False information continuous detection method fusing preference modeling and representative memory bank

Abstract

The invention relates to a false information continuous detection method fusing preference modeling and a representative memory library, and the method comprises the steps: modeling a news propagation process into a propagation graph containing a root node, a user node and a directed edge, extracting user preference and propagation structure features through a graph neural network and a pre-training language model, and constructing dual-perception fusion embedding. On the basis, representative samples are selected under a fixed storage budget by adopting a DP-kcenter strategy to construct a memory library, and efficient condensation and storage of historical knowledge are realized. Through a lightweight playback training mechanism TiM-lite, a mixed memory sample and current task data are subjected to joint optimization, catastrophic forgetting is effectively relieved, and the cross-time and cross-domain generalization ability and detection stability of the model in a dynamic data stream are improved. The method gives consideration to detection precision, storage efficiency and privacy friendliness, and is suitable for continuous identification of social platform false information.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a method for persistent detection of false information that integrates preference modeling and a representative memory. Background Technology

[0002] In the digital information age, social media has become a core platform for information dissemination. However, the rapid evolution and disguise of misinformation pose a serious threat to public opinion security. Misinformation exhibits significant concept drift characteristics. Traditional misinformation detection models are mostly based on static graph assumptions, assuming that the distribution of training and testing data is consistent. In dynamic data streams, they face a dilemma: relying on old data cannot capture new dissemination patterns, while fine-tuning new data can lead to catastrophic amnesia, meaning the model quickly loses the knowledge it has learned.

[0003] With the widespread use of social media, the rapid spread of fake news poses a serious threat to public opinion security. Existing fake news detection methods are mostly based on static data assumptions, using news text content, dissemination network structure, or user context features for modeling. However, the social media environment is highly dynamic; news topics, dissemination patterns, and user behavior are constantly evolving, leading to concept drift in data distribution. If traditional static models are directly applied to dynamic data streams, the model will quickly forget old knowledge when learning new knowledge, resulting in the "catastrophic forgetting" problem.

[0004] Continuous learning aims to enable models to continuously learn a range of tasks without forgetting previously learned material. Existing methods include regularization (such as EWC and MAS), architecture extension, memory replay-based methods (such as GEM and ER), and knowledge distillation. In the field of fake information detection, research has attempted to incorporate continuous learning.

[0005] However, existing technologies still have the following limitations in practical applications:

[0006] Traditional replay methods require storing a large number of real samples, resulting in high storage costs. Furthermore, directly storing subgraphs in graph data may compromise structural integrity.

[0007] Existing methods mostly focus on the dynamic changes in the propagation structure, failing to fully utilize the relatively stable discriminative features inherent in user behavior preferences over time, resulting in insufficient robustness of the model in dealing with concept drift.

[0008] Therefore, there is an urgent need for a continuous learning detection method that can efficiently utilize limited storage resources while integrating dynamic propagation structures and stable user preference features, so as to achieve long-term, stable and effective detection of false information in dynamic environments. Summary of the Invention

[0009] The purpose of this invention is to provide a method for continuous detection of false information that integrates preference modeling and a representative memory bank, so as to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for persistent detection of false information that integrates preference modeling and a representative memory bank, comprising the following steps:

[0011] S1. Construct a news dissemination graph: Obtain news data streams in sequence, and model each news event and its dissemination process as a news dissemination graph. The news dissemination graph includes a root node representing news, several user nodes representing users, and directed edges representing dissemination relationships.

[0012] S2. Encoding User Preferences and Propagation Structure Features: Based on the news propagation graph, construct and train a graph neural network model; generate user preference feature vectors by encoding user historical posting records through a pre-trained language model, encode the text features of the news root node, combine the graph neural network to aggregate neighbor information to generate node embedding representations, and then obtain a joint graph-level representation containing user preferences and propagation structure information by reading out operations and linear mapping.

[0013] S3. Representative Preference Graphs (CPGs) Generation and Memory Construction: For each news dissemination graph in the current task batch, extract structural embeddings and preference embeddings, and generate dual-perception fusion embeddings through fusion weights and normalization operations; based on the DP-kcenter strategy and cosine distance similarity metric, select representative samples to form CPGs under a fixed storage budget and store them in the memory. The memory stores graph structure, node features and graph-level labels, and is dynamically updated according to sample representativeness and time order.

[0014] S4, Preference Replay and Lightweight Consolidation Optimization TiM-lite: includes three stages: warm-up training, memory replay, and memory update; warm-up training independently trains the model on new task batches to obtain initial parameters; memory replay stage mixes memory bank samples and current task samples proportionally to construct a mixed training set, and updates model parameters based on the total loss function; memory update stage re-evaluates representative samples and replaces redundant samples to maintain distribution coverage.

[0015] S5. Dynamic False Information Detection: The continuously updated graph neural network model is used to classify and detect false information in the input news dissemination graph.

[0016] Preferably, the news dissemination graph in step S1 is a directed graph G = (V, E);

[0017] The node set V includes the root node V0 representing the news itself and the set of user nodes {v1, v2, ..., v...} representing the participants in the dissemination of the news. N}, where N represents the total number of participating users;

[0018] Directed edges (v) in edge set E i ,v j This indicates that user i forwarded the news or replied to user j;

[0019] The topological structure of the graph is determined by the adjacency matrix A∈{0,1}. (N+1)×(N+1) express.

[0020] Preferably, the user preference feature extraction in step S2 specifically involves: collecting the historical post set H of user u. u ={t1,t2,…,t M Each post is mapped to a vector using a pre-trained language model, and then aggregated using average pooling to obtain user preference feature vectors.

[0021] Where Enc(·) is the pre-trained language model encoder, M is the number of historical posts, the feature vector of the news root node V0 is obtained by encoding the original news text through the pre-trained language model, and the initial features of all nodes constitute the feature matrix X. (0) ∈R (N +1)×d d is the feature dimension;

[0022] The pre-trained language model is the BERT model.

[0023] Preferably, the node embedding update rule of the graph neural network model in step S2 is as follows:

[0024]

[0025] Where N(v) is the set of neighbors of node v, W (k) Let be the learnable weight matrix of the k-th layer, AGG(·) be the aggregation function, σ(·) be the non-linear activation function, and k be the number of graph convolutional layers.

[0026] Preferably, in step S3:

[0027] The formula for generating dual-sensory fusion embedding is:

[0028] z sel,i = Norm([(1-α)·z str,i ⊕α·z pref,i ]);

[0029] Where z str,i For structural embedding, z pref,i For preference embedding, α∈[0,1] is the fusion weight of preference and structure, and Norm(·) is the normalization operation;

[0030] The goal of the DP-kcenter strategy is to select a representative set of size K. Minimize the maximum distance from any sample in the sample set to the nearest representative sample, i.e. in The cosine distance similarity metric is used, where S is the set of dual-sensory fusion embeddings corresponding to the news dissemination graph of the current task batch.

[0031] Preferably, in step S4:

[0032] The hybrid training set is constructed as follows:

[0033] B mix =B mem ∪(γ·B cur );

[0034] Among them B mem B is a sample for the memory bank. cur This is a sample for the current task;

[0035] The total loss function is:

[0036]

[0037] Where θ represents the parameters of the graph neural network model. For memory classification loss, γ is the classification loss for the current task sample, and γ∈[0,1] is the mixing coefficient for controlling the weights of the memory and the current task sample, which is used to control the proportion of the current task sample introduced.

[0038] Preferably, the update rule for the model parameter θ based on the overall optimization objective is as follows:

[0039]

[0040] Where η is the learning rate.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. This invention performs representative memory construction in a dual-perception embedding space of "user preferences and propagation structure". Under a fixed storage budget, it selects representative propagation graph samples with a balance between categories and domains to approximate the historical propagation distribution, thereby effectively mitigating catastrophic forgetting in continuous learning with less storage overhead. By adopting a representative selection strategy based on maximizing coverage through the representative memory bank, it significantly reduces storage and training costs while preserving key semantics and topological patterns, avoiding the problem of traditional playback requiring the retention of a large amount of historical data. By explicitly modeling user preference features, the model provides stable preference anchors in topic drift and cross-task scenarios, complementing the propagation structure features, thereby improving the generalization ability across time and domains. At the same time, in the model update stage, a lightweight consolidation training mechanism (TiM-lite) based on the memory bank is adopted. Through multiple rounds of training with memory samples and proportional mixing with current task data, it achieves continuous retention of historical knowledge and rapid adaptation to new tasks, ensuring the detection accuracy and stability of the model in dynamic environments. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the overall structure of the persistent false information detection method that integrates preference modeling and a representative memory bank, as described in this invention.

[0044] Figure 2 This is a schematic diagram of the propagation tree formed by constructing a news propagation graph for the continuous detection method of false information that integrates preference modeling and representative memory bank, as described in this invention. Detailed Implementation

[0045] 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.

[0046] Please see Figure 1-2 This invention provides a technical solution: a method for persistent detection of false information that integrates preference modeling and a representative memory bank, comprising the following steps:

[0047] S1. Constructing a News Dissemination Graph: Sequentially acquire news data streams and model each news event and its dissemination process as a news dissemination graph G = (V, E), where each graph G has a corresponding real label y. G ∈{0,1}, defined as follows:

[0048]

[0049] That is, 1 represents fake news and 0 represents real news;

[0050] The set of nodes V = {v0} ∪ {v1, v2, ..., v...} N}, where v0 represents the root node of the news item.

[0051] {v1,…,v N} represents the user nodes that participated in forwarding or interacting with this news, and N represents the total number of participating users;

[0052] Let the edge set E represent the news dissemination path or social relationships between users. If user i forwards news or replies to user j, then there exists a directed edge (v... i ,v j )∈E;

[0053] Based on the edge list of the original propagation records, construct a binary adjacency matrix A∈{0,1}. (N+1)×(N+1) Let represent the topological structure of the graph, and let the matrix elements satisfy: If That is, if v i ,v j ∈E then A ij =1, otherwise A ij =0.

[0054] S2. Encoding User Preferences and Propagation Structure Characteristics: Based on the news propagation graph, construct and train a graph neural network model;

[0055] A user's historical posting records reveal their long-term political leanings, emotional preferences, and linguistic style. For each user node u, we collect its historical posting set H. u ={t1,t2,…,t M The pre-trained language model (BERT) is used as the encoder Enc(·) to map each text into a vector, and then aggregated into a user's preference feature vector x through average pooling. u The formula is:

[0056]

[0057] For the news root node v0, we directly use Enc(·) to encode the original news text to obtain its feature vector. Ultimately, the initial features of all nodes constitute the feature matrix X. (0) ∈R (N+1)×d d is the feature dimension;

[0058] To capture high-order propagation patterns of news on social networks, the aforementioned graph neural network model is used to encode the graph structure. Given a propagation graph G and an initial feature matrix X... (0) Aggregate its neighbor information to obtain the node representation of the k-th layer. The specific update rules are as follows:

[0059]

[0060] Where N(v) is the neighbor aggregation of node v, W (k) Let be the learnable weight matrix of the k-th layer, AGG(·) be the aggregation function, and σ(·) be the non-linear activation function;

[0061] After K layers of graph convolution, the final node embedding representation is obtained.

[0062] To generate a graph-level representation z for misinformation classification G For classifying false information, the node features are first pooled through readout operations to obtain the propagation structure representation g. struct The root node text of the news is encoded to obtain the root node representation r. root The two are then concatenated and fused through a linear mapping layer to obtain a joint graph-level representation z that includes user preferences and propagation structure information. G .

[0063] This level represents z. G It implicitly preserves user preferences and propagation structure information, providing a unified semantic foundation for subsequent memory generation and continuous learning.

[0064] S3. Representative Preference Graph (CPG) Generation and Memory Construction: In continuous news detection tasks, the model needs to constantly process new propagation graphs that arrive over time. Directly storing all historical propagation data would lead to enormous computational and storage resource consumption and could easily cause catastrophic forgetting. To address this, this invention proposes a representative memory construction mechanism based on "dual-sensory fusion." By fusing user preference features and propagation structure features, and dynamically selecting representative propagation graph samples within a fixed storage budget, high-information-density representative preference graphs (CPGs) are formed and stored in a memory for knowledge retention and lightweight replay training in subsequent tasks. These subsequent methods will be used to simultaneously retain long-term user preference patterns, thereby effectively mitigating the forgetting problem during dynamic updates.

[0065] Faced with large-scale graph data, traditional training processes are often limited by computational resources and storage overhead, making it difficult to improve processing efficiency. This invention constructs a small but highly information-dense representative memory graph set that can represent the overall task distribution by selecting samples within a structure-preference fusion embedding space. Unlike replay methods that directly store all historical data or node-level behaviors, this method implicitly preserves the structural features of user preferences during the selection process. This achieves efficient and continuous learning while also improving data privacy, specifically:

[0066] S31. Dual-sensory embedding construction and fusion:

[0067] For the current task batch D t Each news dissemination graph G in i First, a graph neural network encoder is used to extract the structural embedding z. str,i This embedding reflects the propagation topology and interaction patterns; simultaneously, based on the user node's historical posting records, the preference embedding z is obtained through a language model encoder. pref,i This embed captures users' long-term emotional and content preferences.

[0068] To retain both types of information simultaneously, this invention introduces a dual-sensor fusion representation:

[0069] z sel,i = Norm((1-α)·z str,i +α·z pref,i )

[0070] Where α∈[0,1] are the fusion weights of preferences and structure, and Norm(·) represents the normalization operation used to constrain the feature scale. The resulting fusion embedding set... This serves as the input space for selecting representative samples in the future.

[0071] This ensures that CPGs can simultaneously preserve user preferences, propagation structure, and the distinguishing characteristics between true and false news, thereby forming a representational basis in the joint embedding space that combines semantic consistency and structural discriminativeness.

[0072] S32. In order to maximize representation coverage under limited storage budget, this invention selects several of the most representative samples from the current batch propagation graph based on the DP-kcenter (Dual-Perceptionk-center) strategy.

[0073] The similarity measure between samples is defined as cosine distance:

[0074]

[0075] The goal is to select a representative set of size K. Minimize the maximum distance from any sample in the sample set to the nearest representative sample:

[0076]

[0077] The algorithm employs a greedy "farthest-first" selection method to iteratively construct a set C, ensuring that the selected samples have the maximum distribution coverage in the embedding space. It can also balance the quota for each class based on category or domain information, where S is the dual-sensory fusion embedding set corresponding to the news dissemination graph of the current task batch.

[0078] S33. Construction and Update of Representative Memory: After the representative sample selection is completed, the obtained CPGs (Coreset Preference Graphs) are written into the representative memory M. The memory is constrained by a fixed storage budget B, that is, |M|≤B.

[0079] When the propagation graph data of a new task arrives and generates new CPGs, the system will update the memory bank according to the representativeness and time order of the samples, and prioritize retaining representative samples with high coverage and balanced category and domain distribution in the dual-sensory embedding space, so as to maintain an approximation of the distribution of historical tasks with limited storage.

[0080] The information stored in the representative memory bank consists of three parts:

[0081] (1) Graph Structure A i Used to represent the node connections in news dissemination;

[0082] (2) Node features X i Used to encode user preferences, content features, and interaction semantics;

[0083] (3) Graph-level tags y i ∈{0,1} is used to indicate the truth or falsehood category of news.

[0084] The above three parts of information together constitute the long-term memory unit of the model, which is used for replay training and continuous optimization in subsequent tasks.

[0085] S4, Preference Replay and Lightweight Consolidation Optimization (TiM-lite):

[0086] A core challenge in persistent fake news detection is that as new news events and user groups constantly change, the model's memory of earlier dissemination patterns, user preference features, and semantic distributions gradually decays, leading to a significant drop in detection accuracy—a typical "catastrophic forgetting" phenomenon. Furthermore, news topics, active user groups, and language features often drift across different time periods. If the model relies solely on current data for training, it is prone to losing previously learned cross-topic dissemination knowledge, thus failing to maintain stable long-term classification performance.

[0087] To address the aforementioned issues, this invention proposes a lightweight consolidation optimization mechanism (TiM-lite) based on a representative memory bank. By mixing and replaying historical representative samples (CPGs) during new task training, it achieves a balanced absorption and dynamic consolidation of new and old knowledge while maintaining learning efficiency. Unlike traditional full replay, TiM-lite only requires a small number of high-information-density samples from the memory bank to reconstruct the memory of historical preferences and propagation structures during parameter updates, significantly reducing storage and computational overhead.

[0088] S41, Warm-up Phase:

[0089] When the new task batch G new Upon arrival, a graph neural network model is first trained independently on this batch to obtain the initial parameters θ0 for the task. This stage enables the model to quickly capture the propagation topology, semantic features, and user group feature distribution of the new task, providing a stable initial representation for subsequent fusion training.

[0090] S42, Memory Replay Stage:

[0091] Subsequently, the system enters a lightweight playback phase based on memory bank M. The system samples memory batch B from the memory bank. mem and with the current batch sample B cur Mix them proportionally γ to construct a mixed training set: B mix =B mem ∪(γ·B cur Jointly optimize the following total loss function on the mixed samples.

[0092] in: The classification loss is used to remember samples (from CPGs) to preserve historical knowledge; γ is the classification loss for the current task sample, used to adapt to the new task; γ∈[0,1] is the mixing coefficient of the control memory and the current sample weight.

[0093] The model parameters are updated according to the following rules:

[0094]

[0095] Where η is the learning rate. This mechanism ensures that the gradient update direction strikes a balance between preserving historical knowledge and learning new information, enabling the model to maintain stable performance even when facing task switching or topic drift.

[0096] S43, Memory Update Phase:

[0097] After each new task is completed, representative samples are re-evaluated based on the embedding distribution of the latest task. The memory bank M is updated using the DP-kcenter selection strategy, and redundant samples are replaced under a fixed budget B, ensuring that the memory covers the semantic range of the current and historical distributions. Through this cyclical process of "replay-update-rebalancing," the model achieves a balance between continuous adaptation and long-term memory.

[0098] S5. Dynamic False Information Detection: Utilizes a continuously updated graph neural network model to classify and detect false information in the input news dissemination graph.

[0099] Example 1

[0100] The present invention also provides a specific embodiment and verification, specifically as follows:

[0101] To comprehensively evaluate the performance of our model in the fake information detection task, especially its ability to resist catastrophic forgetting in a continuous learning environment, we designed a systematic experimental procedure. The experiments are based on two widely used real-world social media datasets—PolitiFact and GossipCop—covering two typical scenarios: Temporal Incremental Learning (T-IL) and Cross-Domain Incremental Learning (D-IL). The following sections will describe the selection and preprocessing of the datasets, the experimental setup and baseline comparisons, and the evaluation metrics. Through comparative experiments with mainstream methods and ablation analysis, we will further verify the role of graph condensation and preference replay mechanisms in the model.

[0102] 4.1 Dataset

[0103] This study selects two subsets of the FakeNewsNet benchmark dataset: PolitiFact and GossipCop. These two datasets together contain approximately 20 million historical tweets related to the spread of fake news, with each news item represented as a propagation tree.

[0104] In the propagation graph, each news item serves as a central node, reflecting its complete propagation process. User nodes connect to news nodes through their tweets, forming "user-news" relationship edges. The edges in the propagation network are constructed by the forwarding behavior between users.

[0105] Table 1 Dataset Description

[0106]

[0107] 4.2 Experimental Setup

[0108] To evaluate the stability and anti-forgetting ability of the model in continuous learning scenarios, we designed two incremental learning task construction methods: time-incremental learning (T-IL) and cross-domain incremental learning (D-IL).

[0109] 4.2.1 Time Increment Learning (T-IL)

[0110] Incremental learning focuses on dividing tasks according to the time sequence of news releases to test the model's ability to retain knowledge over time. Based on the timestamps of news releases, we divide the entire dataset into multiple consecutive and non-overlapping task stages. Each stage contains news released and disseminated within the corresponding time period and its dissemination structure, thus constructing a continuous learning sequence that progresses over time.

[0111] In the GossipCop and PolitiFact datasets, each node (user or tweet) is accompanied by a UNIX timestamp, recording the specific time of its posting or retweeting. We perform the following preprocessing steps based on these timestamps:

[0112] Node timestamp mapping: Extract the timestamp of each node from the raw data (e.g., node_id=1, timestamp=1231344102, corresponding to UTC time 2009-01-07 03:41:42). For news nodes, if there is no direct timestamp, the publication time of the news is taken as its most recent forwarding time (i.e., the minimum value of the timestamps of related user nodes).

[0113] News sorting and task partitioning: Each news dissemination graph is sorted in ascending order according to its start time and divided into three consecutive tasks (T1, T2, T3) according to time windows. Considering the uneven distribution of actual data over time (for example, the GossipCop dataset covers the period from 19:41:21 on 2008-05-12 to 15:05:04 on 2018-12-17, but the vast majority of samples are concentrated in the last two years), we adopted a partitioning strategy that divides the data roughly equally according to the number of samples, so as to maintain a relative balance in the data scale of each stage.

[0114] The specific division is as follows:

[0115] PolitiFact dataset:

[0116] Task 1 (April 2008 to November 2014): Covers early political news, with a relatively simple dissemination structure.

[0117] Task 2 (November 2014 to August 2017): Key data before and after the 2016 US presidential election cycle, showing a significant increase in the density of disinformation dissemination.

[0118] Task 3 (August 2017 to December 2018): Corresponds to the latest propagation event and is used to test the model's ability to adapt to rapid concept drift.

[0119] GossipCop dataset:

[0120] Task 1 (May 2008 to August 2017): Spanning nine years, this task reflects the long-term accumulation of fake news in the entertainment category.

[0121] Task 2 (August 2017 to February 2018): Although it lasted only six months, it vividly demonstrated the typical characteristics of explosive dissemination on the Twitter platform.

[0122] Task 3 (February 2018 to December 2018): Includes recent propagation samples to examine the model's response to the latest trends.

[0123] In each task, the model is trained using only the news dissemination graph of the current task. After training, the model moves on to learn the next time period task. This setup naturally leads to a catastrophic forgetting problem—if the model lacks an effective memory mechanism when learning new tasks, it will quickly lose its ability to discern earlier news patterns.

[0124] Through time-increment learning experiments, we simulated a dynamic environment on a real social platform where news continuously emerges and the model needs to be constantly updated. We focused on examining how this invention maintains its ability to continuously recognize past knowledge when historical data cannot be accessed repeatedly.

[0125] 4.2.2 Cross-Domain Incremental Learning (D-IL)

[0126] In cross-domain incremental learning, we treat datasets from different content domains as independent task sequences, aiming to evaluate the model's transferability and generalization performance when faced with cross-platform and cross-topic data.

[0127] In FakeNewsNet, PolitiFact and GossipCop correspond to two distinct content areas: politics and entertainment.

[0128] PolitiFact focuses on political news, with a relatively concentrated user base and high level of interaction; GossipCop mainly covers entertainment news, with a wider reach and a more dispersed user structure.

[0129] The two differ significantly in node feature distribution, user behavior patterns, and propagation structure, making them suitable as experimental vehicles for cross-domain incremental learning.

[0130] In D-IL, we designed two task sequences to comprehensively examine model performance:

[0131] PolitiFact→GossipCop: a shift from political news to entertainment news; GossipCop→PolitiFact: a shift from entertainment news to political news.

[0132] In each order, the previous dataset is used as Task1 for training, and the next dataset is introduced as Task2. The model cannot access the original data of Task1 when learning Task2.

[0133] The specific learning process is as follows:

[0134] 1) Train the model on domain A (such as PolitiFact) to learn the propagation structure, user preferences and news characteristics in that domain.

[0135] 2) Use the propagation graph of domain B (such as GossipCop) as input for the new task. In this case, the model must be trained in the new news domain and cannot access the original data of domain A.

[0136] 3) After completing the training for the new task, re-evaluate the model performance on the test set in domain A. If the performance drops significantly compared to stage one, it indicates that the model has experienced catastrophic forgetting.

[0137] This setup effectively simulates the real-world application scenario of cross-platform news detection: the model needs to continuously adapt to emerging domains or topics while maintaining its ability to recognize knowledge from already learned domains. By introducing a continuous learning mechanism, this invention aims to achieve knowledge retention during cross-domain transfer, thereby obtaining stable detection performance across multiple domains.

[0138] Therefore, cross-domain incremental learning is not only used to test the transfer and generalization ability of the model, but also to evaluate its robustness in resisting forgetting and maintaining historical knowledge when facing propagation structures with large distribution differences.

[0139] 4.3 Experimental Verification and Evaluation

[0140] After training for each incremental task, we comprehensively evaluate the model's performance on both the current and historical tasks. To simultaneously consider both the classification performance of false information detection and the resistance to forgetting during continuous learning, this paper uses the following two types of metrics:

[0141] 1) Classification performance indicators

[0142] Accuracy: Reflects the overall classification accuracy of the model across various tasks and is the most intuitive performance metric.

[0143] Macro average F1 score (F1Macro): By calculating the F1 score for each category separately and then averaging it, it better addresses the evaluation bias caused by category imbalance (such as an uneven distribution of fake news and real news).

[0144] Average AP (Average Precision): This measures the overall discriminative ability of a model at different decision thresholds. Especially in scenarios with imbalanced positive and negative samples, such as fake news detection, AP can more accurately reflect the model's performance in identifying fake news (positive class).

[0145] mAP(finalmeanAPovertasks): Only considers the model after the last training iteration (AfterT3).

[0146] In both the time-incremental learning (T-IL) and cross-domain incremental learning (D-IL) experimental settings, we will report the accuracy, F1Macro, and AP / AUC for each task, and calculate the average values ​​(e.g., AverageAccuracy, AverageAP) for all tasks as a summary metric for overall classification performance.

[0147] 2) Continuous learning ability indicators

[0148] To assess the extent to which the model retains historical knowledge during the learning process, we introduce evaluation metrics commonly used in the field of continuous learning:

[0149] Backward transfer (BWT): This reflects the change in a model's performance on earlier tasks relative to its initial learning performance after completing training on all tasks. A significantly negative BWT indicates that the model exhibits significant catastrophic forgetting of old tasks when learning new ones; a BWT close to zero indicates that the model can effectively maintain or even improve its ability to recognize learned tasks.

[0150] 4.3.1 T-IL Detection Results

[0151] On the PolitiFact and GossipCop datasets, we comprehensively verified the effectiveness of this invention in time-incremental and cross-domain incremental scenarios from two dimensions: classification performance and continuous learning ability, using the two types of metrics mentioned above.

[0152] To fully demonstrate the advantages of our method, we selected the following two types of methods as comparative baselines: ER-GNN and SSM. ER-GNN is a graph-based continuous learning method based on experience replay, while SSM is a static fine-tuning method. Neither of them explicitly models user preferences or performs graph condensation, so they can serve as the basis for comparison in verifying the anti-forgetting ability of our invention.

[0153] Under the time-incremental learning setting, we divided each dataset into three sequential tasks based on the start time of news dissemination, and evaluated the performance on each task after training on all tasks. Table 2 shows the task-specific AP of the model on the GossipCop and PolitiFact datasets; Table 3 further summarizes the Average AP, mAP, and BWT metrics for the two datasets.

[0154] As shown in Table 2 of the GossipCop task results, the performance of ER-GNN and SSM decreased to varying degrees in early tasks as the task number increased, indicating a significant catastrophic forgetting phenomenon in time-incremental scenarios. For example, in GossipCop Task 1, the AP of SSM after completing all tasks was approximately 0.9369, still showing a certain gap compared to subsequent tasks, reflecting performance fluctuations during continuous learning; ER-GNN maintained a high AP in early tasks, but also fluctuated to some extent as tasks progressed. In contrast, the AP of this invention remained at a high level across all tasks, and the performance degradation was smaller as tasks progressed, indicating that the model can better preserve early news dissemination patterns through representative memory and preference replay. Traditional methods such as EWC only constrain parameter weights and are difficult to cope with drastic changes in graph topology; while the strategy of this invention not only preserves node features but also retains key propagation structure subgraphs in the representative memory, enabling continuous replay and consolidation of user interaction structure patterns, making it more robust in the face of time drift.

[0155] On the smaller PolitiFact dataset, traditional methods also exhibit significant forgetting: ER-GNN and SSM both show varying degrees of AP reduction for earlier tasks in subsequent tasks. In contrast, the AP of this invention remains approximately 0.9492-0.9957 across the three tasks, significantly outperforming ER-GNN and SSM overall. Due to the limited number of graphs and higher data noise in PolitiFact, and the limited data volume for each task after time partitioning, this scenario more readily exposes the stability of continuous learning methods. The fact that this invention maintains relatively stable performance under these conditions demonstrates its robustness to insufficient data and time drift.

[0156] Table 2. AP for T-IL Task Sub-tasks

[0157]

[0158] The overall metric (Table 3) is the mean AP across all task test sets after training; AverageAP (AP_cl) is the average AP of the current task after each task training is completed, and then averaged over the entire training process. It is used to measure the overall performance of continuous learning.

[0159] In the overall metrics of time-incremental learning (Table 3), the BWT of this invention is closer to 0 on both datasets, significantly outperforming ER-GNN and SSM, indicating that the model covers less existing knowledge and forgets less when learning new tasks. Combining mAP and AverageAP (AP_cl), it can be seen that this invention maintains relatively stable detection performance evolution throughout the training process, and can continuously maintain effective recognition ability under adverse temporal drift and changes in propagation structure, thus significantly improving the reliability and practical value of continuous fake news detection.

[0160] This indicates that the present invention introduces only very slight forgetting when learning new tasks, while traditional methods significantly overwrite early knowledge in time-increment learning scenarios. In summary, the present invention demonstrates better resistance to forgetting and more stable fake news detection performance in time-increment learning tasks.

[0161] Table 3 Overall metrics for time-incremental learning (AverageAP, mAP, and BWT)

[0162]

[0163] 4.3.2 D-IL Detection Results

[0164] In the cross-domain incremental learning (D-IL) setting, we consider two task sequences: PolitiFact → GossipCop and GossipCop → PolitiFact. In each sequence, the model is first trained on the source domain (Task 1), and then trained on the target domain (Task 2), with access to the original source domain data denied during the second training phase. Table 4 summarizes the detection performance (AP, AverageAP, and BWT) of different methods under the two transfer orders for cross-domain incremental learning (D-IL).

[0165] In the PolitiFact→GossipCop transfer direction, all three methods achieved high detection performance in the target domain GossipCop, but their ability to retain the source domain PolitiFact differed after the transfer. ER-GNN's source domain AP dropped to 0.9535 and BWT to -0.1875 after the transfer, showing some degree of forgetting; SSM-like's source domain AP was 0.9617 and BWT to -0.1562; this invention maintained an AP of 0.9840 in the source domain and achieved an AP of 0.9706 in the target domain, with an Average AP of 0.9773 and a BWT of -0.0625. Overall, this invention achieved higher cross-domain average performance in this direction.

[0166] In the GossipCop→PolitiFact transfer model, the contradiction between maintaining the source domain and adapting to the target domain becomes more prominent due to the significant differences in scale, topic distribution, and propagation structure between the two datasets. After the transfer, ER-GNN achieves an AP of 0.9869 for the source domain GossipCop, but only 0.7042 for the target domain PolitiFact, with a BWT of -0.3876, demonstrating significant instability and forgetting effects in cross-domain transfer. In comparison, the SSM-like Average AP (mean AP of the two domains in Table 4) is 0.9572 and the BWT is -0.0640. The present invention achieves APs of 0.9378 and 0.9967 in the source domain and target domain, respectively, with an Average AP of 0.9673 and a BWT of -0.0494 (the smallest absolute value). This indicates that the preference replay and representative memory mechanism introduced in the present invention can adapt to new domains while covering old domain knowledge less and making the cross-domain training process more stable, effectively alleviating the forgetting of old domain knowledge.

[0167] Combining the two migration directions reveals significant differences between PolitiFact (political) and GossipCop (entertainment) in user behavior patterns, propagation structures, and topic semantics. Simply relying on experience replay or local structure replay can easily cause model parameters to shift towards the new domain, leading to a degradation of capabilities in the old domain. This invention, through a representative memory and lightweight consolidation mechanisms, preserves the distribution anchors of the old domain in the feature space and continuously reawakens key propagation patterns from the old domain during training in the new domain. This results in stronger cross-domain generalization and anti-forgetting capabilities in cross-domain average performance metrics (AverageAP in Table 4 or AP_overall in Tables 5 / 6) and BWT metrics, validating the effectiveness of this invention in cross-platform fake news detection tasks.

[0168] Table 4 Summary of Cross-Domain Incremental Learning (D-IL) Results

[0169]

[0170] 4.3.3D-IL Comparative Experiment: Cross-Domain Incremental Learning

[0171] In the cross-domain incremental learning (D-IL) setting, we consider two transfer orders: PolitiFact → GossipCop (pf2gc) and GossipCop → PolitiFact (gc2pf). In each order, the model is first trained on the source domain (Task1), and then trained on the target domain (Task2); the original data from the source domain is not accessed during the Task2 training phase. After completing Task2 training, the model is evaluated on both domain test sets; Table 4 reports the AP and its mean AP for both domains, and Tables 5 / 6 further report the overall summary metrics AP_overall and BWT. We also report BackwardTransfer (BWT) to characterize the impact of learning the target domain on the source domain performance (a BWT closer to 0 indicates less forgetting).

[0172] To ensure fairness in the comparison, all methods involving memory mechanisms operate under the same storage budget (e.g., all replay methods in the table are run under a unified storage budget). It is important to emphasize that, under the same storage budget, replay strategies that directly store the true historical graph often only cover a limited sample distribution; while this invention, through representative memory construction, retains the key distribution information required for cross-domain discrimination in a more compact form, thereby improving replay efficiency under limited storage.

[0173] (1) Analysis of PolitiFact→GossipCop(pf2gc) results (see Table 5)

[0174] As can be seen from Table 5, the Fine-tune (lower bound) shows significant forgetting after cross-domain sequential training, with AP_overall being 0.8567 and BWT reaching -0.4375, indicating that the model suffers severe performance regression on the PolitiFact task after adapting to GossipCop.

[0175] Regularization methods (such as EWC, MAS, and SI) can improve AP_overall to some extent (approximately 0.8494–0.8733), but BWT is still a significantly negative value (approximately -0.4375 to -0.3438), indicating that simply retaining old knowledge through parameter constraints is still insufficient to stably maintain the source domain capability during cross-domain migration.

[0176] For the distillation class LwF (α=1, T=4), the BWT is relatively closer to 0 (-0.0938) in this order, but its AP_overall is 0.8823, which is still lower than that of the strong playback / structure playback class methods.

[0177] In the replay-type methods, ER-GNN and SSM further improved their overall performance, reaching 0.9160 and 0.9215 respectively; among them, the BWT of ER-GNN was -0.1875 and the BWT of SSM was -0.1562.

[0178] This invention achieves the highest AP_overall (0.9247) in the pf2gc scenario while keeping the BWT at -0.0625, indicating that it can effectively retain the source domain discrimination capability while adapting to the target domain. Overall, this invention achieves a higher AP_overall in the pf2gc scenario and keeps the BWT closer to 0, demonstrating superior cross-domain comprehensive detection capability and anti-forgetting ability.

[0179] (2) Analysis of GossipCop→PolitiFact(gc2pf) results (see Table 6)

[0180] In the gc2pf direction, due to the greater differences in dataset size and propagation structure between the two datasets, some methods exhibit more severe degradation after transfer (e.g., ER-GNN AP_overall is 0.8542, BWT is -0.3876). Regularization and distillation methods show significant differences in results in this direction: for example, MAS (λ=10) AP_overall is 0.9103, but BWT is -0.1280; SI (λ=50) AP_overall is 0.8351, BWT is -0.1645; LwF (α=0.5, T=2) AP_overall is 0.8650, BWT is -0.2815.

[0181] In contrast, the Structure Replay (SSM) class combines a higher AP_overall (0.8923) with a BWT closer to 0 (-0.0640) in this direction. This invention further compresses the BWT to -0.0494 in this direction, exhibiting more stable source domain preservation and less forgetting during cross-domain training.

[0182] Table 5 Comparison of different continuous learning methods for D-IL (PolitiFact→GossipCop)

[0183] Table5.Performance of differentcontinuallearningmethodsunderthecross-domai nsetting(PolitiFact→GossipCop)

[0184]

[0185] Table 6 compares the results of different continuous learning methods in D-IL (GossipCop→PolitiFact).

[0186] Table6.Performance of different continual learning methods under the cross-domain setting(GossipCop→PolitiFact)

[0187]

[0188] 4.4 Ablation Experiment

[0189] To deeply analyze the contributions of each core component in this invention framework to detection performance and anti-forgetting capabilities, this section sets up comparative experiments focusing on key aspects such as representative memory bank construction, preference modeling, and lightweight consolidation optimization (TiM-lite). The experiments are conducted under a cross-domain incremental learning framework (GossipCop→PolitiFact), ensuring that all methods involving replay / memory mechanisms operate under consistent memory storage budget constraints. Specifically, the following variant is constructed and compared with the complete model:

[0190] 1) DP-kcenter (w / o TiM-lite): Retain the representative memory construction and playback process of DP-kcenter, but remove the TiM-lite lightweight consolidation module to verify the role of TiM-lite in maintaining and stabilizing the old domain in cross-domain sequential training.

[0191] 2) w / RandomMemory+TiM-lite(RealReplay): Remove the representative selection strategy of DP-kcenter, randomly sample and store the real historical propagation graph for replay under the same memory budget, and keep the rest of the training process unchanged. Compare the replay efficiency difference between "random real replay" and "representative memory replay".

[0192] 3) DP-kcenter+TiM-lite(w / o Preference): The user preference encoding module is removed, and the node representation no longer contains preference-related features. Only the propagation structure and other training settings are retained to verify the necessity of user preference modeling in cross-domain incremental learning and its contribution to cross-domain stable discrimination.

[0193] 4) DP-kcenter+TiM-lite (structure-only selection): In the representative memory construction stage, selection / condensation is performed only based on propagated structural information, without introducing preference signals to participate in the memory selection process, in order to verify the gain of "preference-driven representative selection" relative to "pure structure selection".

[0194] Table 7 Ablation Experiment Results of the Core Components of this Invention (GossipCop→PolitiFact)

[0195] Table7.Ablation study on key components of the proposed method(GossipCop→PolitiFact)

[0196]

[0197] Compared to the RandomMemory+TiM-lite (RealReplay) variant that stores the real historical graph, the full model DP-kcenter+TiM-lite (FullModel) achieves superior overall performance under the cross-domain incremental learning (GossipCop→PolitiFact) setting:

[0198] Its Accuracy / F1_macro / AP_cl reached 0.8571 / 0.8535 / 0.8573 respectively, which is an improvement of 0.0491 / 0.0689 / 0.0312 compared to RealReplay's 0.8080 / 0.7846 / 0.8261; at the same time, BWT improved from -0.1296 to -0.0494 (an improvement of 0.0802), indicating that under the same storage budget constraint, the representative memory construction and playback mechanism of this invention has higher playback efficiency and stronger anti-forgetting ability. The reason is that randomly sampled real graphs are prone to insufficient coverage and loss of key distribution information under limited budget, while the representative selection of DP-kcenter can more centrally retain the propagation structure and feature patterns that are more critical to cross-domain discrimination, thereby achieving better old domain preservation and overall detection performance under limited storage overhead.

[0199] When the user preference encoding module (DP-kcenter+TiM-lite (w / o Preference)) was removed, the model's performance and stability simultaneously degraded: its AP_cl decreased from 0.8573 to 0.8349 (a decrease of 0.0224), while BWT worsened from -0.0494 to -0.2188 (a difference of 0.1694), indicating that the preference signal provided a more stable discriminative basis and distribution anchor point in cross-domain sequential training. This phenomenon suggests that during cross-domain transfer, there are significant differences in content / topic distribution and propagation structure. When relying solely on structure and general semantic features, model parameters are more likely to shift towards the new domain distribution and cover old domain knowledge; while user preference modeling can provide relatively stable cross-domain discriminative cues, thereby reducing the degradation of source domain capabilities and improving the robustness of the training process.

[0200] Furthermore, when representative memory construction is based solely on structural information (DP-kcenter+TiM-lite (structure-only selection)), the model's AP_cl is 0.8245 and BWT is -0.1645, representing a decrease of 0.0328 and a deterioration of 0.1151 compared to the Full Model's AP_cl = 0.8573 and BWT = -0.0494, respectively. This result indicates that while relying solely on topological structure for representative selection can preserve some diffusion patterns, introducing preference signals into the representative selection process can further improve the cross-domain discriminative density of the selected memories, thus achieving better overall performance and less forgetting under the same budget.

[0201] Furthermore, after removing the lightweight consolidation optimization module (DP-kcenter (w / o TiM-lite)), the model's AP_cl decreased from 0.8573 to 0.8442 (a decrease of 0.0131), and BWT deteriorated from -0.0494 to -0.1755 (a difference of 0.1261), indicating that TiM-lite significantly contributes to the continuous consolidation and retention of the old domain during the cross-domain sequential training phase. This module can suppress excessive perturbation of the discrimination boundary of the old domain while adapting to the new domain, thereby reducing forgetting and improving overall training stability.

[0202] In summary, the ablation results in Table 7 validate the effectiveness of this invention through three key aspects: "representative memory construction, preference modeling, and lightweight consolidation optimization." Under the same storage budget, DP-kcenter's representative memory has higher playback efficiency than random real playback; preference modeling significantly improves the stability of cross-domain sequential training and reduces forgetting; TiM-lite further suppresses source domain capability degradation, enabling the model to achieve higher comprehensive detection performance and stronger anti-forgetting ability in cross-domain incremental learning scenarios.

[0203] In summary, this invention effectively mitigates catastrophic forgetting caused by data distribution evolution during continuous learning through a collaborative mechanism of user preference modeling and representative memory replay. Specifically, user preference representations exhibit relative stability across time and domain scenarios, serving as discriminative anchors for the model to resist phased fluctuations in news content, topic semantics, and dissemination patterns. Simultaneously, this invention constructs representative memory units within a limited storage budget, preserving key dissemination structures and feature distributions from historical tasks in a more compact form. This allows the model to continuously recall old domain discriminative patterns when learning new domain / new stage data, thereby achieving effective preservation of historical knowledge and rapid adaptation to new knowledge.

[0204] In terms of storage efficiency, the representative memory construction strategy of this invention can save and replay large-scale historical propagation graphs in a compressed form. Compared with the playback method of directly storing real historical samples, it can cover more discriminative structural and feature information under the same budget constraints, reducing storage overhead and playback burden in continuous learning. Furthermore, by integrating user preference features and propagation structure features, this invention takes into account both cross-domain general discrimination criteria and scene-specific diffusion rules, enabling the model to achieve more stable detection performance and less forgetting under both time-incremental learning and cross-domain incremental learning settings, thereby improving robustness and generalization ability in complex dynamic scenarios.

[0205] In terms of process implementation, this invention employs a closed-loop mechanism of "data stream input—graph construction / encoding—memory update—joint training / lightweight consolidation" to support incremental updates of the model as data continuously arrives, avoiding the computational and engineering overhead caused by frequent full retraining. Furthermore, in some embodiments, representative memory units can be obtained by optimizing historical distributions, rather than simply copying and storing real user data. This reduces direct dependence on sensitive behavioral data while retaining key discriminative information, improving data security and compliance friendliness in actual deployment.

[0206] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0207] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for persistent detection of false information that integrates preference modeling and a representative memory, characterized by: Includes the following steps: S1. Construct a news dissemination graph: Obtain news data streams in sequence, and model each news event and its dissemination process as a news dissemination graph. The news dissemination graph includes a root node representing news, several user nodes representing users, and directed edges representing dissemination relationships. S2. Encoding User Preferences and Propagation Structure Features: Based on the news propagation graph, construct and train a graph neural network model; generate user preference feature vectors by encoding user historical posting records through a pre-trained language model, encode the text features of the news root node, combine the graph neural network to aggregate neighbor information to generate node embedding representations, and then obtain a joint graph-level representation containing user preferences and propagation structure information by reading out operations and linear mapping. S3. Representative Preference Graphs (CPGs) Generation and Memory Construction: For each news dissemination graph in the current task batch, extract structural embeddings and preference embeddings, and generate dual-perception fusion embeddings through fusion weights and normalization operations; based on the DP-kcenter strategy and cosine distance similarity metric, select representative samples to form CPGs under a fixed storage budget and store them in the memory. The memory stores graph structure, node features and graph-level labels, and is dynamically updated according to sample representativeness and time order. S4, Preference Replay and Lightweight Consolidation Optimization TiM-lite: Includes three stages: warm-up training, memory replay, and memory update; warm-up training independently trains the model on new task batches to obtain initial parameters; During the memory replay phase, the memory bank samples are proportionally mixed with the current task samples to construct a mixed training set, and the model parameters are updated based on the total loss function; during the memory update phase, representative samples are re-evaluated and redundant samples are replaced to maintain distribution coverage. S5. Dynamic False Information Detection: The continuously updated graph neural network model is used to classify and detect false information in the input news dissemination graph.

2. The method for persistent detection of false information based on the fusion of preference modeling and representative memory as described in claim 1, characterized in that: In step S1, the news dissemination graph is a directed graph G = (V, E); The node set V includes the root node V0 representing the news itself and the set of user nodes {v1, v2, ..., v...} representing the participants in the dissemination of the news. N }, where N represents the total number of participating users; Directed edges (v) in edge set E i ,v j This indicates that user i forwarded the news or replied to user j; The topological structure of the graph is determined by the adjacency matrix A∈{0,1}. (N+1)×(N+1) express.

3. The method for persistent detection of false information based on the fusion of preference modeling and representative memory as described in claim 2, characterized in that: The user preference feature extraction in step S2 specifically involves: collecting the historical posting set H of user u. u ={t1,t2,…,t M Each post is mapped to a vector using a pre-trained language model, and then average pooling is used to aggregate the vectors to obtain user preference feature vectors. Where Enc(·) is the pre-trained language model encoder, M is the number of historical posts, the feature vector of the news root node V0 is obtained by encoding the original news text through the pre-trained language model, and the initial features of all nodes constitute the feature matrix X. (0) ∈R (N+1)×d d is the feature dimension; The pre-trained language model is the BERT model.

4. The method for persistent detection of false information by fusing preference modeling and representative memory as described in claim 3, characterized in that: The node embedding update rule for the graph neural network model in step S2 is as follows: Where N(v) is the set of neighbors of node v, and W (k) Let be the learnable weight matrix of the k-th layer, AGG(·) be the aggregation function, σ(·) be the non-linear activation function, and k be the number of graph convolutional layers.

5. The method for persistent detection of false information based on the fusion of preference modeling and representative memory as described in claim 4, characterized in that: In step S3: The formula for generating dual-sensory fusion embedding is: Where z str,i For structural embedding, z pref,i For preference embedding, α∈[0,1] is the fusion weight of preference and structure, and Norm(·) is the normalization operation; The goal of the DP-kcenter strategy is to select a representative set of size K. Minimize the maximum distance from any sample in the sample set to the nearest representative sample, i.e. in The cosine distance similarity metric is used, where S is the set of dual-sensory fusion embeddings corresponding to the news dissemination graph of the current task batch.

6. The method for persistent detection of false information based on the fusion of preference modeling and representative memory as described in claim 5, characterized in that: In step S4: The hybrid training set is constructed as follows: B mix =B mem ∪(γ·B cur ); Among them B mem B is a sample for the memory bank. cur This is a sample for the current task; The total loss function is: Where θ represents the parameters of the graph neural network model. For memory classification loss, γ is the classification loss for the current task sample, and γ∈[0,1] is the mixing coefficient for controlling the weights of the memory and the current task sample, which is used to control the proportion of the current task sample introduced.

7. The method for persistent detection of false information based on the fusion of preference modeling and representative memory as described in claim 6, characterized in that: The update rule for the model parameter θ based on the overall optimization objective is as follows: Where η is the learning rate.

Images