Aigc content generation method and system adaptive to user feedback

Abstract

The application provides an AIGC content self-adaptive method and system based on user feedback, comprising: acquiring and preprocessing a multi-modal user feedback dataset, event modeling and causal discovery, acquiring and performing feedback event quality evaluation according to a first feedback event dataset and a causal influence dataset, generating a second feedback event dataset, synchronously acquiring event exploration intensity, performing counterfactual reasoning based on a preset neural jump differential equation and the causal influence dataset, acquiring a corrected event exploration intensity and a third feedback event dataset, performing AIGC self-adaptive generation based on a preset multi-scale stochastic differential equation, synchronously combining the second and third feedback event datasets and the corrected event exploration intensity, acquiring an original AIGC content dataset, performing quality evaluation on the original AIGC content dataset, and combining a Bayesian optimization algorithm to perform optimization fine-tuning to generate an optimized AIGC content dataset, thereby improving the accuracy of the generated content.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an adaptive method and system for AIGC-generated content based on user feedback. Background Technology

[0002] In recent years, generative artificial intelligence has made significant progress in the creation of content such as images, text, and audio / video. However, existing systems still face significant challenges in continuously and adaptively optimizing based on real-time, multimodal user feedback.

[0003] On the one hand, traditional methods mostly treat feedback from different modalities as independent, discrete signals, often lacking a unified dynamic modeling of user interaction event flows on a continuous time scale. This results in insufficient exploration and utilization of the potential, time-evolving causal dependencies between feedback events, leading to a lack of understanding of the user's complex intentions.

[0004] On the other hand, traditional methods mostly use fixed, pre-trained generative model architectures. In complex, multi-round iterative creation tasks, static models often lack sufficient adaptability in the face of complex and ever-changing user feedback, resulting in insufficient accuracy of the generated content. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an adaptive method for AIGC-generated content based on user feedback, the method comprising:

[0006] The multimodal user feedback dataset is acquired and preprocessed, and event modeling is performed using a neural Hawkes process to obtain a first feedback event dataset. Causal discovery is then performed on the multimodal user feedback dataset to generate a causal impact dataset.

[0007] Based on deep kernel learning, and based on the first feedback event dataset and the causal impact dataset, the quality of feedback events is evaluated, and the results of the feedback event quality evaluation are combined to generate a second feedback event dataset, while simultaneously obtaining the event exploration intensity;

[0008] Counterfactual reasoning is performed based on a preset neural jump differential equation and a causal influence dataset. The intensity of event exploration is then corrected based on the counterfactual reasoning results, and the corrected event exploration intensity and a third feedback event dataset are obtained.

[0009] Based on the preset multi-scale stochastic differential equation, AIGC adaptive generation is performed simultaneously by combining the second and third feedback event datasets and the exploration intensity of the correction event to obtain the original AIGC content dataset;

[0010] The original AIGC content dataset is evaluated based on a preset content quality analysis mechanism, and then optimized and fine-tuned using a Bayesian optimization algorithm to generate an optimized AIGC content dataset.

[0011] Furthermore, embodiments of the present invention also provide an AIGC-generated content adaptive system based on user feedback, including:

[0012] The data processing module is used to acquire and process the multimodal user feedback dataset to acquire the first feedback event dataset and the causal influence dataset.

[0013] The feedback evaluation module is used to evaluate the quality of user feedback events and generate a second feedback event dataset and event exploration intensity based on the results of the quality evaluation.

[0014] The data optimization module is used to perform counterfactual reasoning based on a preset neural jump differential equation and a causal influence dataset to obtain the intensity of the modified event exploration and the third feedback event dataset.

[0015] The content generation module is used to obtain the original AIGC content dataset based on the preset multi-scale stochastic differential equation, the second and third feedback event datasets, and the intensity of the modified event exploration.

[0016] The content optimization module is used to perform quality assessment on the original AIGC content dataset and perform fine-tuning by combining Bayesian optimization algorithm to generate an optimized AIGC content dataset.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] By introducing neural Hawkes processes, discrete multimodal interaction event flows are uniformly modeled as continuous-time jump processes, and differentiable causal adjacency matrices between event types are learned simultaneously. This enables explicit characterization and comprehensive utilization of the temporal dynamics and potential causal dependencies of user interactions, enhancing the deep understanding of users' complex intentions.

[0019] Meanwhile, based on the real-time generation status and feedback context, parameters adapted to the current task requirements are dynamically generated to ensure that the generated AIGC content can be flexibly adjusted according to changes in the creation stage and feedback characteristics, thereby improving the accuracy of the generated content and complex and ever-changing user intentions. Attached Figure Description

[0020] Figure 1 This is a flowchart of the steps of the AIGC-generated content adaptive method based on user feedback in this invention;

[0021] Figure 2This is a schematic diagram of the AIGC-generated content adaptive system based on user feedback according to the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart of the steps of the AIGC content generation adaptive method based on user feedback according to the present invention. The following is a detailed introduction to the AIGC content generation adaptive method based on user feedback.

[0023] Step S1: Obtain and preprocess the multimodal user feedback dataset, use neural Hawkes processes for event modeling, obtain the first feedback event dataset, and perform causal discovery on the multimodal user feedback dataset to generate a causal impact dataset.

[0024] Specifically, in this embodiment, step S1 includes:

[0025] Step S1-1: Use neural Hawkes processes to model events and obtain the first feedback event dataset.

[0026] Specifically, firstly, a multimodal user feedback dataset is collected, which includes at least the original user feedback event sequence, multimodal signal data, and session context data, and then the data is preprocessed.

[0027] The original user feedback event sequence is represented as a sequence of specific operation instructions executed by the user on the generated content, and each user feedback event is represented as a tuple, i.e., an atomic event tuple. ,in, For the timestamp corresponding to the user feedback event; Enumerate event types, such as click, drag start / end, text insertion / deletion, slider adjustment, and area selection; The structured dictionary for event parameters, the contents of which depend on the event type. For example, for a drag-to-end event, the dictionary parameters may contain the start and end coordinates of the drag event and the code of the dragged content. Hardware or logical components used to identify the source of an event, such as a mouse, keyboard, or touchscreen.

[0028] Multimodal signal data refers to the feedback signals corresponding to the user's subjective feedback and intent. It includes at least the natural language description generated by the user in the input box or after speech-to-text conversion; the preset emotion icon clicked by the user; the sequence of user facial images captured at a fixed frequency when the device camera is enabled; and audio segments recorded in voice interaction mode.

[0029] Session context data represents the background information necessary to understand the current interaction, and includes at least a task identifier that uniquely identifies the current content generation task; a summary of all previous inputs and outputs in this session, such as a sequence of past user feedback events and their processing results.

[0030] Next, a preset multimodal encoder is used to perform event encoding on the multimodal signal data and session context data, generating a corresponding feedback event vector for each user feedback event.

[0031] Specifically, for the original user feedback event sequence, each atomic event tuple Input an event encoder based on the Transformer architecture, which is categorized by event type. Perform learnable embedding lookup for timestamps. Apply sinusoidal position encoding and convert the parameter dictionary These sub-embeddings are projected onto a fixed-dimensional vector using a parameter-specific multilayer perceptron. After being concatenated, they are then fused and context-modeled through a shared Transformer encoding layer, outputting a fixed-dimensional event semantic embedding vector. .

[0032] For unstructured multimodal signals, feature extraction and cross-modal alignment are performed on each modality. Specifically, for text comments, sentence Transformers such as Sentence-BERT are used for encoding; for pre-defined emotional icons such as Emojis, encoding lookup is performed using a predefined semantic vector table; for facial image frames, facial expression recognition networks such as MobileNet-V2 are used to extract the intensity vector of facial expression action units; for speech segments, prosody and emotion feature vectors are extracted using a speech emotion recognition model pre-trained based on historical data; and a contrastive learning-based cross-modal alignment method is adopted, that is, by comparing positive and negative samples, the distance between semantically consistent cross-modal samples is shortened, and the distance between irrelevant samples is widened, thereby aligning the feature encodings corresponding to the above modalities.

[0033] Embedding all atomic events Embedded with aligned multimodal signals Sort the events according to their corresponding timestamps and merge them into a unified event embedding sequence S in ascending order of timestamps. The event embedding sequence S is the original user feedback event sequence, i.e., S = [ , ,..., ],in for Moment Atomic Event Embedding Embedded with aligned multimodal signals Fusion embedding, i.e. The feedback event vector at any given moment.

[0034] Next, continuous-time modeling is performed on the original user feedback event sequence based on neural Hawkes processes to obtain the intensity function of each user feedback event, which is used to describe the probability of occurrence of the user feedback event on the time axis.

[0035] Specifically, a neural Hawkes process is used to model the event sequence S, which defines a conditional strength function. , indicating that in a given historical event Under the conditions, at time The instantaneous probability density of the event, and this intensity function is parameterized as a neural network, i.e. ;

[0036] in, It is the softplus activation function;

[0037] For each historical event For the current moment The effect decays exponentially over time, i.e. ,in, It is a multilayer perceptron used to embed events This is mapped to the strength of its impact on the future; This is the attenuation rate parameter.

[0038] By solving the above function, we can obtain any continuous time point. Intensity value and historical context representation This means obtaining a continuous-time event intensity function.

[0039] Finally, the original user feedback event sequence and the continuous-time event intensity function are used as the first feedback event dataset.

[0040] Step S1-2: Perform causal discovery on the multimodal user feedback dataset to generate a causal impact dataset.

[0041] Specifically, causal mining is performed on the original user feedback event sequence based on neural relationship reasoning to obtain a differentiable causal adjacency matrix, which is the causal influence dataset.

[0042] Specifically, the encoder of the neural relational reasoning model takes the original user feedback event sequence S corresponding to historical user feedback events as input, for The encoder outputs one of the event types. continuous adjacency matrix The continuous adjacency matrix This is a causal adjacency matrix, where the elements are... Represented as event type The occurrence of the event type The intensity of future causal effects;

[0043] The decoder of the neural relational reasoning model is used to... Gumbel-Softmax sampling is used to obtain discrete or continuous samples directly. Given potential noise, the dynamic or correlation statistics of the original user feedback event sequence S are reconstructed.

[0044] For example, in an image editing session, a neural relational inference model might output a... The matrix shows that the weights of event type A: {"brightness down", slider adjustment} to event type B: {"contrast up", slider adjustment} are 0.8, and the weights of event type A to event type C: {input text, "too dark"} are 0.1. This indicates that in historical data, after users reduce brightness, they are more likely to immediately increase contrast to compensate, rather than complaining immediately, thus representing a causal pattern corresponding to a user's operating habit.

[0045] Step S2: Based on deep kernel learning, the quality of feedback events is evaluated according to the first feedback event dataset and the causal impact dataset. The second feedback event dataset is generated by combining the results of the feedback event quality evaluation, and the event exploration intensity is obtained simultaneously.

[0046] Specifically, in this embodiment, step S2 includes:

[0047] Step S2-1: Perform feedback event quality assessment based on the first feedback event dataset and the causal impact dataset to obtain the quality assessment dataset.

[0048] Specifically, firstly, a directed graph G=(V,E,X,W) with node features and edge weights is constructed based on the first feedback event dataset and the causal influence dataset. This directed graph G is the user feedback event graph.

[0049] Specifically, each node in the user feedback event graph ∈V corresponds to an event type, node features This can be achieved by examining all elements belonging to the type in this session. event embedding The aggregated statistics show that, To determine the number of event types in the system, aggregation operations can include calculating the mean vector, variance vector, maximum and minimum vectors, and non-empty counts of the event embedding set.

[0050] For example, if there are three "brightness adjustment" events in a session, their embeddings are respectively , , The characteristic of the node "brightness adjustment" is:

[0051] The concatenation of each node's features integrates the intensity, semantic consistency, and frequency of this type of event in the current feedback.

[0052] Directed edges in the user feedback event graph ∈E represents a slave node Pointing to node Edges used to represent type Event type The potential causal effects of an event, and the weights of its edges. Directly taken from the causal adjacency matrix Corresponding element in ;

[0053] like If the value is below a preset threshold, the causal relationship is considered negligible, and the edge is not created to maintain the sparsity of the graph.

[0054] Besides from the causal adjacency matrix The derived causal edges can be further enhanced by adding additional edges based on the temporal relationships of the original user feedback event sequence S, thereby improving the graph's connectivity. For example, "temporally co-occurring edges" can be added between temporally adjacent event pairs, with their weights set to... ,in, The time difference between sequentially adjacent events. The preset time decay constant;

[0055] In addition, "co-pointing edges" can be added, which means that if two events point to the same content entity semantically or spatially, for example, if the nodes corresponding to the two events, such as a text description "top left corner" and a region selection box, then an undirected edge can be added between the corresponding nodes.

[0056] For example, suppose a feedback event includes event type T: {text comment: 'colors don't match'}, event type S: {drag the theme color slider from blue to green within 300ms}, and event type D: {select and delete a decorative element}. Then, the graph G constructed based on these three feedback events will have three nodes: T, S, and D. The feature of node T is the embedding statistics of the phrase "colors don't match," and the features of nodes S and D are the embedding statistics of their corresponding operations, respectively. Assuming these are learned from historical data... [T][S]=0.6, If [T][D]=0.4, then there exists a causal edge from T to S with a weight of 0.6 and a causal edge from T to D with a weight of 0.4 in the graph.

[0057] Meanwhile, since the S and D operations are close in time, if the time difference between the two is 50ms < the preset threshold of 500ms, an undirected temporal co-occurrence edge with a weight of w=exp(-50 / 500)≈0.9 is added between them.

[0058] If the deletion operation happens to target an element that is "incompatible" implied in the text comment, a co-reference edge between T and D may be added through co-reference analysis. This graph structure reflects the causal tendency that "color criticism" is more likely to lead to "color adjustment" rather than "deletion", and captures the co-occurrence relationship of "color adjustment" and "deletion" as parallel responses in this feedback.

[0059] Next, the following types of global and local graph theory features are calculated for the constructed user feedback event graph G to serve as auxiliary inputs for subsequent quality assessment. Specifically:

[0060] Overall topological characteristics: graph density (actual number of edges / maximum possible number of edges), number of connected components, average path length, and graph diameter;

[0061] Centrality characteristics: Calculate the degree centrality, proximity centrality, betweenness centrality, and out-degree centrality of each node;

[0062] Weight distribution characteristics: Calculate the mean, variance, and entropy of all edge weights;

[0063] Pattern Subgraph Detection: Search the graph for predefined subgraph patterns associated with high-quality or low-quality feedback. For example, detect whether there are chains from "negative emotion nodes" to "specific action nodes", or whether there are a large number of "leaf nodes" with only in-degree and no out-degree, which may indicate isolated feedback that has not received a response.

[0064] Then, using a preset depth kernel function, the user feedback event graph is combined synchronously to perform feedback event quality assessment and obtain a quality assessment dataset;

[0065] A pre-trained graph neural network maps a graph G to a low-dimensional, task-relevant feature space. Then, a standard kernel function, such as the radial basis function (RBF), or a deep kernel function, is applied to this feature space. Among them, graph neural networks Graph convolutional networks, graph attention networks, or graph isomorphic network architectures can be used, and their parameters... Learn from the data along with kernel parameters.

[0066] The input to a graph neural network is the feature matrix of the graph nodes. and adjacency matrix After multiple layers of message passing, the graph representation vector is finally obtained through global pooling. ;

[0067] The deep kernel function is trained on a large amount of labeled historical feedback graph data, and its training data is as follows: ,in, This is a historical feedback chart. It is a true label of feedback quality. The labels can come from post-event manual annotation or from implicit signals such as the user's adoption rate of subsequently generated content or the frequency of secondary modifications. The training objective is set to maximize the marginal likelihood of a Gaussian process under given observation data, thereby jointly optimizing the GNN parameters. Hyperparameters such as amplitude and length scale of the RBF kernel.

[0068] The current user feedback event graph G is input into a trained deep kernel function for confidence evaluation. Specifically:

[0069] First, the current graph representation is calculated based on the depth kernel function. With all of the training set A graphical representation of a sample Similarity between them, generate kernel vectors ,in ;

[0070] Next, the following data, pre-calculated during the training phase, is loaded from memory:

[0071] Training kernel matrix ,in ;

[0072] Training label vectors It contains the true confidence scores of all training samples;

[0073] Training kernel matrix Cholesky decomposition of the inverse matrix ,in It is the observation noise variance, which is a model hyperparameter. It is an identity matrix.

[0074] Calculate the mean of the posterior distribution of the confidence level using the Gaussian process regression formula. and variance Specifically:

[0075] ;

[0076] ,in It is self-similarity; for example, for the RBF kernel, its value is always 1.

[0077] And represent vectors in graphs Input into a separate, pre-trained multilayer perceptron regressor to obtain a specificity score. ;

[0078] and the mean ,variance and specificity score As a quality assessment dataset;

[0079] The training data for the multilayer perceptron regressor comes from historical interaction records. Its input is a graph representation vector extracted from the historical feedback event graph by the same GNN encoder. The label is the specific score of the corresponding feedback, which is obtained through subsequent manual annotation or rule-based automatic methods. For example, by analyzing whether the feedback contains specific location coordinates, clear attribute modification values, precise replacement text, and other quantifiable indicators, and normalizing them to a scalar in the [0,1] interval. During training, the mean squared error loss function is used, and the weight parameters of the MLP are optimized through the backpropagation algorithm to make the network output value as close as possible to the true label.

[0080] Step S2-2: Generate a second feedback event dataset based on the quality assessment dataset, and obtain the event exploration intensity according to the preset exploration intensity mapping function.

[0081] Specifically, the user's response is modeled according to a preset decision scheduling mechanism to obtain a decision tuple, which is used to constrain the generation of AIGC content;

[0082] Based on the aforementioned quality assessment dataset, stochastic optimal control modeling is performed, and deep reinforcement learning is used to solve the problem to obtain the optimal clarification strategy.

[0083] The decision tuples and the optimal clarification strategy are encapsulated into a second feedback event dataset.

[0084] Based on a preset exploration intensity mapping function, the event exploration intensity is obtained, and the event exploration intensity is used to constrain the generation of AIGC content.

[0085] In one possible embodiment, firstly, based on the feedback quality assessment results and information such as remaining time budget and task stage, the optimal action of the system at this moment is determined: whether to wait for possible further input from the user, generate a clarifying follow-up question, or directly generate exploratory AIGC content based on the current information. This process is formalized as a stochastic optimal control problem, specifically:

[0086] Define continuous-time state ,in It is the current confidence level belief distribution, such as a Gaussian distribution. ; It is a specificity score; It is the remaining decision-making time budget. The preset maximum duration; It is a task pattern embedding vector.

[0087] Define Action ∈A, where A is a preset action space, and includes at least:

[0088] Actively wait for users Second, Choose from a discrete set and set in Within a given time period, users' cumulative response time distribution function at time points... The function value at the given point is used to determine the response probability. If a response is received, the state transitions to a new feedback evaluation state; otherwise, the state remains unchanged. reduce , It also reduces the attenuation coefficient by combining it with the preset attenuation coefficient;

[0089] The generated question type is a predefined clarification question of type i, such as type 1: ("Can you point out what is wrong?"), type 2: ("Is it a color problem, a layout problem, or a content problem?"), and type 3: ("Can you describe the ideal effect?").

[0090] And set The action will consume a fixed time cost. and with probability Upon obtaining effective clarification, the system will be simultaneously upgraded according to a preset coefficient. and The value, or in terms of probability Received invalid or irrelevant replies, among which For the predefined first The probability of answering a clarification question;

[0091] Immediately initiate exploratory generation based on current understanding, and plan to generate multiple candidates.

[0092] Set the total expected cost ,in, Preset operating costs; The negative effect of the expected final content quality can be estimated through a value network that uses final confidence, specificity, and time margin as inputs to predict user satisfaction.

[0093] With the goal of finding the optimal policy function, an algorithm based on an actor-critic architecture is used to solve the above stochastic control problem. The policy network corresponding to the actor adopts the policy through parameterization and outputs the probability or deterministic action of each action, while the value network corresponding to the critic evaluates the current value. The network is trained by collecting a large number of trajectories in a simulated environment and the network parameters are updated by combining the policy gradient algorithm, so that the policy network parameters gradually approach the optimum, thereby obtaining a decision tuple and the optimal clarification policy.

[0094] The decision tuple includes at least whether clarification is needed, the type of clarification question, the maximum number of clarification candidate questions, and the waiting time; the optimal clarification strategy is the policy network with optimal parameters after training.

[0095] For example, the current state of the system is After forward computation, the probability distribution of the output action of the policy network is: P( )=0.6,P( )=0.25,P( The value is 0.15, which means that the system evaluation believes that asking for details is most likely to achieve the best results in the long run. Therefore, the optimal clarification strategy is to ask for details, and the system generates the corresponding decision tuple based on the preset rule base. The corresponding specific query statement;

[0096] Similarly, if the probability distribution of the current output action indicates that we should actively wait for the user's response, i.e., P( If the value is the largest, then its decision tuple is {no clarification needed, wait 6 seconds}, where the waiting time is determined by... *The scaling factor is obtained; here, the scaling factor is set to 3.

[0097] Subsequently, the top user feedback event graph G with the highest out-degree centrality was selected. There are 10 nodes, which are taken as the key node set. For each key node in the key node set, its node feature vector, i.e., the corresponding feedback event vector, is extracted. Multiply by its normalized out-degree centrality weight, and then sum the weighted features of all key nodes to obtain a key node aggregated feature vector.

[0098] Combine the key node aggregate feature vector with the graph representation vector. After concatenation, a pre-trained multilayer perceptron is injected. This multilayer perceptron consists of 2-3 fully connected layers, with the intermediate layers using the ReLU activation function, and outputs a fixed-dimensional vector. The vector This refers to the intensity of the event exploration.

[0099] The multilayer perceptron is trained in a supervised manner on historical interaction data. It extracts the generated content that the user is ultimately satisfied with after the user feedback event from the historical interaction data, calculates the ideal driving vector of the content in the state space, uses mean squared error as the loss function, and is trained by the Adam optimizer.

[0100] For example, assuming that the ideal driving vector corresponding to the "brightness increase" feedback in historical data has a positive value in the 5th dimension of the state space, the multilayer perceptron will aggregate the feature vector of the key node with the graph representation vector. The concatenated joint vector is mapped to a vector that is positive in the 5th dimension, i.e. =[0,0,0,0,0.8,0,...].

[0101] Step S3: Perform counterfactual reasoning based on the preset neural jump differential equation and causal influence dataset. Adjust the event exploration intensity according to the counterfactual reasoning results and obtain the adjusted event exploration intensity and the third feedback event dataset.

[0102] Specifically, in this embodiment, step S3 includes:

[0103] Step S3-1: Initialize the preset neural jump differential equation based on the causal influence dataset and the second feedback event dataset.

[0104] Specifically, firstly, the corresponding parameters of the state space of the preset neural jump differential equation are initialized based on the causal influence dataset and the second feedback event dataset;

[0105] Specifically, the preset neural jump differential equation is expressed as:

[0106] ;

[0107] in, This represents the state corresponding to the state space.

[0108] In one possible embodiment, for Define the system in time The complete state is a five-dimensional vector ;

[0109] in ;

[0110] Represented as the latent state of content, used to characterize the generated content in... The instantaneous state in the latent space, its evolution directly corresponds to the process of content generation and modification;

[0111] It is represented as a causal parameter, which is a time-varying parameter. A dimensional vector encodes the causal structure between variables within the system, and its value is taken from the causal adjacency matrix.

[0112] Represented as architectural parameters, these parameters characterize the specific neural network architecture used by the current generative model. The dimensional parameterized representation is generated by the isovariant graph hypernetwork in step S4.

[0113] Represented as user agent parameters, these parameters characterize the current distribution of beliefs about user preferences. The dimensional parameter takes the mean and variance of a Gaussian distribution.

[0114] Represented as policy parameters, used to encode the system's current decision-making strategy. A dimensional vector, which is the parameter or characteristic representation of the optimal clarification strategy.

[0115] Regarding the The initialization is performed in five dimensions, specifically:

[0116] set up This is used to indicate the evolution starting from the current latent state of the content;

[0117] The causal adjacency matrix obtained in step S1 Flattened into a vector, and used as a causal parameter. initial value ;

[0118] Will Initialize to a default architecture, such as an all-zero vector or a predefined baseline architecture vector;

[0119] Initialize it as the mean vector of the user's historical posterior distribution of preferences; if there is no history, initialize it as the prior mean.

[0120] The activation of the first few layers of the network weights in the optimal clarification strategy is used as... This allows for the injection of optimal decision-making knowledge into the system.

[0121] For example, in an image editing task, if the user reports "the sky is too bright and cloud details are lost," then this task should be reworked. It is a 256-dimensional latent representation of the current overexposed sky image; It is a 400-dimensional vector, where the element corresponding to the event from "brightness criticism" to "exposure adjustment" has a value of 0.8 and the element corresponding to "contrast adjustment" has a value of 0.6, thus reflecting the causal knowledge learned from historical data; The strategy of "prioritizing exploration when there is ambiguous feedback" is encoded; This encodes the user's preference for a "high dynamic range" style.

[0122] This is the drift term, which is generated by the neural network. Parameterization, with the current state as the input. and time Output a value that is the same as... A vector of the same dimension is used to describe the evolution trend of the system state in the absence of external disturbances and random perturbations.

[0123] This is the diffusion term, which is generated by the neural network. Parameterization outputs a diagonal covariance matrix or scalar variance to describe the random disturbances affecting the system state. This is the standard Wiener process.

[0124] This introduces uncertainty into state evolution, thereby simulating stochastic exploration during the generation process. The magnitude can be dynamically adjusted based on the quality assessment dataset in step S2, such as increasing it when the confidence level is low.

[0125] For jump terms, jump kernel It is a neural network whose input includes the system state one instant before the jump occurs. Feedback event vector of feedback events and the intensity of event exploration ;

[0126] As a conditional input, it directly controls the amplitude and direction of the jump. Output a jump vector ΔS in the state space to describe the instantaneous change in the system state caused by the user feedback event.

[0127] For the counting process of user feedback events, when When there are feedback events at all times, =1, otherwise 0.

[0128] The and Pre-training is performed on a large amount of data on the evolution of content without intervention. The system is then trained on feedback-state change pairing data to learn how to predict state changes caused by feedback.

[0129] For example, the instant you receive feedback that "the sky is getting dark," Set to 1, jump core Based on the current state Event characteristics and Calculate a jump vector S, S in The component may be a negative offset, in The causal weights may be slightly adjusted in terms of quantity, while The system updates the user's brightness preference, and at the same time, the status changes abruptly. = + S.

[0130] Next, a structural causal model is constructed based on the causal influence dataset, and the structural causal model is injected as a drift term into the preset neural jump differential equation. Specifically:

[0131] Causal parameters Interpreted as a structural causal model (SCM) defined on a subset of state variables, where The high weight values ​​in the equation directly define the strong causal dependency between variables;

[0132] For each structural equation in the SCM, a conditional score function is learned. This score function represents the logarithmic gradient of the distribution that the child variables should satisfy given the parent variables in the diffusion model. The score functions of all variables are then concatenated to form a causal score function. ;

[0133] and causal scoring function As an additional causal guiding force, it is added to the drift term of the neural jump differential equation to introduce soft constraints based on causal rules.

[0134] in, Training is performed on data that satisfies causal constraints using maximum likelihood or fractional matching, for example, on data containing causal pairs such as "sky brightness affects cloud contrast". You will learn that when sky brightness increases, the score function for cloud contrast should guide it to change in a downward direction to maintain physical plausibility.

[0135] For example, in image generation, a causal influence dataset might indicate that "global exposure" is a cause of both "sky brightness" and "ground shadows," in which case the scoring function... Will ensure: When When the feedback was that the "sky brightness" was darkened, This will simultaneously generate a force in the drift settings, brightening the "ground shadows" to maintain consistent global exposure, thereby preventing the generation of unreasonable images.

[0136] Step S3-2: Perform parameter correction and solve the accompanying process according to the preset neural jump differential equation to obtain the third feedback event dataset.

[0137] Specifically, based on the preset neural jump differential equation, the stochastic maximum principle is used to solve the adjoint process to obtain the event exploration intensity correction gradient and adjoint process parameters;

[0138] The event exploration intensity is corrected based on the event exploration intensity correction gradient to obtain the corrected event exploration intensity;

[0139] After extracting and solving the adjoint process, the corresponding parameters of the state space of the neural jump differential equation are preset to generate a state parameter set.

[0140] Specifically, defining the terminal time loss function This is used to measure the degree of undesirability of the final generated content. This may include: discrepancies with explicit user instructions; user dissatisfaction; penalties for violating pre-set physical or aesthetic rules, with the aim of minimizing expected loss. For the goal;

[0141] At the same time, the jumping item Event exploration intensity Consider it as an optimizable control input, explore Optimal correction , making Minimize The goal.

[0142] Define the Hamiltonian function:

[0143] ;

[0144] in, For the system state vector, corresponding to The possible values ​​of ;

[0145] As a state-accompanying variable, related to the state Same dimension, its physical meaning is the marginal value of the state to the objective function;

[0146] As a concomitant variable of diffusion, its dimension is similar to that of the diffusion term. Dimensional matching is used to handle the effects of random noise. Represents the trace of a matrix;

[0147] Represents the transpose of a vector or matrix;

[0148] To introduce the drift term of the neural jump differential equation after introducing soft constraints based on causal rules;

[0149] This is a jump kernel function used to describe the jump between control variables. Under the influence of feedback events, state jumps are caused;

[0150] According to the principle of random maximum, the optimal adjoint process It satisfies a backward stochastic differential equation, i.e. Its terminal conditions are: ;

[0151] in, Represents the state variable The partial derivatives, It is the differential of the Wiener process.

[0152] Backward stochastic differential equations describe the adjoint variables From the terminal time Initially, the dynamics of reverse time evolution;

[0153] Terminal conditions Terminal loss In the final state The gradient at that point provides a starting point for the backward solution.

[0154] The principle of stochastic maximum states that at the optimal solution, the gradient of the Hamiltonian with respect to the control variable should be zero. From this, it can be deduced that the expected terminal loss... Regarding control variables gradient It can be derived from the accompanying process and jump kernel function Express;

[0155] Right now ;

[0156] in, For the jump kernel function For control variables Jacobian matrix;

[0157] Represented as from 0 to Expected result of integration over the time interval.

[0158] The gradient is obtained by solving the above equation. ,according to The method to obtain the exploration intensity of the correction event. ,in This is the learning rate.

[0159] Accompanying process trajectory As an adjoint process parameter, and a causal parameter that will be updated during the solution of the neural jump differential equation. With user agent parameters As a set of state parameters;

[0160] Finally, the set of state parameters and the accompanying process parameters are encapsulated into a third feedback event dataset.

[0161] Step S4: Based on the preset multi-scale stochastic differential equation, AIGC adaptive generation is performed simultaneously by combining the second and third feedback event datasets and the exploration intensity of the correction event to obtain the original AIGC content dataset.

[0162] Specifically, the AIGC adaptive generation process parameters are initialized based on the state parameter set within the third feedback event dataset, and the content generation parameter set is obtained by combining the preset isomorphic graph hypernetwork and style mapping network.

[0163] Based on the preset multi-scale stochastic differential equation, AIGC content is generated cyclically in conjunction with the content generation parameter set to generate initial AIGC content.

[0164] The preset multi-scale stochastic differential equation includes event, diffusion, and causal scales, where the event scale is represented as follows: ,in For the cumulative state of events, It is the derivative of a Poisson process, which takes the value of 1 when a user feedback event occurs, and 0 otherwise;

[0165] Event metrics are used to characterize when user feedback arrives. A size of will occur instantly The jump was recorded simultaneously, capturing the instantaneous impact of the event.

[0166] The diffusion scale can be expressed as ,in, For the latent state of content, For dynamic architecture parameters, For causal parameters, This is the drift term in the neural jump differential equation. It is a diffusion term used to introduce random noise to ensure the diversity of generated content. It is the Wiener process.

[0167] The causal scale can be expressed as The ordinary differential equation form, where the function Used to describe how causal knowledge is based on current knowledge and expectations for generated content This function can be updated by analyzing historical data.

[0168] The equations at the three scales mentioned above are coupled with each other through variables to form a coupled set of stochastic differential equations, which is the pre-defined multi-scale stochastic differential equation.

[0169] Candidate content is generated based on the second feedback event dataset and the exploration intensity of the correction event, resulting in a candidate AIGC content set.

[0170] In one possible implementation, a noise vector is independently sampled from a multivariate standard normal distribution N(0,I). ,in This vector represents the dimension of the content's latent space. This represents the pure noise state at the beginning of the diffusion process, which is the starting point of the generation process;

[0171] The updated causal parameter vector within the state parameter set of the third feedback event dataset is directly set as the initial value of the causal scale state. Similarly, the updated user agent parameter vector is set to the initial value of the user preference state. ;

[0172] The intensity of the modified event exploration will be accumulated throughout the entire generation cycle. The lasting impact Initialize to zero, that is .

[0173] The initialized parameter variables are encapsulated into a structured environment state tuple. ,Right now This environment state tuple is the content generation parameter set.

[0174] Run a multi-scale generation loop. At each step of this loop, dynamically generate the architecture, extract styles, perform denoising, and update the multi-scale state based on the current state. If the total number of denoising steps is... Then it will be executed in a loop. Next, specifically:

[0175] First, the latent state of the content at the current moment. Causal parameters User agent parameters Concatenate them into a state feature vector ,Will The pre-trained isomorphic graph hypernetwork HG is input for forward propagation, and a parameter tensor is output. This parameter tensor is then defined as the parameters of the denoising network. Its dimensions match the total number of parameters in the basic denoising network.

[0176] HG is a graph neural network whose nodes correspond to different dimensions or semantic groups of the system state, and edges represent the interaction relationships between dimensions. The training objective of this network is to optimize the dynamic architecture parameters generated by the network through end-to-end optimization so that the user satisfaction score of the final generated content on historical feedback data can be maximized. The training data comes from the state sequences recorded in the system's historical interactions and the corresponding quality labels of the generated results.

[0177] Next, calculate mean vector ,like If the parameters are Gaussian distribution parameters, then the mean is taken directly; if it is a point estimate, then the mean is set to itself.

[0178] Will Input the pre-trained style mapping network and output the style vector for the current step. The dimension of this vector matches the dimension of the conditional injection layer of the generative model.

[0179] The style mapping network is a multilayer perceptron that is trained on a large amount of content-preference pairing data with style labels such as "oil painting" and "cool color". The network aims to minimize the distance between its output style vector and the target style in the generation space.

[0180] Next, the fixed computational graph skeleton of the basic denoising model is loaded, and... The parameter tensor is filled into the corresponding weight positions in the skeleton, and a parameter is temporarily instantiated. Denoising network;

[0181] Style vectors are processed using adaptive grouping normalization. Injected into each layer of this denoising network, for example, using AdaLN based on style vectors The scaling factor and offset factor for each set of normalization operations are dynamically generated to modulate the distribution of the feature map and achieve style control.

[0182] Invoke the instantiated denoising network and input the current latent state. and time step embedding Obtain the noise direction predicted by the network. ;

[0183] according to Obtaining correction items ,in, and These are two learnable scalar coefficients. To adjust the intensity of event exploration, These are parameters associated with the accompanying process;

[0184] According to the aforementioned correction item Correcting the noise direction predicted by the network The corrected denoising direction is obtained. ,Right now ;

[0185] Based on the update rule of the selected diffusion model sampler, combined with the correction noise direction Update the latent state to obtain an updated, clearer latent state. .

[0186] Subsequently, update the environment state tuple. The value of, i.e. ,in, It is the normalized time step;

[0187] Similarly, based on the preset parameter update rules, the values ​​of the causal parameters and user agent parameters are updated, and the updated parameters are obtained. and .

[0188] Ultimately, when from When the value decreases to 1, the loop ends, and the terminal latent state is obtained. .

[0189] Will The input is decoded by the decoder that is paired with the basic generative model. Preset post-processing operations are applied to the decoded original content to generate initial AIGC content. For example, automatic level / contrast adjustment, sharpening, color correction, and high-resolution restoration are performed on the image, and automatic typesetting, punctuation standardization, removal of duplicate phrases, and grammar checking and correction are performed on the text content.

[0190] Based on the decision tuples in the second feedback event dataset and the intensity of the modified event exploration, repeat the above multi-scale generation loop process until candidate content with the maximum number of clarification candidate questions is generated, and encapsulate all generated candidate content into a candidate AIGC content set.

[0191] Finally, the initial AIGC content, the candidate AIGC content set, and the corresponding content generation parameter set are encapsulated into the original AIGC content dataset.

[0192] Step S5: Based on the preset generated content quality analysis mechanism, the original AIGC content dataset is evaluated for quality, and then optimized and fine-tuned using a Bayesian optimization algorithm to generate an optimized AIGC content dataset.

[0193] Specifically, the preset generated content quality analysis mechanism includes at least content self-consistency analysis and deep kernel reverse simulation;

[0194] Content self-consistency analysis involves using an isovariant graph neural network to detect whether the original AIGC content dataset violates preset self-consistency constraints, obtaining the content self-consistency analysis results, and performing local repairs based on these results. Specifically:

[0195] An isotropic graph neural network is used to parse the initial AIGC content and each candidate AIGC content in the original AIGC content dataset into a weighted directed graph. , where nodes This is represented by dividing the content into meaningful units; for example, for images, using superpixel segmentation or object detection boxes; for text, using sentences or semantic blocks, and each node... It contains a feature vector extracted by the content encoder;

[0196] side Used to connect nodes that are related in space, time, or semantics;

[0197] edge weight Represents a node and The strength of the relationship between them.

[0198] Encode the pre-defined self-consistency constraints into pairs The topological conditions are determined, and the persistent homology features of the graph are obtained. These persistent homology features are used to measure topological features of the graph under different connection thresholds, such as the number of connected components, cycles, and holes.

[0199] By tracing the set of edges or nodes that cause topological anomalies, singular regions that violate constraints can be located. For example, a "shade" node should form a connected subgraph with its corresponding "light source" and "receiving surface" nodes. If the subgraph is not connected, it is identified as a singularity.

[0200] A conditional diffusion instigation model is used to instigate content regions marked as singularities. Specifically, the latent representation of the conditional region is used as an additional guide during the denoising process of the diffusion model, with the normal content surrounding the singularity region as a condition, and local instigation is performed with the goal of minimizing the topological anomalies of the singularity region.

[0201] The deep kernel inverse simulation is represented by combining a preset deep kernel function to predict user feedback to corresponding content in the original AIGC content dataset, obtaining the predicted user feedback distribution, and using a Bayesian optimization algorithm to optimize and fine-tune the original AIGC content dataset based on the predicted user feedback distribution to generate an optimized AIGC content dataset.

[0202] Specifically, the original AIGC content dataset, after self-consistency analysis and repair, is mapped to the latent space. The optimal perturbation vector is searched in the neighborhood of the latent space using a Bayesian optimization framework. The optimization aims to maximize the expected improvement in EI. The optimized initial AIGC content and each candidate AIGC content are then packaged into an optimized AIGC content dataset for output.

[0203] Figure 2 This is a schematic diagram of the AIGC-generated content adaptive system based on user feedback according to the present invention.

[0204] Specifically, the AIGC-generated content adaptive system based on user feedback includes:

[0205] The data processing module is used to acquire and process the multimodal user feedback dataset to acquire the first feedback event dataset and the causal influence dataset.

[0206] The feedback evaluation module is used to evaluate the quality of user feedback events and generate a second feedback event dataset and event exploration intensity based on the results of the quality evaluation.

[0207] The data optimization module is used to perform counterfactual reasoning based on a preset neural jump differential equation and a causal influence dataset to obtain the intensity of the modified event exploration and the third feedback event dataset.

[0208] The content generation module is used to obtain the original AIGC content dataset based on the preset multi-scale stochastic differential equation, the second and third feedback event datasets, and the intensity of the modified event exploration.

[0209] The content optimization module is used to perform quality assessment on the original AIGC content dataset and perform fine-tuning by combining Bayesian optimization algorithm to generate an optimized AIGC content dataset.

[0210] The specific usage and function of this embodiment are explained below:

[0211] First, by employing neural Hawkes processes, discrete and heterogeneous user interaction events are encoded into a continuous-time structured representation. Then, causal discovery is performed by combining neural relation reasoning to obtain a differentiable causal adjacency matrix, providing a richer and more accurate data foundation for subsequent steps.

[0212] Subsequently, the confidence and specificity of the feedback are quantified based on the deep kernel function, and a decision strategy of clarifying, waiting or exploring is generated based on stochastic optimal control theory.

[0213] Next, based on the causal influence dataset, causal constraints are injected into the preset neural jump differential equation, and counterfactual reasoning is performed by solving the adjoint equation to obtain the optimal correction gradient for the original feedback signal. This transforms the user's ambiguous or conflicting feedback into precise state-space driving instructions, thereby improving the deep understanding and execution accuracy of the user's potential intentions.

[0214] Next, a neural network architecture adapted to the current generation state is generated in real time through an isovariable graph hypernetwork, and content generation is performed in combination with a preset multi-scale stochastic differential equation, thereby enhancing the alignment capability and flexibility of the generated content.

[0215] Finally, through self-consistent topology repair and Bayesian optimization based on deep kernel inverse simulation, the initial content is locally enhanced and globally optimized, thereby generating final content with higher overall quality.

[0216] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via limited means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0217] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0218] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0219] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An adaptive method for AIGC-generated content based on user feedback, characterized in that, It includes the following steps: The multimodal user feedback dataset is acquired and preprocessed, and event modeling is performed using a neural Hawkes process to obtain a first feedback event dataset. Causal discovery is then performed on the multimodal user feedback dataset to generate a causal impact dataset. Based on deep kernel learning, and based on the first feedback event dataset and the causal impact dataset, the quality of feedback events is evaluated, and the results of the feedback event quality evaluation are combined to generate a second feedback event dataset, and the event exploration intensity is obtained simultaneously. Counterfactual reasoning is performed based on a preset neural jump differential equation and a causal influence dataset. The intensity of event exploration is then corrected based on the counterfactual reasoning results, and the corrected event exploration intensity and a third feedback event dataset are obtained. Based on the preset multi-scale stochastic differential equation, AIGC adaptive generation is performed simultaneously by combining the second and third feedback event datasets and the exploration intensity of the correction event to obtain the original AIGC content dataset; The original AIGC content dataset is evaluated based on a preset content quality analysis mechanism, and then optimized and fine-tuned using a Bayesian optimization algorithm to generate an optimized AIGC content dataset.

2. The AIGC-generated content adaptive method based on user feedback according to claim 1, characterized in that, The multimodal user feedback dataset is acquired and preprocessed. Event modeling is performed using a neural Hawkes process to obtain a first feedback event dataset. Causal discovery is then performed on the multimodal user feedback dataset to generate a causal impact dataset, including: The multimodal user feedback dataset includes at least the original user feedback event sequence, multimodal signal data, and session context data; A preset multimodal encoder is used to perform event encoding on the multimodal signal data and session context data, generating a corresponding feedback event vector for each user feedback event; The original user feedback event sequence is modeled continuously over time based on the neural Hawkes process to obtain the intensity function of each user feedback event; The intensity function is used to describe the probability of user feedback events occurring on the time axis; Based on neural relationship reasoning, causal mining is performed on the original user feedback event sequence to obtain a differentiable causal adjacency matrix, which is the causal influence dataset. The original user feedback event sequence, along with the corresponding feedback event vector and intensity function, are used as the first feedback event dataset.

3. The AIGC-generated content adaptive method based on user feedback according to claim 1, characterized in that, Based on deep kernel learning, and using the first feedback event dataset and the causal impact dataset to perform feedback event quality assessment, a second feedback event dataset is generated by combining the results of the feedback event quality assessment, and the event exploration intensity is obtained simultaneously, including: A user feedback event graph is constructed based on the first feedback event dataset and the causal impact dataset; A preset depth kernel function is used to simultaneously combine the user feedback event graph to perform feedback event quality assessment and obtain a quality assessment dataset. The quality assessment dataset shall include at least a confidence distribution and a specificity score; Based on a preset exploration intensity mapping function, the event exploration intensity is obtained, and the event exploration intensity is used to constrain the generation of AIGC content; The user's response is modeled according to a preset decision scheduling mechanism to obtain a decision tuple, which is used to constrain the generation of AIGC content; Based on the aforementioned quality assessment dataset, stochastic optimal control modeling is performed, and deep reinforcement learning is used to solve the problem to obtain the optimal clarification strategy. The decision tuples and the optimal clarification strategy are encapsulated into a second feedback event dataset.

4. The AIGC-generated content adaptive method based on user feedback according to claim 1, characterized in that, Counterfactual reasoning is performed based on a preset neural jump differential equation and a causal influence dataset. The intensity of event exploration is then adjusted based on the counterfactual reasoning results. This process yields the adjusted event exploration intensity and a third feedback event dataset, including: The corresponding parameters of the state space of the preset neural jump differential equation are initialized based on the causal influence dataset and the second feedback event dataset. A structural causal model is constructed based on the aforementioned causal influence dataset, and the structural causal model is injected as a drift term into the preset neural jump differential equation. Based on the preset neural jump differential equation, the adjoint process is solved using the stochastic maximum principle to obtain the event exploration intensity correction gradient and adjoint process parameters; The event exploration intensity is corrected based on the event exploration intensity correction gradient to obtain the corrected event exploration intensity; After extracting and solving the adjoint process, the corresponding parameters of the state space of the neural jump differential equation are preset to generate a state parameter set. The state parameter set and the accompanying process parameters are encapsulated into a third feedback event dataset.

5. The AIGC-generated content adaptive method based on user feedback according to claim 4, characterized in that, The predefined neural jump differential equations include: The preset neural jump differential equation is expressed as: ； in, The corresponding state in the state space; This is the drift term, used to describe the evolution trend of the system state under the condition of no external disturbances and random perturbations; The diffusion term describes the random disturbances that affect the system state. For standard Wiener process; These are jump items, used to describe transient changes in the system state caused by user feedback events; The system state just moments before the jump occurs; This refers to the feedback event vector corresponding to the user feedback event in the first feedback event dataset; For the counting process of user feedback events, when When there are feedback events at all times, =1, otherwise 0.

6. The AIGC-generated content adaptive method based on user feedback according to claim 1, characterized in that, Based on a preset multi-scale stochastic differential equation, AIGC adaptive generation is performed simultaneously by combining the second and third feedback event datasets with the exploration intensity of the correction event, resulting in the original AIGC content dataset, including: The AIGC adaptive generation process parameters are initialized based on the state parameter set within the third feedback event dataset, and the content generation parameter set is obtained by combining the preset isomorphic graph hypernetwork and style mapping network. Based on the preset multi-scale stochastic differential equation, AIGC content is generated cyclically in conjunction with the content generation parameter set to generate initial AIGC content. Based on the second feedback event dataset and the exploration intensity of the correction event, candidate content is generated to produce a candidate AIGC content set. The initial AIGC content, the candidate AIGC content set, and the corresponding content generation parameter set are encapsulated into the original AIGC content dataset.

7. The AIGC-generated content adaptive method based on user feedback according to claim 1, characterized in that, The original AIGC content dataset is evaluated for quality based on a pre-defined content quality analysis mechanism, and then optimized and fine-tuned using a Bayesian optimization algorithm to generate an optimized AIGC content dataset, including: The preset generated content quality analysis mechanism includes at least content self-consistency analysis and deep kernel reverse simulation; The content self-consistency analysis is expressed as using an equivariant graph neural network to detect whether the original AIGC content dataset violates a preset self-consistency constraint, and obtaining the content self-consistency analysis results; The deep kernel inverse simulation is represented by combining a preset deep kernel function to predict the user's feedback to the corresponding content in the original AIGC content dataset, and obtaining the predicted user feedback distribution. Based on the content self-consistency analysis results, the corresponding content in the original AIGC content dataset is repaired, and the original AIGC content dataset is optimized and fine-tuned simultaneously by combining the Bayesian optimization algorithm according to the predicted user feedback distribution, so as to generate an optimized AIGC content dataset.

8. An AIGC-generated content adaptive system based on user feedback, used to implement the method described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to acquire and process the multimodal user feedback dataset to acquire the first feedback event dataset and the causal influence dataset. The feedback evaluation module is used to evaluate the quality of user feedback events and generate a second feedback event dataset and event exploration intensity based on the results of the quality evaluation. The data optimization module is used to perform counterfactual reasoning based on a preset neural jump differential equation and a causal influence dataset to obtain the intensity of the modified event exploration and the third feedback event dataset. The content generation module is used to obtain the original AIGC content dataset based on the preset multi-scale stochastic differential equation, the second and third feedback event datasets, and the intensity of the modified event exploration. The content optimization module is used to perform quality assessment on the original AIGC content dataset and perform fine-tuning by combining Bayesian optimization algorithm to generate an optimized AIGC content dataset.

Images