemination on social networks.

[0017] The technical solution to achieve the purpose of this invention is a method for predicting user influence by rapidly evaluating the effect of information dissemination on social networks. The method includes the following steps:

[0018] Step A: Training data and construction with offline annotation. Several training sub-images are obtained by sampling from the target large image in a corresponding manner, and the supervised labels are estimated for each sample pair.

[0019] Step B: Seed set-graph joint encoding, which simultaneously encodes interaction relationships in the representation and aggregates the messages generated by all neighbors;

[0020] Step C: Contrastive learning, which brings samples with similar influence or consistent ranking closer together in the representation space, while separating samples with significant differences in influence or opposite ranking.

[0021] Step D: Joint training of influence value regression and ranking consistency;

[0022] Step F, the intelligent optimization solution assisted by the surrogate model, determines the overall solution efficiency and the reliability of the search method.

[0023] A further preferred embodiment is that: a step E is provided between step D and step E, which involves uncertainty estimation and active correction, and active correction is performed on a small number of real assessments driven by uncertainty.

[0024] A further preferred embodiment is that step A also includes the following steps:

[0025] Step A1: Sample several training subgraphs Gi from the target large graph G according to time window, community division, random walk or k-hop neighborhood expansion method;

[0026] Step A2: Construct multiple seed sets Sij of different sizes and structures for each Gi;

[0027] Step A3: For each sample pair (G) i ,S ij ), and estimate σ(G) using an offline, affordable method. i ,S ij ), obtain the supervised label y ij ≈σ(G i ,S ij );

[0028] Step A4: In addition to the regression label yij, select several pairs (S) within the same Gi. ia ,S ib If y ia > y ib + ε, then it is denoted as ordered pair a ≻ b; ε is the tolerance threshold, used to resist MC noise. With a fixed basis set S, the marginal gain labels of different candidate nodes v are compared to form training pairs (v1 ≻ v2 | S).

[0029] A further preferred embodiment is that step B includes the following steps:

[0030] Step B1, Adjacency relation E and edge probability matrix / edge feature p uv；

[0031] Step B2: Set initial node features Xv that include in-degree / out-degree, PageRank, k-core, community tags, user attributes, activity level, or content preferences;

[0032] Step B3: Determine the seed mask m v If v∈S, then m v =1, otherwise 0, and used as a one-dimensional or multi-dimensional embedding of node features;

[0033] Step B4: The encoder uses L-layer message passing, and the edge probabilities are modulated and aggregated at each layer.

[0034]

[0035] in gl ϕ is the probability modulation function; l For non-linear updates; function F represents "aggregation operation on neighbor messages, that is, performing a set-level aggregation on the messages generated by all neighbors u";

[0036] And it adopts layered readout:

[0037] ;

[0038] ;

[0039] , |S|, Statistical characteristics;

[0040] σ(S) = MLP(z);

[0041] Different S values ​​on the same graph will change m. v This changes the "activation source" in the message passing path, enabling the model to learn the cascading and overlapping effects of diffusion.

[0042] A further preferred embodiment is that step C includes the following steps:

[0043] Step C1: The object is defined as the representation vector z = Enc(G, S) of the sample (G, S), where Enc is the intermediate representation between the encoder and the readout, i.e. z before the MLP.

[0044] Step C2: Construct positive and negative samples;

[0045] Influence nearest neighbor positive samples: in the same graph G i Above, if |y ia -y ib | ≤ δ, then (G i , S ia ) and (G i ,S ib ) Forming a direct pair;

[0046] Augmented consistent positive samples: for the same (G) i , S) Perform graph augmentation and seed augmentation to obtain (G i If the label changes are controlled, it is considered a positive match;

[0047] Graph augmentation includes: edge dropping / perturbation based on probability thresholds, subgraph pruning, edge probability noise injection, and node attribute occlusion;

[0048] Difficult negative samples: Select samples with y differences greater than Δ as negative pairs, or select samples that "represent similarity but have large label differences" as difficult negatives to enhance the discrimination ability;

[0049] Marginal gain comparison: Under a fixed basis set S, construct samples (G,S∪{v}) for candidate nodes v, and form a positive / negative relationship based on the true / approximate marginal gain, so that the model learns a more stable Δ(v | S) ranking.

[0050] Step C3: Compare the losses using the InfoNCE format:

[0051]

[0052] Where sim can be the cosine similarity and τ is the temperature coefficient.

[0053] A further preferred embodiment is that step D includes the following steps:

[0054] Step D1, Regression Loss, Or Huber loss, used to fit the expected value of influence;

[0055] Step D2: Ranking Consistency Loss (Pairwise Ranking)

[0056] In the same graph G i Within, apply the following to all ordered pairs (Sa ≻ Sb):

[0057]

[0058] Where m is the sorting interval;

[0059] Step D3, Joint Objective

[0060] Where λ1, λ2, and λ3 are weighting coefficients.

[0061] A further preferred embodiment is that step F includes the following steps:

[0062] Step F1: Use a proxy-assisted greedy algorithm to initialize S = ∅; repeat k times: calculate S for all v ∈ V\S

[0063] Δ(v|S), select the largest value and add it to S; output S;

[0064] Step F2: Agent-assisted evolution, individual encoding: a node sequence of length k, or an n-dimensional 0 / 1 vector with constraint |S|=k, fitness function: Fit(S)=σ(S), or multi-objective Fit(S)=(σ(S), cost(S), risk(S)), genetic operations: selection, crossover, mutation, repair, and ensuring |S|=k and deduplication, elite retention: retain the top r high-fitness solutions, termination condition: upper limit of generations or convergence threshold, output optimal / Pareto front solution;

[0065] Step F3: Hybrid Realistic Evaluation and Calibration. Every T generations, q candidate solutions are drawn from the current elite set for realistic MC evaluation to obtain y. true ; with (S, y true Update / calibrate the agent model or perform ranking re-ranking; continue to use the calibrated agent in subsequent iterations.

[0066] Compared with the prior art, the present invention has the following positive effects: the present invention significantly reduces the time overhead and computing power consumption of a single influence assessment, enabling the influence maximization method to be stably implemented in large-scale networks and real-time or near-real-time business systems;

[0067] This invention constructs samples with similar influence, samples with significant differences, and samples ranked by marginal gain, and applies constraints to both the representation space and the output space. This enables the surrogate model to maintain a stable and reliable relative ranking relationship between candidate seed sets even with a certain numerical estimation error, significantly improving the overall convergence quality, solution stability, and consistency of repeated runs.

[0068] The encoding method of this invention can characterize complex nonlinear effects such as multi-hop propagation, influence overlap and cascading decay, thereby more accurately approximating the true combination structure of the seed set influence function σ(S);

[0069] This invention utilizes mechanisms such as subgraph sampling, joint augmentation of graph structure and seed set, and contrastive learning to focus on learning local propagation rules and structural invariance features during the training phase, thereby significantly improving the model's cross-scale inductive ability and engineering reuse value.

[0070] This invention not only covers a variety of implementation forms and application scenarios, but also facilitates the rapid adjustment of the solution strategy according to different business needs during the productization process, without the need to reconstruct the core influence assessment module, thereby significantly improving the overall system's engineering flexibility and patent protection scope. Attached Figure Description

[0071] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0072] Figure 1 This is a block diagram of the training structure of the present invention;

[0073] Figure 2 This is a block diagram of the solution structure of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Example

[0076] See Figure 1 and Figure 2 As shown, a method for predicting user influence for rapid evaluation of the dissemination effect of information on social networks includes the following steps:

[0077] Step A involves training data and constructing offline annotations. Several training sub-images are sampled from the target large image in a corresponding manner, and the supervised labels are estimated for each sample pair. Step A also includes the following steps:

[0078] Step A1: Sample several training subgraphs Gi from the target large graph G according to time window, community division, random walk or k-hop neighborhood expansion method;

[0079] Step A2: Construct multiple seed sets Sij of different sizes and structures for each Gi (|Sij| can be 1..k or perturbed around k); the seed sets can be generated by a mixture of random, centrality heuristic, historical delivery logs, or existing output candidate sets to increase distribution coverage.

[0080] Step A3: For each sample pair (G) i ,S ij ), and estimate σ(G) using an offline, affordable method. i ,S ij ), obtain the supervised label y ij ≈σ(G i ,S ij Offline annotation can be performed on small-scale subgraphs or scaled graphs, and after training, it can be generalized to larger-scale graphs through inductive reasoning.

[0081] Step A4: In addition to the regression label yij, select several pairs (S) within the same Gi. ia ,S ib If y ia > yib + ε, then it is denoted as ordered pair a ≻ b; ε is the tolerance threshold, used to resist MC noise. With a fixed basis set S, the marginal gain labels of different candidate nodes v are compared to form training pairs (v1 ≻ v2 | S).

[0082] Step B, seed set-graph joint encoding, encodes interaction relationships simultaneously in the representation and aggregates messages generated by all neighbors; Step B includes the following steps:

[0083] Step B1, Graph Structure and Edge Probability: Adjacency Relationship E and Edge Probability Matrix / Edge Feature p uv；

[0084] Step B2, Initial Node Features Xv: Set initial node features Xv that include in-degree / out-degree, PageRank, k-core, community tags, user attributes, activity level, or content preferences;

[0085] Step B3: Determine the seed mask m v If v∈S, then m v =1, otherwise 0, and used as a one-dimensional or multi-dimensional embedding of node features;

[0086] Step B4, Optional "diffusion prior features": such as prior propagation potential obtained based on DMP upper bound, RR coverage count, or historical diffusion statistics.

[0087] The encoder uses L-layer message passing, and each layer modulates and aggregates edge probabilities.

[0088]

[0089] in g l ϕ is the probability modulation function; l For non-linear updates; function F represents "aggregation operation on neighbor messages, that is, performing a set-level aggregation on the messages generated by all neighbors u";

[0090] And it adopts layered readout:

[0091] ;

[0092] ;

[0093] , |S|, Statistical characteristics;

[0094] σ(S) = MLP(z);

[0095] Different S values ​​on the same graph will change m. v This changes the "activation source" in the message passing path, enabling the model to learn the cascading and overlapping effects of diffusion.

[0096] Step C: Contrastive learning, which brings samples with similar influence or consistent ranking closer together in the representation space, while separating samples with significant differences in influence or opposite ranking.

[0097] To improve the robustness of the surrogate model in scenarios involving "cross-graph scale, cross-distribution, and cross-noise labels", a contrastive learning mechanism oriented towards IM is introduced. This mechanism makes samples with "similar influence / consistent ranking" closer in the representation space, while samples with "significantly different influence / opposite ranking" are pulled apart, thus directly serving the ranking stability of subsequent optimization search.

[0098] Step C includes the following steps:

[0099] Step C1: The object is defined as the representation vector z = Enc(G, S) of the sample (G, S), where Enc is the intermediate representation between the encoder and the readout, i.e. z before the MLP.

[0100] Step C2: Construct positive and negative samples (IM specialization);

[0101] Influence nearest neighbor positive samples: in the same graph G i Above, if |y ia -y ib | ≤ δ, then (G i , S ia ) and (G i ,S ib ) Forming a direct pair;

[0102] Augmented consistent positive samples: for the same (G) i , S) Perform graph augmentation and seed augmentation to obtain (G i If the label changes are controlled, it is considered a positive match;

[0103] Graph augmentation includes: edge dropping / perturbation based on probability thresholds, subgraph pruning, edge probability noise injection, and node attribute occlusion; seed augmentation includes: adding / replacing "low marginal gain nodes" or deleting "redundant nodes" in S while keeping the influence approximately unchanged;

[0104] Difficult negative samples: Select samples with y differences greater than Δ as negative pairs, or select samples that "represent similarity but have large label differences" as difficult negatives to enhance the discrimination ability;

[0105] Marginal gain comparison: Under a fixed basis set S, construct samples (G,S∪{v}) for candidate nodes v, and form a positive / negative relationship based on the true / approximate marginal gain, so that the model learns a more stable Δ(v | S) ranking.

[0106] Step C3: Compare the losses using the InfoNCE format:

[0107]

[0108] Where sim can be the cosine similarity and τ is the temperature coefficient.

[0109] Step D: Joint training of influence value regression and ranking consistency; the training objective of the surrogate model is designed as a combination of "numerical fitting + ranking consistency + structure regularization" to match the real needs of IM.

[0110] Step D includes the following steps:

[0111] Step D1: Regression Loss (Numerical Fitting) Or Huber loss, used to fit the expected value of influence;

[0112] Step D2: Ranking Consistency Loss (Pairwise Ranking)

[0113] In the same graph G i Within, apply the following to all ordered pairs (Sa ≻ Sb):

[0114]

[0115] Where m is the sorting interval;

[0116] Step D3, Joint Objective

[0117] Where λ1, λ2, and λ3 are weighting coefficients.

[0118] Step E is set between step D and step E, which involves uncertainty estimation and active correction, and active correction of a small number of real assessments driven by uncertainty.

[0119] Step F, the intelligent optimization solution assisted by the surrogate model, determines the overall solution efficiency and the reliability of the search method. The trained fθ is used as a pluggable IE engine, supporting multiple solution strategies:

[0120] Step F includes the following steps:

[0121] Step F1: Surrogate Greedy, initialize S=∅; repeat k times: calculate Δ(v|S) for all v∈V\S, select the largest and add it to S; output S; lazy updates such as CELF can be used to speed up the process, but the true σ is replaced with σ̂, thus turning each evaluation from MC into a forward inference.

[0122] Step F2, Surrogate-assisted Evolutionary Search: Individual encoding: a node sequence of length k, or an n-dimensional 0 / 1 vector constrained to |S|=k; Fitness function: Fit(S)=σ(S), or multi-objective Fit(S)=(σ(S), cost(S), risk(S)); Genetic operations: selection, crossover, mutation, repair, ensuring |S|=k and deduplication; Elite retention: retaining the top r high-fitness solutions; Termination condition: upper bound on generation or convergence threshold; Output optimal / Pareto front solution. In this framework, f... θ The high throughput characteristic directly determines the overall solution efficiency; contrastive learning + consistent sorting training directly determines whether the search direction is reliable.

[0123] Step F3: Hybrid Realistic Evaluation and Calibration. Every T generations, q candidate solutions are drawn from the current elite set for realistic MC evaluation to obtain y. true ; with (S, y true Update / calibrate the surrogate model or perform ranking reordering; continue using the calibrated surrogate in subsequent iterations. This strategy significantly reduces the risk of "surrogate misdirection" when computational budget is manageable.

[0124] The algorithm can be implemented in the following steps: Step 1: Data Acquisition and Graph Construction Step 1.1 Acquire social network data, construct a directed graph G=(V,E), and assign a diffusion probability p_uv to each edge (u,v); Step 1.2 Construct initial features Xv for nodes and determine the budget k. Step 2: Offline Training Set Generation and Labeling Step 2.1 Sample several training subgraphs G. i Step 2.2 In each G i Generate multiple seed sets S ij Step 2.3 For each (G) i ,S ij Perform offline dissemination of MC / RIS to obtain the influence tag y. ijStep 2.4 Construct ranking pairs and marginal gain ranking pairs based on yij. Step 3: Contrastive learning pre-training and supervised joint training Step 3.1 Initialize seed mask conditional encoder Enc and prediction head fθ; Step 3.2 Construct positive and negative pairs through graph augmentation / seed augmentation and influence nearest neighbor rules, and minimize L con Step 3.3 Jointly minimize L reg L rank The final proxy model f is obtained. θ Step 4: Online Influence Estimation and Marginal Gain Calculation Step 4.1 Input the target graph G and the candidate seed set S, and generate the mask m v Step 4.2 Run Enc(G,S) to obtain the representation z, and then use f θ Output σ(S); Step 4.3 Calculate Δ(v|S) for candidate node v (can be batched and parallelized as needed). Step 5: Solving the agent-assisted influence maximization problem Step 5.1 Select the solver: agent-assisted greedy algorithm / agent-assisted evolutionary optimization;

[0125] Step 5.2 In the solver iteration, σ(S) is used as the fitness / evaluation function for high-frequency evaluation; Step 5.3 A small number of real evaluations are triggered based on uncertainty, and the calibration is re-boosted; Step 5.4 After reaching the termination condition, the optimal seed set S* is output. Step 6: Result Output and Deployment Step 6.1 Output S* and its predicted influence σ(S*); Step 6.2 Output node-level explanation information: such as the predicted marginal contribution of each seed node, the high-confidence sub-communities covered by the influence, etc., for business-side auditing and secondary constraints.

[0126] Compared with the prior art, the present invention has the following positive effects: the present invention significantly reduces the time overhead and computing power consumption of a single influence assessment, enabling the influence maximization method to be stably implemented in large-scale networks and real-time or near-real-time business systems;

[0127] This invention constructs samples with similar influence, samples with significant differences, and samples ranked by marginal gain, and applies constraints to both the representation space and the output space. This enables the surrogate model to maintain a stable and reliable relative ranking relationship between candidate seed sets even with a certain numerical estimation error, significantly improving the overall convergence quality, solution stability, and consistency of repeated runs.

[0128] The encoding method of this invention can characterize complex nonlinear effects such as multi-hop propagation, influence overlap and cascading decay, thereby more accurately approximating the true combination structure of the seed set influence function σ(S);

[0129] This invention utilizes mechanisms such as subgraph sampling, joint augmentation of graph structure and seed set, and contrastive learning to focus on learning local propagation rules and structural invariance features during the training phase, thereby significantly improving the model's cross-scale inductive ability and engineering reuse value.

[0130] This invention not only covers a variety of implementation forms and application scenarios, but also facilitates the rapid adjustment of the solution strategy according to different business needs during the productization process, without the need to reconstruct the core influence assessment module, thereby significantly improving the overall system's engineering flexibility and patent protection scope.

[0131] The standard parts used in this embodiment can be purchased directly from the market, and the non-standard structural parts described in the instruction manual can also be processed without any doubt based on existing technical common sense. At the same time, the connection methods of each component adopt mature conventional methods in the existing technology, and the machinery, parts and equipment all adopt conventional models in the existing technology, so they will not be described in detail here.

[0132] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, these obvious variations or modifications derived from the essential spirit of the present invention still fall within the scope of protection of the present invention.

Claims

1. A method for rapidly evaluating the effect of information dissemination on social networks and predicting user influence, characterized in that: The method includes the following steps: Step A: Training data and construction with offline annotation. Several training sub-images are obtained by sampling from the target large image in a corresponding manner, and the supervised labels are estimated for each sample pair. Step B: Seed set-graph joint encoding, which simultaneously encodes interaction relationships in the representation and aggregates the messages generated by all neighbors; Step C: Contrastive learning, which brings samples with similar influence or consistent ranking closer together in the representation space, while separating samples with significant differences in influence or opposite ranking. Step D: Joint training of influence value regression and ranking consistency; Step F, the intelligent optimization solution assisted by the surrogate model, determines the overall solution efficiency and the reliability of the search method.

2. The method for predicting user influence for rapid evaluation of social network information dissemination effects according to claim 1, characterized in that: Step E is set between step D and step E, which involves uncertainty estimation and active correction, and active correction of a small number of real assessments driven by uncertainty.

3. The method for predicting user influence for rapid evaluation of social network information dissemination effects according to claim 2, characterized in that: Step A also includes the following steps: Step A1: Sample several training subgraphs Gi from the target large graph G according to time window, community division, random walk or k-hop neighborhood expansion method; Step A2: Construct multiple seed sets Sij of different sizes and structures for each Gi; Step A3: For each sample pair (G) i ,S ij ), and estimate σ(G) using an offline, affordable method. i ,S ij ), obtain the supervised label y ij ≈σ(G i ,S ij ); Step A4: In addition to the regression label yij, select several pairs (S) within the same Gi. ia ,S ib If y ia > y ib + ε, then it is denoted as ordered pair a ≻ b; ε is the tolerance threshold, used to resist MC noise. With a fixed basis set S, the marginal gain labels of different candidate nodes v are compared to form training pairs (v1 ≻ v2 | S).

4. The method for predicting user influence for rapid evaluation of social network information dissemination effects according to claim 3, characterized in that: Step B includes the following steps: Step B1: Adjacency relation E and edge probability matrix / edge feature p uv； Step B2: Set initial node features Xv that include in-degree / out-degree, PageRank, k-core, community tags, user attributes, activity level, or content preferences; Step B3: Determine the seed mask m v If v∈S, then m v =1, otherwise 0, and is used as a one-dimensional or multi-dimensional embedding of node features; Step B4: The encoder uses L-layer message passing, and the edge probabilities are modulated and aggregated at each layer. ,in g l ϕ is the probability modulation function; l For non-linear updates; function F represents "aggregation operation on neighbor messages, that is, performing a set-level aggregation on the messages generated by all neighbors u"; And it adopts layered readout: ； ； , |S|, Statistical characteristics; σ(S) = MLP(z); Different S values ​​on the same graph will change m. v This changes the "activation source" in the message passing path, enabling the model to learn the cascading and overlapping effects of diffusion.

5. The method for predicting user influence for rapid evaluation of social network information dissemination effects according to claim 4, characterized in that: Step C includes the following steps: Step C1: The object is defined as the representation vector z = Enc(G, S) of the sample (G, S), where Enc is the intermediate representation between the encoder and the readout, i.e. z before the MLP. Step C2: Construct positive and negative samples; Influence nearest neighbor positive samples: in the same graph G i Above, if |y ia -y ib | ≤ δ, then (G i , S ia ) and (G i , S ib ) form a positive pair; Augmented consistent positive samples: for the same (G) i , S) Perform graph augmentation and seed augmentation to obtain (G i If the label changes are controlled, it is considered a positive match; Graph augmentation includes: edge dropping / perturbation based on probability thresholds, subgraph pruning, edge probability noise injection, and node attribute occlusion; Difficult negative samples: Select samples with y differences greater than Δ as negative pairs, or select samples that "represent similarity but have large label differences" as difficult negatives to enhance the discriminative ability; Marginal gain comparison: Under a fixed basis set S, construct samples (G,S∪{v}) for candidate nodes v, and form a positive / negative relationship based on the true / approximate marginal gain, so that the model learns a more stable Δ(v | S) ranking; Step C3: Compare the losses using the InfoNCE format: ， Where sim can be the cosine similarity and τ is the temperature coefficient.

6. The method for predicting user influence for rapid evaluation of social network information dissemination effects according to claim 5, characterized in that: Step D includes the following steps: Step D1, Regression Loss, Or Huber loss, used to fit the expected value of influence; Step D2: Ranking Consistency Loss (Pairwise Ranking) In the same graph G i Within, apply the following to all ordered pairs (Sa ≻ Sb): , where m is the sorting interval; Step D3, Joint Objective , Where λ1, λ2, and λ3 are weighting coefficients.

7. The method for predicting user influence for rapid evaluation of the dissemination effect of social network information according to claim 6, characterized in that: Step F includes the following steps: Step F1: Use a proxy-assisted greedy algorithm to initialize S = ∅; repeat k times: calculate S for all v ∈ V\S Δ(v|S), select the largest value and add it to S; output S; Step F2: Agent-assisted evolution, individual encoding: a node sequence of length k, or an n-dimensional 0 / 1 vector with constraint |S|=k, fitness function: Fit(S)=σ(S), or multi-objective Fit(S)=(σ(S), cost(S), risk(S)), genetic operations: selection, crossover, mutation, repair, and ensuring |S|=k and deduplication, elite retention: retain the top r high-fitness solutions, termination condition: upper limit of generations or convergence threshold, output optimal / Pareto front solution; Step F3: Hybrid Realistic Evaluation and Calibration. Every T generations, q candidate solutions are drawn from the current elite set for realistic MC evaluation to obtain y. true ; with (S, y true Update / calibrate the agent model or perform ranking re-ranking; continue to use the calibrated agent in subsequent iterations.

Images