Small sample anti-fraud classification method and device, medium and product

Abstract

The embodiment of the invention provides a small sample anti-fraud classification method and device, a medium and a product, and relates to the technical field of artificial intelligence. The method comprises the steps of constructing intra-class semantic features and inter-class difference features based on a preset task framework of small sample learning; determining an anti-fraud classification network model according to the intra-class semantic features and the inter-class difference features; and according to the network model, performing anti-fraud classification identification on a to-be-classified text. According to the scheme, the preset task framework based on small sample learning does not need large-scale labeling, when a category is newly added, only a small number of samples need to be supplemented according to the framework to be fused into an existing model, an overall data system does not need to be reconstructed, the expansibility is greatly improved, through intra-class and inter-class feature learning, the distinction degree of similar fraud classes is enhanced, wrong classification is reduced, and the classification efficiency is improved. The classification accuracy is improved, general fraud features can be migrated to a new scene, and the generalization ability of the new fraud scene is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a small sample anti-fraud classification method, device, medium and product. BACKGROUND

[0002] The anti-fraud classification method based on text data has applications in rule library, machine learning and deep learning, but there are obvious bottlenecks in actual scenarios: first, the dynamic fraud scene has poor scalability, the rule library method relies on artificial preset rules and cannot cover real-time fraud methods, and the rule maintenance cost is extremely high; second, the accuracy of similar fraud categories is lacking, existing models focus on text feature learning and do not explicitly distinguish the differences between similar fraud labels, and as the label level is refined, the classification effect of similar labels is poor, affecting the accuracy of anti-fraud governance; third, the generalization of new fraud scenes is weak, machine learning requires artificial design of features and deep learning relies on a large amount of labeled data, when encountering new fraud categories outside the training data, the model prediction ability drops sharply, and the construction cost of new category samples is high and the cycle is long, which is difficult to meet the real-time needs of anti-fraud. SUMMARY

[0003] At least one embodiment of the present application provides a small sample anti-fraud classification method, device, medium and product, which is used to solve the problems of poor scalability, lack of accuracy and weak generalization of new fraud scenes in the prior art anti-fraud classification method based on text data.

[0004] To solve the above technical problems, the present application is implemented as follows:

[0005] In a first aspect, the embodiments of the present application provide a small sample anti-fraud classification method, comprising:

[0006] Based on the preset task framework of small sample learning, the intra-class semantic features and the inter-class difference features are constructed;

[0007] According to the intra-class semantic features and the inter-class difference features, a network model for anti-fraud classification is determined;

[0008] According to the network model, the text to be classified is identified for anti-fraud classification.

[0009] Optionally, based on the preset task framework of small sample learning, the intra-class semantic features and the inter-class difference features are constructed, comprising:

[0010] Based on the preset task framework of small sample learning, the intra-class semantic features are constructed;

[0011] The intra-class semantic features are learned for inter-class difference, and the inter-class difference features are constructed.

[0012] Optionally, the preset task framework based on small sample learning constructs the intra-class semantic features, including:

[0013] The description information supporting each relationship category is used to generate a relationship vector; the support set is used to represent a set of labeled samples for learning the features of the abnormal behavior category in each training, verification or test task;

[0014] Based on the text sequence features, the features of the head entity and the tail entity are extracted according to the pre-embedded entity position markers to generate entity feature vectors with the same dimension;

[0015] The preset translation mechanism is used to construct the relationship semantic vector between the feature vectors of the head entity and the tail entity by using the entity feature vectors corresponding to the head entity and the tail entity;

[0016] The relationship vector, the entity feature vector and the relationship semantic vector are fused through the attention mechanism to determine the intra-class semantic features of the support set.

[0017] Optionally, in the model training of the preset task framework based on small sample learning, each training step corresponds to an independent task including the support set and the query set;

[0018] The relationship categories in different training steps are the same or different, and the corresponding instances are non-repetitive.

[0019] Optionally, the intra-class semantic features are subjected to inter-class difference learning to construct inter-class difference features, including:

[0020] For a target relationship category, the instance feature vector corresponding to the target relationship category is extracted from the intra-class semantic features, and an interaction matrix including instance indexes is established; the indexes in the interaction matrix correspond to the positions of the instances in the similar categories of the target relationship category; the instances in the interaction matrix are constructed based on the instance features of the similar categories other than the target relationship category in the intra-class semantic features;

[0021] According to the interaction matrix, the interaction features between the target relationship category and the similar categories in the support set are constructed;

[0022] The inter-class difference learning is performed on the interaction features, the difference features of each category are aggregated, the sentence with the largest information amount in the target relationship category is selected, and the inter-class difference representation is generated;

[0023] According to the inter-class difference representation, the inter-class difference features of the target relationship category are constructed.

[0024] Optionally, the network model for anti-fraud classification is determined according to the intra-class semantic features and the inter-class difference features, and the network model includes:

[0025] The intra-class semantic features and the inter-class difference features are spliced and fused to determine fusion features.

[0026] Based on a preset query set formula and the fusion features, probabilities that the query set belongs to each relationship class in the support set are determined.

[0027] When a preset training model is trained and optimized by using a cross-entropy loss function, a loss value is calculated based on the probabilities, the model parameters of the preset training model are adjusted by minimizing the loss value, and in the model training process, the model parameters are optimized and trained by using an optimizer to determine the network model for anti-fraud classification.

[0028] Optionally, before the intra-class semantic features and the inter-class difference features are constructed based on a preset task framework of small sample learning, the method further includes:

[0029] For any text, user vocabulary in the text is determined as a head entity, and business vocabulary is determined as a tail entity.

[0030] An abnormal behavior class between the head entity and the tail entity is used to generate a structured triple including the head entity, the abnormal behavior class, and the tail entity.

[0031] In a second aspect, an embodiment of the present application provides a small sample anti-fraud classification device, including:

[0032] A first construction module is configured to construct intra-class semantic features and inter-class difference features based on a preset task framework of small sample learning.

[0033] A first determination module is configured to determine a network model for anti-fraud classification according to the intra-class semantic features and the inter-class difference features.

[0034] A first processing module is configured to perform anti-fraud classification and identification on a text to be classified according to the network model.

[0035] In a third aspect, an embodiment of the present application provides a computer readable storage medium, which stores a program. When the program is executed by a processor, the steps of the method according to any one of the first aspect are implemented.

[0036] In a fourth aspect, an embodiment of the present application provides a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, the steps of the method according to any one of the first aspect are implemented.

[0037] Compared with the prior art, the small sample anti-fraud classification method, device, medium and product provided by the embodiment of the application, based on the preset task framework of small sample learning, construct intra-class semantic features and inter-class difference features; according to the intra-class semantic features and the inter-class difference features, determine a network model for anti-fraud classification; according to the network model, perform anti-fraud classification and identification on the text to be classified. The scheme of the application, based on the preset task framework of small sample learning, does not need large-scale labeling, and when a new class is added, only a small amount of samples need to be supplemented according to the framework to integrate into the existing model, without the need to reconstruct the overall data system, greatly improving the expansibility, and through intra-class and inter-class feature learning, enhancing the discrimination degree of similar fraud classes, reducing false classification, improving classification accuracy, migrating general fraud features to new scenarios, and improving the generalization ability to new fraud scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0038] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not intended to limit the application thereto. Moreover, like reference numerals in the drawings designate similar parts throughout the several views. In the drawings:

[0039] Figure 1 A flowchart of a small sample anti-fraud classification method provided by the embodiment of the application is shown in FIG. 1.

[0040] Figure 2 A structure diagram of a small sample anti-fraud classification device provided by the embodiment of the application is shown in FIG. 2. DETAILED DESCRIPTION

[0041] The terms "first", "second", and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second" are usually a class, and are not limited to the number of objects, for example, the first object can be one or more. In addition, "or" in the present application means at least one of the connected objects. For example, "A or B" covers three schemes, namely, scheme one: including A and not including B; scheme two: including B and not including A; scheme three: including A and including B. The character " / " generally represents that the objects before and after are in an "or" relationship.

[0042] The term "indication" in this application can be a direct indication (or explicit indication) or an indirect indication (or implicit indication). The direct indication can be understood as the sender explicitly informing the receiver of specific information, operations to be performed or requested results, etc. in the sent indication. The indirect indication can be understood as the receiver determining the corresponding information according to the indication sent by the sender, or judging and determining the operations to be performed or the requested results according to the judgment result.

[0043] As described in the background, the text data-based anti-fraud classification method can be mainly divided into three categories. The first category is a rule-based method, which takes rules, standards, risk measurement systems, etc. as the core, is widely used in Internet finance such as banks, and performs fraud evaluation according to whether the customer behavior triggers the rule base. However, with the changing methods of fraud, this method relying on statistical matching has poor scalability and cannot adapt to actual security needs; the second category is a machine learning model-based method, including logistic regression, random forest, gradient boosting machine, etc., which provides a probability score to assist in determining whether to mark the data as high risk. This method usually needs manual design of feature engineering to extract abnormal information in the data. In addition, there is no correlation between different features designed, and it is impossible to mine deep semantics to deal with more complex text scenarios; the third category is a deep learning-based method, which usually trains a neural network model combined with supervised learning, automatically analyzes and extracts features of text data, but it needs to rely on a large amount of labeled data to achieve good performance, and cannot well handle new fraud type data, which seriously affects the user experience. To solve at least one of the above problems, the embodiments of the present application provide a small sample anti-fraud classification method, device, medium and product, which can reduce or avoid the above situations, and solve the problems of insufficient scalability, lack of accuracy and weak generalization of new fraud scenarios in the existing anti-fraud classification method based on text data.

[0044] The embodiments of the present application provide a small sample anti-fraud classification method, device, medium and product. The method and device are based on the same application concept. Since the principles of solving problems are similar, the implementation of the device and the method can be mutually referred to, and the repeated parts will not be described again.

[0045] Please refer to Figure 1 The small sample anti-fraud classification method provided by the embodiments of the present application comprises:

[0046] Step 11, based on the preset task framework of small sample learning, constructing intra-class semantic features and inter-class difference features.

[0047] In the embodiments of the present application, step 11 is the feature construction stage of the method, and under the premise of a small amount of labeled samples, two types of key features supporting fraud classification are accurately extracted to provide effective input for subsequent models. The preset task framework of small sample learning can optionally adopt an N-way-K-shot task setting, N is the number of fraud categories contained in each task, and K is the number of labeled samples required for each category, K is usually 5 to 10, and the fraud classification is abstracted as a multi-label relation extraction task.

[0048] For example, in any text data, the customer in the text is marked as a head entity, and the business is marked as a tail entity. The goal is to mine the fraud relationship between the two, such as abnormal machine switching and repeated card opening, without relying on large-scale labeled data of traditional methods, and adapting to the actual situation of sample scarcity in the fraud scene.

[0049] The intra-class semantic feature focuses on the essence feature extraction of the same fraud category. Using the text description of each fraud category in the support set (a small set of samples used for learning in each task), a relationship description vector is first constructed to guide relationship identification; then the feature vector of the head / tail entity is extracted from the text encoding feature, and the relationship semantic vector with text context is modeled by combining the entity pair association mechanism; finally, through the attention mechanism, the text part with large contribution to classification is highlighted, and the intra-class semantic feature of each fraud category is finally formed, ensuring that the model captures the essence of the category rather than individual sample details.

[0050] The construction of inter-class difference features is used to solve the problem that similar fraud categories are easily confused. First, analyze the fraud label system, identify similar label pairs with the same parent class, and determine the similarity; then, the text instances of similar categories are modeled interactively, the instance vector in the intra-class feature is extracted to construct an interaction matrix, the instance-level similarity and difference are calculated, and the most different instance is selected by the minimum function, and the class distinction is enhanced by comparison, to avoid misjudgment of similar categories.

[0051] Step 12, determining a network model for fraud classification according to the intra-class semantic feature and the inter-class difference feature.

[0052] This step 12 is the model construction and optimization stage of the method, which fuses the two types of features extracted in the previous step to construct and optimize a network model that adapts to the small sample fraud scene. The intra-class semantic feature (reflecting the essence of the category) and the inter-class difference feature (reflecting the difference between categories) are spliced and fused to form a complete feature representation of each fraud category. Both the common semantics of the same category and the distinction from other categories are highlighted to avoid classification bias caused by single features.

[0053] Optionally, the infrastructure can use a large language model as a text encoder to ensure effective extraction of text features; and a classification layer is subsequently accessed to input the fused category features to calculate the probability that the text to be classified belongs to each fraud category.

[0054] Step 13: According to the network model, the text to be classified is classified and identified for fraud.

[0055] Step 13 of the present application is the model application and classification stage of the method, which applies the trained model to the actual anti-fraud scene to realize accurate category determination of the text to be classified. The text to be classified, such as customer business handling records and customer service dialogue texts, is preprocessed in the same way as in the training stage, and the customer (head entity) and business (tail entity) in the text are marked. The text sequence features and entity features are extracted by the same large language model. The features of the text to be classified are input into the trained network model. The model first judges whether the text meets the essential features of a certain fraud category based on the intra-class semantic features, and then excludes the interference of similar categories through the inter-class difference features. Finally, the probability that the text belongs to each fraud category is output. According to the maximum probability, the fraud category corresponding to the text is determined, and if the probability of all categories is lower than a preset threshold, it is determined as non-fraud. The output classification result can directly support anti-fraud governance. It is clear whether there is fraud between the customer and the business in the text to be classified and the specific fraud category, which provides decision basis for subsequent risk control. At the same time, because the model has small sample generalization ability, when facing new fraud texts, only a small amount of samples need to be supplemented to update the features, so that the model can quickly adapt to new scenes.

[0056] In the embodiments of the present application, the logic of extracting key features through a small sample framework, constructing an adaptive model through feature fusion, and realizing accurate classification through model application solves the problems of traditional anti-fraud classification, such as dependence on large-scale labeled data and weak generalization in new scenes.

[0057] Optionally, before step 11, the method further comprises:

[0058] For any text, the user vocabulary in the text is determined as the head entity, and the business vocabulary is determined as the tail entity.

[0059] An abnormal behavior category between the head entity and the tail entity is used to generate a structured triple containing the head entity, the abnormal behavior category, and the tail entity.

[0060] In the embodiments of the present application, the anti-fraud classification task is mainly focused on identifying the category of abnormal behavior between the customer and the mobile service contained in the text data. For example, given the text "the user is a high-value customer in city X, and has frequently applied for a new number in the past two months, suspected of malicious card opening, and needs to be monitored for behavior", in this sentence, the object of attention is the user, and the business involved is the application for a new number. The anti-fraud classification task needs to capture the text segment "suspected of malicious card opening", and then determine that the category of abnormal behavior between the customer and the business in the text is "repeated card opening". Therefore, the present application further increases the constraint condition on the basis of the classification task, and refines the anti-fraud identification into a relation extraction task. Specifically, the above text is abstracted, and for any text, the "customer" in the text is marked as a head entity (Sub), and the "business" is marked as a tail entity (Obj). The goal is to mine the relationship between the head and tail entities in the text, i.e., the abnormal behavior category, and generate a structured triple <customer, abnormal behavior category, business>.

[0061] Specifically, in order to adapt to the inherent training mode of few-shot learning, the present application uses an N-way-K-shot setting to construct data training sets, validation sets, and test sets. The model under the N-way-K-shot setting uses a single task instead of a single data as a training step, and each task contains a support set S and a query set Q. Specifically, the support set contains N different relationship classes, each of which represents an abnormal behavior type, such as abnormal machine replacement, repeated card opening, card cancellation, etc. Each class has K instances, which are annotated text data belonging to the current category. The task goal is to predict the abnormal behavior type contained in the text data in the query set according to the support set. For the training process, n (n∈N) x k (k∈K) instances are randomly selected as the support set S={ ;i=1,…,N,k=1,…,K}, representing the kth instance of the ith relationship class in the support set. Then, M samples are extracted from the remaining instances of the N relationship classes to construct the query set Q={ ;m=1,…,M}, representing the mth instance in the query set. Different training steps may have the same relationship, but the instances cannot be repeated. Few-shot learning requires that each category in the support set contains very few instances (5-10) to get rid of the dependence on large-scale labeled data; the model test also follows the N-way-K-shot setting, but the relationship categories in the test set are not allowed to appear in the training set to improve the generalization ability of the model in new anti-fraud classification scenarios.

[0062] Optionally, the input to the model in this application includes three parts: relations in the support set, instances in the support set, and instances in the query set. All three parts of data rely on a large language model for encoding and transformation. Here, the encoding process is illustrated using text data from the support set as an example. The text is abstracted as follows: = { , , …, }, in, It supports centralized ownership of the first The first relation category One instance; yes The first in One word; These represent the head and tail entities, respectively. Subsequently, a large language model is used to encode the text sequence, yielding basic text feature representations. It should be noted that no specific large language model is specified here; it can be adjusted according to available memory and task size. The final encoded representation is as follows: = Encoder({ , , …, }), where Encoder( ) represents a large language model.

[0063] Furthermore, step 11 above includes:

[0064] Based on a pre-defined task framework for few-shot learning, construct intra-class semantic features;

[0065] Inter-class difference learning is performed on the intra-class semantic features to construct inter-class difference features.

[0066] In the embodiments of the present application, relying on the N-way-K-shot preset task framework of small sample learning, each task contains N fraud categories, and only K labeled samples are required for each category. Based on the concentrated text instances, first, the relationship vector for guiding relationship identification is constructed using relationship description information, then the feature vector of the customer (head entity) and the business (tail entity) in the text is extracted, the relationship semantic vector with text context is modeled combining the entity pair association mechanism, finally the attention mechanism is used to fuse the entity, relationship and sentence three-level information, highlight the key text part for classification, form the intra-class semantic features of each fraud category, and capture the core commonality of the same category. In view of the problem that deep similar labels (belonging to the same parent class) in the fraud label system are easy to confuse, first, similar label pairs are analyzed and identified, and the similarity is determined; then, an interaction matrix is constructed for the instance feature vectors of similar categories, the instance-level similarity and difference are calculated, and the class distinction degree is enhanced by comparison and the most different instances are selected, so as to strengthen the unique difference between different fraud categories, complete the inter-class difference learning of intra-class semantic features, generate inter-class difference features, and avoid misjudgment of similar categories.

[0067] Specifically, based on the preset task framework of small sample learning, the intra-class semantic features are constructed, including:

[0068] The relationship vector is generated by using the description information of each relationship category in the support set; the support set is used to represent the labeled sample set for learning the features of abnormal behavior categories in each training, verification or test task;

[0069] Based on the text sequence features, the features belonging to the head entity and the tail entity are extracted according to the pre-embedded entity position markers, and the entity feature vectors with the same dimension are generated;

[0070] The relationship semantic vector between the feature vectors of the head entity and the tail entity is constructed by using the preset translation mechanism and the entity feature vectors corresponding to the head entity and the tail entity;

[0071] The relationship vector, the entity feature vector and the relationship semantic vector are fused by the attention mechanism to determine the intra-class semantic features of the support set.

[0072] In the embodiments of the present application, a support set relationship category description is used to generate a relationship vector, and for each fraud relationship category, such as abnormal machine replacement and repeated card opening, a category identity is established to clearly define the core semantic definition of the category. The support set is a small set of labeled samples for model learning in each task, containing N fraud categories and K samples per category. From the support set, the description information of each relationship category is extracted, such as the description of "abnormal machine replacement" being "the customer changes the bound mobile device multiple times in a short period of time"; the description information is encoded by a text encoding model to convert the literal semantics into a fixed-dimensional vector, i.e., a relationship vector. The vector directly corresponds to the semantic label of the fraud category, avoiding the model relying only on the surface text of the sample and ignoring the essence of the category, and providing clear category semantic guidance for subsequent feature learning.

[0073] An entity position marker is used to extract entity feature vectors to accurately extract the features of the head entity and the tail entity from the support set text, and to clearly define the core focus of the classification. The support set text is first encoded to obtain the overall text sequence features; then, according to the pre-embedded entity position marker, the feature fragments at the corresponding positions are intercepted from the text sequence features; and the feature fragments of the head entity and the tail entity are processed into vectors of the same dimension, i.e., the head entity feature vector and the tail entity feature vector. The core of anti-fraud classification is to judge the relationship between the customer and the business, and this step locks the core object through the entity position marker to ensure that the subsequent feature learning focuses on the customer-business interaction rather than irrelevant text content, and improves the relevance of the features.

[0074] A preset translation mechanism is used to construct a relationship semantic vector to model the associated semantics between the head "entity-tail entity" and capture the fraud behavior features implied in their interaction. The preset translation mechanism is introduced to generate a new vector by performing a preset mathematical operation on the head entity feature vector and the tail entity feature vector; the vector not only contains the entity's own features, but also integrates the interaction semantics between the two, i.e., the relationship semantic vector. This solves the problem of only extracting entity features, which cannot reflect the relationship between entities, and makes the features directly related to the interaction behavior of "customer-business", which meets the core needs of judging the relationship between the two for anti-fraud classification.

[0075] An attention mechanism is used to fuse multiple vectors to determine the intra-class semantic features, integrate the relationship vector, the entity feature vector, and the relationship semantic vector, highlight the key feature information for classification, and form complete intra-class semantic features. The above three types of vectors are input into an attention mechanism module; the attention mechanism calculates the contribution weight of each vector to "distinguish fraud categories" and weights and integrates the three types of vectors; the vector obtained after integration is the intra-class semantic feature of the support set corresponding to the fraud category. This avoids the feature redundancy or key information being covered caused by different vectors having the same weight, and makes the intra-class semantic features contain category definitions, core objects, and interaction semantics, comprehensively and accurately reflecting the common features of the same fraud category.

[0076] Optionally, in the model training of the preset task framework based on small sample learning, each training step corresponds to an independent task including the support set and the query set.

[0077] The relationship classes in different training steps are the same or different, and the corresponding instances are non-repeated.

[0078] In the embodiment of the application, the small sample anti-fraud classification model training is through independent task division and instance non-repeated constraint, so that the model learns the general fraud class feature extraction ability under a small amount of samples, rather than memorizing the samples. Each training step corresponds to an independent task of the support set and the query set. The model training process is divided into multiple continuous independent tasks, and each completed task is a training step. In each task, the data is explicitly divided into a support set and a query set, which jointly serve the class learning and verification in a small sample scenario.

[0079] The support set is used to represent a set of labeled samples for teaching the model to learn class features in each task, which follows the N-way-K-shot setting (including N different fraud relationship classes, and K labeled instances for each class, K is usually 5-10), for example, the support set of a certain task includes 2 classes of abnormal machine change and repeated card opening (N=2), and each class has 5 labeled texts (K=5).

[0080] The query set is used to represent a sample set for verifying the learning effect of the model in each task, which is extracted from the remaining instances of the N classes to which the support set belongs, for example, 3 instances are extracted from the non-support set samples of abnormal machine change and repeated card opening, and the model needs to predict the class of the query set sample based on the features learned from the support set to complete a task learning. The query set learns features through the support set and verifies the effect of the query set, which simulates a real small sample scenario, and only a small amount of labeled samples are needed for classification, so that the model focuses on learning the class rule from a small amount of samples at each step, rather than relying on large-scale data stacking.

[0081] The application allows different training steps to include the same relationship class, such as training step 1 and training step 3 both including the abnormal machine change class, the purpose is to let the model contact different instances of the same class multiple times, and strengthen the learning of the core semantics of the class, such as the expression difference of abnormal machine change in different texts, to avoid insufficient single learning. Different training steps are allowed to include different relationship classes, such as training step 1 including abnormal machine change and repeated card opening, and training step 2 including frequent card cancellation and illegal transfer, the purpose is to let the model contact more fraud classes and learn the general class feature extraction logic, rather than just adapt to a specific class.

[0082] Whether the relationship classes of different training steps are the same or not, the text instances (labeled samples) used must not be repeated. Instance non-repetition can avoid the model memorizing samples, i.e., only remembering the seen instances, rather than learning the category rules, ensuring that the model learns based on new samples each time, and truly mastering the ability to identify categories from a small number of new samples, laying the foundation for subsequent generalization to new fraud scenarios; the same or different relationship classes balance category deep learning and category breadth coverage, ensuring that the model fully understands the core categories, and avoiding the model being limited to a few categories, improving the adaptability to a variety of fraud scenarios.

[0083] It should be noted that the text features modeled by the encoder can be directly used for anti-fraud classification, however, this simple modeling method cannot highlight the important features in each text that can reflect fraudulent behavior. Taking the above text as an example, the focus is on the semantic feature learning of the segment "frequently applies for new numbers, suspected of malicious card opening, and needs to be monitored for behavior." Specifically, the present application proposes a relationship representation method based on intra-class semantic learning, and the specific implementation steps are as follows: 1. Relationship description vector construction: Use the description information of each relationship category in the support set to construct a relationship vector. This vector is used to guide the discovery of relationship information in the text. 2. Entity representation modeling: According to the pre-embedded entity position markers, the features belonging to the head entity and the tail entity are extracted from the text sequence features obtained through data encoding, generating feature vectors with the same dimension. 3. Relationship semantic vector construction: Combine the "translation" mechanism r≈{ - }, indirectly model the relationship representation through entity pair feature vectors, from the perspective of the association between the head and tail entities, and model the relationship semantic vector with the current text context. 4. Multi-level information fusion: Consider the information at the entity level, relationship level and sentence level, and realize relationship representation learning of the instance through the attention mechanism, highlighting the most important part of the sentence for relationship representation. For the i-th instance belonging to the j-th relationship category in the support set, the present application uses the following method to obtain its representation vector:

[0084] = ×( - );

[0085] =softmax(tanh( · ));

[0086] = · ;​​

[0087] wherein, is an n-row matrix, according to the above formula, for the i-th relation class in the support set, the intra-class representation is: = , , ... }; repeat the above steps to obtain the intra-class feature representation of the entire support set: = , , ... }.

[0088] wherein, i represents the number of the relation class in the support set, such as the i-th fraud relation class, the value range is 1, 2,..., N, N is the total number of relation classes contained in the support set. k represents the number of samples under the i-th relation class, such as the k-th labeled text instance under the class, the value range is 1, 2,..., K, K is the number of samples of each relation class, that is, K in K-shot. n represents the dimension or position index of the feature vector or matrix, such as the n-th token in the text sequence, or the n-th dimension of the feature vector. represents the text sequence feature of the k-th sample in the i-th relation class, which contains the complete semantic information of the sample. represents the head entity feature vector of the k-th sample in the i-th relation class. represents the tail entity feature vector of the k-th sample in the i-th relation class. represents the relation vector of the i-th relation class. represents the entity relation adjustment feature of the k-th sample in the i-th relation class. It is obtained by multiplying the text sequence feature Ski and the difference between the tail entity-head entity feature, which is used to strengthen the interaction information of “customer-business” in the text feature. represents the attention weight of the n-th dimension or position. It is obtained by taking the dot product of the adjustment feature and the relation vector , and then doing softmax normalization after tanh activation, which is used to measure the importance of different positions in the text sequence to the “i-th relation class”. represents the n-th component of the adjustment feature . represents the text sequence length (or feature dimension) of the k-th sample in the i-th relation class, that is, contains the total number of components. represents the attention fusion feature of the k-th sample in the i-th relation class, which is obtained by weighting and summing each component of the adjustment feature with the attention weight , highlighting the most critical text information for the relation class. an intra-class representation set representing an i-th relation category, composed of attention fusion features of all samples under the category, , reflecting common semantic features of all samples in the category. an intra-class feature representation set representing the entire support set, composed of intra-class representations of all relation categories, , providing a complete intra-class feature basis for subsequent anti-fraud classification.

[0089] Optionally, inter-class difference learning is performed on the intra-class semantic features to construct inter-class difference features, including:

[0090] For a target relation category, an instance feature vector corresponding to the target relation category is extracted from the intra-class semantic features, and an interaction matrix containing instance indexes is established; the indexes in the interaction matrix correspond to the positions of instances in similar categories of the target relation category; the instances in the interaction matrix are constructed based on the instance features of similar categories other than the target relation category in the intra-class semantic features;

[0091] According to the interaction matrix, interaction features between the target relation category and similar categories in the support set are constructed;

[0092] Inter-class difference learning is performed on the interaction features, the difference features of each category are aggregated, the sentence with the largest information quantity in the target relation category is selected, and an inter-class difference representation is generated;

[0093] According to the inter-class difference representation, an inter-class difference feature of the target relation category is constructed.

[0094] It should be noted that through intra-class feature learning, the present application realizes the capture of instance-level abnormal information. However, in actual research and development, as the fraud label system is continuously improved, there is usually a high similarity between deep-level labels due to the existence of the same parent category label, which needs to be specially paid attention to in order to avoid false classification. Intra-class feature learning only focuses on the representation learning of the category itself and does not process this label similarity scenario, which inevitably causes false recognition results and affects the effectiveness of anti-fraud governance. Therefore, the present application proposes inter-class difference learning to further capture the diversity features between different categories. In order to obtain more fine-grained information, the present application adopts instance-level similarity and difference learning.

[0095] ​​In the embodiments of the present application, the target relationship category is used to represent a specific fraudulent behavior or normal behavior category that needs to be focused on and analyzed in detail in the anti-fraud task scenario. For example, if the focus is on studying the fraud type of "suspected malicious card opening" at this moment, "suspected malicious card opening" is the target relationship category. After the target category is determined, the instance feature vectors of all samples of the target category are extracted from the in-class semantic features. Then, other fraud or normal categories similar in semantics to the target category are found, and the features of the instances of these similar categories are taken into account to establish an interaction matrix according to the position information of their respective instances. The index of the matrix corresponds to the position of the instance of the similar category, and the element in the matrix is the instance feature of the corresponding similar category. This matrix builds the basis for the interaction of the features of the target category and the similar categories. According to the generated interaction matrix, the correlation and differences between the features of the target relationship category and the similar categories are mined through matrix operations, feature fusion, etc. For example, the dot product and Euclidean distance between the elements of the matrix can be calculated, or a specific neural network layer can be used to convolve the matrix to obtain a new set of feature vectors. This set of features can show the interaction between the target category and the similar categories, which is the interaction feature.

[0096] For the interaction feature obtained in the above step, a deep learning training mechanism such as back propagation optimization is used to carry out inter-class difference learning, so that the model can clearly distinguish the differences between the target category and the similar categories. After learning, the features of each category in terms of differences are aggregated, which can be directly superimposed or weighted summed. Then, through information entropy calculation, sentence importance scoring, etc., the sentences that best reflect the differences between the target category and other categories, i.e., the sentences with the largest amount of information, are selected from the text corresponding to the target relationship category samples, and a compact vector representation is formed based on these information, which is the inter-class difference representation.

[0097] Finally, based on the generated inter-class difference representation, further feature extraction, transformation, etc. are performed, such as passing through a fully connected layer, normalizing the vector, etc., to output a set of refined features with strong discrimination ability, which is the inter-class difference feature of the target relationship category. It can make the model more accurately divide the target category and other categories in subsequent anti-fraud classification and judgment.

[0098] Specifically, first, label similarity analysis is performed. For the similarity problem between deep-level labels in the fraud label system, label pairs with the same parent category label are analyzed and identified to determine their similarity degree.

[0099] Furthermore, instance-level interaction modeling is performed. Interaction modeling is conducted for all sentence pairs within each category that exhibit high similarity, calculating instance-level similarity and dissimilarity. Specific steps include: feature extraction for each sentence pair, calculation of similarity scores between sentence pairs, and further calculation of dissimilarity features based on the similarity scores. Technically, for the i-th class, feature vectors of relevant class instances are extracted, and then an interaction matrix is ​​constructed. ,in Indicates the first The first in the class The index of each instance. The instances in the interaction matrix are represented as follows: = .in, This represents the set of intra-class feature representations of the entire support set obtained in the preceding steps; (N−1,K) denotes the dimensional constraint, representing the range from... After removing the intra-class features of the target relation category i, the remaining set of intra-class features of N-1 similar categories is obtained, and each similar category still contains features of K instances. The essence is a set matrix of intra-class features of similar categories. The rows of the matrix correspond to the similar category number j, the columns correspond to the instance number k within the similar category, and the elements are the intra-class fusion features of the kth instance of the jth category, providing a feature carrier for subsequent interaction modeling between the target class and similar classes.

[0100] Subsequently, modeling supports the centralization of the first Interactions between one category and other similar categories: = · ;in, The intra-class feature representation of the target relation category i; Representing instance-level interaction features between target relation category i and all similar categories, through (Target class features) and The dot product operation of (the representation of instances in the interaction matrix) is essentially used to quantify the semantic association between each instance of the target class and each instance of the similar class. The dimension is [K, (N−1)×K] (each row corresponds to one instance of the target class, each column corresponds to one instance of the similar class, and the elements are the association values ​​between the two). =min(1, ); where min(1,t) represents selecting the smallest element in the second dimension of tensor t. This application uses a minimum function to select the most distinct instance in each class, avoiding excessive confusion between similar classes; where, indicates the filtered difference instance features, each target class instance corresponds to an instance feature with the minimum association degree with the similar class, the smaller the association degree, the greater the difference between the similar class instance and the target class instance, through the function, the instance that can best reflect the difference between the target class and the similar class can be filtered out, avoiding excessive confusion between similar classes, and the dimension is [K, 1].

[0101] Finally, inter-class difference learning is performed. Based on the similarity and difference calculation results at the instance level, the diversity features between different classes are captured. The specific steps include: aggregating the difference features of each class, enhancing the distinguishability between different classes through comparison, and generating inter-class difference representation by selecting the sentence with the largest information quantity in the i th relationship class: . . = max(0, ). Wherein, indicates the attention fusion feature of the k th instance in the target relationship class i; indicates the average feature representation of the target relationship class i, which is obtained by taking the arithmetic mean of the fusion features of the K instances in the class, and the essence is to extract the core common semantic of the target class, providing the target class benchmark feature for subsequent difference comparison. The symbol “;” represents the feature splicing operation; indicates the preliminary inter-class difference feature set, which contains the core common feature of the target class and the most different instance feature of the target class and the similar class, providing complete input for subsequent enhancement of class distinguishability. The function max(0, t) represents the ReLU activation function, which filters the negative difference values in the feature and only retains the non-negative effective difference values; indicates the final inter-class difference feature of the target relationship class i, the dimension is consistent with , and the feature only retains the target class commonness and “target-similar class effective difference”, which can be directly used for subsequent anti-fraud classification model, and enhances the distinguishability of the model to similar classes.

[0102] The above steps are repeated to obtain the inter-class difference vector of each class.

[0103] Optionally, the step 13 described above comprises:

[0104] Splicing and fusing the intra-class semantic features and the inter-class difference features to determine the fusion features;

[0105] Based on the preset query set formula and the fusion features, determining the probability that the query set belongs to each relationship class in the support set;

[0106] ​​​When training and optimizing a preset training model using the cross-entropy loss function, the loss value is calculated based on the probability. The model parameters of the preset training model are adjusted by minimizing the loss value. During the model training process, an optimizer is used to optimize the model parameters to determine the anti-fraud classification network model.

[0107] In this embodiment of the application, for the class-level semantic reconstruction of fraudulent text data, intra-class semantic features and inter-class difference features are concatenated to determine fusion features, which are used to represent the final abnormal behavior category representation: = ; This indicates a fusion feature that supports the i-th relation class in the set (such as a specific abnormal behavior category); This indicates the intra-class feature representation supporting the i-th relation class in the set; This represents the inter-class difference feature representation of the i-th relation class in the support set.

[0108] Furthermore, the preset query set formula can be expressed as: ={ - ); };in, Represents the first in the query set The basic semantic feature vector of each instance is a feature representation directly generated by the large language model without any additional processing. This indicates the m-th instance in the query set (denoted as m). ) target feature representation; Represents a query instance Feature vectors of mid-head entities; Indicates a query instance The feature vector of the mid-tail entity. - The value represents the difference between the tail entity feature vector and the head entity feature vector, i.e., the semantic difference feature between entities.

[0109] The objective of this application is to analyze customer data to determine whether abnormal behavior exists between customers and mobile services in the text, and the specific categories of such abnormal behavior, thereby supporting anti-fraud governance. Since the categories of abnormal behavior and the objects of interest in the text have been identified, this application abstracts this process into a multi-label relationship extraction task, and trains it using few-shot learning. Specifically, based on a preset query set formula and the fusion features, the probability that the query set belongs to each relationship class in the support set is determined. It can be expressed by the formula:

[0110] ;

[0111] in, denotes the probability that the query instance belongs to the i-th relation class in the support set, which is the core result of the final output of the formula. denotes the query instance to be classified, i.e. the text data that needs to be predicted by the model in the small sample task. In the N-way-K-shot task setting, the query set is extracted from the remaining instances of the N relation classes to which the support set belongs, and m is the number of query instances in the query set, such as the mth text to be predicted. denotes the feature representation vector of the query instance , which is obtained after the same feature processing procedure is performed on the query instance text as the support set instance. i or n denotes the number of the relation class in the support set, and the value range is 1, 2,..., N, N is the total number of relation classes contained in the support set, i.e. “N” in N-way-K-shot, corresponding to N kinds of abnormal behavior categories in the anti-fraud scene, such as “abnormal machine change” and “repeated card opening”. i represents the target relation class number corresponding to the probability being calculated. n represents the loop variable for traversing all relation classes in the support set, which is used to calculate the probability normalization term of all classes in the denominator. or denotes the class feature representation vector of the i-th or n-th relation class in the support set, which is the aggregation result of the features of all instances in the class, and is used to represent the core semantics of the entire class.

[0112] Subsequently, the cross-entropy loss function is used for model training and optimization: .

[0113] The model uses the AdamW optimizer to optimize the hyperparameters in the model. There are two data scenarios: 5-way-5-shot and 10-way-5-shot. During the training process, the present application uses BERT as the encoder to improve the extraction accuracy of the model by minimizing the loss function, and all hyperparameter tuning is completed on the data validation set. Specifically, the training step of the model iteration is set to 30000; the learning rate (Learning Rate), linear warm-up (Linearwarm-up), temperature coefficient (Temperature) and other parameters of the model are fine-tuned according to the performance of the model in the gradient update process.

[0114] To sum up, the application constructs a class-level semantic reconstruction network model. In the class semantic feature learning stage, the text key features are mined by combining the class features with the translation mechanism to generate the representation vector of each fraud class. In the inter-class difference learning stage, the similarity between instances is modeled in a fine-grained manner to enhance the model's ability to identify similar fraud labels and improve the model's robustness, thereby improving the fraud governance effect. The anti-fraud classification is abstracted as a relation extraction task, and the algorithm and large language model are fine-tuned on professional data to make up for the lack of professional knowledge of general large language models, which can effectively enhance the professional ability of the model and improve the data analysis efficiency.

[0115] The application proposes a new network model to implement anti-fraud classification, which improves the identification ability and robustness of similar abnormal behavior classes by mining the semantic features of the class itself and the difference features between classes, and enables anti-fraud governance. By training the model through small sample learning, compared with the model based on machine learning and supervised deep learning, the dependence of the model on large-scale data is effectively alleviated, and through the iteration mode with task as the minimum training step, the anti-fraud classification with high generalization ability is realized when facing new fraud classes.

[0116] The above introduces various methods of embodiments of the application. The following will further provide a device for implementing the above method.

[0117] Please refer to Figure 2 The embodiment of the application provides a small sample anti-fraud classification device, which comprises:

[0118] The first construction module 21 is configured to construct the intra-class semantic features and the inter-class difference features based on the preset task framework of small sample learning.

[0119] The first determination module 22 is configured to determine the network model of anti-fraud classification according to the intra-class semantic features and the inter-class difference features.

[0120] The first processing module 23 is configured to perform anti-fraud classification and identification on the text to be classified according to the network model.

[0121] Optionally, the first construction module 21 comprises:

[0122] The first construction unit is configured to construct the intra-class semantic features based on the preset task framework of small sample learning.

[0123] The second construction unit is configured to perform inter-class difference learning on the intra-class semantic features to construct the inter-class difference features.

[0124] Optionally, the first construction unit is specifically configured to:

[0125] Generate a relation vector by using support set which supports description information of each relation category; the support set is used to represent a set of labeled samples for learning features of abnormal behavior categories in each training, verification or test task;

[0126] Based on the text sequence features, the features of the head entity and the tail entity are extracted according to the pre-embedded entity position markers to generate entity feature vectors with the same dimension;

[0127] A relationship semantic vector between the feature vectors of the head entity and the tail entity is constructed by using a preset translation mechanism and the entity feature vectors corresponding to the head entity and the tail entity.

[0128] The relation vector, the entity feature vector and the relationship semantic vector are fused by an attention mechanism to determine the intra-class semantic features of the support set.

[0129] Optionally, in the model training based on the preset task framework of small sample learning, each training step corresponds to an independent task including the support set and the query set.

[0130] The relation categories in different training steps are the same or different, and the corresponding instances are non-repetitive.

[0131] Optionally, the second construction unit is specifically used for:

[0132] For a target relation category, an instance feature vector corresponding to the target relation category is extracted from the intra-class semantic features, and an interaction matrix including instance indexes is established; the indexes in the interaction matrix correspond to the positions of instances in similar categories of the target relation category; the instances in the interaction matrix are represented based on the instance features of similar categories other than the target relation category in the intra-class semantic features.

[0133] According to the interaction matrix, an interaction feature between the target relation category and similar categories in the support set is constructed.

[0134] The inter-class difference learning is performed on the interaction feature, the difference features of each category are aggregated, the sentence with the largest information amount in the target relation category is selected, and an inter-class difference representation is generated.

[0135] According to the inter-class difference representation, an inter-class difference feature of the target relation category is constructed.

[0136] Optionally, the first determination module 22 includes:

[0137] A first determination unit is configured to splice and fuse the intra-class semantic features and the inter-class difference features to determine a fused feature.

[0138] The second determining unit is configured to determine a probability that the query set belongs to each relationship class in the support set based on the preset query set formula and the fused features.

[0139] The first processing unit is configured to calculate a loss value based on the probability when training and optimizing the preset training model by using a cross-entropy loss function, adjust model parameters of the preset training model by minimizing the loss value, and optimize the model parameters by using an optimizer in a model training process to determine a network model for anti-fraud classification.

[0140] Optionally, the device of the present application further comprises:

[0141] The second processing module is configured to determine a user vocabulary in a text as a head entity and a business vocabulary as a tail entity for any text.

[0142] The third processing module is configured to generate a structured triple containing the head entity, the abnormal behavior class, and the tail entity by using the abnormal behavior class between the head entity and the tail entity.

[0143] It should be noted that the device in this embodiment corresponds to the small sample anti-fraud classification method described above, and the implementation modes in the above embodiments are applicable to the embodiments of the device and can achieve the same technical effects. The device described above provided in the embodiments of the present application can implement all method steps achieved by the method embodiments and can achieve the same technical effects. Therefore, the same parts and beneficial effects in the method embodiments will not be described in detail here.

[0144] The embodiments of the present application further provide a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement each process of the small sample anti-fraud classification method embodiments described above and can achieve the same technical effects. To avoid repetition, the computer readable storage medium will not be described here. The computer readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0145] The embodiments of the present application further provide a computer program product, which includes computer instructions. The computer instructions are executed by a processor to implement each process of the small sample anti-fraud classification method embodiments described above and can achieve the same technical effects. To avoid repetition, the computer program product will not be described here.

[0146] It should be noted that, in the present document, the terms "comprises / comprising" or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises... a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0147] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and a necessary general hardware platform, and of course, they can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product in essence or in the form of a part that contributes to the prior art. The computer software product is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk) and includes a plurality of instructions for causing a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device) to execute the methods described in the various embodiments of the present application.

[0148] The embodiments of the present application are described above in combination with the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative rather than limiting. Those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope of protection of the claims.

Claims

1. A small-sample anti-fraud classification method, characterized in that, include: Based on a pre-defined task framework for few-shot learning, we construct intra-class semantic features and inter-class differential features. Based on the intra-class semantic features and the inter-class difference features, a network model for anti-fraud classification is determined; Based on the network model, anti-fraud classification and recognition are performed on the text to be classified.

2. The method according to claim 1, characterized in that, Based on a pre-defined task framework for few-shot learning, intra-class semantic features and inter-class differential features are constructed, including: Based on a pre-defined task framework for few-shot learning, in-class semantic features are constructed. Inter-class difference learning is performed on the intra-class semantic features to construct inter-class difference features.

3. The method according to claim 2, characterized in that, Based on a pre-defined task framework for few-shot learning, intra-class semantic features are constructed, including: Relation vectors are generated using the descriptive information of each relation category in the support set; the support set is used to represent the set of labeled samples used by the model to learn the features of abnormal behavior categories in each training, validation or testing task. Based on text sequence features, and according to pre-embedded entity location markers, features belonging to the head entity and the tail entity are extracted respectively, and entity feature vectors of the same dimension are generated. Using a preset translation mechanism, a semantic vector relating the feature vectors of the head entity and the tail entity is constructed by using the entity feature vectors corresponding to the head entity and the tail entity. The relation vector, the entity feature vector, and the relation semantic vector are fused using an attention mechanism to determine the intra-class semantic features of the support set.

4. The method according to claim 3, characterized in that, In the model training based on the preset task framework of few-shot learning, each training step corresponds to an independent task containing the support set and query set. The relation classes in different training steps are either the same or different, and the corresponding instances are non-repeating.

5. The method according to claim 2, characterized in that, Inter-class difference learning is performed on the intra-class semantic features to construct inter-class difference features, including: For a target relation category, an instance feature vector corresponding to the target relation category is extracted from the intra-class semantic features, and an interaction matrix containing instance indices is established; the indices in the interaction matrix correspond to the positions of each instance in the similar categories of the target relation category; the instance representations in the interaction matrix are constructed based on the instance features of similar categories other than the target relation category in the intra-class semantic features. Based on the interaction matrix, construct interaction features that support the target relationship category and similar categories in the set; Inter-class difference learning is performed on the interaction features, the difference features of each category are aggregated, and the sentence with the most information in the target relationship category is selected to generate an inter-class difference representation; Based on the inter-class difference representation, construct the inter-class difference features of the target relationship category.

6. The method according to claim 2, characterized in that, Based on the intra-class semantic features and the inter-class difference features, a network model for anti-fraud classification is determined, including: The intra-class semantic features and the inter-class difference features are concatenated and fused to determine the fused features; Based on the preset query set formula and the fusion features, the probability that the query set belongs to each relation class in the support set is determined; When training and optimizing a preset training model using the cross-entropy loss function, the loss value is calculated based on the probability. The model parameters of the preset training model are adjusted by minimizing the loss value. During the model training process, an optimizer is used to optimize the model parameters to determine the anti-fraud classification network model.

7. The method according to claim 1, characterized in that, Based on a pre-defined task framework for few-shot learning, before constructing intra-class semantic features and inter-class differential features, the method further includes: For any given text, user terms are defined as head entities, and business terms are defined as tail entities. By utilizing the abnormal behavior categories between the head entity and the tail entity, a structured triplet containing the head entity, the abnormal behavior category, and the tail entity is generated.

8. A small-sample anti-fraud classification device, characterized in that, include: The first building module is used to construct intra-class semantic features and inter-class differential features based on a pre-defined task framework for few-shot learning. The first determining module is used to determine the network model for anti-fraud classification based on the intra-class semantic features and the inter-class difference features; The first processing module is used to perform anti-fraud classification and recognition on the text to be classified based on the network model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images