Important mail identification method and system, intelligent mailbox system

Abstract

The present application is applicable to the field of email technology, and provides an important email identification method and system, and an intelligent mailbox system, the method comprising: determining the behavioral influence weight of positive feedback behavior based on user behavior data; constructing a contact social graph based on social relationship data to determine the social influence weight of the target sender; extracting core tags from email content data to determine the semantic matching weight of the target email; obtaining a comprehensive email score based on the behavioral influence weight, social influence weight, and semantic matching weight; and determining whether the target email is an important email based on a dynamic judgment threshold and a comprehensive score. The present application determines whether the current email is important to the user by integrating recent multi-dimensional dynamic data such as user behavior, email content, and social relationships, as well as a dynamic judgment threshold. It can adapt to changes in user needs in real time, significantly reduces the misjudgment rate, and can accurately and individually intelligently identify important emails.

Description

Technical Field

[0001] The present application belongs to the field of email technology, and in particular relates to an important email identification method and system, and an intelligent mailbox system. Background Art

[0002] With the rapid development of internet technology, email has become an indispensable information exchange tool for important tasks such as corporate task scheduling. Traditional email systems are time-consuming and laborious to find, requiring users to spend considerable time identifying important emails from a flood of incoming emails. While several email recognition methods exist, these are relatively simple, resulting in inaccurate and inflexible results and prone to misidentification.

[0003] Existing email recognition methods include rule-based filtering, which primarily relies on static rules like sender whitelists or keyword matching. This approach is unable to adapt to dynamic needs, suffers from poor flexibility, and is prone to misjudgment. Another example is user behavior analysis, which primarily uses single behavioral signals like star marks and unread status to determine importance. This approach fails to consider multi-dimensional behavioral correlations and user social relationships. Semantic matching, which primarily matches email content with pre-set key topics based on TF-IDF or word frequency statistics, lacks deep semantic understanding, making it difficult to identify implicit importance and resulting in low accuracy. In situations where high accuracy and personalized service are required, these methods fail to meet practical application requirements, resulting in a poor user experience. Summary of the Invention

[0004] The embodiments of the present application provide a method and system for identifying important emails, and an intelligent mailbox system, which can solve the problems of low accuracy, high misjudgment rate and lack of personalization when identifying important emails.

[0005] In a first aspect, an embodiment of the present application provides an important email identification method, including:

[0006] Determining the influence weight of the target user's positive feedback behavior on the current behavior based on the user behavior data in the previous time period; wherein the user behavior data includes at least one feedback behavior of the target user on the target email, and the feedback behavior includes positive feedback behavior and negative feedback behavior;

[0007] Calculating the interaction strength between any two contacts based on social relationship data from the previous time period and constructing a contact social graph. Then, using the contact social graph and an improved PageRank algorithm, deriving the target sender's current social influence weight. The interaction strength includes the base relationship weight and / or interaction frequency between the two contacts, where the contacts represent the sender and recipient of the email.

[0008] Using a preset large model, extract at least one core tag from the email content data within the previous time period, then calculate the semantic similarity between each core tag and the email content of the target email, and determine the maximum semantic similarity calculated as the semantic matching weight of the target email;

[0009] Obtaining a comprehensive email score for the target email based on the behavioral influence weight, the social influence weight, and the semantic matching weight;

[0010] Whether the target email is currently an important email to the target user is determined based on a dynamic determination threshold and the email comprehensive score; wherein, when the email comprehensive score is greater than the dynamic determination threshold, the target email is currently an important email to the target user, otherwise it is an ordinary email; the dynamic determination threshold is determined based on the email comprehensive scores of all emails in the previous time period.

[0011] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0012] Dynamic weights such as behavioral influence weight, social influence weight and semantic matching weight are determined respectively by multi-dimensional dynamic data such as recent user behavior data, email content data and social relationship data. The impact of user behavior, email content and social relationships on the importance of emails is analyzed from different dimensions, which significantly reduces the misjudgment rate of email importance. The dynamic weight and dynamic judgment threshold are used to determine whether the current email is important to the user. The dynamic weight and dynamic judgment threshold are adaptively updated based on multi-dimensional dynamic data, which can adapt to changes in user needs in real time. They can accurately and individually intelligently identify important emails, thereby helping users quickly locate important emails, saving processing time, and helping to improve the quality and efficiency of information exchange.

[0013] In a possible implementation of the first aspect, the step of determining the current behavior influence weight of the target user's positive feedback behavior based on the user behavior data in the previous time period includes:

[0014] Determine, based on the user behavior data, the number of occurrences of each feedback behavior and a first time length from the occurrence of each feedback behavior to the current time;

[0015] The current behavior influence weight of the positive feedback behavior is calculated according to a preset time decay factor, the basic weight of the positive feedback behavior, the number of occurrences and the first time length.

[0016] In the above solution, positive feedback behavior reflects the user's recognition of the importance of emails in one dimension. At the same time, the time decay factor is introduced to give more recent user behavior data a higher weight, making the calculated behavior impact weight more consistent with the user's cognitive logic, thereby improving the accuracy of subsequent email importance predictions.

[0017] In a possible implementation of the first aspect, the step of calculating the interaction intensity between any two contacts based on the social relationship data in the previous time period and constructing a social graph of the contacts includes:

[0018] The contact social graph is defined as a directed weighted graph; wherein the nodes of the directed weighted graph represent contacts, and the edges of the directed weighted graph represent the relationship between two contacts;

[0019] Acquire, based on the social relationship data, the number of interactions between any two contacts in a previous time period, a second time length of the previous time period, and a basic relationship between the two contacts;

[0020] Calculating the interaction frequency between the two contacts based on the number of interactions and the second time length, and obtaining a basic relationship weight corresponding to the basic relationship from a preset relationship weight comparison table;

[0021] The interaction strength between any two contacts is determined according to the interaction frequency and the basic relationship weight, and the interaction strength is represented as an edge weight of the directed weighted graph.

[0022] In the above scheme, the interaction intensity between two contacts is quantified by combining interaction frequency and basic relationships. For email recipients, emails sent by email senders with higher interaction intensity are more important. Constructing a contact social graph based on a directed weighted graph can not only clarify the direction of information flow, but also accurately express complex relationships, thereby intuitively clarifying the interaction intensity between different contacts, which is conducive to the subsequent analysis of the influence of contacts.

[0023] In a possible implementation of the first aspect, the step of obtaining the current social influence weight of the target sender based on the contact social graph and the improved PageRank algorithm includes:

[0024] The PageRank algorithm is improved by using the directed weighted graph and the time decay factor. The improved PageRank algorithm is expressed as follows:

[0025] ;

[0026] in, Indicates that at time step Contact Person PageRank value, represents the damping coefficient, represents the total number of all contacts in the directed weighted graph, Represents a directed weighted graph with contacts Other directly connected contacts, Represents a directed weighted graph with contacts Other directly connected contacts, Indicates contact person and Contacts The length of time since the last interaction. represents the time decay factor, Indicates contact person Point to Contact The edge weights of Indicates contact person Point to Contact The edge weight represents the interaction strength. Indicates that at time step Contact Person The PageRank value represents the social influence weight;

[0027] The social relationship data related to the target sender is obtained, and the social influence weight of the target sender is determined by the improved PageRank algorithm.

[0028] In the above scheme, the influence of the email sender reflects the importance of the email in one dimension. By comprehensively considering the interaction intensity and time decay factor, and using the improved PageRank algorithm to quantify the sender's influence in the contact social graph, it is possible to capture the dynamic changes of influence in real time, making the calculated social influence weight more consistent with the user's cognitive logic, thereby improving the accuracy of subsequent predictions of email importance.

[0029] In a possible implementation of the first aspect, the step of extracting at least one core tag from email content data within a previous time period using a preset large model includes:

[0030] Using the BERT model to encode the email content data, and outputting semantic representations of multiple keywords based on the hidden layer;

[0031] Performing weighted fusion processing on the semantic representation through an attention mechanism to extract multiple keyword tags;

[0032] The keyword tags are vectorized using the BGE-M3 model;

[0033] Use the DBSCAN clustering algorithm to perform cluster analysis on the vectorized keyword tags and identify all core tags;

[0034] The positive feedback frequency of each core tag is counted, and the core tags are sorted from high to low according to the positive feedback frequency to obtain the top K core tags; wherein K is a positive integer, and the positive feedback frequency of the core tag represents the number of positive feedback behaviors of the target user on the email containing the core tag in the previous time period.

[0035] In the above scheme, analyzing the semantics of email content is conducive to identifying the importance of emails. The BERT model is used in combination with the attention mechanism to extract keyword tags of email content. The keyword semantics are deeply understood based on the associated context, and the implicit importance of email content can be identified. The BGE-M3 model is used to vectorize keyword tags to obtain vector representations that are more in line with the label clustering task and semantic similarity calculation requirements. The core label set is screened through DBSCAN clustering. In the subsequent process, the importance of the email can be intuitively quantified by obtaining the degree of match between the email content and the core label set.

[0036] In a possible implementation of the first aspect, the important mail identification method further includes:

[0037] Regularly constructing a training data set based on the user behavior data and the email content data in the previous time period; the training data set includes multiple email samples, each email sample corresponds to multiple behavior labels and multiple real semantic labels;

[0038] Inputting the email sample into a behavior prediction model and a semantic matching model, obtaining a predicted probability of the user performing the feedback behavior on the email sample based on the behavior prediction model, and obtaining a matching probability between the email sample and the core tag based on the semantic matching model;

[0039] A current joint loss is obtained according to the behavior label, the true semantic label, the predicted probability, the matching probability, and a joint loss function, and the model parameters of the behavior prediction model and the model parameters of the semantic matching model are updated by minimizing the current joint loss; wherein the expression of the joint loss function is:

[0040] ;

[0041] in, represents the joint loss, m represents the email sample, represents the training dataset, represents the loss function of the behavior prediction model, represents the loss function of the semantic matching model, represents L2 regularization, Indicates a behavior label. Both represent true semantic labels, represents the predicted probability of the behavior prediction model, Represents the matching probability of the semantic matching model.

[0042] In the above scheme, the models involved in the entire important email identification method are jointly optimized to balance the weights of user behavior and email content in email importance and prevent overfitting of a single model. By designing a loss function that can jointly optimize the behavior prediction model and the semantic matching model, and using more recent training data to update the model parameters, the model can adapt to the dynamic changes of user behavior and semantic labels in a timely manner, which is conducive to improving the accuracy of email importance identification.

[0043] In a possible implementation of the first aspect, the current behavior influence weight of the positive feedback behavior is determined by the following formula:

[0044] ;

[0045] ;

[0046] in, represents the behavior influence weight, represents the i-th feedback behavior, represents the basic weight of the i-th feedback behavior, represents the number of occurrences of the i-th feedback behavior, It represents the time length from the kth occurrence of the i-th feedback behavior to the present time, represents the time decay factor, N represents the number of feedback behaviors, represents the feedback behavior set, represents the positive feedback behavior set, Represents a set of negative feedback behaviors.

[0047] In a possible implementation of the first aspect, the interaction strength between two contacts is calculated using the following formula:

[0048] ;

[0049] in, represents the interaction strength, and Represents contacts respectively and Contacts , Indicates contact person and Contacts The number of interactions in the previous time period, Indicates the length of the previous time period. Indicates contact person and Contacts The interaction frequency, Indicates contact person and Contacts The basic relationship weight.

[0050] In a possible implementation of the first aspect, the loss function of the behavior prediction model is a cross entropy loss function, and the expression of the loss function of the behavior prediction model is:

[0051] ;

[0052] in, represents the behavior prediction loss, Indicates the Behavior label for each feedback behavior, Indicates the The predicted probability of a feedback behavior.

[0053] In a possible implementation manner of the first aspect, the loss function of the semantic matching model is a contrast loss function, and the loss function of the semantic matching model is expressed as:

[0054] ;

[0055] in, represents the semantic matching loss, Indicates the The real semantic labels that match the email content, Indicates the The semantic similarity between the email content and the core tag is expressed as m, and m represents the margin value.

[0056] In a second aspect, an embodiment of the present application provides an important mail identification system, including:

[0057] A behavior analysis module is configured to determine the influence weight of the target user's positive feedback behavior on the current behavior based on user behavior data in the previous time period; wherein the user behavior data includes at least one feedback behavior of the target user on the target email, and the feedback behavior includes positive feedback behavior and negative feedback behavior;

[0058] A social analysis module is configured to calculate the interaction strength between any two contacts based on social relationship data from the previous time period and construct a contact social graph. The module then uses the contact social graph and an improved PageRank algorithm to determine the target sender's current social influence weight. The interaction strength includes the base relationship weight and / or interaction frequency between two contacts, where the contacts represent the sender and recipient of an email.

[0059] a semantic analysis module, configured to extract at least one core tag from the email content data within the previous time period using a preset large model, then calculate the semantic similarity between each core tag and the target email, and determine the maximum semantic similarity calculated as the semantic matching weight of the target email;

[0060] A scoring module, configured to obtain a comprehensive email score for the target email based on the behavioral influence weight, the social influence weight, and the semantic matching weight;

[0061] A determination module is configured to determine whether the target email is currently an important email to the target user based on a dynamic determination threshold and the email comprehensive score; wherein, when the email comprehensive score is greater than the dynamic determination threshold, the target email is currently an important email to the target user, otherwise it is an ordinary email; the dynamic determination threshold is determined based on the email comprehensive scores of all emails in the previous time period.

[0062] In a third aspect, an embodiment of the present application provides an intelligent mailbox system, including the important mail identification system and mail push system described in the second aspect above;

[0063] The important email identification system is used to determine whether the target email is currently an important email to the target user;

[0064] The mail push system is used to push reminders or send ordinary emails to the target emails according to the identification results of the important email identification system.

[0065] In a fourth aspect, an embodiment of the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the important email identification method described in any one of the first aspects above is implemented.

[0066] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and is characterized in that when the computer program is executed by a processor, it implements the important email identification method described in any one of the above-mentioned first aspects.

[0067] In a sixth aspect, an embodiment of the present application provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute the important email identification method described in any one of the above-mentioned first aspects.

[0068] It can be understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0070] Figure 1 This is a flowchart of an important email identification method provided by an embodiment of the present application;

[0071] Figure 2 This is a flowchart of step S11 in the important mail identification method provided in one embodiment of the present application;

[0072] Figure 3 This is a flowchart of step S12 in the important mail identification method provided in one embodiment of the present application;

[0073] Figure 4 This is a flowchart of step S13 in the important mail identification method provided in one embodiment of the present application;

[0074] Figure 5 This is a flowchart of an important mail identification method provided by another embodiment of the present application;

[0075] Figure 6 This is a schematic diagram of the structure of an important mail identification system provided in one embodiment of the present application;

[0076] Figure 7 This is a schematic diagram of the structure of the smart mailbox system provided by one embodiment of the present application;

[0077] Figure 8 It is a structural diagram of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0078] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0079] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0080] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0081] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0082] See also Figure 1 The present application embodiment provides a flowchart of an important mail identification method. As an example and not a limitation, the method may include the following steps:

[0083] S11. Determine the influence weight of the target user's positive feedback behavior on the current behavior based on the user behavior data in the previous time period.

[0084] The user behavior data includes at least one target user's feedback behavior on the target email, and the feedback behavior includes positive feedback behavior and negative feedback behavior.

[0085] In one possible implementation, some operations performed by users on emails are classified, where positive feedback behaviors include performing operations such as "marking unread", "starring", "pinning", "clicking to browse", and "setting to process later" on emails, and negative feedback behaviors include performing operations such as "unstarring", "unpinning", "marking read", and "deleting" on emails.

[0086] In this embodiment, all types of positive feedback behaviors are defined as a positive feedback behavior set, and all types of negative feedback behaviors are defined as a negative feedback behavior set. The union of the positive feedback behavior set and the negative feedback behavior set can be defined as a feedback behavior set.

[0087] S12. Calculate the interaction intensity between any two contacts based on the social relationship data in the previous time period and construct a contact social graph. Then, obtain the current social influence weight of the target sender based on the contact social graph and the improved PageRank algorithm.

[0088] The interaction strength includes the basic relationship weight and / or interaction frequency between two contacts, the contacts represent the sender and the recipient of the email, and the target sender refers to the sender of the target email.

[0089] In this embodiment, contact social interaction refers to email exchanges. For any two contacts p and q, whether contact p sends an email to contact q or contact q sends an email to contact p, it will be counted as an interaction. The more interactions a contact has with other contacts, the greater the influence of the contact in the information dissemination process, and the more important the emails sent by the contact are.

[0090] S13. Extract at least one core tag from the email content data in the previous time period using a preset large model, then calculate the semantic similarity between each core tag and the target email, and determine the calculated maximum semantic similarity as the semantic matching weight of the target email.

[0091] S14. Obtain a comprehensive email score for the target email based on the behavioral influence weight, the social influence weight, and the semantic matching weight.

[0092] S15: Determine whether the target email is currently important to the target user based on the dynamic determination threshold and the email comprehensive score.

[0093] When the comprehensive score of the email is greater than the dynamic judgment threshold, the target email is currently an important email to the target user; otherwise, it is an ordinary email.

[0094] In this embodiment, the dynamic determination threshold is determined based on the comprehensive email scores of all emails in the previous time period.

[0095] This embodiment provides a method for identifying important emails. Dynamic weights such as behavioral influence weights, social influence weights, and semantic matching weights are determined by using multi-dimensional dynamic data such as recent user behavior data, email content data, and social relationship data. The impact of user behavior, email content, and social relationships on email importance is analyzed from different dimensions, significantly reducing the misjudgment rate of email importance. Whether the current email is important to the user is determined based on dynamic weights and dynamic judgment thresholds. Both dynamic weights and dynamic judgment thresholds are adaptively updated based on multi-dimensional dynamic data, can adapt to changes in user needs in real time, and can accurately and individually intelligently identify important emails, thereby helping users quickly locate important emails, saving processing time, and helping to improve the quality and efficiency of information exchange.

[0096] Alternatively, see Figure 2 One implementation of step S11 may include:

[0097] S111. Determine the number of occurrences of each feedback behavior and a first time length from the occurrence of each feedback behavior to the present based on the user behavior data.

[0098] In this embodiment, feedback actions within user behavior data within a time period are categorized and counted. For example, if a user stars an email three times, the star action is recorded as occurring three times. Furthermore, user behavior decays over time, meaning that more recent feedback actions have a greater impact, while older ones have a smaller impact. Therefore, the time elapsed between each feedback action is also recorded, typically in days.

[0099] S112: Calculate the current behavior impact weight of the positive feedback behavior according to a preset time decay factor, the basic weight of the positive feedback behavior, the number of occurrences, and the first time length.

[0100] This embodiment provides a method for identifying important emails. Positive feedback behavior reflects the user's recognition of the importance of the email in one dimension. At the same time, a time decay factor is introduced to give more recent user behavior data a higher weight, so that the calculated behavior impact weight is more consistent with the user's cognitive logic, thereby improving the accuracy of subsequent email importance predictions.

[0101] Optionally, an implementation of step S112 includes determining the current behavior influence weight of the positive feedback behavior using the following formula:

[0102] ;

[0103] ;

[0104] in, represents the behavior influence weight, represents the i-th feedback behavior, represents the basic weight of the i-th feedback behavior, represents the number of occurrences of the i-th feedback behavior, It represents the time length from the kth occurrence of the i-th feedback behavior to the present time, represents the time decay factor, N represents the number of feedback behaviors, represents the feedback behavior set, represents the positive feedback behavior set, Represents a set of negative feedback behaviors.

[0105] In one possible implementation, ,in, Indicates the time since the action occurred. is an adjustable parameter used to control the decay speed. The larger the value of , the faster the weight of the old behavior decreases. In this embodiment, A value of 0.01 means that the weight decays by about 1% every day.

[0106] In this embodiment, Specifically represents the behavioral impact weight of a positive feedback behavior. A user may implement multiple different positive feedback behaviors on a certain email. Therefore, when calculating the behavioral impact weight of the positive feedback behavior on the email, the behavioral impact weights of different types of positive feedback behaviors can be accumulated and combined to obtain .

[0107] Alternatively, see Figure 3 One implementation of step S12 may include:

[0108] S121. Define the contact social graph as a directed weighted graph.

[0109] The nodes of the directed weighted graph represent contacts, and the edges of the directed weighted graph represent the relationship between two contacts. Specifically, we define the directed weighted graph , where V represents a set of nodes and nodes represent contacts; represents a set of directed edges. A directed edge indicates the link between a sender and a recipient. In this embodiment, a contact can be both a sender and a recipient, and there may be directed edges in both directions between two nodes. P represents a set of node weights. Node weights represent the influence of a contact and are the values ​​that need to be calculated and quantified in this embodiment. Represents a set of edge weights. The edge weight represents the interaction strength between two contacts and is also a value that needs to be calculated and quantified in this embodiment. At the same time, the edge weight also affects the value of the node weight.

[0110] S122: Acquire, based on the social relationship data, the number of interactions between any two contacts in a previous time period, a second time length of the previous time period, and a basic relationship between the two contacts.

[0111] In this embodiment, the basic relationship between two contacts represents the relationship type of the sender to the recipient. As an example, assuming it is a mailbox used for work, for the user, the basic relationships involved may include superiors, customers, partners, and colleagues, among which emails sent by superiors and customers are usually more important and have a larger weight value.

[0112] S123: Calculate the interaction frequency between the two contacts according to the number of interactions and the second time length, and obtain the basic relationship weight corresponding to the basic relationship from a preset relationship weight comparison table.

[0113] S124. Determine the interaction strength between any two contacts based on the interaction frequency and the basic relationship weight, and represent the interaction strength as an edge weight of a directed weighted graph.

[0114] S125. Improve the PageRank algorithm using directed weighted graphs and time decay factors.

[0115] Among them, the improved PageRank algorithm is expressed as:

[0116] ;

[0117] in, Indicates that at time step Contact Person PageRank value, represents the damping coefficient, represents the total number of all contacts in the directed weighted graph, Represents a directed weighted graph with contacts Other directly connected contacts, Represents a directed weighted graph with contacts Other directly connected contacts, Indicates contact person and Contacts The length of time since the last interaction. represents the time decay factor, Indicates contact person Point to Contact The edge weight represents the interaction strength. Indicates contact person Point to Contact The edge weights of Indicates that at time step Contact Person The PageRank value represents the social influence weight.

[0118] S126. Acquire social relationship data related to the target sender, and determine the social influence weight of the target sender using an improved PageRank algorithm.

[0119] This embodiment provides a method for identifying important emails. It combines interaction frequency and underlying relationships to quantify the interaction intensity between two contacts. For email recipients, emails from senders with higher interaction intensity are more important. Constructing a contact social graph based on a directed weighted graph not only clarifies the direction of information flow but also accurately expresses complex relationships, thereby intuitively clarifying the interaction intensity between different contacts and facilitating subsequent analysis of the contact's influence. Furthermore, the influence of the email sender reflects the importance of the email in one dimension. By comprehensively considering interaction intensity and time decay factors, and utilizing an improved PageRank algorithm to quantify the sender's influence in the contact social graph, this method can capture the dynamic changes in influence in real time, making the calculated social influence weight more consistent with user cognitive logic, thereby improving the accuracy of subsequent predictions of email importance.

[0120] Optionally, an implementation of step S124 may include calculating the interaction strength between two contacts using the following formula:

[0121] ;

[0122] in, represents the interaction strength, and Represents contacts respectively and Contacts , Indicates contact person and Contacts The number of interactions in the previous time period, Indicates the length of the previous time period. Indicates contact person and Contacts The interaction frequency, Indicates contact person and Contacts The basic relationship weight.

[0123] In this embodiment, the log function is used to make the influence weight of each additional interaction gradually decrease as the number of interactions increases, that is, to achieve the diminishing marginal effect, thereby preventing high-frequency interactions from causing Of course, this embodiment can also use square root function, exponential decay function, inverse proportional function and logistic function to achieve marginal effect reduction.

[0124] In one possible implementation, the damping coefficient in step S125 is typically set to 0.85, which means that after receiving an email from the target sender, the user has a 0.85 probability of continuing to view or process the next email from the target sender, while the user has a 1-d = 0.15 probability of stopping viewing or processing the email from the target sender and randomly selecting another email from another sender. Of course, the damping coefficient in this embodiment may also take other values ​​between 0 and 1.

[0125] As an example, for a work mailbox, the relationship weight comparison table in step S124 may be as shown in Table 1.

[0126]

[0127] Of course, this embodiment is not limited to the above basic relationships and basic relationship weights, and can be set according to actual needs. When there is no basic relationship in the relationship weight comparison table between the sender and the user, the basic relationship weight The value of is zero.

[0128] Alternatively, see Figure 4 One implementation of step S13 specifically includes:

[0129] S131. Use the BERT model to encode the email content data and output semantic representations of multiple keywords based on the hidden layer.

[0130] In a possible implementation, before encoding the email content data, a word segmenter is used to segment the email content data to obtain multiple semantic information words. The semantic representation refers to the semantic information word vector of each semantic information word.

[0131] For each semantic information word, the BERT model outputs a series of hidden layer representations, recorded as ,in, represents the hidden layer output of the mth semantic information word, L represents the length of the context contained in the semantic information word, It represents the semantic vector representation at the nth position, capturing the semantic information of the semantic information word in the context.

[0132] S132. Perform weighted fusion processing on the semantic representation through the attention mechanism to extract multiple keyword tags.

[0133] In one possible implementation, the formula To calculate the output of each hidden layer The attention weight ,in, represents the query vector, which is a learnable parameter used to interact with the hidden layer output; and are the weight matrix and bias vector, which are also learnable parameters and are used to adjust the hidden layer output; It is a nonlinear activation function, specifically the hyperbolic tangent function, which is used to increase the expressive power of the model; The function is used to convert the dot product result into a probability distribution, ensuring that the sum of all attention weights is 1.

[0134] In one possible implementation, the attention weights calculated are , output to the hidden layer Perform weighted summation to obtain the semantic representation of semantic information words , specifically:

[0135] ;

[0136] In one possible implementation, the semantic representation of multiple semantic information words can be obtained based on the calculation results of the attention mechanism. Multiple keyword tags are screened out, and the tags are label vectors, that is, semantic representation vectors.

[0137] S133. Use the BGE-M3 model to vectorize keyword tags.

[0138] In this embodiment, the BGE-M3 model is optimized for specific tasks (such as semantic similarity calculation or label clustering). The BGE-M3 model further maps keyword labels into 768-dimensional vectors, which can better capture semantic information and generate vector representations that better meet task requirements.

[0139] S134. Use the DBSCAN clustering algorithm to perform cluster analysis on the vectorized keyword tags and identify all core tags.

[0140] In this embodiment, the DBSCAN clustering algorithm is expressed as:

[0141] ;

[0142] Among them, represents the core tag, for the keyword tag , if in its Neighborhood (specifically referring to As the center, With radius Euclidean distance Less than or equal to Keyword tags The number of is greater than or equal to min_sample, then the keyword label For the core tag.

[0143] S135 , counting the positive feedback frequency of each core tag, and sorting the core tags from high to low according to the positive feedback frequency, to obtain the top K core tags.

[0144] Where K is a positive integer, and the positive feedback frequency of the core tag represents the number of positive feedback behaviors of the target user on the email with the core tag in the previous time period.

[0145] S136 : Calculate the semantic similarity between each core tag and the email content of the target email, and determine the calculated maximum semantic similarity as the semantic matching weight of the target email.

[0146] In one possible implementation, K core tags are used as a core tag set. The cosine similarity is used to calculate the semantic similarity between each core tag and the target email content. The core tag with the highest similarity to the email content is found in the core tag set, and its semantic similarity is obtained as the semantic matching weight. Specifically:

[0147] ;

[0148] in, Represents email D and core tag set The semantic matching weight of is the semantic representation of the email content, It is the semantic representation of the core tag.

[0149] This embodiment provides a method for identifying important emails. Analyzing the semantics of email content is beneficial for identifying the importance of emails. The BERT model is used in combination with the attention mechanism to extract keyword tags from the email content. The keyword semantics are deeply understood based on the associated context, and the implicit importance of the email content can be identified. The BGE-M3 model is used to vectorize the keyword tags to obtain a vector representation that is more in line with the requirements of tag clustering tasks and semantic similarity calculations. The core tag set is screened through DBSCAN clustering, and the importance of the email is intuitively quantified by obtaining the maximum match between the email content and the core tag set.

[0150] Optionally, an implementation of step S14 may include calculating the email comprehensive score using the following formula, specifically:

[0151] ;

[0152] in, This means the sender of email D. 、 、 The calculation details of are described in the above embodiment and will not be repeated here.

[0153] In this embodiment, , , All of them are adjustable hyperparameters, which can be used to meet It is set according to actual needs under certain conditions to balance the weight values ​​of the three different dimensions of semantics, behavior, and sociality.

[0154] As an example, Usually the value is 0.5, Usually the value is 0.3, The value is usually 0.2.

[0155] Optionally, an implementation of step S15 may include adaptively adjusting the dynamic decision threshold every time period based on the target user's comprehensive email score in the previous time period to avoid misjudgment caused by a fixed threshold. Specifically:

[0156] ;

[0157] It should be noted that t does not represent a specific unit of time, but is equivalent to the number of iterative updates. It can mean that the dynamic judgment threshold is updated once a week, that is, there are 7 days between t and t-1, or it can mean that the dynamic judgment threshold is updated once a month, that is, there are 30 days between t and t-1.

[0158] In this embodiment, Indicates the current dynamic judgment threshold, Indicates the average comprehensive score of emails in the previous time period. Indicates the standard deviation of the comprehensive score of emails in the previous time period. represents the standard normal distribution, The value corresponding to the quantile set according to actual needs, for example, the 0.9 quantile corresponds to approximately 1.28, which means that the dynamic judgment threshold is guaranteed to be higher than 90% of the historical email comprehensive scores. Similarly, the 0.95 quantile corresponds to approximately 1.645, and the 0.99 quantile corresponds to approximately 2.326. It should be noted that The higher the threshold, the stricter the threshold. The value of can flexibly control the strictness of the threshold in different scenarios, thereby better meeting the personalized needs of users.

[0159] In one possible implementation, see Figure 5 , the important mail identification method also includes:

[0160] S16. Regularly construct a training data set based on user behavior data and email content data in the previous time period.

[0161] The training data set includes multiple email samples, each of which corresponds to multiple behavioral labels and multiple real semantic labels.

[0162] S17. Input the email sample into the behavior prediction model and the semantic matching model, obtain the predicted probability of the user's feedback behavior on the email sample based on the behavior prediction model, and obtain the matching probability of the email sample and the core tag based on the semantic matching model.

[0163] S18. Obtain the current joint loss based on the behavior label, the true semantic label, the predicted probability, the matching probability, and the joint loss function, and update the model parameters of the behavior prediction model and the model parameters of the semantic matching model by minimizing the current joint loss.

[0164] Among them, the expression of the joint loss function is:

[0165] ;

[0166] in, represents the joint loss, m represents the email sample, represents the training dataset, represents the loss function of the behavior prediction model, represents the loss function of the semantic matching model, represents L2 regularization, Indicates a behavior label. Both represent true semantic labels, represents the predicted probability of the behavior prediction model, Represents the matching probability of the semantic matching model.

[0167] This embodiment provides a method for identifying important emails, which jointly optimizes the models involved in the entire important email identification method, balances the weights of user behavior and email content in email importance, prevents overfitting of a single model, and designs a loss function that can jointly optimize the behavior prediction model and the semantic matching model, and uses more recent training data to update the model parameters, so that the model can adapt to the dynamic changes of user behavior and semantic labels in a timely manner, thereby helping to improve the accuracy of email importance identification.

[0168] In one possible implementation, the loss function of the behavior prediction model is a cross entropy loss function. Specifically, the loss function of the behavior prediction model is expressed as:

[0169] ;

[0170] in, represents the behavior prediction loss, Indicates the Behavior label for each feedback behavior, Indicates the The predicted probability of a feedback behavior.

[0171] In one possible implementation, the loss function of the semantic matching model is a contrast loss function. Specifically, the loss function of the semantic matching model is expressed as:

[0172] ;

[0173] in, represents the semantic matching loss, Indicates the The real semantic labels that match the email content, Indicates the The semantic similarity between the email content and the core label is expressed as follows: m represents the margin value, which is used to control the contribution of negative samples to the loss and is a constant greater than 0.

[0174] In this embodiment, every time window ,use The new data generated in the update model parameters, specifically:

[0175] ;

[0176] in, is the dynamic learning rate, represents the updated model parameters, represents the model parameters before updating, express New data generated within express About parameters gradient.

[0177] In this embodiment, The value is usually 30 days, and the learning rate The value ranges from 0.01 to 0.1, but this embodiment does not limit this.

[0178] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0179] Corresponding to the important mail identification method described in the above embodiment, Figure 6 A structural block diagram of an important mail identification system provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0180] Reference Figure 6 , this important mail identification system includes:

[0181] The behavior analysis module 11 is used to determine the current behavior influence weight of the target user's positive feedback behavior based on the user behavior data in the previous time period; wherein the user behavior data includes at least one feedback behavior of the target user on the target email, and the feedback behavior includes positive feedback behavior and negative feedback behavior.

[0182] The social analysis module 12 is used to calculate the interaction strength between any two contacts based on the social relationship data in the previous time period and construct a contact social graph, and then obtain the current social influence weight of the target sender based on the contact social graph and the improved PageRank algorithm; wherein the interaction strength includes the basic relationship weight and / or interaction frequency between the two contacts, and the contacts represent the sender and recipient of the email.

[0183] The semantic analysis module 13 is used to extract at least one core tag from the email content data in the previous time period using a preset large model, then calculate the semantic similarity between each core tag and the target email, and determine the calculated maximum semantic similarity as the semantic matching weight of the target email.

[0184] The scoring module 14 is configured to obtain a comprehensive email score for the target email based on the behavioral influence weight, the social influence weight, and the semantic matching weight.

[0185] Determination module 15 is configured to determine whether the target email is currently important to the target user based on a dynamic determination threshold and the email comprehensive score. If the email comprehensive score is greater than the dynamic determination threshold, the target email is currently important to the target user; otherwise, it is an ordinary email. The dynamic determination threshold is determined based on the comprehensive scores of all emails in the previous time period.

[0186] In some embodiments of the present application, the behavior analysis module 11 is specifically used to determine the number of occurrences of each feedback behavior and the first time length from the occurrence of each feedback behavior to the current time based on user behavior data; and calculate the current behavior impact weight of the positive feedback behavior based on a preset time attenuation factor, the basic weight of the positive feedback behavior, the number of occurrences and the first time length.

[0187] In some embodiments of the present application, the social analysis module 12 is specifically used to define the contact social graph as a directed weighted graph; wherein, the nodes of the directed weighted graph represent contacts, and the edges of the directed weighted graph represent the relationship between two contacts; based on the social relationship data, the number of interactions between any two contacts in the previous time period, the second time length of the previous time period, and the basic relationship between the two contacts are obtained; the interaction frequency of the two contacts is calculated according to the number of interactions and the second time length, and the basic relationship weight corresponding to the basic relationship is obtained from a preset relationship weight comparison table; the interaction strength between any two contacts is determined according to the interaction frequency and the basic relationship weight, and the interaction strength is represented as the edge weight of the directed weighted graph.

[0188] In some embodiments of the present application, the social analysis module 12 is further specifically used to improve the PageRank algorithm using a directed weighted graph and a time decay factor; obtain social relationship data related to the target sender, and determine the social influence weight of the target sender through the improved PageRank algorithm.

[0189] In some embodiments of the present application, the semantic analysis module 13 is specifically used to encode the email content data using the BERT model, and output the semantic representation of multiple keywords based on the hidden layer; perform weighted fusion processing on the semantic representation through the attention mechanism to extract multiple keyword tags; use the BGE-M3 model to vectorize the keyword tags; use the DBSCAN clustering algorithm to perform cluster analysis on the vectorized keyword tags to identify all core tags; count the positive feedback frequency of each core tag, and sort the core tags from high to bottom according to the positive feedback frequency to obtain the top K core tags; wherein K is a positive integer, and the positive feedback frequency of the core tag represents the number of positive feedback behaviors of the target user on the email containing the core tag in the previous time period.

[0190] In some embodiments of the present application, the important email identification system also includes an online learning module, which is specifically used to periodically construct a training data set based on user behavior data and email content data in the previous time period; the training data set includes multiple email samples, each email sample corresponds to multiple behavior labels and multiple true semantic labels; the email sample is input into the behavior prediction model and the semantic matching model, and the predicted probability of the user's feedback behavior on the email sample is obtained based on the behavior prediction model, and the matching probability of the email sample and the core label is obtained based on the semantic matching model; the current joint loss is obtained according to the behavior label, the true semantic label, the predicted probability, the matching probability and the joint loss function, and the model parameters of the behavior prediction model and the model parameters of the semantic matching model are updated by minimizing the current joint loss.

[0191] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment section and will not be repeated here.

[0192] The important mail identification system provided in the embodiment of the present application can be applied in the intelligent mailbox system. Figure 7 The following is a schematic block diagram of an intelligent mailbox system provided by the present application. As an example and not a limitation, the intelligent mailbox system includes:

[0193] An important email identification system 1 is used to determine whether a target email is currently important to a target user;

[0194] The mail push system 2 is used to push reminders or send ordinary emails to target emails according to the identification results of the important email identification system.

[0195] In some embodiments of the present application, the email push system 2 reminds users of important emails by placing the title and / or content of the email at the top.

[0196] In some embodiments of the present application, the email push system 2 may also be associated with a text message system to push important emails via text messages.

[0197] It should be noted that the important mail identification system 1 is not limited to being used in an intelligent mailbox system, but can also be integrated into an OA system, etc.

[0198] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0199] Figure 8 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present application. Figure 8 As shown, the electronic device 3 of the embodiment includes: at least one processor 30 ( Figure 8 Only one is shown in the figure) a processor, a memory 31, and a computer program 32 stored in the memory 31 and executable on the at least one processor 30. When the processor 30 executes the computer program 32, the steps in the above-mentioned embodiments of the method for detecting abnormalities in each power grid operation and maintenance image are implemented.

[0200] The electronic device 3 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The electronic device 3 may include, but is not limited to, a processor 30 and a memory 31. Those skilled in the art will understand that Figure 8 This is merely an example of the electronic device 3 and does not constitute a limitation on the electronic device 3 . The electronic device 3 may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device 3 may also include input and output devices, network access devices, etc.

[0201] The processor 30 may be a central processing unit (CPU), or may be another general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, etc.

[0202] In some embodiments, the memory 31 may be an internal storage unit of the electronic device 3, such as the hard drive or memory of the electronic device 3. In other embodiments, the memory 31 may also be an external storage device of the electronic device 3, such as a plug-in hard drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. equipped on the electronic device 3. Furthermore, the memory 31 may include both an internal storage unit of the electronic device 3 and an external storage device. The memory 31 is used to store an operating system, application programs, a boot loader, data, and other programs, such as the program code of a computer program. The memory 31 may also be used to temporarily store data that has been output or is about to be output.

[0203] An embodiment of the present application further provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the steps in each of the above-mentioned important email identification method embodiments can be implemented.

[0204] An embodiment of the present application provides a computer program product. When the computer program product is run on a mobile terminal, the mobile terminal can implement the steps in the above-mentioned important email identification method embodiments when executing the computer program product.

[0205] If the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned important mail identification method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a camera / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, removable hard drives, magnetic disks, or optical disks. In some jurisdictions, based on legislation and patent practice, computer-readable media cannot be electric carrier signals or telecommunication signals.

[0206] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0207] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0208] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A method for identifying important mail, characterized in that: include: Determining the influence weight of the target user's positive feedback behavior on the current behavior based on the user behavior data in the previous time period; wherein the user behavior data includes at least one feedback behavior of the target user on the target email, and the feedback behavior includes positive feedback behavior and negative feedback behavior; Calculating the interaction strength between any two contacts based on social relationship data from the previous time period and constructing a contact social graph. Then, using the contact social graph and an improved PageRank algorithm, deriving the target sender's current social influence weight. The interaction strength includes the base relationship weight and / or interaction frequency between the two contacts, where the contacts represent the sender and recipient of the email. Extracting at least one core tag from the email content data within the previous time period using a preset large model, then calculating the semantic similarity between each core tag and the target email, and determining the maximum semantic similarity obtained by calculation as the semantic matching weight of the target email; Obtaining a comprehensive email score for the target email based on the behavioral influence weight, the social influence weight, and the semantic matching weight; Whether the target email is currently an important email to the target user is determined based on a dynamic determination threshold and the email comprehensive score; wherein, when the email comprehensive score is greater than the dynamic determination threshold, the target email is currently an important email to the target user, otherwise it is an ordinary email; the dynamic determination threshold is determined based on the email comprehensive scores of all emails in the previous time period.

2. The important mail identification method according to claim 1, characterized in that: The step of determining the influence weight of the positive feedback behavior of the target user on the current behavior based on the user behavior data in the previous time period includes: Determine, based on the user behavior data, the number of occurrences of each feedback behavior and a first time length from the occurrence of each feedback behavior to the current time; The current behavior influence weight of the positive feedback behavior is calculated according to a preset time decay factor, the basic weight of the positive feedback behavior, the number of occurrences and the first time length.

3. The important mail identification method according to claim 1, characterized in that: The steps of calculating the interaction intensity between any two contacts based on the social relationship data in the previous time period and constructing the contact social graph include: The contact social graph is defined as a directed weighted graph; wherein the nodes of the directed weighted graph represent contacts, and the edges of the directed weighted graph represent the relationship between two contacts; Acquire, based on the social relationship data, the number of interactions between any two contacts in a previous time period, a second time length of the previous time period, and a basic relationship between the two contacts; Calculating the interaction frequency between the two contacts based on the number of interactions and the second time length, and obtaining a basic relationship weight corresponding to the basic relationship from a preset relationship weight comparison table; The interaction strength between any two contacts is determined according to the interaction frequency and the basic relationship weight, and the interaction strength is represented as an edge weight of the directed weighted graph.

4. The important mail identification method according to claim 3, characterized in that: The step of obtaining the current social influence weight of the target sender based on the contact social graph and the improved PageRank algorithm includes: The PageRank algorithm is improved by using the directed weighted graph and the time decay factor. The improved PageRank algorithm is expressed as follows: ； in, Indicates that at time step Contact Person PageRank value, represents the damping coefficient, represents the total number of all contacts in the directed weighted graph, Represents a directed weighted graph with contacts Other directly connected contacts, Represents a directed weighted graph with contacts Other directly connected contacts, Indicates contact person and Contacts The length of time since the last interaction. represents the time decay factor, Indicates contact person Point to Contact The edge weights of Indicates contact person Point to Contact The edge weight represents the interaction strength. Indicates that at time step Contact Person The PageRank value represents the social influence weight; The social relationship data related to the target sender is obtained, and the social influence weight of the target sender is determined by the improved PageRank algorithm.

5. The important mail identification method according to claim 1, characterized in that: The step of extracting at least one core tag from the email content data within the previous time period using the preset large model includes: Using the BERT model to encode the email content data, and outputting semantic representations of multiple keywords based on the hidden layer; Performing weighted fusion processing on the semantic representation through an attention mechanism to extract multiple keyword tags; The keyword tags are vectorized using the BGE-M3 model; Use the DBSCAN clustering algorithm to perform cluster analysis on the vectorized keyword tags and identify all core tags; The positive feedback frequency of each core tag is counted, and the core tags are sorted from high to low according to the positive feedback frequency to obtain the top K core tags; wherein K is a positive integer, and the positive feedback frequency of the core tag represents the number of positive feedback behaviors of the target user on the email containing the core tag in the previous time period.

6. The important mail identification method according to claim 1, characterized in that: The important mail identification method further includes: Regularly constructing a training data set based on the user behavior data and the email content data in the previous time period; the training data set includes multiple email samples, each email sample corresponds to multiple behavior labels and multiple real semantic labels; Inputting the email sample into a behavior prediction model and a semantic matching model, obtaining a predicted probability of the user performing the feedback behavior on the email sample based on the behavior prediction model, and obtaining a matching probability between the email sample and the core tag based on the semantic matching model; A current joint loss is obtained according to the behavior label, the true semantic label, the predicted probability, the matching probability, and a joint loss function, and the model parameters of the behavior prediction model and the model parameters of the semantic matching model are updated by minimizing the current joint loss; wherein the expression of the joint loss function is: ； in, represents the joint loss, m represents the email sample, represents the training dataset, represents the loss function of the behavior prediction model, represents the loss function of the semantic matching model, represents L2 regularization, Indicates a behavior label. Both represent true semantic labels, represents the predicted probability of the behavior prediction model, Represents the matching probability of the semantic matching model.

7. An important mail identification system, characterized in that: include: A behavior analysis module is configured to determine the influence weight of the target user's positive feedback behavior on the current behavior based on user behavior data in the previous time period; wherein the user behavior data includes at least one feedback behavior of the target user on the target email, and the feedback behavior includes positive feedback behavior and negative feedback behavior; A social analysis module is configured to calculate the interaction strength between any two contacts based on social relationship data from the previous time period and construct a contact social graph. The module then uses the contact social graph and an improved PageRank algorithm to determine the target sender's current social influence weight. The interaction strength includes the base relationship weight and / or interaction frequency between two contacts, where the contacts represent the sender and recipient of an email. a semantic analysis module, configured to extract at least one core tag from the email content data within the previous time period using a preset large model, then calculate the semantic similarity between each core tag and the target email, and determine the maximum semantic similarity calculated as the semantic matching weight of the target email; A scoring module, configured to obtain a comprehensive email score for the target email based on the behavioral influence weight, the social influence weight, and the semantic matching weight; A determination module is configured to determine whether the target email is currently an important email to the target user based on a dynamic determination threshold and the email comprehensive score; wherein, when the email comprehensive score is greater than the dynamic determination threshold, the target email is currently an important email to the target user, otherwise it is an ordinary email; the dynamic determination threshold is determined based on the email comprehensive scores of all emails in the previous time period.

8. An intelligent mailbox system, characterized in that: Including the important mail identification system and mail push system as claimed in claim 7; The important email identification system is used to determine whether the target email is currently an important email to the target user; The mail push system is used to push reminders or send ordinary emails to the target emails according to the identification results of the important email identification system.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the important mail identification method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the important mail identification method according to any one of claims 1 to 6 is implemented.

Images