A device for sparse label-oriented graph few-shot node classification

Abstract

The application discloses a kind of facing sparse label's graph few sample node classification device, it is related to graph neural network and meta-learning technical field.The application proposes an innovative technical solution, by designing a novel two-stage hybrid enhancement strategy, effectively alleviate the problem of double scarcity of node sample and meta-training task.The strategy adopts the technical means that in-task mixing is combined with cross-task mixing, data expansion is carried out in feature space and task space simultaneously, avoids the problem that the generalization performance of traditional graph meta-learning method sharply declines when task is scarce.In addition, the application creatively introduces a prototype super-resolution optimization strategy based on external objects (query set), uses the information of unlabeled query set to correct and support set calculation initial class prototype, significantly improves the estimation accuracy of class prototype.In the model training stage, the distance-based prototype classifier is trained using the enhanced task set, effectively improves the node classification accuracy and generalization ability of the model in the sparse labeling scene.

Description

Technical Field

[0001] This invention belongs to the field of image classification, specifically relating to a few-sample node classification device for sparsely labeled graphs. Background Technology

[0002] Graph data (such as social networks and citation networks) are widespread in modern society. Although Graph Neural Networks (GNNs) have achieved great success in some tasks where data is readily available, such as node classification, their effectiveness is limited by a large amount of labeled data, making them difficult to apply in many practical applications, such as new drug discovery and rare graph data classification. This is because obtaining a large number of labeled samples in these fields is not only expensive but also time-consuming. Therefore, Graph Few-Shot Learning (GFSL) was developed, with the aim of identifying new categories using a very small number of labeled samples (Support Set).

[0003] Although Graph Few Shot Learning (GFSL) has achieved good results, it still faces severe challenges under the extreme conditions of both task and sample scarcity. The first is the insufficient utilization of structural information (Neglect of Structural Priors). Existing methods do not focus on important prior information in the graph topology (such as node degree centrality). In most cases, nodes located at the edge of the graph contain more invalid information, so we need to distinguish them to ensure the stability of the prototype quality.

[0004] Then there's the scarcity of meta-tasks. Existing meta-learning frameworks require a large number of meta-training tasks to extract transferable knowledge. Therefore, when the number of available meta-tasks is limited, overfitting is prone to occur during training, and the test results are difficult to apply to new tasks in the testing phase. Finally, there's prototype estimation bias. Because the support set samples are small (e.g., 5-shot), the class prototype obtained by averaging them is often biased. This is like using a few low-resolution mosaic images to infer the whole picture of an object, losing a lot of detailed features, leading to blurred classification boundaries. Summary of the Invention

[0005] The purpose of this invention is to provide a graph node classification device for sparsely labeled graphs with few samples, aiming to solve the problem of poor classification performance of existing technologies for sparsely labeled small sample datasets.

[0006] The first aspect of this invention provides a novel graph few-shot learning paradigm—a two-layer Mixup—addressing the dual scarcity of "task-samples." In each meta-training iteration, the intra-task Mixup "upsamples" the node space by performing linear interpolation of the same class on the support set of the current task, generating a batch of "pseudo-node" features and their corresponding labels (preserving hard labels). These pseudo-nodes, together with the real nodes, constitute the augmented support set, effectively expanding the 5-shot to 5+Ns′-shot. Since the interpolation process is only performed within the same class, it does not violate the assumption of local graph smoothness and allows the classification boundary to obtain more "support points" in the feature space, directly alleviating the problem of excessive variance in prototype estimation caused by node scarcity.

[0007] Mixing tasks within a single task can only solve the problem of scarce nodes within the dataset, but it cannot solve the problem of too few task categories. Therefore, mixing tasks between different tasks is also necessary to address the issue of too few categories.

[0008] A second aspect of this invention provides a super-resolution technique based on external objects. Because the Support Set contains very few samples (e.g., only 5 points), the "class centers" calculated using these 5 points are extremely sparse and biased. If the true class centers are likened to a high-resolution image, then the prototype calculated from only 5 samples is a blurry mosaic (low-resolution representation). It loses many fine-grained features of the categories, resulting in unclear classification boundaries. To solve this problem, this invention designs C... 2 The module introduces the concept of super-resolution for feature reconstruction.

[0009] It includes the following steps:

[0010] Step 1: Use the unlabeled information in the QuerySet as an "auxiliary high-frequency information source". Although the queries are the objects we want to predict, they contain potential distribution information for that category.

[0011] Step 2: Input the basic prototype calculated by the Support Set (biased, low information content).

[0012] Step 3: Collect auxiliary information (Query Mean): Feature mean of the Ouery Set (supplementary distribution information).

[0013] Step 4: Dynamically determine how much information to extract from the Query Mean to "fill in" the missing details of the Support prototype using an adaptive neural network (Adaptive AlphaGenerator).

[0014] Step 5: Output a Refined Prototype. This is equivalent to "clarifying" the vague class centers, making them closer to the actual class distribution centers.

[0015] The beneficial effects of this invention are as follows:

[0016] High utilization rate: By incorporating super-resolution technology and two-layer Mixup technology into the small sample classification method, the information of each sample can be utilized more efficiently, which is significantly better than traditional methods.

[0017] High versatility: This invention mainly creates a brand-new small sample classification framework, which can be applied to many scenarios.

[0018] Low computational power requirement: Due to the high utilization rate of information for each sample, this invention has a low computational power requirement, which is suitable for most classification scenarios with small sample datasets to reduce costs. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the graph classification device with few samples provided in the embodiment of the present invention; it includes three main parts: ① node representation learning; ② two-layer Mixup; ③ super-resolution technology based on external objects. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0021] Example 1: Node representation learning.

[0022] like Figure 1 As shown in the node representation learning section, the initial step is generally to encode the nodes in the graph into a latent space, thus transforming them into low-dimensional hidden vectors. Graph Neural Networks (GNNs) have become the primary choice for node embeddings due to their powerful representational capabilities on graphs. They follow a message-passing mechanism, continuously aggregating messages from neighboring nodes to iteratively update the embedding of the target node. Based on this simple concept, we employ the SGC model to learn node embeddings. Specifically, it can be defined as:

[0023] in = - 1 / 2 −1 / 2 It is a symmetric normalized adjacency matrix with self-loops added, i.e. =A+I. = i,jThis represents the corresponding degree matrix. W ∗ This is a compressed weight matrix. After performing graph convolution, we can obtain a node vector H∈R that simultaneously encodes node features and topological structure. n×d .

[0024] Given that few-shot models are highly sensitive to noise, it is necessary to incorporate more prior knowledge to refine the representation. This prior knowledge is typically reflected in node degree, i.e., the popularity and importance of a node. Therefore, we consider explicitly incorporating it to evaluate each node. Specifically, we first employ another SGC to derive the interaction weights κ∈R for all nodes. n×1 Then, based on the node degree information, we obtain the node centrality α∈Rn×1 to adjust the degree normalization of κ. Finally, we use the adjusted score β∈R n×1 Obtain refined node representations X∈R n×d The above process can be described as follows:

[0025]

[0026] Example 2: Double-layer Mixup.

[0027] like Figure 1 As shown in the double-layer Mixup section, the double-layer Mixup in this embodiment includes the following steps:

[0028] Step S1: In-task blending.

[0029] Intra-task mixing. Since the number of sampling nodes in the support and query sets for each task is extremely limited during the meta-training phase, the efficiency of meta-training is significantly impacted. Therefore, we propose using an intra-task mixing strategy to generate more samples and increase data diversity. Specifically, for a given meta-training task T... t We support set S t and query set Q t Random sampling is performed, and two samples i and j from the same class k are selected and linearly interpolated to generate a new sample r. The above process can be described as follows:

[0030] Step S2: Task blending.

[0031] Simply performing intra-task mixing cannot solve the problem of a limited number of tasks. Therefore, we utilize cross-task mixing to directly create new tasks, making the task distribution more dense. Specifically, the first step is to start from the given meta-training task Dorg ={T t} Torg Two tasks T are randomly selected at t=1.i and T j The second step is to start from T. i Support set S i Randomly sample a class k from T j Support set S j We randomly sample a category k′ and then compute category-specific support prototypes. This process can be represented as:

[0032] Where II(·) is the indicator function, when Y i The value is 1 when ϱ = k, and 0 otherwise. Similarly, by applying Formula 4 to the query set Q of Ti... i and T j Q j We can obtain the query prototype Cq i;k of category k and Cqj;k′ of category k′.

[0033] Example 3: Super-resolution technology based on external objects.

[0034] like Figure 1 The super-resolution technology module in the image is shown below:

[0035] Information from samples used as part of the query set was used to make full use of the sample information. In the overall experiment, unlabeled information was used as an auxiliary information source. Then, an adaptive neural network (Adaptive AlphaGenerator) was used to dynamically determine the decision problem of extracting information from the information source. Finally, a modified prototype was given to make it closer to the real class distribution center.

[0036] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. A graph classification device for sparsely labeled nodes with few samples, characterized in that, include: By introducing a graph encoder with node degree prior, the representation vectors of nodes in the target graph are learned; Based on the node representation vectors, a meta-training task set is constructed using the N-way K-shot paradigm, where each task includes a support set and a query set; A two-stage hybrid augmentation operation is performed on the meta-training task set to simultaneously increase sample diversity and task quantity; The two-level blending enhancement operation includes intra-task blending and cross-task blending; For the enhanced task, the initial class prototype supporting set computation is super-resolution optimized using its query set information to obtain a high-precision class prototype.

2. The graph classification device for sparsely labeled nodes according to claim 1, characterized in that, This invention proposes a node representation learning mechanism that incorporates structural priors. Specifically, firstly, an initial node embedding representation is obtained by simplifying a graph convolutional network. Then, node degree information is introduced as a centrality prior and combined with interaction weights learned through another graph convolutional layer to generate weight coefficients for adjusting the initial embedding. These coefficients are used to weight the initial node embedding, resulting in a refined representation that better reflects the importance of nodes in the graph structure. The specific formula is shown below: Where Z is the node feature matrix, Â is the normalized adjacency matrix after adding self-loops, ℓ is the number of convolutional layers, W* and W are learnable parameter matrices, D̂i is the degree of node i, κ is the interaction weight, α is the node degree centrality prior, δ(·) is the activation function, β is the weight coefficient, X is the refined node representation of the output, and ⊙ represents element-wise multiplication.

3. The graph few-sample node classification device for sparse annotation as described in claim 1, characterized in that, The intra-task blending operation specifically involves: for a single meta-training task, randomly selecting two node samples from the same category in its support set or query set, linearly interpolating their representation vectors to generate new synthetic samples to augment the task. The cross-task blending operation specifically involves: randomly selecting two different tasks from the meta-training task set, calculating the class prototype of a certain category in their support sets for each, linearly interpolating the two class prototypes from different tasks to generate an interpolation task with a new category, thereby increasing the total number of meta-training tasks. The blending coefficient λ of the linear interpolation is sampled from the Beta distribution.

4. The graph classification device for sparsely labeled nodes according to claim 1, characterized in that, This invention proposes a class prototype super-resolution optimization strategy based on external objects. Specifically, firstly, an initial class prototype is calculated based on the task support set. Then, the mean feature value of all nodes in the same task query set is extracted as auxiliary context information. Through a learnable adaptive parameter generator, the initial class prototype and the auxiliary information are fused to reconstruct a higher-resolution optimized class prototype that more closely approximates the true class distribution. This process effectively utilizes unlabeled query set information to correct prototype bias caused by a small number of support samples. Specifically, it can be described as... init Optimization: Where F is the optimization function and Q is the query set. Model training employs a distance-based prototype loss function. For each augmentation task, the distance between its query set node and the optimized prototypes of each class is calculated, and the classification loss of the query node is minimized. The loss function is as follows: in, For the enhanced task set, For the task The query set, f(·) is the graph encoder, and d(·, ·) is the distance metric function (such as Euclidean distance). This is an optimized prototype for the true category y of node x.

5. A system for implementing the method of claim 1, characterized in that, include: The node representation learning module is used to learn the representation vectors of nodes in the target graph; The meta-task construction module is used to construct a meta-training task set in the N-way K-shot paradigm; A two-stage hybrid enhancement module is used to simultaneously expand sample diversity and task quantity; The prototype super-resolution optimization module is used to perform super-resolution optimization on the initial class prototype that supports set computation using its query set information, so as to obtain a high-precision class prototype.

Images