Mail classification method and device, mail classification model adjustment method and device and electronic equipment

Abstract

The invention provides a mail classification method and device, a mail classification model adjustment method and device and electronic equipment. The method comprises the steps of obtaining target mail data and a vocabulary unit; generating a plurality of sub-data sets based on the vocabulary units; calculating an information gain value of each element in the sub-data set; dividing elements in the sub-data sets to generate a plurality of new sub-data sets; stopping dividing the sub-data sets until the number of samples in the sub-data sets generated after division is smaller than a preset threshold value; and obtaining candidate classifications corresponding to all the finally generated sub-data sets, and determining a target classification of the target mail based on the candidate classifications. According to the scheme, based on the information gain, the most distinctive feature is selected, the subset purity is gradually improved, meanwhile, the sample size threshold value is set, excessive splitting on a small sample is avoided, the model generalization ability is improved, accurate classification of the target mail is achieved, the method has good classification performance, over-fitting can be effectively prevented, and the classification efficiency is improved. And the practicability is high.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a method, apparatus, and electronic device for email classification, email classification model adjustment. Background Technology

[0002] With the rapid development of internet technology and the widespread use of email in personal communication, corporate office work, and business activities, the number of emails users receive daily is growing exponentially. These emails are complex and diverse, covering multiple categories such as work communication, customer service, marketing promotions, and private exchanges. Faced with this massive amount of email information, how to efficiently and accurately identify and categorize email types has become a key issue for improving user work efficiency and optimizing information processing workflows.

[0003] Traditional email classification methods primarily rely on rule-based filtering systems or simple keyword matching mechanisms. For example, they use preset keywords such as "invoice," "order," and "promotion" to determine whether an email is business or advertising. However, these methods have significant limitations: firstly, the rule base requires continuous manual maintenance and struggles to adapt to the diversity and dynamic changes in language; secondly, relying solely on keywords fails to understand semantic context, easily leading to misjudgments or omissions, resulting in limited classification accuracy.

[0004] In recent years, machine learning techniques have been widely applied in text classification, with methods such as decision trees, Naive Bayes, support vector machines, and deep learning models significantly improving classification performance. In particular, decision tree algorithms based on information gain are favored in practical applications due to their good interpretability, robustness to noisy data, and the fact that they do not require parameter tuning. This type of method recursively selects the features with the highest information gain to partition the dataset, gradually building a classification model to predict unknown samples. Summary of the Invention

[0005] This disclosure aims to at least partially address one of the technical problems in the related art.

[0006] Therefore, one objective of this disclosure is to propose a method for classifying emails.

[0007] The second objective of this disclosure is to propose a method for adjusting an email classification model.

[0008] The third objective of this disclosure is to propose a mail sorting device.

[0009] The fourth objective of this disclosure is to propose a mail classification model adjustment device.

[0010] The fifth objective of this disclosure is to provide an electronic device.

[0011] The sixth objective of this disclosure is to provide a non-transitory computer-readable storage medium.

[0012] The seventh objective of this disclosure is to provide a computer program product.

[0013] To achieve the above objectives, a first aspect of this disclosure proposes an email classification method, comprising: acquiring target email data of target emails to be classified, and generating lexical units of the target emails based on the email data; generating multiple subsets based on the lexical units, each subset including multiple elements; calculating the information gain value of each element in any subset; dividing the elements in the subset based on the information gain value to generate multiple new subsets; repeating the above steps of calculating the information gain value of each element in any subset and subsequent steps until the number of samples in the subsets generated after division is less than a preset threshold, at which point the division of elements in the corresponding subsets is stopped; and continuing the division until all elements in all subsets are no longer divided, obtaining candidate classifications corresponding to each of the finally generated subsets, and determining the target classification of the target email based on the candidate classifications.

[0014] According to one embodiment of this disclosure, generating the vocabulary units of the target email based on the email data includes: performing word segmentation processing on the email data to generate the vocabulary units of the target email.

[0015] According to one embodiment of this disclosure, generating multiple sub-datasets based on the lexical units includes: calculating the information gain value of any lexical unit; and processing all lexical units based on the information gain value to generate multiple sub-datasets.

[0016] According to one embodiment of this disclosure, the step of processing all lexical units based on the information gain value to generate multiple sub-data sets includes: determining multiple optimal feature data based on the information gain value; and dividing all lexical units based on the optimal feature data to generate multiple sub-data sets.

[0017] According to one embodiment of this disclosure, determining multiple optimal feature data based on the information gain value includes: sorting the information gain values ​​corresponding to all lexical units in order of approximate smaller values ​​to generate a lexical gain sequence; selecting the first N target lexical units from the lexical gain sequence based on a preset selection rule, and determining N optimal feature data based on the first N target lexical units, where N is an integer greater than 1.

[0018] According to one embodiment of this disclosure, calculating the information gain value of a target object, where the target object is each element in the subset or the lexical unit, includes: performing feature segmentation on the target dataset where the target object is located based on the target object to obtain a segmented dataset corresponding to the target object, wherein when the target object is an element, the target dataset is the subset containing the element, or when the target object is a lexical unit, the target dataset is a set composed of all lexical units; calculating a first information entropy of the target dataset where the target object is located, and calculating a second information entropy of the segmented dataset; and calculating the information gain value of the target object based on the first information entropy, the second information entropy, the number of samples corresponding to the target dataset, and the number of samples corresponding to the segmented dataset.

[0019] According to one embodiment of this disclosure, the step of dividing the elements in the subset based on the information gain value to generate multiple new subsets includes: sorting the information gain values ​​corresponding to all elements in order of approximate smaller values ​​to generate an element gain sequence; selecting the first M target elements from the element gain sequence based on a preset selection rule, and determining M optimal features based on the first M target elements, where M is an integer greater than 1; and dividing the elements in the subset based on the M optimal features to generate M new subsequences.

[0020] To achieve the above objectives, a second aspect of this disclosure proposes a method for adjusting an email classification model, comprising: obtaining an initial email classification model to be trained, wherein the initial email classification model is used to calculate information gain values ​​for subsets of target emails to be classified, and to divide the elements in the subsets based on the information gain values ​​to generate multiple new subsets; then repeating the above process of calculating information gain values ​​for subsets of target emails to be classified until the number of samples in the generated subsets after division is less than a preset threshold, stopping the division of elements in the corresponding subsets, and stopping the division of elements in all subsets; obtaining candidate classifications corresponding to each of the finally generated subsets; determining the target classification of the target email based on the candidate classifications; performing performance analysis on the initial email classification model to obtain comprehensive parameters of the initial email classification model; and adjusting a preset threshold of the initial email classification model based on the comprehensive parameters to generate a target email classification model.

[0021] According to one embodiment of this disclosure, adjusting the preset threshold of the initial email classification model based on the comprehensive parameters includes: obtaining a preset value of the comprehensive parameters; comparing the comprehensive parameters with the preset value of the comprehensive parameters; and adjusting the preset threshold of the initial email classification model in response to the comparison result satisfying the adjustment condition.

[0022] According to one embodiment of this disclosure, the method further includes: adjusting a preset threshold of the initial email classification model using a large model, and obtaining the comprehensive parameters of the adjusted initial email classification model; repeating the above steps until the comprehensive parameters of the adjusted initial email classification model no longer meet the adjustment conditions, and outputting the target email classification model.

[0023] According to one embodiment of this disclosure, the comprehensive parameters include one or more of the following: email classification accuracy parameter, email classification precision parameter, email classification processing speed parameter, system daily maximum email throughput parameter, email open rate parameter, email click-through rate parameter, email reply rate parameter, system CPU utilization parameter, system memory usage parameter, and system crash count parameter.

[0024] To achieve the above objectives, a third aspect of this disclosure provides an email classification apparatus, comprising: an acquisition module for acquiring target email data of a target email to be classified, and generating lexical units of the target email based on the email data; a generation module for generating multiple subsets based on the lexical units, the subsets including multiple elements; a calculation module for calculating the information gain value of each element in any subset; a partitioning module for partitioning the elements in the subsets based on the information gain values ​​to generate multiple new subsets; a looping module for repeating the above steps of calculating the information gain value of each element in any subset and subsequent steps until the number of samples in the subsets generated after partitioning is less than a preset threshold, at which point the partitioning of elements in the corresponding subsets is stopped; and a classification module for obtaining candidate classifications corresponding to each of the finally generated subsets until the partitioning of elements in all subsets is stopped, and determining the target classification of the target email based on the candidate classifications.

[0025] To achieve the above objectives, a fourth aspect of this disclosure provides an email classification model adjustment device, comprising: a receiving module, configured to acquire an initial email classification model to be trained, wherein the initial email classification model is used to calculate information gain values ​​for subsets of target emails to be classified, and to divide the elements in the subsets based on the information gain values ​​to generate multiple new subsets, and then repeat the above process of calculating information gain values ​​for subsets of target emails to be classified until the number of samples in the generated subsets after division is less than a preset threshold, at which point the division of elements in the corresponding subsets is stopped, until the division of elements in all subsets is stopped, and candidate classifications corresponding to each of the finally generated subsets are acquired, and the target classification of the target email is determined based on the candidate classifications; an analysis module, configured to perform performance analysis on the initial email classification model to obtain comprehensive parameters of the initial email classification model; and an adjustment module, configured to adjust a preset threshold of the initial email classification model based on the comprehensive parameters to generate a target email classification model.

[0026] To achieve the above objectives, a fifth aspect of this disclosure provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to implement the mail classification method as described in the first aspect of this disclosure.

[0027] To achieve the above objectives, a sixth aspect of this disclosure provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to implement the mail classification method as described in the first aspect of this disclosure.

[0028] To achieve the above objectives, a seventh aspect of this disclosure provides a computer program product including a computer program that, when executed by a processor, implements the mail classification method as described in the first aspect of this disclosure.

[0029] This disclosed scheme selects the most discriminative features based on information gain, gradually improves subset purity, and sets a sample size threshold to avoid excessive splitting on small samples, thereby improving the model's generalization ability and achieving accurate classification of target emails. This method not only has good classification performance but also effectively prevents overfitting and has strong practicality. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of an email classification method according to one embodiment of the present disclosure; Figure 2 This is a schematic diagram of another mail classification method according to one embodiment of the present disclosure; Figure 3 This is a schematic diagram of another mail classification method according to one embodiment of the present disclosure; Figure 4 This is a schematic diagram of an email classification model adjustment method according to one embodiment of the present disclosure; Figure 5 This is a schematic diagram of a mail sorting device according to one embodiment of the present disclosure; Figure 6 This is a schematic diagram of an email classification model adjustment device according to one embodiment of the present disclosure; Figure 7 This is a schematic diagram of an electronic device according to one embodiment of the present disclosure. Detailed Implementation

[0031] Embodiments of this disclosure are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting this disclosure.

[0032] The acquisition, storage, use, and processing of data in this disclosed technical solution all comply with the relevant provisions of relevant laws and regulations.

[0033] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0034] Current technologies using machine learning for email classification still face several challenges in practical applications. First, when the dataset is large and the feature dimensions are high, decision trees are prone to overfitting without setting reasonable termination conditions—that is, the model focuses excessively on a few exceptional cases in the training samples, leading to a decline in generalization ability. Second, some sub-branches may contain only a very small number of samples during the splitting process. Continuing to split in this case is not only computationally expensive, but the splitting results also lack statistical significance, severely affecting classification stability.

[0035] To address the aforementioned problems, this disclosure proposes an email classification method. Figure 1 This is a schematic diagram of an email classification method according to one embodiment of the present disclosure, such as... Figure 1 As shown, this email classification method includes the following steps: S101, Obtain the target email data of the target emails to be classified, and generate the vocabulary units of the target emails based on the email data.

[0036] The email classification method of the embodiments of the present application can be applied to scenarios of email classification and automatic processing. The execution subject of the email classification of the embodiments of the present application can be the email classification device of the embodiments of the present application, and this email classification device can be set on an electronic device.

[0037] In the embodiments of the present disclosure, the target email to be classified can be processed to generate the lexical units of the target email.

[0038] It should be noted that there can be multiple methods for processing the target email to be classified to generate the lexical units of the target email, and no limitation is made here.

[0039] In a possible implementation manner, the target email to be classified can be processed by a word segmentation processing model to obtain the lexical units of the target email. This word segmentation processing model can be pre-trained and stored in the storage space of the electronic device for convenient retrieval and use when needed.

[0040] In another possible implementation manner, the email text of the target email to be classified can also be scanned from left to right. Each time, the longest possible word is taken to match the words in the dictionary. If the match is successful, it is segmented out; otherwise, one character is removed and the attempt continues, and the continuous text is segmented into meaningful lexical units.

[0041] It should be noted that before generating the lexical units of the target email based on the email data, the text corresponding to the target email to be classified can also be preprocessed.

[0042] In a possible implementation manner, the HTML tags can be removed from the target email to be classified: Since the email content often contains text in HTML format, these tags are irrelevant to the analysis of the text content. Therefore, tools such as regular expressions are needed to remove these HyperText Markup Language Tags (HTMLTags).

[0043] Special symbols and stop words can be removed from the target email to be classified: Special symbols such as punctuation marks and numbers usually do not help substantially in text analysis, and stop words such as "de" (of), "shi" (is) and other high-frequency but meaningless words also need to be removed to reduce noise data.

[0044] The case conversion can be performed on the target email to be classified: All text is uniformly converted to lowercase or uppercase to avoid duplicate words caused by case differences.

[0045] Duplicate removal can be performed on the target email to be classified: If there are duplicate emails in the email dataset, duplicate removal processing needs to be performed to avoid unnecessary impacts on model training.

[0046] You can standardize the format of emails to be categorized: ensure that all emails have a consistent format, such as the consistency of date, time, and number formats.

[0047] Noise can be removed from emails that are not relevant to the email category: such as advertising links and image links in the email, which need to be removed.

[0048] S102 generates multiple sub-datasets based on vocabulary units, and each sub-dataset includes multiple elements.

[0049] In the embodiments of this disclosure, there are various methods for generating multiple subset datasets based on lexical units, and no limitation is made here.

[0050] In one possible implementation, natural language processing can be performed on lexical units to generate semantic features for each lexical unit, and then clustering can be performed based on the semantic features to generate multiple subsets of data.

[0051] In another possible implementation, lexical units can be divided into combinations of N consecutive words to generate richer semantic features. Then, clustering is performed based on these combined semantic features to generate multiple subsets.

[0052] In another possible approach, lexical units are grouped by part of speech (noun, verb, adjective) or semantic category (time words, personal names, organization names, project names).

[0053] In this embodiment of the disclosure, before generating multiple subsets based on lexical units, the lexical units can be vectorized, thereby transforming unstructured text information into structured numerical vectors, providing a data foundation for subsequent information gain calculation, feature selection, and subset partitioning.

[0054] In one possible implementation, the vocabulary units in the email dataset can be vectorized using a TF-IDF model to obtain a vector set s; The specific calculation method for TF-IDF model values ​​is as follows: Term frequency (TF) represents how often a particular word appears in a document. This number is usually normalized by dividing by the total number of words in the document to prevent it from being biased towards long documents. The formula for calculating TF is: = ,in Frequency of words This refers to the total number of lexical units r in this email (w). This refers to the total number of words in this email; The formula for calculating inverse document frequency is: = ,in N refers to the inverse document frequency, N refers to the total number of all emails in the email database, and N(r) refers to the total number of emails in the email database containing the word r. The formula for calculating TF-IDF value is as follows: = × ,in Refers to TF-IDF value, Frequency of words Inverse document frequency.

[0055] S103, for any subset of data, calculate the information gain value of each element in the subset of data.

[0056] It's important to note that information gain refers to the degree to which uncertainty is reduced after segmenting a dataset using a particular feature. It's a concept based on information entropy, which measures the degree of uncertainty or disorder in a dataset. Specifically, information gain equals the information entropy of the original dataset minus the weighted sum of the information entropies of the segmented subsets. Finally, features are ranked based on their calculated information gain values; features with higher information gain values ​​contribute more to the classification task. Based on a preset threshold, features with information gains greater than the preset threshold are selected as the final feature set.

[0057] In this embodiment, the information gain value of each element in the subset can be calculated using a preset algorithm. This algorithm can be modified according to actual design needs, and no limitations are imposed here.

[0058] S104, divide the elements in the subset based on the information gain value to generate multiple new subsets.

[0059] In this embodiment of the disclosure, after calculating and obtaining the information gain value, the feature with the highest information gain can be selected as the segmentation basis for the current node.

[0060] For example, the information gain of the feature "online" is 0.42.

[0061] The information gain of the feature “promotion” is 0.38.

[0062] The information gain of the feature “dear” is 0.35.

[0063] Based on the different values ​​of the "online" feature (present / non-present), the current dataset $D$ is split into two new sub-datasets: D1 contains emails that say "Going live" → mainly work emails (project-related).

[0064] D2 does not contain emails with "online" in them → Mix other types.

[0065] Generate a new subset of datasets: {D1, D2}.

[0066] S105, repeat the above steps for any subset of data, calculate the information gain value of each element in the subset and follow up, until the number of samples in the subset generated after partitioning is less than a preset threshold, then stop partitioning the elements in the corresponding subset.

[0067] It should be noted that for each subset, the above logic is repeated—the optimal segmentation feature for that subset is selected again using the information gain feature selection method, and then further subdivided into finer subsets according to different values ​​of that feature. Each segmentation aims to minimize data uncertainty (i.e., select the feature with the highest information gain), ensuring that the class purity of each subset gradually increases (e.g., the proportion of spam mixed in with the work email subset continuously decreases).

[0068] In this embodiment of the disclosure, in order to avoid overfitting of the decision tree (the model focuses too much on the special features of a few samples and ignores the overall data distribution), a preset threshold for the sample size is set: when the number of samples in a certain subset is less than the preset threshold, further segmentation of the subset is stopped - at this time it is considered that the samples in the subset are sufficiently homogeneous (high class consistency) and there is no need to continue splitting.

[0069] It should be noted that this preset threshold is set based on industry experience. Its core function is to balance classification accuracy and model generalization ability, and to prevent the segmentation results from losing representativeness due to insufficient sample size.

[0070] S106, until the elements in all subsets are no longer divided, obtain the candidate categories corresponding to each of the final generated subsets, and determine the target category of the target email based on the candidate categories.

[0071] When all divisible subsets stop because the sample size is less than the threshold, each final subset (the subset that has stopped splitting) corresponds to a specific email category (such as spam / work emails, project communication emails, personal emails, and emails for family and friends). After preprocessing and vectorization, new emails to be classified are assigned to a final subset according to the hierarchical feature rules of the decision tree (matching feature values ​​sequentially from the top to the bottom). This subset can then be directly mapped to the category corresponding to that subset, thus achieving the final classification from feature matching to category determination.

[0072] In this embodiment, target email data is first acquired, and vocabulary units for the target emails are generated based on the email data. Then, multiple subsets are generated based on these vocabulary units, each subset containing multiple elements. For any subset, the information gain value of each element is calculated. Next, the elements in the subset are partitioned based on the information gain value to generate multiple new subsets. This process of calculating the information gain value of each element in any subset and subsequent steps is repeated until the number of samples in the resulting subsets is less than a preset threshold. At this point, partitioning the elements in the corresponding subsets stops. Finally, partitioning all elements in all subsets stops, and candidate classifications are obtained for each of the final generated subsets. The target classification of the target email is determined based on these candidate classifications. Therefore, by selecting the most discriminative features based on information gain, gradually improving subset purity, and setting a sample size threshold to avoid excessive splitting on small samples, this method improves the model's generalization ability and achieves accurate classification of target emails. This method not only has good classification performance but also effectively prevents overfitting and has strong practicality.

[0073] In the above embodiments, multiple sub-datasets are generated based on lexical units, and can also be generated through... Figure 2 To further explain, the method includes: S201, For any given lexical unit, calculate the information gain value of the lexical unit.

[0074] In this embodiment of the disclosure, the information gain value of the target object is calculated. The target object is each element or lexical unit in the subset. First, the target dataset in which the target object is located is segmented based on the target object to obtain the segmented dataset corresponding to the target object. When the target object is an element, the target dataset is the subset in which the element is located, or when the target object is a lexical unit, the target dataset is the set of all lexical units. Then, the first information entropy of the target dataset in which the target object is located is calculated, and the second information entropy of the segmented dataset is calculated. Finally, the information gain value of the target object is calculated based on the first information entropy, the second information entropy, the number of samples in the target dataset, and the number of samples in the segmented dataset.

[0075] In this embodiment of the disclosure, the formula for calculating information gain is as follows: = - H

[0076] in This represents the information gain of the target dataset s with respect to the target object w. The information entropy of the target dataset s is represented by T, where T is the segmented dataset corresponding to the target object. It represents the number of samples corresponding to the segmentation of the dataset. H is the number of samples corresponding to the target dataset. It is a subset Information entropy.

[0077] In this embodiment of the disclosure, the information entropy is calculated as follows: Information entropy is calculated based on the distribution of categories in a dataset. Specifically, information entropy measures the degree of uncertainty or disorder in a dataset. Given a dataset s containing t distinct categories, with the number of samples for each category being |s1|, |s2|, ..., |st|, the total number of samples is |st| = |s1| + |s2| + ... + |st|.

[0078] =

[0079] in The information entropy of dataset s is represented by its information entropy. It is the probability of category st appearing in dataset s, i.e. = , The number of samples in st. This refers to the number of samples in dataset s.

[0080] S202, all lexical units are processed according to the information gain value to generate multiple subsets of the dataset.

[0081] In this embodiment of the disclosure, multiple optimal feature data can be determined based on the information gain value, and then all lexical units can be divided based on the optimal feature data to generate multiple subset datasets.

[0082] In one possible implementation, the information gain values ​​corresponding to all lexical units can be sorted in order of approximate smaller values ​​to generate a lexical gain sequence. Then, the top N target lexical units are selected from the lexical gain sequence based on a preset selection rule, and N optimal feature data are determined based on the top N target lexical units, where N is an integer greater than 1.

[0083] In this embodiment, the information gain value of any given lexical unit is first calculated. Then, all lexical units are processed based on the information gain value to generate multiple subsets. This allows for the selection of the most discriminative keywords through information gain, focusing on core features, avoiding interference from noisy words in classification decisions, reducing computational complexity in subsequent segmentation steps, and accelerating inference.

[0084] In the above embodiments, elements in the subset are divided based on information gain values ​​to generate multiple new subsets. This can also be achieved through... Figure 3 To further explain, the method includes: S301, sort the information gain values ​​corresponding to all elements in order of approximate smaller values ​​to generate an element gain sequence.

[0085] It should be noted that the calculation of the information gain value corresponding to all elements can be referred to the content in the above embodiments, and will not be repeated here.

[0086] S302, select the first M target elements from the element gain sequence based on the preset selection rules, and determine M optimal features based on the first M target elements, where M is an integer greater than 1.

[0087] S303, divides the elements in the subset based on M optimal features to generate M new subsequences.

[0088] In this embodiment, the information gain values ​​of all elements are first sorted in order of approximate smaller values ​​to generate a gain sequence. Then, based on a preset selection rule, the top M target elements are selected from the element gain sequence, and M optimal features are determined based on these top M target elements, where M is an integer greater than 1. Finally, the elements in the subset are partitioned based on the M optimal features to generate M new subsequences. This disclosure selects M high-value features in parallel for partitioning, significantly reducing the number of tree layers and recursions, improving partitioning efficiency, reducing recursion depth, and simultaneously using multiple high-gain features to capture various semantic patterns, avoid classification bias, and enhance classification diversity and coverage.

[0089] Figure 4 This is a schematic diagram of an email classification model adjustment method according to one embodiment of this disclosure, as shown below. Figure 4 As shown, the method includes the following steps: S401, Obtain the initial email classification model to be trained. The initial email classification model is used to calculate the information gain value of the subset of target emails to be classified, and to divide the elements in the subset based on the information gain value to generate multiple new subsets. Then, repeat the above process of calculating the information gain value of the subset of target emails to be classified until the number of samples in the subset generated after division is less than a preset threshold. Stop dividing the elements in the corresponding subset until the elements in all subsets are divided. Obtain the candidate classification corresponding to each of the finally generated subsets, and determine the target classification of the target email based on the candidate classification.

[0090] It should be noted that the initial email classification model in this embodiment is used to perform, for example... Figures 1-3 The logic of the method shown in the embodiment.

[0091] S402, Perform performance analysis on the initial email classification model to obtain comprehensive parameters of the initial email classification model.

[0092] In this embodiment, the comprehensive parameters refer to email classification accuracy, email classification precision, email classification processing speed, system daily maximum email throughput, email open rate, email click-through rate, email reply rate, system central processing unit (CPU) utilization, system memory usage, and system crash count.

[0093] Among them, the email classification accuracy parameter refers to the proportion of emails that are correctly classified out of the total number of emails.

[0094] Email classification accuracy refers to the proportion of emails that are classified as belonging to a certain category out of the total number of emails.

[0095] The email classification and processing speed parameter refers to the time tc required for the system to process each email.

[0096] The system's maximum daily email throughput parameter refers to the system's maximum daily email throughput (ttl).

[0097] The email open rate parameter refers to the percentage of emails that are opened by the recipient.

[0098] The email click-through rate parameter refers to the email click-through rate (djl).

[0099] The email response rate parameter refers to the email response rate (hfl).

[0100] The system CPU utilization parameter refers to the system CPU utilization rate.

[0101] The system memory usage parameter refers to the system memory usage rate (zyl).

[0102] The system crash count parameter refers to the number of system crashes, bkcs. It should be noted that there are various methods for obtaining parameters such as email classification accuracy, email classification precision, email classification processing speed, maximum daily email throughput, email open rate, email click-through rate, email reply rate, system CPU usage, system memory usage, and system crash count. For example, these parameters can be obtained by querying computer logs. Therefore, this embodiment does not impose any specific limitations on these methods. In one possible implementation, the performance of the email classification model can be analyzed using a mathematical model of a comprehensive system performance evaluation index. The specific mathematical model of the comprehensive system performance evaluation index is as follows: = + +

[0103] in The system performance evaluation metrics are as follows: ZQL (zql) represents the percentage of correctly categorized emails out of the total emails; JQL (jql) represents the percentage of emails actually belonging to a specific category out of the total emails; TC (tc) represents the time required for the system to process each email; TTL (ttl) represents the system's maximum daily email throughput; DKL (dkl) represents the percentage of emails opened by recipients; DJL (djl) represents the email click-through rate; HFL (hfl) represents the email reply rate; CPU (cpu) represents the system's CPU utilization; ZYL (zyl) represents the system's memory usage; and BKCS (bkcs) represents the number of system crashes. , , These are the weighting coefficients.

[0104] In this embodiment, it is necessary to specifically explain the weighting coefficient. , , It is calculated using the entropy method, and the specific calculation steps of the entropy method are as follows: Data standardization: Since the units of measurement for various indicators may differ, it is necessary to standardize the collected data to eliminate the influence of dimensions. Common standardization methods include Z-score standardization and range standardization. Calculate the weight of the indicator: Calculate the weight Pij of the i-th sample value under the j-th indicator relative to the sum of all sample values ​​for that indicator. This step is the basis for subsequent calculation of information entropy; Calculate information entropy: Based on the definition and calculation formula of information entropy, calculate the information entropy ej of the j-th indicator. The smaller the information entropy, the greater the degree of variation of the indicator, and the greater the amount of information it provides. Calculate information entropy redundancy: Information entropy redundancy dj is the complement of information entropy, i.e., dj = 1 - ej. The larger the information entropy redundancy, the more information the indicator contains, and the greater its impact on the overall evaluation. Calculate the weights: Based on the information entropy redundancy of each indicator, calculate its weight in the comprehensive evaluation. The weight coefficient is the ratio of the information entropy redundancy of each indicator to the sum of the information entropy redundancy of all indicators.

[0105] S403, adjust the preset threshold of the initial email classification model based on comprehensive parameters to generate the target email classification model.

[0106] In this embodiment of the disclosure, a preset value of the comprehensive parameter can be obtained, and then the comprehensive parameter and the preset value of the comprehensive parameter can be compared. In response to the comparison result meeting the adjustment condition, the preset threshold of the initial email classification model is adjusted.

[0107] It should be noted that the preset values ​​of the comprehensive parameters are warning values ​​obtained from industry practice to evaluate the efficiency and stability of the email classification system. When the comprehensive performance evaluation index of the system is lower than the preset value, it indicates that the email classification system has poor classification effect, low work efficiency and low stability, and corresponding measures need to be taken to reduce the risk.

[0108] In one possible implementation, the preset threshold of the initial email classification model can be adjusted using a large model, and the comprehensive parameters of the adjusted initial email classification model can be obtained. The above steps are repeated until the comprehensive parameters of the adjusted initial email classification model no longer meet the adjustment conditions, and the target email classification model is output.

[0109] It should be noted that this large model can be considered a large model for Artificial Intelligence (AI).

[0110] It should be noted that when the comparison result is lower than the preset value, the system will upload the comparison result to the AI ​​big model and use AI to analyze the reasons affecting the comprehensive evaluation index of system performance. The system will then be adjusted by AI, such as by using an exhaustive method to adjust the preset value of feature selection, until the comparison result of the comprehensive evaluation index of system performance is higher than the preset value.

[0111] Corresponding to the email classification methods provided in the above embodiments, one embodiment of this disclosure also provides an email classification device. Since the email classification device provided in this disclosure corresponds to the email classification methods provided in the above embodiments, the implementation methods of the above email classification methods are also applicable to the email classification device provided in this disclosure, and will not be described in detail in the following embodiments.

[0112] Figure 5 This is a schematic diagram of an email sorting device according to one embodiment of the present disclosure. As shown in FIG5, the email sorting device 500 includes: The acquisition module 510 is used to acquire target email data of the target emails to be classified, and generate vocabulary units of the target emails based on the email data.

[0113] The generation module 520 is used to generate multiple sub-datasets based on the vocabulary unit, and the sub-datasets include multiple elements.

[0114] The calculation module 530 is used to calculate the information gain value of each element in any subset of the dataset.

[0115] The partitioning module 540 is used to partition the elements in the subset based on the information gain value to generate multiple new subsets.

[0116] The loop module 550 is used to repeat the above steps for any subset of data, calculating the information gain value of each element in the subset and subsequent steps, until the number of samples in the subset generated after partitioning is less than a preset threshold, at which point the partitioning of the elements in the corresponding subset is stopped.

[0117] The classification module 560 is used to obtain the candidate classification corresponding to each of the finally generated sub-datasets until the elements in all sub-datasets are divided, and to determine the target classification of the target email based on the candidate classification.

[0118] According to one embodiment of this disclosure, generating the vocabulary units of the target email based on the email data includes: performing word segmentation processing on the email data to generate the vocabulary units of the target email.

[0119] According to one embodiment of this disclosure, generating multiple sub-datasets based on the lexical units includes: calculating the information gain value of any lexical unit; and processing all lexical units based on the information gain value to generate multiple sub-datasets.

[0120] According to one embodiment of this disclosure, the step of processing all lexical units based on the information gain value to generate multiple sub-data sets includes: determining multiple optimal feature data based on the information gain value; and dividing all lexical units based on the optimal feature data to generate multiple sub-data sets.

[0121] According to one embodiment of this disclosure, determining multiple optimal feature data based on the information gain value includes: sorting the information gain values ​​corresponding to all lexical units in order of approximate smaller values ​​to generate a lexical gain sequence; selecting the first N target lexical units from the lexical gain sequence based on a preset selection rule, and determining N optimal feature data based on the first N target lexical units, where N is an integer greater than 1.

[0122] According to one embodiment of this disclosure, calculating the information gain value of a target object, where the target object is each element in the subset or the lexical unit, includes: performing feature segmentation on the target dataset where the target object is located based on the target object to obtain a segmented dataset corresponding to the target object, wherein when the target object is an element, the target dataset is the subset containing the element, or when the target object is a lexical unit, the target dataset is a set composed of all lexical units; calculating a first information entropy of the target dataset where the target object is located, and calculating a second information entropy of the segmented dataset; and calculating the information gain value of the target object based on the first information entropy, the second information entropy, the number of samples corresponding to the target dataset, and the number of samples corresponding to the segmented dataset.

[0123] According to one embodiment of this disclosure, the step of dividing the elements in the subset based on the information gain value to generate multiple new subsets includes: sorting the information gain values ​​corresponding to all elements in order of approximate smaller values ​​to generate an element gain sequence; selecting the first M target elements from the element gain sequence based on a preset selection rule, and determining M optimal features based on the first M target elements, where M is an integer greater than 1; and dividing the elements in the subset based on the M optimal features to generate M new subsequences.

[0124] This disclosed scheme selects the most discriminative features based on information gain, gradually improves subset purity, and sets a sample size threshold to avoid excessive splitting on small samples, thereby improving the model's generalization ability and achieving accurate classification of target emails. This method not only has good classification performance but also effectively prevents overfitting and has strong practicality.

[0125] Corresponding to the email classification model adjustment methods provided in the above embodiments, an embodiment of this disclosure also provides an email classification model adjustment device. Since the email classification model adjustment device provided in this disclosure corresponds to the email classification model adjustment methods provided in the above embodiments, the implementation methods of the above email classification model adjustment methods are also applicable to the email classification model adjustment device provided in this disclosure, and will not be described in detail in the following embodiments.

[0126] Figure 6 This is a schematic diagram of an email classification model adjustment device according to one embodiment of the present disclosure. As shown in FIG6, the email classification model adjustment device 600 includes: The receiving module 610 is used to acquire an initial email classification model to be trained. The initial email classification model is used to calculate the information gain value of a subset of the target emails to be classified, and to divide the elements in the subset based on the information gain value to generate multiple new subsets. Then, the above process of calculating the information gain value of the subset of the target emails to be classified is repeated until the number of samples in the subset generated after division is less than a preset threshold. The division of the elements in the corresponding subset is stopped until the division of the elements in all subsets is stopped. The candidate classification corresponding to each of the finally generated subsets is obtained, and the target classification of the target email is determined based on the candidate classification. Analysis module 620 is used to perform performance analysis on the initial email classification model in order to obtain comprehensive parameters of the initial email classification model; The adjustment module 630 is used to adjust the preset threshold of the initial email classification model based on the comprehensive parameters in order to generate the target email classification model.

[0127] According to one embodiment of this disclosure, adjusting the preset threshold of the initial email classification model based on the comprehensive parameters includes: obtaining a preset value of the comprehensive parameters; comparing the comprehensive parameters with the preset value of the comprehensive parameters; and adjusting the preset threshold of the initial email classification model in response to the comparison result satisfying the adjustment condition.

[0128] According to one embodiment of this disclosure, the method further includes: adjusting a preset threshold of the initial email classification model using a large model, and obtaining the comprehensive parameters of the adjusted initial email classification model; repeating the above steps until the comprehensive parameters of the adjusted initial email classification model no longer meet the adjustment conditions, and outputting the target email classification model.

[0129] According to one embodiment of this disclosure, the comprehensive parameters include one or more of the following: email classification accuracy parameter, email classification precision parameter, email classification processing speed parameter, system daily maximum email throughput parameter, email open rate parameter, email click-through rate parameter, email reply rate parameter, system CPU utilization parameter, system memory usage parameter, and system crash count parameter.

[0130] To implement the above embodiments, this disclosure also proposes an electronic device 700. Figure 7 This is a schematic diagram of an electronic device according to one embodiment of the present disclosure, such as... Figure 7 As shown, the electronic device 700 includes: a processor 701 and a memory 702 communicatively connected to the processor. The memory 702 stores instructions executable by at least one processor. The instructions are executed by at least one processor 701 to implement the functions described in this disclosure. Figures 1-3The email classification method in the embodiment, or as shown in the example Figure 4 The email classification model adjustment method in the embodiment.

[0131] To implement the above embodiments, this disclosure also proposes a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to implement the present disclosure. Figures 1-3 The email classification method in the embodiment, or as shown in the example Figure 4 The email classification model adjustment method in the embodiment.

[0132] To implement the above embodiments, this disclosure also proposes a computer program product, including a computer program, which, when executed by a processor, implements the features of this disclosure. Figures 1-3 The email classification method in the embodiment, or as shown in the example Figure 4 The email classification model adjustment method in the embodiment.

[0133] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.

[0134] This application is intended to provide an implementation scheme for users to selectively prevent the use or access to their personal information data. Specifically, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information is de-identified to protect user privacy.

[0135] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0136] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0137] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0138] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that contains, stores, communicates, propagates, or transmits programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0139] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0140] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0141] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0142] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for classifying emails, characterized in that, include: Obtain target email data for the target emails to be classified, and generate vocabulary units for the target emails based on the email data; Multiple sub-datasets are generated based on the lexical units, and each sub-dataset includes multiple elements; For any subset of data, calculate the information gain value of each element in the subset of data; The elements in the subset are divided based on the information gain value to generate multiple new subsets; Repeat the above steps for any subset of data, calculating the information gain value of each element in the subset and subsequent steps, until the number of samples in the subset generated after partitioning is less than a preset threshold, then stop partitioning the elements of the corresponding subset. The process continues until all elements in all subsets are divided, and candidate categories corresponding to each of the final generated subsets are obtained. The target category of the target email is then determined based on the candidate categories.

2. The method according to claim 1, characterized in that, The generation of the vocabulary units for the target email based on the email data includes: The email data is segmented to generate vocabulary units for the target email.

3. The method according to claim 1, characterized in that, The generation of multiple sub-datasets based on the lexical units includes: For any given lexical unit, calculate the information gain value of that lexical unit; All lexical units are processed based on the information gain value to generate multiple sub-datasets.

4. The method according to claim 3, characterized in that, The process of processing all lexical units based on the information gain value to generate multiple subset datasets includes: Multiple optimal feature data are determined based on the information gain value; All lexical units are divided based on the optimal feature data to generate multiple sub-datasets.

5. The method according to claim 4, characterized in that, The determination of multiple optimal feature data based on the information gain value includes: The information gain values ​​corresponding to all lexical units are sorted in order of approximate smaller values ​​to generate a lexical gain sequence; Based on a preset selection rule, the first N target word units are selected from the word gain sequence, and N optimal feature data are determined based on the first N target word units, where N is an integer greater than 1.

6. The method according to claim 1 or 3, characterized in that, Calculating the information gain value of a target object, where the target object is each element in the subset of data or the lexical unit, includes: Based on the target object, feature segmentation is performed on the target dataset in which the target object is located to obtain the segmented dataset corresponding to the target object. Wherein, when the target object is an element, the target dataset is the sub-dataset in which the element is located, or when the target object is a word unit, the target dataset is a set composed of all word units. Calculate the first information entropy of the target dataset where the target object is located, and calculate the second information entropy of the segmented dataset; The information gain value of the target object is calculated based on the first information entropy, the second information entropy, the number of samples corresponding to the target dataset, and the number of samples corresponding to the segmented dataset.

7. The method according to any one of claims 1-5, characterized in that, The step of partitioning the elements in the subset based on the information gain value to generate multiple new subsets includes: The information gain values ​​corresponding to all elements are sorted in order of approximate smaller values ​​to generate an element gain sequence; Based on a preset selection rule, the first M target elements are selected from the element gain sequence, and M optimal features are determined based on the first M target elements, where M is an integer greater than 1. The elements in the subset are divided based on the M optimal features to generate M new subsequences.

8. A method for adjusting an email classification model, characterized in that, include: An initial email classification model to be trained is obtained. The initial email classification model is used to calculate the information gain value of a subset of the target emails to be classified, and to divide the elements in the subset based on the information gain value to generate multiple new subsets. Then, the above process of calculating the information gain value for the subset of the target emails to be classified is repeated until the number of samples in the subset generated after division is less than a preset threshold. The division of the elements in the corresponding subset is stopped until the division of the elements in all subsets is stopped. The candidate classifications corresponding to all the finally generated subsets are obtained, and the target classification of the target email is determined based on the candidate classifications. The initial email classification model is subjected to performance analysis to obtain the comprehensive parameters of the initial email classification model; The preset threshold of the initial email classification model is adjusted based on the comprehensive parameters to generate the target email classification model.

9. The method according to claim 8, characterized in that, The adjustment of the preset threshold of the initial email classification model based on the comprehensive parameters includes: Obtain the preset values ​​of comprehensive parameters; The comprehensive parameters are compared with the preset values ​​of the comprehensive parameters. If the comparison result meets the adjustment conditions, the preset threshold of the initial email classification model is adjusted.

10. The method according to claim 8 or 9, characterized in that, The method further includes: The preset threshold of the initial email classification model is adjusted using a large model, and the comprehensive parameters of the adjusted initial email classification model are obtained. Repeat the above steps until the overall parameters of the adjusted initial email classification model no longer meet the adjustment conditions, and then output the target email classification model.

11. The method according to claim 8, characterized in that, The comprehensive parameters include one or more of the following: email classification accuracy, email classification precision, email classification processing speed, system's maximum daily email throughput, email open rate, email click-through rate, email reply rate, system CPU utilization, system memory usage, and number of system crashes.

12. A mail sorting device, characterized in that, include: The acquisition module is used to acquire target email data of the target emails to be classified, and generate vocabulary units of the target emails based on the email data; A generation module is used to generate multiple sub-datasets based on the lexical unit, wherein each sub-dataset includes multiple elements; The calculation module is used to calculate the information gain value of each element in any subset of the dataset. A partitioning module is used to partition the elements in the subset based on the information gain value to generate multiple new subsets; The loop module is used to repeat the above steps for any subset of data, calculating the information gain value of each element in the subset and subsequent steps, until the number of samples in the subset generated after partitioning is less than a preset threshold, at which point the partitioning of the elements in the corresponding subset is stopped. The classification module is used to obtain the candidate classification corresponding to each of the finally generated sub-datasets until the elements in all sub-datasets are divided, and to determine the target classification of the target email based on the candidate classification.

13. A mail classification model adjustment device, characterized in that, include: The receiving module is used to acquire an initial email classification model to be trained. The initial email classification model is used to calculate the information gain value of a subset of the target emails to be classified, and to divide the elements in the subset based on the information gain value to generate multiple new subsets. Then, the above process of calculating the information gain value for the subset of the target emails to be classified is repeated until the number of samples in the subset generated after division is less than a preset threshold. The division of the elements in the corresponding subset is stopped until the division of the elements in all subsets is stopped. The candidate classification corresponding to each of the finally generated subsets is obtained, and the target classification of the target email is determined based on the candidate classification. The analysis module is used to perform performance analysis on the initial email classification model in order to obtain the comprehensive parameters of the initial email classification model. The adjustment module is used to adjust the preset threshold of the initial email classification model based on the comprehensive parameters in order to generate the target email classification model.

14. An electronic device, characterized in that, Including memory and processor; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the email classification method as described in any one of claims 1-7, or to implement the email classification model adjustment method as described in any one of claims 8-11.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the email classification method as described in any one of claims 1-7, or to implement the email classification model adjustment method as described in any one of claims 8-11.

Images