A document processing recommender method based on user profile collaborative analysis

Abstract

This invention provides a method for recommending document processors based on collaborative analysis of user profiles in the field of text recommendation technology, comprising the following steps: Step S10, collecting a large number of document samples; Step S20, preprocessing the collected document samples to obtain a document dataset; Step S30, constructing user profiles of document processors based on the document dataset; Step S40, clustering document processors based on the user profiles to obtain a clustering matrix; Step S50, constructing a document recommendation model based on the clustering matrix; Step S60, validating the document recommendation model based on the F1 score; Step S70, recommending document processors using the validated document recommendation model. The advantages of this invention are: significantly improving the accuracy and efficiency of document recommendation.

Description

Technical Field

[0001] This invention relates to the field of text recommendation technology, and in particular to a method for recommending document processors based on collaborative analysis of user profiles. Background Technology

[0002] With the development of information technology, OA (Office Automation) systems have been widely used in various departments and enterprises, and electronic documents have gradually replaced traditional paper documents. Currently, most commercially available OA systems only facilitate document flow based on workflows, without effectively utilizing processed document data and lacking auxiliary office tools. During the office process, problems arise such as low-quality processing and high return rates due to staff lacking experience in handling such documents, thus affecting document flow efficiency. Furthermore, as the number of documents increases, the workload also rises rapidly. If staff cannot process them promptly, document backlog is likely to occur, significantly impacting work efficiency.

[0003] Today, artificial intelligence (AI) technology is constantly developing and improving our lives. The application of various user information modeling and personalized recommendation technologies in portal websites represents a major application direction of AI. Recommendation technology is a product of the development of the internet and electronic information technology. It is an advanced intelligent platform built on data mining, used to provide users with information services and decision support. The recommendation algorithm is the soul of the system, and its quality largely determines the performance of recommendations.

[0004] Traditional text recommendation technologies are mainly divided into the following three categories: 1. Knowledge-based text recommendation: This technology relies on a predefined knowledge base. The knowledge base establishes association rules based on user characteristics and text characteristics. For new text, it queries matching users through association rules to make recommendations. It can directly associate needs with recommendation information, so there is no cold start problem. 2. Content-based text recommendation: This technology uses text and user history information for recommendations. It extracts features from the user's historical text information to build a user profile model. After extracting features from the text to be recommended, it calculates the degree of matching between the text and the user to make recommendations of similar texts. 3. Collaborative filtering-based text recommendation: The core idea of ​​this technology is that people of the same type will be interested in similar texts. It is mainly divided into two algorithms: user-based collaborative filtering recommendation algorithm and text-based collaborative filtering recommendation algorithm.

[0005] However, traditional text recommendation technologies have the following drawbacks: knowledge-based text recommendation involves complex knowledge base construction, requiring the organization of numerous prior information into standardized association rules for representation; content-based text recommendation and collaborative filtering-based text recommendation cannot adequately understand the semantics of official document topics and fail to take into account some user characteristics, resulting in low recommendation accuracy.

[0006] Therefore, how to provide a document processing recommender method based on user profile collaborative analysis to improve the accuracy and efficiency of document recommendation has become an urgent technical problem to be solved. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a method for recommending document processors based on collaborative analysis of user profiles, thereby improving the accuracy and efficiency of document recommendation.

[0008] This invention is implemented as follows: A method for recommending document processors based on collaborative analysis of user profiles, comprising the following steps:

[0009] Step S10: Collect a large number of official document samples;

[0010] Step S20: Preprocess the collected document samples to obtain a document dataset;

[0011] Step S30: Construct a user profile of the document processor based on the document dataset;

[0012] Step S40: Cluster the document processors based on the similarity matrix of the user profiles to obtain a clustering matrix;

[0013] Step S50: Construct a document recommendation model based on the clustering matrix;

[0014] Step S60: Validate the document recommendation model based on the F1 score;

[0015] Step S70: Recommend document processors using the validated document recommendation model.

[0016] Further, step S10 specifically includes:

[0017] A large number of official document samples were collected based on the dimensions of document type, issuance time, and department.

[0018] The document types in the document type dimension include at least departmental outbound documents, departmental inbound documents, business orders, and work orders; the document issuance time dimension includes peak time periods, low-peak time periods, and off-peak time periods, and document samples are collected in a ratio of 3:1:1 respectively; the department dimension is used to set the collection weight of each part based on the number of outbound and inbound documents.

[0019] Further, step S20 specifically includes:

[0020] Step S21: Remove noise data from each of the collected document samples; the noise data includes at least URL addresses, special symbols, images, and zero-width characters;

[0021] Step S22: Convert the traditional Chinese characters contained in each of the document samples after removing noise data into simplified Chinese characters;

[0022] Step S23: Construct a vocabulary list containing a number of sensitive words, and filter the sensitive words in each document sample based on the vocabulary list;

[0023] Step S24: Construct a stop word library containing a number of stop words, and filter the stop words in each document sample based on the stop word library;

[0024] Step S25: Perform part-of-speech tagging on each of the document samples, and retain the document samples with the part of speech as nouns to construct a document dataset.

[0025] Further, step S30 specifically includes:

[0026] Step S31: Construct a tag system for user profiles based on static and dynamic tag information; the static tag information includes at least gender, age, job title, and department; the dynamic tag information includes the area of ​​responsibility and processing habits; the area of ​​responsibility includes the record of the next person in charge and the specific content of the document; the processing habits include the frequency of processing and the time of processing.

[0027] Step S32: Construct user profiles based on the aforementioned tag system and official document dataset;

[0028] Step S33: Perform multidimensional similarity calculation on each user profile to obtain the user similarity matrix US.

[0029] Further, step S32 specifically includes:

[0030] The user profile constructed based on the aforementioned tag system is as follows:

[0031] P={Info, Keyword, Frequency, Time};

[0032] Where P represents the user profile; Info = {Gender, Age, Post, Department}, where Gender, Age, Post, and Department represent gender, age, job title, and department, respectively; Frequency represents the processing frequency; and Time represents the processing time.

[0033] Further, step S40 specifically includes:

[0034] Step S41: Set the initial user class set C to have a capacity of 0, the document processing user set to be clustered to be U, and the number of clusters to be m;

[0035] Step S42: Determine whether set U is empty. If not, obtain the most similar pair of users (u) based on the similarity matrix US. i u j If the value is true and the corresponding similarity value is maximum, set the corresponding position of the similarity matrix US to 0 and proceed to step S43; if so, proceed to step S46.

[0036] Step S43: Determine whether this is the first clustering. If yes, proceed to step S44; otherwise, proceed to step S45.

[0037] Step S44: Determine if the current total number of clusters is equal to m. If yes, select the cluster with the fewest users from all existing clusters to add new users and proceed to step S42. If no, create a new cluster, add the two current users to the newly created cluster, delete the corresponding users from set U, and proceed to step S45.

[0038] Step S45: Traverse the categories c in set C. i , if u i and u j There is a user who belongs to category c i Then, simply add another user to category c. i Remove the corresponding user from set U and proceed to step S42; if u i and u j If it does not belong to set C, proceed to step S44;

[0039] Step S46: After traversing set C, determine whether the total number of clusters is equal to m. If yes, obtain the clustering matrix CM based on each cluster; otherwise, proceed to step S47.

[0040] Step S47: Sort the user groups in descending order of the number of users contained in the cluster, traverse the sorted user groups to perform group segmentation, and proceed to step S46.

[0041] Further, in step S50, the document recommendation model is used to perform the following steps:

[0042] Step S51: Filter the document processor u that contains the current stage from the clustering matrix CM. i User groups;

[0043] Step S52, based on the u i Search left within the user group to find all users located in the specified u. i The users on the left represent a set of similar users.

[0044] Step S53: Obtain the preference set SU = (SU1, SU2, ..., SU3) of each user in the similar user set. n); where SU n This represents the set of preferred documents of the nth similar user;

[0045] Step S54: Use the clustering matrix CM to weight the documents in the preference set SU, and then sort them in descending order of similarity:

[0046] Step S55: Recall the recommended results for document processors based on the similarity ranking results.

[0047] Further, in step S60, the formula for calculating the F1 value is as follows:

[0048] F1 = (2 * P * R) / (P + R);

[0049] Where P represents precision, P = TP / (TP + FN); R represents recall, R = TP / (TP + FP); TP represents positive samples predicted as positive by the model; FN represents positive samples predicted as negative by the model; and FP represents negative samples predicted as positive by the model.

[0050] The advantages of this invention are:

[0051] By collecting official document samples based on document type, issuance time, and department, a realistic business scenario is constructed. This involves learning document characteristics from the perspectives of type, time, and department, making the system more robust and improving the accuracy of subsequent recommendations. User profiles are built through multi-dimensional similarity calculations, mitigating the cold start problem common in collaborative filtering by relying on users' multi-dimensional basic information features. Deeper feature extraction of users fully utilizes text content to enrich user profile features, avoiding the matrix sparsity problem common in collaborative filtering. Users are rationally grouped through multi-dimensional similarity calculations and improved clustering algorithms, alleviating the problem of excessive computational load caused by searching the entire user set. Ultimately, this significantly improves the accuracy and efficiency of document recommendations. Attached Figure Description

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0053] Figure 1 This is a flowchart of a document processing recommender method based on user profile collaborative analysis, according to the present invention. Detailed Implementation

[0054] Please refer to Figure 1 As shown, a preferred embodiment of the document processing personnel recommendation method based on user profile collaborative analysis of the present invention includes the following steps:

[0055] Step S10: Collect a large number of official document samples;

[0056] Step S20: Preprocess the collected document samples to obtain a document dataset;

[0057] Step S30: Construct a user profile of the document processor based on the document dataset;

[0058] Step S40: Cluster the document processors based on the similarity matrix of the user profiles to obtain a clustering matrix;

[0059] Step S50: Construct a document recommendation model based on the clustering matrix;

[0060] Step S60: Validate the document recommendation model based on the F1 score;

[0061] Step S70: Recommend document processors using the validated document recommendation model.

[0062] Step S10 specifically involves:

[0063] A large number of official document samples were collected based on the dimensions of document type, issuance time, and department.

[0064] The document types in the document type dimension include at least departmental outbound documents, departmental inbound documents, business orders, and work orders; the document issuance time dimension includes peak time periods, low-peak time periods, and off-peak time periods, and document samples are collected in a ratio of 3:1:1 respectively; the department dimension is used to set the collection weight of each part based on the number of outbound and inbound documents.

[0065] Traditional sample collection methods involve directly exporting all data within a specific time span. While this ensures a sufficient number of training and testing sets and eliminates concerns about poor training results due to insufficient data, it is too simplistic and crude. Furthermore, there are significant differences between different document types and from different years. Using such data directly without filtering may affect the accuracy of subsequent recommendations. Therefore, this invention collects document samples based on different dimensions.

[0066] Regarding the document type dimension, since different document types have their own characteristics, such as meeting minutes which record the meeting and are usually distributed to all participants and non-participants, and joint departmental documents which are generally distributed to the personnel of the departments involved, after analyzing these different types of documents, the document types to be collected are mainly departmental documents, departmental receipts, business orders, work orders, etc.

[0067] Regarding the document issuance time dimension, scatter plots and piecewise function fitting were used to analyze the document issuance and processing times. The results showed that the number of work orders began to rise around 8:00 AM, peaking around 10:00 AM, then declining to a small trough around 12:00 PM, followed by slow fluctuations. A small peak was reached at 4:00 PM, followed by a significant drop after 6:00 PM, with the lowest point of the day occurring between midnight and 6:00 AM. Overall, the daily distribution of official documents aligns with people's work, life, and daily routines. Therefore, data was collected in time periods, with peak, trough, and off-peak periods collected in a 3:1:1 ratio to ensure sample authenticity.

[0068] When collecting data at the departmental level, since most official documents involve only a few departments, and other departments each have a certain number of official documents, but the number is relatively small, the sample collection will focus more on departments with a high volume of issued and received documents, and the sample size collected from departments with a smaller amount of data will be reduced accordingly.

[0069] The original document samples are complex and have many features, making them unsuitable for direct use as training data. Some data may be irrelevant to subsequent analysis operations or contain interference items. Therefore, it is necessary to preprocess the data before building the model.

[0070] Step S20 specifically includes:

[0071] Step S21: Remove noise data from each of the collected document samples; the noise data includes at least URL addresses, special symbols, images, and zero-width characters;

[0072] Since this useless data not only greatly increases the workload of text processing, but also directly affects a series of subsequent operations, thereby reducing the efficiency of text processing, removing this noisy data is the primary task of data preprocessing.

[0073] Step S22: Convert the traditional Chinese characters contained in each of the document samples after removing noise data into simplified Chinese characters;

[0074] Since the official document samples may contain some traditional Chinese characters, if they are not processed, the semantics of the text may be affected when the model performs semantic feature engineering. Therefore, it is necessary to convert the traditional Chinese characters in the official document samples.

[0075] Step S23: Construct a vocabulary list containing a number of sensitive words, and filter the sensitive words in each document sample based on the vocabulary list;

[0076] Step S24: Construct a stop word library containing a number of stop words, and filter the stop words in each document sample based on the stop word library;

[0077] After word segmentation, the keywords will contain some words that are not helpful for reflecting the actual meaning of official documents, such as words like "de" (的) and "le" (了). These are called stop words. In this invention, it is preferred to use the Harbin Institute of Technology stop word library to perform stop word removal processing on the word segmentation.

[0078] Step S25: Perform词性标注 on each of the official document samples. Since nouns can better represent the characteristics of official documents, the official document samples with the词性标注 of nouns are retained to construct an official document dataset.

[0079] The traditional recommendation method is to extract the features of the official documents involved by the user and generate recommendation results. Essentially, it is a content-based recommendation method. Generating personalized result sets in this way for a long time will make the types of recommendations too single. This invention comprehensively mines the behavior preferences of users after integrating different domain features to increase the richness of the recommendation results, thus making up for the above problems.

[0080] The specific steps of step S30 include:

[0081] Step S31: Construct a label system for the user portrait based on static label information and dynamic label information; the static label information at least includes gender, age, position, and department; the dynamic label information includes the responsible field and handling habits; the responsible field includes the record of the next link handler and the specific content of the official document; the handling habits include the handling frequency and handling time.

[0082] A user portrait is to label the user-related information. The premise of depicting a user portrait is to establish a label system for the user portrait; static information labels refer to the basic attribute information of the user; dynamic information labels refer to the feature labels mined through the user's behavior.

[0083] Step S32: Construct a user portrait based on the label system and the official document dataset.

[0084] A comprehensive user portrait can more accurately describe the characteristics and interest preferences of the user. This invention uses the official document text as the basic information for depicting the portrait and proposes a user portrait with multi-feature fusion, that is, fusing four dimensions of user features.

[0085] Step S33: Perform multi-dimensional similarity calculation on each of the user portraits to obtain a similarity matrix US of the users.

[0086] For a user portrait P = {Info, Keyword, Frequency, Time} with four-dimensional features, a similarity matrix InfoS of the basic information, and for the remaining three-dimensional features Keyword, Frequency, Time, the cosine similarity calculation method is directly used for joint similarity calculation. For example, for users U i and U j It should be noted that the "词性标注" in step S3 needs to be further accurately translated according to the specific context, as it may be a specific term in the field of natural language processing. Here, a general placeholder is used for translation.The user profile PT, represented as a vector, is obtained from the above steps. i and PT j The calculation formula is as follows:

[0087]

[0088] The n×n user similarity matrix PTS is obtained. The basic information similarity matrices InfoS and PTS, which are both n×n, are added together to obtain the final user similarity matrix US.

[0089] Step S32 specifically involves:

[0090] The user profile constructed based on the aforementioned tag system is as follows:

[0091] P={Info, Keyword, Frequency, Time};

[0092] Where P represents the user profile; Info = {Gender, Age, Post, Department}, where Gender, Age, Post, and Department represent gender, age, job title, and department, respectively; Frequency represents the processing frequency; and Time represents the processing time.

[0093] User profiles include basic attribute information, document subject, processing frequency, and processing time.

[0094] Basic attribute information:

[0095] User basic attribute information includes gender, age, job title, and department, as this information helps in the accuracy of personalized recommendations. However, directly measuring user similarity based on these dimensions is difficult. Therefore, user tag information for each dimension is standardized and constructed beforehand, and finally, the basic attribute information is represented in a vectorized form. The specific dimensions of each piece of information are as follows:

[0096] Gender: Male / Female

[0097] Age: Youth / Middle-aged / Elderly

[0098] Job title: Sales / Technical Staff / Department Manager…

[0099] Department: Human Resources / Finance / Technology...

[0100] To facilitate similarity calculation of basic attribute information, features in each dimension are represented using binary sequences. Gender is assigned a flag bit: 1 for male, 0 otherwise. Age group is assigned three flag bits: 1 for the specified age group and 0 for the rest. Job title and department are assigned flag bits based on the data: 1 for the current position and 0 for the rest. This yields the binary representation of the user's basic information, Info, which is then used to calculate similarity using the binary Jaccard similarity method, specifically for user U. i and U j The similarity calculation method is as follows:

[0101]

[0102] Among them, B i and B j For user U i and U j The binary representation of is given, where l represents the length of the binary sequence, and the final similarity result is within the interval [0,1]. The basic information similarity between all users is calculated sequentially, ultimately yielding the similarity matrix InfoS.

[0103] Document Body Subject:

[0104] Official documents involve projects and areas handled by office staff and are closely related to business operations. Therefore, it is necessary to conduct in-depth analysis of these documents to extract the characteristics between office staff and business work. Traditional analysis uses collaborative filtering methods, typically constructing a user-document feature matrix. If a user interacts with a document, the value is 1; otherwise, it is 0. Then, the similarity between each user's binary sequence is calculated for comparison. However, this method does not utilize the feature information of the document to characterize users and calculate similarity; it only represents user interest sequences in the form of 0 and 1. The resulting similarity results are often not comprehensive or accurate. Therefore, this invention uses topic modeling to characterize user features and perform similarity comparisons. The steps are as follows:

[0105] Step 1: Collect user U i D is a collection of official document data containing operational actions within the past week. u ={d u,1 ,d u,2 ,...,d u,n}, where d u,k This represents the k-th document in which the user recently performed an action;

[0106] Step 2: Perform text preprocessing on the articles used as training corpus;

[0107] Step 3: Use the official documents in the corpus to train the LDA topic model. The total number of categories m in the document corpus is used as the topic number parameter. After training, the topic model of the training set is obtained.

[0108] Step 4: Input the official documents in the document set as the validation set into the topic model to obtain the topic distribution vector T for each document. d ={p d,1 ,p d,2 ,...,p d,k ,...,p d,m}, where p d,k This vector represents the probability that document d belongs to topic k, and is used to represent each document.

[0109] Step 5: From D u The topic distribution of all official documents in the database yields the user's m×n dimensional topic profile matrix P. u P u D represents u The joint feature of all document subject characteristics in the set;

[0110] Step 6: Apply a time decay mechanism to weight each document, as shown in formula (1), and then sum the topic vectors of each document to obtain a 1×m-dimensional topic-based user feature vector F. u As shown in formula (2), where n represents the total number of preference documents;

[0111]

[0112]

[0113] Employee processing frequency calculation:

[0114] Processing frequency, as a manifestation of dynamic user behavior, can effectively reflect a user's work habits. Since different users have different processing frequencies, using processing frequency as a one-dimensional feature of a user profile can effectively distinguish users and contribute to the accuracy of personalized recommendations. The formula for calculating processing frequency R is as follows:

[0115]

[0116] Where k represents the total number of days with recorded transactions, and R l This represents the total number of official documents processed by the user on day 1. The average number of documents processed by the user can be obtained through this calculation formula.

[0117] Employee processing time statistics:

[0118] Processing time, as a form of behavioral information, can also reflect a user's document processing habits. People with similar situations are more likely to have similar identities or pay attention to the same categories of information. To facilitate the measurement of an individual's processing time distribution, this invention divides the day into time periods. When obtaining the processing time distribution vector representation, the first step is to obtain the set of articles in which the user has processed the document. The documents are then categorized according to the time divisions in the table, and the number of reads TN for each time period is calculated. Then, the processing time distribution vector is calculated based on the distribution of the number of reads, using the following formula:

[0119]

[0120] Because official document data is constantly being updated and the data volume is increasing rapidly, the computational workload also increases rapidly. Therefore, in actual production environments, it is necessary to apply appropriate clustering methods to improve the efficiency of recommendation calculations. General clustering methods result in an object belonging to only one class after clustering. However, for user clustering, a user can have multiple different characteristics and belong to multiple categories. Therefore, this invention proposes an optimized ordered clustering method to cluster users, resulting in the same user existing in different user groups.

[0121] Step S40 specifically includes:

[0122] Step S41: Set the initial user class set C to have a capacity of 0, the document processing user set to be clustered to be U, and the number of clusters to be m;

[0123] Step S42: Determine whether set U is empty. If not, obtain the most similar pair of users (u) based on the similarity matrix US. i u j If the value is true and the corresponding similarity value is maximum, set the corresponding position of the similarity matrix US to 0 and proceed to step S43; if so, proceed to step S46.

[0124] Step S43: Determine whether this is the first clustering. If yes, proceed to step S44; otherwise, proceed to step S45.

[0125] Step S44: Determine if the current total number of clusters is equal to m. If yes, select the cluster with the fewest users from all existing clusters to add new users and proceed to step S42. If no, create a new cluster, add the two current users to the newly created cluster, delete the corresponding users from set U, and proceed to step S45.

[0126] Step S45: Traverse the categories c in set C. i , if u i and u j There is a user who belongs to category c iThen, simply add another user to category c. i Remove the corresponding user from set U and proceed to step S42; if u i and u j If it does not belong to set C, proceed to step S44;

[0127] Step S46: After traversing set C, determine whether the total number of clusters is equal to m. If yes, obtain the clustering matrix CM based on each cluster; otherwise, proceed to step S47.

[0128] Step S47: Sort the user groups in descending order of the number of users contained in the cluster, traverse the sorted user groups to perform group segmentation, and proceed to step S46.

[0129] Since the pair of users with the highest similarity is selected each time during the clustering process, there may be cases where the same user appears in different groups at the same time. In addition, since the similarity value is decreasing, for a specific user in a group, his right neighbor is more similar to him than his left neighbor. That is, the similarity difference between adjacent users in the whole group is constantly decreasing, which makes it easier to find the nearest neighbor user in the subsequent recommendation.

[0130] In step S50, the document recommendation model is used to perform the following steps:

[0131] Step S51: Filter the document processor u that contains the current stage from the clustering matrix CM. i User groups;

[0132] Step S52, based on the u i Search left within the user group to find all users located in the specified u. i The users on the left represent a set of similar users.

[0133] Step S53: Obtain the preference set SU = (SU1, SU2, ..., SU3) of each user in the similar user set. n ); where SU n This represents the set of preferred documents of the nth similar user;

[0134] Step S54: Use the clustering matrix CM to weight the documents in the preference set SU, and then sort them in descending order of similarity:

[0135] Step S55: Recall the recommended results for document processors based on the similarity ranking results.

[0136] In step S60, the formula for calculating the F1 value (F1-score) is as follows:

[0137] F1 = (2 * P * R) / (P + R);

[0138] Where P represents precision, P = TP / (TP + FN); R represents recall, R = TP / (TP + FP); TP represents positive samples predicted as positive by the model; FN represents positive samples predicted as negative by the model; and FP represents negative samples predicted as positive by the model.

[0139] Precision (P) refers to the proportion of samples correctly predicted as having intent A out of all samples correctly predicted as having intent A. Recall (R) is a measure of coverage, referring to the proportion of samples correctly predicted as having intent A out of all samples that actually have intent A. Precision and recall metrics can sometimes contradict each other, so they need to be considered together. The most common method is to use the F1 score, which is a metric that combines precision (P) and recall (R). A higher F1 score indicates better model recognition performance.

[0140] In summary, the advantages of this invention are as follows:

[0141] By collecting official document samples based on document type, issuance time, and department, a realistic business scenario is constructed. This involves learning document characteristics from the perspectives of type, time, and department, making the system more robust and improving the accuracy of subsequent recommendations. User profiles are built through multi-dimensional similarity calculations, mitigating the cold start problem common in collaborative filtering by relying on users' multi-dimensional basic information features. Deeper feature extraction of users fully utilizes text content to enrich user profile features, avoiding the matrix sparsity problem common in collaborative filtering. Users are rationally grouped through multi-dimensional similarity calculations and improved clustering algorithms, alleviating the problem of excessive computational load caused by searching the entire user set. Ultimately, this significantly improves the accuracy and efficiency of document recommendations.

[0142] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for recommending a document processing person based on user portrait collaborative analysis, characterized in that: The method comprises the following steps: Step S10, collecting a large number of official document samples; Step S20, preprocessing each of the collected official document samples to obtain an official document data set; Step S30, constructing a user portrait of an official document processor based on the official document data set; Step S40, clustering official document processors based on a similarity matrix of the user portrait to obtain a clustering matrix; Step S50, constructing an official document recommendation model based on the clustering matrix; Step S60, verifying the official document recommendation model based on an F1 value; Step S70, recommending an official document processor by using the verified official document recommendation model; The step S30 specifically comprises: Step S31, constructing a label system of the user portrait based on static label information and dynamic label information; the static label information at least includes gender, age, position, and department; the dynamic label information includes responsible field and handling habit; the responsible field includes next-link processor record and official document specific content; the handling habit includes handling frequency and handling time; Step S32, constructing the user portrait based on the label system and the official document data set; Step S33, performing multi-dimensional similarity calculation on each of the user portraits to obtain a similarity matrix US of the users; The step S40 specifically comprises: Step S41, setting the capacity of an initial user class set C as 0, the official document processors to be clustered as a set U, and the clustering number as m; Step S42, judging whether the set U is empty, if not, obtaining the most similar pair of users (u i , u j ) and the corresponding similarity value Max based on the similarity matrix US, setting the corresponding position of the similarity matrix US as 0, and entering step S43; if yes, entering step S46; Step S43, judging whether it is the first clustering, if yes, proceeding to step S44; if no, proceeding to step S45; Step S44, judging whether the current clustering total number is equal to m, if yes, selecting a cluster containing the least users from all existing clusters for new user addition, and proceeding to step S42; if no, creating a new class, and adding the current two users to the newly created class, deleting the corresponding users from the set U, and proceeding to step S45; Step S45, traversing the category c in the set C i , if u i and u j have one user belonging to the category c i , directly add the other user to the category c i , delete the corresponding user from the set U, and enter step S42; if u i and u j do not belong to the set C, enter step S44; Step S46, after the set C is traversed, judging whether the current clustering total number is equal to m, if yes, obtaining a clustering matrix CM based on each cluster; if no, proceeding to step S47; Step S47, sorting the users contained in the clusters in descending order, and performing group segmentation on the sorted user groups, and proceeding to step S46; In the step S50, the official document recommendation model is used to perform the following steps: Step S51, filtering the document processing person u containing the current link from the clustering matrix CM i of the user group; Step S52, based on the u i the location searches left in the user group to find all users to the left of the u i user set; Step S53, obtaining the preference set SU= (SU1, SU2, …, SU n ) of each user in the similar user set; wherein, SU n represents the preference document set of the nth similar user. Step S54, weighting the official documents in the preference set SU by using the clustering matrix CM, and then sorting the similarity in descending order: Step S55, recalling the recommendation result of the official document processor based on the sorting result of the similarity.

2. The method of claim 1, wherein the method is characterized by: The step S10 specifically comprises: Collecting a large number of official document samples based on official document type dimension, document sending time dimension, and department dimension; The official document types of the official document type dimension at least include department document sending, department document receiving, business single, and work order; the document sending time of the document sending time dimension includes peak time period, low-estimation time period, and flat-peak time period, and official document samples are collected in a ratio of 3:1:1; the department dimension is used to set the collection weight of each part according to the number of document sending and the number of document receiving. 3.The official document processing person recommendation method based on user portrait collaborative analysis according to claim 1, characterized in that: The step S20 specifically comprises: Step S21, the noise data of each of the collected official document samples is removed; the noise data at least includes URL address, special symbol, picture and zero-width character; Step S22, the traditional Chinese characters contained in each of the official document samples after removing the noise data are converted into simplified Chinese characters; Step S23, a vocabulary table containing a plurality of sensitive words is constructed, and the sensitive words in each of the official document samples are filtered based on the vocabulary table; Step S24, a stop word library containing a plurality of stop words is constructed, and the stop words in each of the official document samples are filtered based on the stop word library; Step S25, each of the official document samples is tagged with a part of speech, and the official document samples with a part of speech of noun are retained to construct an official document data set.

4. The method of claim 1, wherein the method further comprises: determining a user portrait of the user; and determining a user portrait of the candidate; and determining a similarity between the user portrait of the user and the user portrait of the candidate. The step S32 is specifically: Based on the label system, the user portrait is constructed as: P={Info, Keyword, Frequency, Time}; Wherein, P represents the user portrait; Info={Gender, Age, Post, Department}, Gender, Age, Post and Department respectively represent gender, age, post and department; Frequency represents the handling frequency; Time represents the handling time.

5. The method of claim 1, wherein the method further comprises: determining a user portrait of the user; and determining a user portrait of the candidate; and determining a similarity between the user portrait of the user and the user portrait of the candidate. In the step S60, the calculation formula of the F1 value is as follows: F1=(2*P*R) / (P+R); Wherein, P represents the precision, P=TP / (TP+FN); R represents the recall rate, R=TP / (TP+FP); TP represents the positive sample predicted as positive class by the model; FN represents the positive sample predicted as negative class by the model; FP represents the negative sample predicted as positive class by the model.

Images