Bank suspicious transaction feature automatic identification method and device

Abstract

The invention discloses a bank suspicious transaction feature automatic identification method and device. Bank transaction records are collected, standardization processing is carried out on texts in the records, and the processing content comprises common transaction types such as transfer, consumption and ATM withdrawal. Then, word segmentation processing is carried out on the standardized text, key information vocabularies are extracted, a vocabulary list is constructed, and the vocabulary list is used for subsequent Word2Vec model training; the Word2Vec model adopts a Skip-Gram architecture, word vectors are trained and generated by predicting context words, the semantic relation between transaction feature vocabularies is captured, and the model calculation amount is reduced by adopting a negative sampling method. In the training process, a maximum log-likelihood function is used as a loss function, and model parameters are optimized. After training is completed, the recognition precision of the model on suspicious transaction features is evaluated by verifying the generated word vector, and model parameters are further adjusted according to a verification result, so that a more accurate suspicious transaction recognition model is obtained.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method and device for automatically identifying suspicious transaction features of a bank. Background Art

[0002] In the existing technology, banks detect suspicious transaction behaviors by monitoring and analyzing transaction records, which usually includes analyzing factors such as transaction amount, frequency, time and geographical location. Natural language processing technology, especially technology based on word vector models, has been gradually applied to the financial field to identify and extract features related to text descriptions in transaction records. As a commonly used word vector generation method, the Word2Vec model can capture the semantic relationship between words by training large-scale text data, and shows high accuracy in feature extraction and pattern recognition.

[0003] However, the natural language processing models and methods used in the prior art still have some shortcomings. First, the lack of specific processing of financial-specific vocabulary in the standardization and word segmentation process leads to inaccurate extraction of transaction features. Second, the model has low training efficiency when processing massive transaction data, especially when it comes to vocabulary size and context window settings, the computational complexity of the model will increase significantly. In addition, the model lacks a dynamic adjustment mechanism, and the parameters are not optimized in time according to the model performance during the verification phase, which limits the adaptability and flexibility of the model in different transaction scenarios.

[0004] Therefore, there is an urgent need to provide an improved automatic identification method for suspicious transaction features in banks to improve the accuracy of identifying suspicious transaction features. Summary of the invention

[0005] The present application provides a method and device for automatically identifying suspicious transaction features of a bank, so as to improve the recognition accuracy of suspicious transaction features.

[0006] The present application provides a method for automatically identifying suspicious transaction characteristics of a bank, comprising: Collect bank transaction records and standardize the text in the transaction records; wherein the transaction records include transfers, consumption, ATM withdrawals, counter withdrawals, financial management purchases, precious metal purchases, and loan records; the standardization includes standardizing the date format, normalizing the amount unit, and removing non-text symbols; Performing word segmentation on the text in the standardized transaction records to extract key information words; constructing a vocabulary based on the key information words; wherein the vocabulary includes words reflecting transaction characteristics for subsequent model training; The Word2Vec model is trained using the word segmentation results of the text in the transaction record and the constructed vocabulary. The Word2Vec model is implemented using the Skip-Gram model architecture. For a central word in a given transaction record text, the Skip-Gram model is trained to generate a word vector by predicting its context words to capture the semantic relationship between words. The probability of occurrence of each context word is calculated according to the following formula (1):

[0007] in, Indicates the central word, represents the context word, where Indicates the position of the central word in the sentence. Represents context-dependent indexes; Representing context words The word vector of Indicates the central word The word vector of A word vector representing each word in the vocabulary; is the size of the vocabulary; The Word2Vec model is trained using a negative sampling method, wherein the negative sampling method randomly selects words that are not related to the central word as negative samples to reduce the amount of calculation required for the model to predict all words. The loss function provided by the following formula (2) is used during the training process:

[0008] in, represents the loss function, which is used to measure the prediction error of the model during the training process; Represents the total number of samples in the training data; Indicates the window size; Representing context words The word vector of Indicates the central word The word vector of Represents the total number of negative samples; Represents the distribution of negative samples Randomly sample negative samples The expected value of probability; Represents negative samples The word vector of is the sigmoid function; After training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features; According to the verification results of the Word2Vec model, the model parameters of the Word2Vec model are adjusted to obtain an adjusted Word2Vec model that can identify the characteristics of suspicious bank transactions.

[0009] Furthermore, the model parameters include vocabulary size, word vector dimension, context window size, number of negative samples, and learning rate.

[0010] Furthermore, after generating word vectors through the Word2Vec model and verifying the model recognition accuracy, the Spearman coefficient is used to evaluate the adjustment effect of the model parameters, where the Spearman coefficient is calculated according to the following formula (3):

[0011] in, represents the Spearman correlation coefficient, which is used to measure the correlation of model performance before and after model parameter adjustment; represents the square of the difference between the rankings of two parameters; Indicates the training round.

[0012] Furthermore, the text in the standardized transaction record is segmented to extract key information words, including: Use natural language processing tools to perform grammatical analysis on transaction record text, identify the part of speech of each word, and capture the central word and its related context words based on the context window technology to ensure that each word has accurate semantic association in its context; Dynamically filter out irrelevant or noisy words based on the contextual words in the transaction scenario; Combined with the characteristic vocabulary of financial transactions, the words related to capital flow are extracted first after vocabulary filtering; Based on the high-risk characteristic words in the historical suspicious transaction records, the extracted funds flow-related words are weighted to enhance their importance in subsequent model training.

[0013] Furthermore, the weighted processing of the extracted fund flow related words based on the high-risk characteristic words in the historical suspicious transaction records includes: According to the following formula (4), the historical data weights of fund flow related words are calculated. :

[0014] in, Words related to fund flow The number of times the transaction appears in historical suspicious transaction records; Indicates the total number of words in the historical suspicious transaction records; Based on the risk level defined internally by the bank, the modified vocabulary weight is calculated using the following formula (5): :

[0015] in, is the risk correction factor; Vocabulary related to the flow of funds within the bank The assigned risk level, with risk level values ​​ranging from 0 to 1, representing a degree of risk from low risk to high risk; Combined with the time features in the transaction records, the vocabulary weight is further adjusted according to the following formula (6) to obtain the time decay weight :

[0016] in, Indicates the timestamp of the current transaction; The reference time point includes high-risk time periods when suspicious transactions are concentrated; is the time decay coefficient, which is used to measure the effect of time on the importance of vocabulary; After considering historical weight, risk correction and time decay, the final vocabulary weight is calculated using the following formula (7): :

[0017] in: is the adjustment factor for controlling risk and time; It is a correction term used to control the smoothing effect in certain situations and prevent the weight from being too high; is a nonlinear adjustment index used to adjust the weighting effect in extreme cases.

[0018] Furthermore, after the training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features, including: By comparing the word vectors generated by the model with the word vectors of known suspicious transaction feature words, the cosine similarity between the two is calculated to evaluate the recognition accuracy of the model.

[0019] Furthermore, after the training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features, which also includes: If the cosine similarity between the generated word vector and the word vector of the known suspicious transaction feature vocabulary exceeds the first threshold, the model is judged to have a high recognition accuracy; if the cosine similarity is lower than the second threshold, the model parameter adjustment mechanism is triggered.

[0020] Furthermore, the constructed vocabulary includes at least 5,000 transaction-related words, wherein the transaction-related words include banking terms, transaction pattern words, and descriptive words of potential suspicious transaction behaviors.

[0021] Furthermore, the vocabulary is dynamically updated according to the frequency of historical transaction data, and high-frequency transaction feature vocabulary is given priority during each training to ensure that the model has good adaptability to the latest transaction patterns.

[0022] The present application also provides a device for automatically identifying suspicious transaction characteristics of a bank, comprising: A collection unit is used to collect bank transaction records and perform standardization on the text in the transaction records; wherein the transaction records include transfers, consumption, ATM withdrawals, counter withdrawals, purchases of financial products, purchases of precious metals and loan records; the standardization includes standardization of date formats, standardization of amount units and removal of non-text symbols; A processing unit, configured to perform word segmentation processing on the text in the standardized transaction record to extract key information words; and construct a vocabulary based on the key information words; wherein the vocabulary includes words reflecting transaction characteristics for subsequent model training; A construction unit is used to train a Word2Vec model using a word segmentation processing result of the text in the transaction record in combination with the constructed vocabulary; wherein the Word2Vec model is implemented using a Skip-Gram model architecture. For a central word in a given transaction record text, the Skip-Gram model is trained to generate a word vector by predicting its context words to capture the semantic relationship between words; A training unit, used for training the Word2Vec model by using a negative sampling method, wherein the negative sampling method reduces the amount of calculation required for the model to predict all words by randomly selecting words that are not related to the central word as negative samples; The verification unit is used to verify the generated word vectors after training is completed and evaluate the accuracy of the Word2Vec model in identifying transaction features; The adjustment unit is used to adjust the model parameters of the Word2Vec model according to the verification result of the Word2Vec model to obtain an adjusted Word2Vec model that can identify the characteristics of suspicious bank transactions.

[0023] The technical solution proposed in this application has the following beneficial technical effects: (1) By standardizing the text in bank transaction records, including unifying the date format, normalizing the amount unit, and removing non-text symbols, the consistency of transaction data is ensured, the interference of noise data is reduced, and the accuracy of subsequent model recognition is improved.

[0024] (2) By performing word segmentation on the standardized transaction record text and combining it with the constructed vocabulary, the model can accurately extract key information words related to bank transactions, thereby capturing the characteristics that reflect suspicious transaction behaviors and improving the recognition efficiency of the model.

[0025] (3) The Skip-Gram model architecture is combined with the negative sampling method to avoid the model's comprehensive prediction of all words, greatly reducing the amount of calculation, improving the model's training speed and resource utilization, and adapting to the processing needs of large-scale bank transaction data. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of a method for automatically identifying suspicious transaction characteristics of a bank provided in the first embodiment of the present application.

[0027] Figure 2 It is a schematic diagram of a device for automatically identifying suspicious transaction characteristics of a bank provided in the second embodiment of the present application. DETAILED DESCRIPTION

[0028] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0029] The first embodiment of the present application provides a method for automatically identifying suspicious transaction characteristics of a bank. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 A method for automatically identifying suspicious transaction characteristics of a bank is described in detail in the first embodiment of the present application.

[0030] Step S101: Collect bank transaction records and standardize the text in the transaction records; wherein the transaction records include transfers, consumption, ATM withdrawals, counter withdrawals, financial management purchases, precious metal purchases and loan records; the standardization includes standardizing the date format, normalizing the amount unit, and removing non-text symbols.

[0031] In step S101, the bank's transaction records need to be collected first. These transaction records usually include various transaction types, such as transfers, consumption, ATM withdrawals, counter withdrawals, purchases of wealth management products, purchases of precious metals and loan records. Transaction records may contain date, time, amount, transaction type, transaction party account information, etc. These data can come from different bank systems, such as the customer's account management system, the bank's internal transaction record database, etc. Data collection should be as comprehensive as possible to ensure that it covers various types of banking business scenarios. The specific collection method can be derived from the database or regularly obtained through the API interface. The frequency and scope of data collection should be determined according to the design requirements of the system.

[0032] After data collection is completed, these original transaction records must be standardized. The text of transaction records usually has problems such as inconsistent formats and symbols. Directly using these data for processing will result in reduced recognition accuracy of the model. Therefore, the first task of standardization is to ensure the consistency of data formats. First of all, the date information in all transaction records should be unified into the standard "YYYY-MM-DDHH:MM:SS" format, which ensures that all time information can be processed consistently during subsequent processing. For dates that do not conform to the format, the system needs to be able to automatically detect and correct them to ensure that all data conforms to the standard. In addition, if there are different time formats in the transaction records, such as time formats that only contain dates or do not contain seconds, the system should automatically supplement them.

[0033] Normalization of the amount unit is another important step in the standardization process. Different transaction systems may use different currency units or use different currency symbols to represent the same currency. Therefore, all transaction amounts should be uniformly converted into the base currency of the banking system and expressed using a consistent number of decimal places.

[0034] In addition to the standardization of dates and amounts, removing non-text symbols is also a key step in the standardization process. Bank transaction records may often contain various non-text symbols, which are usually meaningless or useless for model training. For example, some transaction notes may contain special characters, emoticons, or too many punctuation marks, which will increase data noise and affect subsequent word segmentation and model training. Therefore, the standardization process must filter or replace these symbols.

[0035] Other special cases may need to be handled during the standardization process. For example, some transaction records may contain mixed languages ​​or abbreviations. In this case, the system should be able to recognize these specific character combinations and perform appropriate processing, such as expanding common abbreviations into complete words, or converting words in different languages ​​into the same language. The system should combine specific banking business scenarios and the language model used to ensure that this processing meets the logical requirements of the business.

[0036] Through the above standardized processing steps, it is possible to ensure that the key information in the transaction records is effectively retained and provide a stable data foundation for subsequent text processing and feature extraction.

[0037] Step S102: Segment the text in the standardized transaction record to extract key information words; construct a vocabulary based on the key information words; wherein the vocabulary includes words that reflect transaction characteristics for subsequent model training.

[0038] In step S102, the core purpose of word segmentation of the standardized bank transaction record text is to extract key information words so as to provide high-quality input data for subsequent model training. First, the starting point of word segmentation is based on the standardized text, which has been formatted and symbol-removed in the previous step S101, so the standardization and consistency of the text content can be ensured. This lays the foundation for the smooth execution of the word segmentation algorithm.

[0039] The specific implementation steps of word segmentation can be completed through existing natural language processing tools, which can include rule-based word segmentation methods or machine learning-based word segmentation algorithms. When processing transaction records, due to the particularity of the financial field, transaction descriptions may contain various types of data such as proper nouns, transaction type abbreviations, account information, time and amount. Therefore, the primary task of word segmentation is to identify and disassemble these mixed text information to ensure that key information words that reflect transaction behavior can be accurately extracted.

[0040] When implementing word segmentation, it is necessary to process specific scenarios of financial transactions. For example, words such as "transfer", "deposit", and "large cash withdrawal" in transaction records are high-frequency words in bank transactions, and the system should be able to identify these words first. This can be achieved by combining a pre-built financial feature vocabulary. The feature vocabulary should include various transaction-related words, especially key words that may indicate suspicious transactions, such as "multiple small transfers" and "cash withdrawals in different places". This kind of word segmentation processing based on domain-specific vocabulary can significantly improve the accuracy of transaction feature extraction.

[0041] In addition, the system needs to have a certain degree of context awareness. Since transaction record texts often contain phrases or sentence structures, simple word segmentation may ignore the semantic relationship between words. Therefore, in the word segmentation process, the system should analyze each word and the words around it in combination with context window technology to capture the association between words. For example, although "distinct" and "large amount" in "distinct large amount cash withdrawal" are independent words, they jointly describe an important transaction feature in the context. This requires the word segmentation process to preserve this association relationship and ensure that the model can recognize the overall meaning of these feature words in subsequent processing.

[0042] After the word segmentation is completed, the system compares the generated vocabulary results with the financial feature vocabulary to screen out key information words. These key information words should be highly representative and can directly reflect the characteristics of transaction behavior. The purpose of this step is to filter out irrelevant and noise words, such as common stop words, punctuation marks or other meaningless characters. In financial transaction scenarios, common words such as "customer" and "bank" are common, but they are not very helpful in identifying transaction behavior characteristics, so they should be regarded as irrelevant words and ignored.

[0043] Once the key information words are extracted, the system will build a vocabulary based on these words. The construction of this vocabulary is not just a simple word list, it needs to include detailed information about each word, such as word frequency, distribution of the word in transaction records, and correlation with other words. With this detailed vocabulary, the system can provide higher quality input data for subsequent model training, so that the model can better capture transaction features based on these key information words, especially when detecting suspicious transactions, it can effectively identify abnormal features that are different from normal transaction patterns.

[0044] When building a vocabulary, you can also consider a dynamic update mechanism. Since bank transaction records will continue to generate new descriptions or features over time, the system should be able to update the vocabulary and its weights in real time based on new transaction data. For example, when some new fraud methods or transaction patterns emerge, new key words may appear in the transaction description. At this time, the system should be able to identify these new words and add them to the vocabulary to improve the model's adaptability to the emerging transaction features.

[0045] Through the above-mentioned word segmentation and vocabulary building process, the system can extract high-quality words that are closely related to transaction features from standardized transaction records, ensuring that subsequent Word2Vec model training can generate accurate word vectors based on these words.

[0046] Furthermore, the text in the standardized transaction record is segmented to extract key information words, including: Use natural language processing tools to perform grammatical analysis on transaction record text, identify the part of speech of each word, and capture the central word and its related context words based on the context window technology to ensure that each word has accurate semantic association in its context; Dynamically filter out irrelevant or noisy words based on the contextual words in the transaction scenario; Combined with the characteristic vocabulary of financial transactions, the words related to capital flow are extracted first after vocabulary filtering; Based on the high-risk characteristic words in the historical suspicious transaction records, the extracted funds flow-related words are weighted to enhance their importance in subsequent model training.

[0047] First, the transaction record text is parsed using natural language processing tools to identify the part of speech of each word. Next, the context window technology is used to capture each central word and its related context words to ensure that each word has accurate semantic associations in its specific context. The system can dynamically filter out irrelevant or noise words by analyzing the context according to the specific transaction scenario, thereby reducing the possibility of misjudgment. Subsequently, combined with the vocabulary of financial transaction characteristics, the words related to capital flow are preferentially extracted, which may better reflect the characteristics and risk behaviors of the transaction. Finally, the system weights the extracted capital flow-related words based on the high-risk characteristic words in the historical suspicious transaction records, so that these words have higher weights in subsequent model training, enhancing the model's ability to identify suspicious transactions.

[0048] Furthermore, the weighted processing of the extracted fund flow related words based on the high-risk characteristic words in the historical suspicious transaction records includes: According to the following formula (4), the historical data weights of fund flow related words are calculated. :

[0049] in, Words related to fund flow The number of times the transaction appears in historical suspicious transaction records; Indicates the total number of words in the historical suspicious transaction records; Based on the risk level defined internally by the bank, the modified vocabulary weight is calculated using the following formula (5): :

[0050] in, is the risk correction factor; Vocabulary related to the flow of funds within the bank The assigned risk level, with risk level values ​​ranging from 0 to 1, representing a degree of risk from low risk to high risk; Combined with the time features in the transaction records, the vocabulary weight is further adjusted according to the following formula (6) to obtain the time decay weight :

[0051] in, Indicates the timestamp of the current transaction; The reference time point includes high-risk time periods when suspicious transactions are concentrated; is the time decay coefficient, which is used to measure the effect of time on the importance of vocabulary; After considering historical weight, risk correction and time decay, the final vocabulary weight is calculated using the following formula (7): :

[0052] in: is the adjustment factor for controlling risk and time; It is a correction term used to control the smoothing effect in certain situations and prevent the weight from being too high; is a nonlinear adjustment index used to adjust the weighting effect in extreme cases.

[0053] In the automatic identification method of suspicious bank transaction features, a complex weighted processing method is used for the extracted fund flow related words to enhance the model's ability to identify suspicious transactions. This method combines historical data, the risk level defined within the bank, and time factors, and dynamically adjusts the weights of fund flow related words through multiple formulas to ensure that the final generated word weight can accurately reflect the importance of the word in suspicious transactions.

[0054] First, formula (4) calculates the historical data weights of capital flow related terms , used to measure the frequency of a certain word appearing in historical suspicious transaction records: in, Words related to fund flow The number of times it appears in historical suspicious transaction records, is the total number of words in the historical suspicious transaction records. This weight reflects the importance of the word in past transactions. By calculating the frequency distribution of words, the system can identify those high-risk words that frequently appear in historical suspicious transactions, and thus give them a higher initial weight.

[0055] Next, formula (5) calculates the modified vocabulary weight by combining the risk level defined within the bank .

[0056] In this formula, is the risk correction factor, which controls the impact of risk level on vocabulary weight; It is a term related to the flow of funds within the bank. The assigned risk level, ranging from 0 to 1, indicates the degree of risk of the term from low risk to high risk. If a term appears frequently in historical data and is assigned a higher risk level, its revised weight will be significantly increased. This step ensures that banks can add additional weight to certain high-risk terms based on their own risk management experience.

[0057] Then, the system combines the time features in the transaction records and uses formula (6) to perform time decay processing on the vocabulary weights to obtain the time-decayed weights .

[0058] In this formula, Indicates the timestamp of the current transaction, The reference time point is usually a high-risk period with a high concentration of suspicious transactions; is the time decay coefficient, which is used to measure the impact of time on word weight. As time goes by, some historically high-frequency words may gradually lose their importance, so the system reduces the weight of these words through the time decay function. If the current transaction time is closer to the reference time, , the smaller the effect of time decay, the higher the weight of the vocabulary remains; conversely, the importance of the vocabulary will decay over time.

[0059] After comprehensively considering the effects of historical weight, risk correction and time decay, formula (7) is used to calculate the final vocabulary weight .

[0060] In this formula, It is the coefficient that controls risk and time adjustment, and determines the relative impact of time decay and risk correction; It is a correction term used to smooth the weights in extreme cases and prevent excessively increasing the weights of certain words; is a nonlinear adjustment index that controls weight changes in extreme cases. The structure of the formula ensures that when the historical weight and time decay are large, the vocabulary weight will increase significantly, but by introducing nonlinear adjustment indexes and correction terms, it prevents some vocabulary weights from increasing excessively and maintains the balance of the model.

[0061] Through the weighted processing method, the system can not only identify frequently occurring high-risk words, but also reasonably and dynamically adjust the final weight of each word based on the bank's internal risk level definition and the time characteristics of the transaction. This method ensures that the model can fully utilize historical data, time information and risk management experience during the training process to improve the accuracy and efficiency of identifying suspicious transactions.

[0062] Step S103: Using the word segmentation results of the text in the transaction record and combining with the constructed vocabulary, the Word2Vec model is trained; wherein the Word2Vec model is implemented using the Skip-Gram model architecture. For a central word in a given transaction record text, the Skip-Gram model is trained to generate a word vector by predicting its context words to capture the semantic relationship between words; wherein the occurrence probability of each context word is calculated according to the following formula (1):

[0063] in, Indicates the central word, represents the context word, where Indicates the position of the central word in the sentence. Represents context-dependent indexes; Representing context words The word vector of Indicates the central word The word vector of A word vector representing each word in the vocabulary; is the size of the vocabulary.

[0064] In step S103, the word segmentation results of the transaction records are combined with the constructed vocabulary and used to train the Word2Vec model. The Word2Vec model is a neural network model that converts text words into word vectors, which can capture the semantic relationship between words in a high-dimensional space. By training the words in the transaction records into word vectors, the model can learn the correlation between different words from these records, thereby helping to identify the characteristics of suspicious transactions.

[0065] The Word2Vec model uses the Skip-Gram model architecture, which means that the core of the model is to learn word vectors by predicting the context words of the central word. In this architecture, given a central word , the system predicts the words that may be associated with the word through the context of the word. Indicates that is a relative position index, indicating the distance between the vocabulary and the central word. The goal of the model is to train the word vector of the central word by predicting the context words, thereby capturing the potential semantic relationship between the words.

[0066] Specifically, the Skip-Gram model learns these relationships by maximizing the probability of the context words appearing. Given the central word The probability of occurrence under the condition is calculated by the following formula:

[0067] In this formula, each variable has a clear meaning. First, Representing context words The word vector is a multi-dimensional vector that captures the semantic features of the word in the training data. Next, Indicates the central word The inner product between these two vectors is The semantic similarity between the center word and the context words was calculated.

[0068] The summation formula in the denominator represents the normalization of the semantic similarity of all words in the vocabulary. Represents the word vector for each word in the vocabulary, is the size of the vocabulary. Therefore, the denominator calculates the similarity between the central word and all the words to ensure that the probability of the context words is relative, not only related to the semantics of the central word, but also affected by other words.

[0069] What this probability formula reflects is that the model gradually adjusts the position of the word vector by observing the relationship between the central word and the context word, so that the word vectors with similar semantics are closer in the high-dimensional space. For example, the system may find that "withdraw cash" and "ATM withdrawal" often appear at the same time in many transaction records, so their word vectors will be trained to be very close. This proximity reflects their semantic similarity, which helps the model identify potential suspicious transaction features.

[0070] Step S104: The Word2Vec model is trained using a negative sampling method, wherein the negative sampling method randomly selects words that are not related to the central word as negative samples to reduce the amount of calculation required for the model to predict all words. The loss function provided by the following formula (2) is used during the training process:

[0071] in, represents the loss function, which is used to measure the prediction error of the model during the training process; Represents the total number of samples in the training data; Indicates the window size; Representing context words The word vector of Indicates the central word The word vector of Represents the total number of negative samples; Represents the distribution of negative samples Randomly sample negative samples The expected value of probability; Represents negative samples The word vector of is the sigmoid function.

[0072] In step S104, negative sampling is used to train the Word2Vec model. Negative sampling is an optimization technique whose main purpose is to reduce the amount of computation required to predict all words during training while ensuring that the model can efficiently capture the semantic relationship between words. Since the size of the vocabulary is usually very large, especially in financial transaction data involving a large number of specific words, directly calculating all words is not only very time-consuming but also unnecessary. Negative sampling greatly improves the computational efficiency of the model by selecting only some negative samples for training.

[0073] First, in order to understand the use of negative sampling, it is necessary to clarify that the goal of the model is to use the center word To predict the context word , and the probability of each context word is given by the following formula:

[0074] here, The word vector representing the context word, Represents the word vector of the central word, and the denominator represents the sum of the similarities of all words. However, due to the huge size of the vocabulary, it is impractical to calculate the probability of all words, so the negative sampling method is introduced.

[0075] The core idea of ​​negative sampling is that for each real context word, the model not only learns how to correctly predict these context words, but also introduces randomly selected negative samples (i.e. words that are not related to the central word) to train the model to identify which words are unlikely to appear in the context. In this way, the model does not have to calculate the similarity of all words, but only needs to calculate the similarity of the actual context words and a small number of negative samples, which greatly reduces the amount of calculation.

[0076] When implementing negative sampling, the model randomly selects negative samples, Usually a small integer, such as 5 or 10. Each negative sample is from the vocabulary based on a probability distribution Randomly selected, this probability distribution is usually related to the frequency of the word, making high-frequency words more likely to be selected as negative samples. Negative sampling ensures that the model can not only identify positive samples (i.e., true context words), but also reduce the probability of wrong predictions by training the model with negative samples.

[0077] During the training process, the loss function of the model is It is used to measure the error of the model in predicting context words. The loss function consists of two parts: one is the prediction for positive samples, and the other is the prediction for negative samples. The specific loss function form is as follows:

[0078] In this formula, Represents the total number of samples in the training data, which means that the model will iterate on all training data; Indicates the size of the context window, which controls how many context words around the center word will be considered. Used to calculate the prediction results of the true context words, where It is a sigmoid function that maps the prediction result to between 0 and 1 to indicate the degree of match between the predicted context word and the central word.

[0079] Part 2 is the prediction for negative samples. Negative samples are from the distribution It is obtained by random sampling from the vocabulary, indicating that those words in the vocabulary should not be related to the central word By maximizing the negative logarithm of the sigmoid of negative samples, the model is trained to minimize the matching probability of negative samples.

[0080] Each negative sample The word vector of The role of the negative sampling method is to introduce irrelevant negative samples and try to reduce the matching probability of these words with the central word, so that the model can better distinguish which words should appear in the context of the central word and which words should not appear. This training mechanism greatly improves the efficiency of the model, especially when dealing with large-scale vocabularies, so that the model can capture complex semantic relationships in transaction records more accurately.

[0081] Through negative sampling, the model can not only complete training faster, but also avoid over-calculating the similarity of irrelevant words, thereby reducing computing costs. This is especially important for identifying suspicious transactions in banks, because the vocabulary involved in transaction records is usually very large, and the correlation between different words varies significantly. Through negative sampling, the system can efficiently train the model so that it can accurately identify abnormal or suspicious behavior in transactions, thereby improving the accuracy of suspicious transaction detection.

[0082] In summary, the negative sampling method provides an efficient training mechanism for the Word2Vec model. By randomly selecting a small number of negative samples, it avoids comprehensive calculations of all words, greatly improving the training efficiency of the model while ensuring the accuracy of the model in semantic capture. This method is particularly suitable for processing large-scale financial transaction data and helps to identify potential suspicious transaction behaviors in practical applications.

[0083] Step S105: After the training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features.

[0084] In step S105, after the training is completed, the generated word vectors need to be verified to evaluate the accuracy of the Word2Vec model in identifying transaction features. The main purpose of this verification step is to ensure that the word vectors learned by the model can accurately reflect the semantic relationship in the bank transaction records, especially in identifying suspicious transaction features with high efficiency and accuracy.

[0085] First, the basic principle of verification is to determine whether the model successfully captures the semantic relationship between words by evaluating the distance between the generated word vector and the words in its corresponding vocabulary in the semantic space. The "distance" here refers to the similarity in the high-dimensional vector space, which is generally quantified by calculating the cosine similarity. Cosine similarity is a commonly used metric in natural language processing to measure the similarity between two word vectors. The formula is as follows:

[0086] In this formula, and Represent the word vectors of two words respectively, is their dot product, and Respectively represent the modulus length of the two vectors. The result of cosine similarity ranges from -1 to 1, where 1 means exactly the same, 0 means irrelevant, and -1 means completely opposite. Therefore, by calculating the cosine similarity of each generated word vector with its corresponding word in the vocabulary, the model's ability to capture the semantics of these words can be evaluated.

[0087] In specific operations, the system will randomly extract some words from the transaction records and compare the generated word vectors with the word vectors in the vocabulary. For each generated word vector, the system will calculate its similarity with other words and sort them according to the cosine similarity value. Ideally, the word vector generated by the model should be ranked in the top few places with the words in the vocabulary that are closest to its semantics. In this way, it can be preliminarily verified whether the model can effectively distinguish the semantics of different words and capture the features in the transaction description.

[0088] In transaction scenarios, especially for feature recognition involving suspicious transactions, the model needs to not only capture the surface vocabulary information of the transaction, but also be able to identify potential semantic associations. For example, when describing "cash withdrawals from other places" or "frequent small transfers", the model should be able to identify the potential connections between these words and correctly identify them as possible suspicious transaction behaviors. Therefore, the accuracy verification of the model is not only to verify the semantic similarity of the word vectors, but also to analyze the transaction features in the actual business scenario.

[0089] To further ensure the accuracy of the model, the system can perform special verification on the model through a specific set of suspicious transaction feature words. This process includes defining a series of known suspicious transaction feature words, such as "multiple transfers" and "large cash withdrawals", and then evaluating the model's recognition of these feature words. By comparing the similarity between the word vectors generated by the model for these feature words and the standard word vectors in the vocabulary, the performance of the model in actual scenarios can be accurately judged.

[0090] In addition to verification based on cosine similarity, the system can also be evaluated through coverage and vocabulary distribution characteristics. Coverage refers to the proportion of matches between the generated word vectors and the words in the vocabulary. If the generated word vectors can cover most of the important words in the vocabulary, it means that the model has a high semantic capture ability. In addition, by analyzing the distribution of word vectors in high-dimensional space, especially whether the suspicious transaction feature words form obvious clusters, the recognition effect of the model can be further evaluated. If the model can concentrate semantically similar word vectors in a certain area of ​​the space, while words with large semantic differences are distributed in different areas, it means that the model training is successful and can accurately distinguish different transaction features.

[0091] Finally, the model verification results can be used as a basis for adjusting model parameters. By analyzing the accuracy and coverage of the generated word vectors, the system can identify deficiencies in model training, such as the low recognition accuracy of certain feature words or the failure to fully capture certain important transaction features. This verification information will be fed back into the model training process, and by adjusting parameters (such as word vector dimension, context window size, number of negative sampling, etc.), the model can be further optimized to make it more accurate in capturing transaction features.

[0092] In summary, the verification process of step S105 ensures that the generated word vector can accurately reflect the semantic relationship in the bank transaction record, especially has a high accuracy in identifying suspicious transaction features.

[0093] Furthermore, after the training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features, including: By comparing the word vectors generated by the model with the word vectors of known suspicious transaction feature words, the cosine similarity between the two is calculated to evaluate the recognition accuracy of the model.

[0094] In the automatic identification method of suspicious transaction features in banks, the generated word vectors need to be verified after training to evaluate the recognition accuracy of the Word2Vec model for transaction features. The verification process is performed by calculating the cosine similarity between the generated word vectors and the word vectors of known suspicious transaction feature words.

[0095] Specifically, the system compares the word vectors generated by the model with the vectors of the known suspicious transaction feature words in the vocabulary. Cosine similarity is used to measure the similarity between two vectors, and the calculation formula is the dot product of the vectors divided by the product of their moduli. The value range of cosine similarity is between -1 and 1, where 1 means that the two vectors are exactly the same, 0 means they are completely unrelated, and -1 means they are completely opposite.

[0096] In this way, the model can quantify the similarity between the generated word vector and known high-risk words, thereby evaluating the accuracy of the model in identifying these suspicious transaction feature words. If the cosine similarity value is high, it means that the model has a good recognition effect on the feature word; on the contrary, if the similarity is low, it means that the model's recognition effect on the word is not ideal, and further optimization of the model or adjustment of parameters may be required.

[0097] Furthermore, after the training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features, which also includes: If the cosine similarity between the generated word vector and the word vector of the known suspicious transaction feature vocabulary exceeds the first threshold, the model is judged to have a high recognition accuracy; if the cosine similarity is lower than the second threshold, the model parameter adjustment mechanism is triggered.

[0098] In the automatic identification method of suspicious transaction features of banks, after the model training is completed, the generated word vectors will be verified to evaluate the recognition accuracy of the Word2Vec model for transaction features. In this verification process, the system evaluates the performance of the model by calculating the cosine similarity between the generated word vectors and the word vectors of known suspicious transaction feature words.

[0099] Specifically, if the cosine similarity between the generated word vector and the known suspicious transaction feature vocabulary exceeds the pre-set first threshold, it means that the model has a high recognition accuracy and can better capture the characteristics of these suspicious transactions. At this time, the system will determine that the performance of the model meets the expected standards.

[0100] However, if the cosine similarity is lower than the second threshold, it means that the model performs poorly in identifying specific suspicious transaction feature words, and there may be problems with improper parameter settings. Therefore, the system will trigger the model parameter adjustment mechanism to further optimize the performance of the model by modifying the model parameters such as vocabulary size, word vector dimension or learning rate. This mechanism ensures the accuracy and robustness of the model in practical applications, enabling it to effectively identify potential suspicious transaction behaviors in different transaction scenarios. Those skilled in the art can optimize the model in real time based on this automated adjustment mechanism to improve the overall transaction feature recognition capability of the system.

[0101] Furthermore, the constructed vocabulary includes at least 5,000 transaction-related words, wherein the transaction-related words include banking business terms, transaction pattern words, and descriptive words of potential suspicious transaction behaviors.

[0102] In the automatic identification method of suspicious bank transaction features, the constructed vocabulary contains at least 5,000 words related to bank transactions. These words include not only common terms in banking business, but also different types of transaction model words and words used to describe potential suspicious transaction behaviors.

[0103] Specifically, banking business terms cover vocabulary involved in daily financial operations, such as deposits, withdrawals, transfers, etc.; transaction pattern vocabulary is used to describe common transaction methods or patterns, such as regular transfers, cross-border transactions, etc.; and descriptive vocabulary for suspicious transaction behaviors is used to describe characteristic vocabulary that often appears in risky transaction analysis, such as "frequent transfers" and "large cash withdrawals". These words are the basis for feature extraction and classification in automatic recognition methods. By building and using this vocabulary containing a wide range of words, the model can more accurately identify and judge potential risk behaviors in banking transactions.

[0104] Furthermore, the vocabulary is dynamically updated according to the frequency of historical transaction data, and high-frequency transaction feature vocabulary is given priority during each training to ensure that the model has good adaptability to the latest transaction patterns.

[0105] In the automatic identification method of suspicious transaction features in banks, the constructed vocabulary is dynamically updated. This dynamic update is adjusted based on the frequency of vocabulary in historical transaction data. Specifically, during each model training, the system analyzes the frequency of vocabulary in transaction records and gives priority to those transaction feature vocabulary that appear frequently to ensure that these vocabulary are better represented and trained in the model.

[0106] Through this dynamic update mechanism, the vocabulary can reflect the latest transaction patterns and trends in a timely manner, so that the model can not only adapt to historical transaction data, but also quickly capture new suspicious transaction behavior characteristics. This update method ensures that the model maintains a high level of recognition ability and accuracy in the ever-changing banking transaction environment, and can effectively respond to new transaction patterns and potential risk behaviors. Those skilled in the art can use this dynamic update mechanism to maintain the real-time and adaptability of the model, thereby improving the overall performance of the system.

[0107] Furthermore, after generating word vectors through the Word2Vec model and verifying the model recognition accuracy, the Spearman coefficient is used to evaluate the adjustment effect of the model parameters, where the Spearman coefficient is calculated according to the following formula (3):

[0108] in, represents the Spearman correlation coefficient, which is used to measure the correlation of model performance before and after model parameter adjustment; represents the square of the difference between the rankings of two parameters; Indicates the training round.

[0109] In this embodiment, after using the Word2Vec model to generate word vectors and verifying the recognition accuracy of the model, in order to further optimize the performance of the model, the parameters of the model need to be adjusted. In order to evaluate the effect of these parameter adjustments, the calculation method of the Spearman coefficient is introduced, which can quantify the correlation of performance changes before and after the model parameter adjustment, and help developers determine whether the parameter adjustment is effective.

[0110] Spearman coefficient It is a non-parametric statistical indicator that measures the correlation between two variables (in this example, the performance results before and after the model parameter adjustment). It is calculated based on the ranking of the variables rather than the actual values, so it is particularly suitable for capturing nonlinear relationships when evaluating the effect of model parameter adjustment.

[0111] In formula (3), each symbol has a specific meaning. First, Represents the Spearman correlation coefficient, and its value range is between -1 and 1. A value close to 1 indicates a high positive correlation, which means that the model performance is significantly improved after parameter adjustment; a value close to -1 indicates a high negative correlation, which means that the model performance decreases after adjustment; and a value close to 0 indicates that there is no significant change in performance before and after adjustment.

[0112] In the formula is the sum of squares of ranking differences, Representative The difference in ranking before and after the model parameter adjustment in the group data. Specifically, for a certain round of training data, the system calculates the performance ranking before and after the model parameter adjustment. If the performance after adjustment is higher than before adjustment, the ranking is higher, and vice versa. By calculating the ranking difference before and after the parameter adjustment for each training round and squaring it, the impact of negative ranking differences on the total can be eliminated and positive and negative differences can be evenly reflected in the results.

[0113] The sum of squares reflects the total amount of ranking differences. A larger ranking difference indicates a greater fluctuation in model performance after parameter adjustment, while a smaller sum of squares means a smaller ranking change, indicating that parameter adjustment has a more stable effect on model performance. The system can intuitively quantify the performance difference between the model before and after adjustment in multiple training rounds.

[0114] Next, is the total number of training rounds. This value reflects the number of ranking results generated by the model during multiple rounds of training. In practical applications, the system evaluates the stability and consistency of the model through multiple training rounds. It provides a more comprehensive assessment, making the results of the Spearman coefficient more reliable.

[0115] In the formula is a multiple of the ranking difference, multiplied by 6, which is a constant in the Spearman coefficient calculation and is used to normalize the final correlation result. It is used to standardize the sum of squares of ranking differences so that the result of the Spearman coefficient remains in the range of -1 to 1.

[0116] Through this formula, the system can effectively evaluate the relevance of the model before and after parameter adjustment. If the Spearman coefficient value is close to 1, it means that the parameter adjustment has a significant effect on the performance of the model, and the system will apply this adjustment method to the subsequent optimization process. If the Spearman coefficient is close to -1, it means that the current parameter adjustment strategy has a negative impact on the performance of the model, and the system needs to further explore and optimize the adjustment method. If the Spearman coefficient is close to 0, it means that the parameter adjustment has little effect on the performance of the model. At this time, it may be necessary to adjust other parameters, or further analyze the overall structure and training strategy of the model.

[0117] In actual operation, the system adjusts the model's vocabulary size, word vector dimension, context window size, number of negative samples, and learning rate through multiple iterations. After each parameter adjustment, the system will analyze the positive and negative impact of parameter adjustment on model performance based on the feedback of the Spearman coefficient, and gradually optimize the overall effect of the model. By introducing the Spearman coefficient as an indicator, the system can evaluate the effect of parameter adjustment in a scientific and quantitative way, and ensure that the final Word2Vec model can efficiently and accurately identify the characteristics of suspicious bank transactions.

[0118] Step S106: According to the verification result of the Word2Vec model, the model parameters of the Word2Vec model are adjusted to obtain an adjusted Word2Vec model capable of identifying the characteristics of suspicious bank transactions.

[0119] In step S106, according to the results obtained in the verification process of the previous step, the model parameters of the Word2Vec model are adjusted to optimize the performance of the model and improve its ability to identify suspicious bank transaction characteristics. The core of this step is to accurately adjust multiple key parameters of the model through evaluation feedback of the model, so that the model can more accurately capture transaction characteristics in different scenarios.

[0120] First, the adjustment of model parameters involves multiple aspects, including word vector dimension, context window size, number of negative samples for negative sampling, learning rate, etc. The adjustment of each parameter has a direct impact on the performance of the model. For example, the dimension of the word vector determines the accuracy with which the vocabulary is represented in the semantic space. Lower dimensions may not be able to fully capture complex semantic relationships, while too high a dimension may increase the computational complexity of the model and introduce overfitting. Therefore, when adjusting the word vector dimension, the model needs to strike a balance between accuracy and computational efficiency. The word vector quality indicators obtained in the verification step, such as cosine similarity, can be used to evaluate the performance of the current word vector and determine whether the dimension needs to be increased or decreased.

[0121] The size of the context window It is also an important parameter. The size of the window determines how many words around the central word will be used as context words for training. If the window is too small, the model may not capture enough semantic information, especially when the transaction description is long; but if the window is too large, the model may introduce too much irrelevant information, resulting in increased noise. Therefore, the adjustment of the window size should be based on the analysis of the semantic relevance between words during the verification process. For example, by observing the impact of the context window size on the vocabulary of specific suspicious transaction features, it can be decided whether the scope of the window needs to be expanded or reduced.

[0122] Number of negative samples is another parameter that needs to be adjusted. The role of negative sampling is to optimize the computational efficiency of the model by selecting a small number of words that are irrelevant to the central word. If the number of negative samples is too small, the model may not be able to fully learn how to distinguish positive samples from irrelevant words; conversely, if the number of negative samples is too large, it may increase the computational burden. Therefore, the number of negative samples can be adjusted by observing how the model handles noisy data. If the model performs poorly when processing noisy words during the verification process, it may mean that the number of negative samples needs to be increased to help the model better learn how to ignore irrelevant information.

[0123] The learning rate is a key parameter that controls the rate at which model parameters are updated. During model training, the learning rate determines the magnitude of the model parameter update at each iteration. If the learning rate is too high, the model may jump too much in the parameter space and fail to converge to the optimal value; if the learning rate is too low, the training process may be too slow or even fall into a local optimum. Therefore, by analyzing the loss function in the verification process The changing trend of can determine whether the learning rate needs to be adjusted. If the loss of the model drops too fast in the early stage and fluctuates greatly in the later stage, it may be necessary to appropriately reduce the learning rate to make the model converge more stably.

[0124] Loss Function The specific formula is as follows:

[0125] In this formula, represents the total number of samples in the training data, and are the word vectors of the context word and the center word respectively, is the number of negative samples. By analyzing the gradual optimization process of the loss function, especially the change in loss under different parameter settings, it can help determine whether the model training is effective. If the value of the loss function continues to decrease after the model is adjusted, it means that the model is moving in the right direction during the optimization process; conversely, if the value of the loss function fluctuates greatly or cannot steadily decrease, it indicates that some parameters need further adjustment.

[0126] In addition, when tuning model parameters, the size of the vocabulary also needs to be considered. Impact on model performance. A larger vocabulary can cover more transaction feature words, but it may also increase the training time of the model. Therefore, by analyzing the coverage of the vocabulary during the validation process, you can determine whether the size of the vocabulary is appropriate. If most of the words in the vocabulary are irrelevant to the training data, you may need to streamline the vocabulary to improve training efficiency.

[0127] Finally, by repeatedly adjusting and verifying these model parameters, the Word2Vec model can more accurately identify suspicious transaction features in different banking transaction scenarios. The adjusted model should show higher word vector similarity and lower loss function value during the verification process, and can accurately capture the key feature words in bank transaction records that reflect potential risks. This optimization process not only ensures the robustness of the model, but also improves its effectiveness in practical applications, allowing banks to detect and respond to potential suspicious transactions earlier.

[0128] Furthermore, the model parameters include vocabulary size, word vector dimension, context window size, number of negative samples, and learning rate.

[0129] The model parameter design of the automatic identification method of suspicious transaction features in banks is crucial. These parameters determine the model training process, the quality of word vectors, and the final effect of identifying suspicious transaction features. In order to ensure that the model can run efficiently in complex bank transaction scenarios and accurately identify potential risky transaction behaviors, the various parameters of the model must be set and adjusted in detail. These parameters mainly include the size of the vocabulary, the dimension of the word vector, the size of the context window, the number of negative samples, and the learning rate.

[0130] First, the size of the vocabulary directly affects the quality of the model's word vector training. The vocabulary is a set of words used by the model to represent transaction records, and contains all the words used in bank transaction descriptions. When building a vocabulary, the size of the vocabulary needs to be reasonably set according to the characteristics of the transaction data. If the vocabulary is too small, the model may ignore some important words, especially when identifying suspicious transactions, some potential risk features may be missed; on the contrary, an overly large vocabulary will increase the model's computational workload, prolong training time, and may even introduce unnecessary noise words. Therefore, the size of the vocabulary should be dynamically adjusted based on the vocabulary distribution of transaction data, the complexity of transaction scenarios, and the frequency of use of specific words.

[0131] Secondly, the word vector dimension is another key parameter that determines how the model represents each word in a high-dimensional space. Word vectors are a fixed-length numerical vector that converts words into words to represent the semantic relationship between words. In bank transaction descriptions, different words may represent different transaction behaviors or characteristics, so the choice of word vector dimension will directly affect the model's ability to capture these transaction features. If the dimension is too low, the model may not be able to accurately represent the subtle semantic differences between words, especially when dealing with complex transaction features, the model may not be able to distinguish similar transaction descriptions; if the dimension is too high, although more complex semantic relationships can be captured, it will also increase the computational complexity and may lead to overfitting.

[0132] The size of the context window is one of the important parameters that affect model training. It determines how the model captures the semantic relationship between the central word and its surrounding words during the training process. The larger the size of the context window, the more context words the model will consider for the central word, which helps the model better understand the relevance between words when the transaction description is long or contains complex semantic relationships. However, a context window that is too large may introduce too much irrelevant information, resulting in a decrease in the prediction accuracy of the model; while a window that is too small may not be able to fully capture the contextual relationship of the central word, especially in the description of certain specific transaction behaviors, the relevance of words may need to be captured through a larger context window.

[0133] The number of negative samples is a crucial parameter in the model optimization process. In the Word2Vec model, the negative sampling method is used to improve the efficiency of model training. The setting of the number of negative samples determines how the model optimizes the prediction results by comparing the real context words and randomly selected irrelevant words during training. If the number of negative samples is too small, the model may not be able to effectively distinguish between positive samples and irrelevant negative samples, resulting in insufficient generalization ability of the model in identifying suspicious transaction features; if the number of negative samples is too large, although the generalization ability of the model will be enhanced, the amount of calculation and training time will also increase significantly.

[0134] Finally, the learning rate is an important factor in controlling the rate at which model parameters are updated. During the model training process, the learning rate determines the magnitude of the model parameter update at each iteration. Too high a learning rate may cause the model to jump too much in the parameter space and fail to converge to the optimal solution stably; while too low a learning rate may cause the model to converge too slowly or even fall into a local optimal solution. Therefore, the setting of the learning rate needs to be dynamically adjusted according to the model training process.

[0135] In the above-mentioned embodiment, a method for automatically identifying suspicious bank transaction features is provided. Correspondingly, the present application also provides a device for automatically identifying suspicious bank transaction features. Figure 2 , which is a schematic diagram of an embodiment of an automatic identification device for suspicious transaction characteristics of a bank of the present application. Since this embodiment, i.e., the second embodiment, is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The embodiment described below is only illustrative.

[0136] The second embodiment of the present application provides a device for automatically identifying suspicious transaction characteristics of a bank, comprising: The collection unit 201 is used to collect bank transaction records and perform standardization processing on the text in the transaction records; wherein the transaction records include transfer, consumption, ATM withdrawal, counter withdrawal, purchase of financial products, purchase of precious metals and loan records; the standardization processing includes standardization of date format, standardization of amount unit and removal of non-text symbols; The processing unit 202 is used to perform word segmentation processing on the text in the standardized transaction record to extract key information words; and construct a vocabulary based on the key information words; wherein the vocabulary includes words reflecting transaction characteristics, which are used for subsequent model training; The construction unit 203 is used to train the Word2Vec model using the word segmentation processing result of the text in the transaction record and the constructed vocabulary; wherein the Word2Vec model is implemented using the Skip-Gram model architecture. For a central word in a given transaction record text, the Skip-Gram model is trained to generate a word vector by predicting its context words to capture the semantic relationship between words; wherein the occurrence probability of each context word is calculated according to the following formula (1):

[0137] in, Indicates the central word, represents the context word, where Indicates the position of the central word in the sentence. Represents context-dependent indexes; Representing context words The word vector of Indicates the central word The word vector of A word vector representing each word in the vocabulary; is the size of the vocabulary; The training unit 204 is used to train the Word2Vec model using a negative sampling method, wherein the negative sampling method randomly selects words that are not related to the central word as negative samples to reduce the amount of calculation required for the model to predict all words; wherein the loss function provided by the following formula (2) is used during the training process:

[0138] in, Represents the loss function, which is used to measure the prediction error of the model during the training process; Represents the total number of samples in the training data; Indicates the window size; Representing context words The word vector of Indicates the central word The word vector of Represents the total number of negative samples; Represents the distribution of negative samples Randomly sample negative samples The expected value of probability; Represents negative samples The word vector of is the sigmoid function; A verification unit 205 is used to verify the generated word vector after training is completed, and evaluate the accuracy of the Word2Vec model in identifying transaction features; The adjustment unit 206 is used to adjust the model parameters of the Word2Vec model according to the verification result of the Word2Vec model to obtain an adjusted Word2Vec model that can identify the characteristics of suspicious bank transactions.

[0139] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for automatically identifying suspicious transaction characteristics of a bank, characterized in that: include: Collect bank transaction records and standardize the text in the transaction records; wherein the transaction records include transfers, consumption, ATM withdrawals, counter withdrawals, financial management purchases, precious metal purchases, and loan records; the standardization includes standardizing the date format, normalizing the amount unit, and removing non-text symbols; Performing word segmentation on the text in the standardized transaction records to extract key information words; constructing a vocabulary based on the key information words; wherein the vocabulary includes words reflecting transaction characteristics for subsequent model training; The Word2Vec model is trained using the word segmentation results of the text in the transaction record and the constructed vocabulary. The Word2Vec model is implemented using the Skip-Gram model architecture. For a central word in a given transaction record text, the Skip-Gram model is trained to generate a word vector by predicting its context words to capture the semantic relationship between words. The Word2Vec model is trained using a negative sampling method, wherein the negative sampling method randomly selects words that are not related to the central word as negative samples to reduce the amount of calculation required for the model to predict all words; After training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features; According to the verification results of the Word2Vec model, the model parameters of the Word2Vec model are adjusted to obtain an adjusted Word2Vec model that can identify the characteristics of suspicious bank transactions.

2. The method for automatically identifying suspicious bank transaction characteristics according to claim 1, characterized in that: The model parameters include vocabulary size, word vector dimension, context window size, number of negative samples, and learning rate.

3. The method for automatically identifying suspicious bank transaction characteristics according to claim 1, characterized in that: After generating word vectors through the Word2Vec model and verifying the model recognition accuracy, the Spearman coefficient is used to evaluate the adjustment effect of the model parameters. The Spearman coefficient is calculated according to the following formula: in, represents the Spearman correlation coefficient, which is used to measure the correlation of model performance before and after model parameter adjustment; represents the square of the difference between the rankings of two parameters; Indicates the training round.

4. The method for automatically identifying suspicious bank transaction characteristics according to claim 1, characterized in that: The text in the standardized transaction record is segmented to extract key information words, including: Use natural language processing tools to perform grammatical analysis on transaction record text, identify the part of speech of each word, and capture the central word and its related context words based on the context window technology to ensure that each word has accurate semantic association in its context; Dynamically filter out irrelevant or noisy words based on the contextual words in the transaction scenario; Combined with the characteristic vocabulary of financial transactions, the words related to capital flow are extracted first after vocabulary filtering; Based on the high-risk characteristic words in the historical suspicious transaction records, the extracted funds flow-related words are weighted to enhance their importance in subsequent model training.

5. The method for automatically identifying suspicious bank transaction characteristics according to claim 4, characterized in that: The weighted processing of the extracted fund flow related words based on the high-risk characteristic words in the historical suspicious transaction records includes: According to the following formula, the historical data weight of fund flow related words is calculated. : in, Words related to fund flow The number of times the transaction appears in historical suspicious transaction records; Indicates the total number of words in the historical suspicious transaction records; Based on the risk level defined internally by the bank, the modified vocabulary weight is calculated using the following formula : in, is the risk correction factor; Vocabulary related to the flow of funds within the bank The assigned risk level, with risk level values ​​ranging from 0 to 1, representing a degree of risk from low risk to high risk; Combined with the time characteristics in the transaction records, the vocabulary weight is further adjusted according to the following formula to obtain the time decay weight : in, Indicates the timestamp of the current transaction; The reference time point includes high-risk time periods when suspicious transactions are concentrated; is the time decay coefficient, which is used to measure the effect of time on the importance of vocabulary; After considering historical weight, risk correction and time decay, the final vocabulary weight is calculated using the following formula : in: is the adjustment factor for controlling risk and time; It is a correction term used to control the smoothing effect in certain situations and prevent the weight from being too high; is a nonlinear adjustment index used to adjust the weighting effect in extreme cases.

6. The method for automatically identifying suspicious bank transaction characteristics according to claim 1, characterized in that: After the training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features, including: By comparing the word vectors generated by the model with the word vectors of known suspicious transaction feature words, the cosine similarity between the two is calculated to evaluate the recognition accuracy of the model.

7. The method for automatically identifying suspicious bank transaction characteristics according to claim 6, characterized in that: After the training is completed, the generated word vectors are verified to evaluate the accuracy of the Word2Vec model in identifying transaction features, which also includes: If the cosine similarity between the generated word vector and the word vector of the known suspicious transaction feature vocabulary exceeds the first threshold, the model is judged to have a high recognition accuracy; if the cosine similarity is lower than the second threshold, the model parameter adjustment mechanism is triggered.

8. The method for automatically identifying suspicious bank transaction characteristics according to claim 1, characterized in that: The constructed vocabulary includes at least 5,000 transaction-related words, wherein the transaction-related words include banking business terms, transaction pattern words, and descriptive words of potential suspicious transaction behaviors.

9. The method for automatically identifying suspicious bank transaction characteristics according to claim 8, characterized in that: The vocabulary is dynamically updated according to the frequency of historical transaction data, and high-frequency transaction feature words are given priority during each training to ensure that the model has good adaptability to the latest transaction patterns.

10. A device for automatically identifying suspicious transaction characteristics in a bank, characterized in that: include: A collection unit is used to collect bank transaction records and perform standardization on the text in the transaction records; wherein the transaction records include transfers, consumption, ATM withdrawals, counter withdrawals, purchases of financial products, purchases of precious metals and loan records; the standardization includes standardization of date formats, standardization of amount units and removal of non-text symbols; A processing unit, configured to perform word segmentation processing on the text in the standardized transaction record to extract key information words; and construct a vocabulary based on the key information words; wherein the vocabulary includes words reflecting transaction characteristics for subsequent model training; A construction unit is used to train a Word2Vec model using a word segmentation processing result of the text in the transaction record in combination with the constructed vocabulary; wherein the Word2Vec model is implemented using a Skip-Gram model architecture. For a central word in a given transaction record text, the Skip-Gram model is trained to generate a word vector by predicting its context words to capture the semantic relationship between words; A training unit, used for training the Word2Vec model by using a negative sampling method, wherein the negative sampling method reduces the amount of calculation required for the model to predict all words by randomly selecting words that are not related to the central word as negative samples; The verification unit is used to verify the generated word vectors after training is completed and evaluate the accuracy of the Word2Vec model in identifying transaction features; The adjustment unit is used to adjust the model parameters of the Word2Vec model according to the verification result of the Word2Vec model to obtain an adjusted Word2Vec model that can identify the characteristics of suspicious bank transactions.

Images