Intelligent retrieval and personalized recommendation system for archives

Abstract

The invention discloses an intelligent file retrieval and personalized recommendation system, and relates to the technical field of file retrieval, and the system comprises a storage module which forms a plurality of task groups for each file according to task types, builds a task flow axis for each task group, and divides the corresponding task group into a plurality of task child nodes; and the storage module collects corresponding archives under each task sub-node, assigns a retrieval value to each archive, and establishes a recording time axis based on archive filing time to form an archive time chain corresponding to the task sub-node. According to the method, a structure modeling mode of'task flow axis-task child node-document time chain 'is introduced into an archive system, so that the recommendation logic can accurately correspond to the current task stage, the context matching property of the recommendation content is improved, and the user can more efficiently obtain the archive required by the current stage.

Description

Technical Field

[0001] The present invention relates to the technical field of archive retrieval, and in particular to an archive intelligent retrieval and personalized recommendation system. Background Art

[0002] With the widespread use of information systems, many enterprises and institutions have accumulated a large number of structured and unstructured archives during project execution, government processes, and management activities. These archives are usually archived according to time and task classification, and need to be repeatedly retrieved and referenced during subsequent task execution. Existing archive management systems often have basic functions such as keyword search, tag classification, and full-text indexing, which can meet users' search needs to a certain extent.

[0003] In recent years, some archive systems have introduced intelligent recommendation mechanisms, providing a certain degree of personalized recommendations based on user click history, access frequency, or keyword matching results. These technologies are mostly based on collaborative filtering or semantic similarity calculations, which can improve the display efficiency of some archives. However, with the increasing complexity of process-based task structures, users have placed higher demands on the process matching and structural pertinence of recommendation results.

[0004] In actual applications, many archives do not exist independently, but are generated based on the phased steps in a specific task process. If the recommendation logic is only based on user behavior or text similarity, it may cause the recommended content to not match the current task node, reducing user efficiency. Combining the archive recommendation strategy with the process structure to form a process-aware recommendation mechanism is one of the important directions for improving the intelligence level of the archive system.

[0005] In addition, when using the archive system, users exhibit behavioral characteristics such as staged concentration and jumping between nodes. If these behaviors can be integrated with the process structure for modeling, it will help the system predict user needs more accurately. Therefore, the present invention proposes an archive intelligent retrieval and personalized recommendation system. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent archive retrieval and personalized recommendation system to solve the problems mentioned in the above background technology.

[0007] The present invention can be implemented through the following technical solutions: an intelligent archive retrieval and personalized recommendation system, comprising: an identification module, a storage module, a retrieval module and a recommendation module;

[0008] The storage module organizes each file into several task groups according to the task type, and the storage module establishes a task flow axis for each task group, dividing the corresponding task group into multiple task sub-nodes;

[0009] And the storage module collects the corresponding files under each task sub-node, assigns a search value to each file, and establishes a collection time axis based on the filing time of the file to form a file time chain corresponding to the task sub-node;

[0010] The identification module is used to identify the search request input by the user;

[0011] Based on the search request, the retrieval module obtains the process stage of the user's current task and extracts the task subnode closest to the current process stage. At the same time, it combines the archive time chain within the task subnode and the impact factor of each archive to calculate the priority score of each archive. Specifically, the impact factor of the archive includes: update time interval, update frequency, archive version order and historical usage frequency;

[0012] The recommendation module generates a search archive group based on the priority score of each archive and displays it visually;

[0013] At the same time, the recommendation module combines the user's browsing behavior trajectory, task sub-node access rhythm and steady-state judgment to predict the task sub-node that the user is about to enter, generate the corresponding recommendation file group in advance and preload it.

[0014] A further technical improvement of the present invention is that the method for calculating the file priority score includes:

[0015] S1. Determine the user's current process stage:

[0016] S2. Extract the nearest task subnode

[0017] S3, build archive time chain and extract search value;

[0018] S4. For file d i , calculate the update time interval factor, update frequency factor, archive version sequence factor and historical usage frequency factor as influencing factors in turn, and perform normalization processing uniformly:

[0019] S5. The search module calculates the search value of each file and each impact factor by normalized linear weighting to obtain the priority score corresponding to the i-th file.

[0020] A further technical improvement of the present invention is that the storage module dynamically adjusts the search value of each version of multiple files of the same type in the same task subnode and belonging to the same version chain, taking into account their update frequency and usage behavior, so that the score better meets the user's actual query needs. Specifically, the adjustment method includes:

[0021] Z1. Analyze each file number V i, and classify them by file type, build a version chain group G under the same type of files, and sort the files in the version chain group G in ascending time order;

[0022] Z2, calculate the update time interval between adjacent files in the version chain group G, and use the sliding window to count the update density;

[0023] If the update density is greater than the preset density threshold, the segment is determined to be a high-frequency update segment;

[0024] Z3. In each high-frequency update segment, the last file with the timestamp d n , set its search value upper limit B n =B max ;

[0025] And for the intermediate version files d in the high-frequency update segment k , set its search value B k for:

[0026]

[0027] Where B k is the final revised retrieval value of the intermediate version file in the kth high-frequency update segment;

[0028] B b is the initial search value of the corresponding file;

[0029] η is the inhibitory regulatory factor, which is used to control the down-regulation amplitude and its value is in the range [0, B b ] interval;

[0030] U k is the usage frequency of the intermediate version files in the kth high-frequency update segment;

[0031] max(u) is the maximum usage frequency value of all files in the current version chain group G;

[0032] Indicates file d k For the "unpopularity" of the most frequently used archives, the closer it is to 1, the lower the usage rate.

[0033] A further technical improvement of the present invention is that the storage module analyzes the update time intervals of different types of files in the same task subnode, including:

[0034] A1. Classify multiple version chain groups G belonging to different file types in the task subnode, where each version chain group G is composed of files with the same file attribute identifier and a version evolution relationship;

[0035] Extract the first archive time of each type of file in each version chain group G and the update frequency of the files in the group;

[0036] A2. Compare the chain groups G of each version as follows:

[0037] a1. Determine whether there is a version chain group G in which the first update time of each file is earlier than that of other types of files;

[0038] a2. Determine whether the update frequency of each file in the corresponding version chain group G in a1 is higher than the preset frequency threshold;

[0039] A3. If a version chain group G satisfies all the judgment conditions in a1 and a2, the storage module marks the file type to which the version chain group G belongs as the dominant type in the task subnode;

[0040] A4. During the scoring and ranking process, the search module performs weighted improvement on the initial search value of each file in the dominant file type version chain group G.

[0041] A further technical improvement of the present invention is that the storage module sets a structural weight for each task sub-node in the same task group, indicating its stage importance in the overall task.

[0042] The recommendation module also records and analyzes the behavior data of the search archive group that the current user is accessing, including:

[0043] Reading dwell time: refers to the time a user spends on the recommended profile of the current task subnode;

[0044] Browsing ratio: refers to the ratio of the number of users actually visited to the total number of recommended archive groups;

[0045] Action count: refers to the number of user interactions such as clicking, collecting, downloading, and marking in recommended profiles;

[0046] The recommendation module performs normalized linear weighted calculation based on reading dwell time, browsing ratio and number of operation behaviors to obtain the behavior recognition score R u ,The higher the behavior recognition score value is, the higher the user recognition degree is;

[0047] The recommendation module extracts the current task subnode number P based on the task flow axis of the current task group c , and find its subsequent task subnode set {P1,P2,...,P n}, each task sub-node corresponds to a subsequent task stage;

[0048] For each subsequent task subnode p j , through the behavior approval score Ru 、P j The historical jump probability H j and subsequent task subnode P j The structural weight W in the entire task process axis j , get the subsequent task subnode P j The incentive score Q is:

[0049] Q=R u ·H j W j ;

[0050] The recommendation module compares the incentive score Q with the preset incentive threshold. When the incentive score Q of any task sub-node is not less than the incentive threshold, the recommendation module will recommend the subsequent task sub-node P j Join the background pre-generation queue;

[0051] And the system pre-generates the subsequent task sub-node P in the background queue j , do the following:

[0052] Extract subsequent task subnode P j The following archive collection;

[0053] Calculate the priority score based on the node's archive time chain, version chain, and historical behavior data;

[0054] Generate a recommended profile group sorted by ratings;

[0055] The generated results are cached in the pre-read pool for quick call when the user jumps to the task subnode.

[0056] A further technical improvement of the present invention is that the recommendation module optimizes the background generation of the recommended archive group based on the process-aware mechanism, specifically including:

[0057] The system records the user's access behavior information in the current task sub-node, including the time spent in the node, the proportion of recommended archives accessed, and the average archive reading time, and constructs a "user browsing rhythm model" for the task sub-node;

[0058] The browsing time of the current node is R t and the node structure weight W c Compare the two. If the difference is lower than the threshold, the user behavior is determined to be consistent with the structural goal, and the system considers that the current user is in a stable process state.

[0059] Browsing time ratio R t = Total stay time T of the current task subnode c / Total time T that the user has used in the current task group g ;

[0060] At the same time, the node structure weights are pre-set by the system or automatically assigned based on historical task execution data;

[0061] When | R t -W c When |≤the preset judgment threshold, the current user is considered to be in a stable process state;

[0062] Based on the current task flow position, historical jump probability, and task flow axis, the system identifies the set of subsequent task subnodes that the user may enter;

[0063] The historical jump probability is the system's record of the current task subnode P in the historical user behavior. c To the subsequent task subnode P j The proportion of jumps;

[0064] The recommendation module sets a jump threshold based on the connection topology and jump probability of the task process axis, and retains the task subnodes whose historical jump probability is not less than the jump threshold as the set of potential subsequent nodes;

[0065] At the same time, based on the steady-state judgment results, low-relevance sub-nodes with high behavioral jumps and irrelevant structures are filtered out;

[0066] The behavior jumpiness is achieved by counting the probability of users switching between non-adjacent process nodes. If the jump behavior of a task subnode has discontinuous characteristics or no obvious sequential correspondence in the history (for example, skipping more than 1-2 nodes), it is judged as a jump node;

[0067] Structural irrelevance determines whether there is a direct path connection through the task process axis. If there is no reachability, the structure is considered irrelevant and filtered;

[0068] The system performs personalized scoring and sorting operations on the selected subsequent task sub-nodes, generates its recommended profile group, and stores it in the cache or preload module, which can be displayed immediately when the user switches the process stage;

[0069] For each selected task sub-node, the system calls the current recommendation system's scoring model, calculates the priority scores of all files in each task sub-node, and generates a sorted recommended file group;

[0070] The recommendation results are cached in the edge cache module or the client preload queue and bound to the task subnode identifier. When the user switches to the subnode, the system can load and display it immediately.

[0071] A further technical improvement of the present invention is that the storage module establishes an initial connection path between archives with similar update times in different task sub-nodes, and counts the jump utilization rate based on the subsequent continuous jump behavior of users. When the jump utilization rate is not less than the preset utilization threshold, the path is retained, otherwise the connection is cancelled, which is used to dynamically optimize the horizontal recommendation relationship between task sub-nodes.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] By introducing the structural modeling method of "task process axis - task sub-node - document time chain" into the archive system, the present invention enables the recommendation logic to accurately correspond to the current task stage, improves the contextual matching of the recommended content, and helps users obtain the archives they need at the current stage more efficiently;

[0074] Furthermore, the present invention proposes a priority scoring model based on multi-dimensional factors, which comprehensively considers the update interval, update frequency, version sequence and historical usage frequency of the document, thus achieving a comprehensive quantification of the importance of the document, effectively avoiding the deviation of the recommended content or the imbalance of a single dimension, and improving the scientificity and practicality of the recommendation ranking.

[0075] At the same time, by building a user browsing behavior rhythm model and combining it with the structural weights of task sub-nodes to perform steady-state judgment, the present invention can accurately identify whether the user is in a normal process rhythm and, based on this, predict the next task stage they may enter, thereby generating recommended content in advance, shortening response time, and improving user experience.

[0076] On the other hand, the present invention further introduces a tentative jump path identification mechanism between archives, establishes horizontal connection relationships based on update time proximity, and dynamically determines whether to solidify the path as a recommendation relationship through user behavior feedback, thereby achieving self-learning and self-optimization of the system structure and improving the adaptive ability of recommendations;

[0077] Moreover, the present invention combines the user behavior recognition score and process structure prediction mechanism. When the recommendation effect is good, the system can automatically pre-generate the recommendation file group of subsequent task nodes and pre-load the cache, which greatly improves the user browsing continuity and recommendation response speed, and enhances the overall intelligence level of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] To facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0079] Figure 1 This is a functional diagram of the system modules of the present invention. DETAILED DESCRIPTION

[0080] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0081] Example 1

[0082] See also Figure 1 As shown, the present invention provides an intelligent archive retrieval and personalized recommendation system, including: an identification module, a storage module, a retrieval module and a recommendation module;

[0083] The storage module organizes each file into several task groups according to the task type, and establishes a task flow axis for each task group, dividing the corresponding task group into multiple task sub-nodes;

[0084] And the storage module collects the corresponding files under each task sub-node, assigns a search value to each file, and establishes a collection time axis based on the filing time of the file to form a file time chain corresponding to the task sub-node;

[0085] The recognition module is used to recognize the search request input by the user. Specifically, the method for recognizing the search request includes:

[0086] Word segmentation and syntactic analysis:

[0087] Perform word segmentation on the search request text, which can be done by using a rule-based word segmenter (such as Jieba word segmenter) or block parsing based on part-of-speech tagging;

[0088] Perform dependency parsing on sentences to identify verb-noun structures or subject-verb-object combinations;

[0089] Feature keyword extraction:

[0090] Extract task-related keywords (such as "performance appraisal", "material procurement", "contract signing", etc.) that appear in the search text;

[0091] Extracting temporal vocabulary that characterizes process stages (e.g., “drafting,” “deliberation,” “execution,” “archiving,” etc.);

[0092] Vector encoding and label comparison:

[0093] Input the extracted keywords into a pre-trained word vector model (such as Word2Vec) to generate semantic vectors;

[0094] Compare the semantic vector with the "task label word vector set" constructed by the system by calculating the cosine similarity;

[0095] Task structure mapping:

[0096] According to the identified task categories and stages, they are mapped to the node numbers in the task flow axis established in the system, which serve as input parameters for the subsequent retrieval and recommendation modules;

[0097] Based on the search request, the retrieval module obtains the process stage of the user's current task and extracts the task subnode closest to the current process stage. At the same time, it combines the archive time chain within the task subnode and the impact factor of each archive to calculate the priority score of each archive. Specifically, the impact factors of archives include: update time interval, update frequency, archive version order and historical usage frequency;

[0098] The recommendation module generates a search archive group based on the priority score of each archive and displays it visually;

[0099] The calculation method of the archive priority score includes:

[0100] S1. Determination of the user's current process stage:

[0101] The retrieval module receives the user search request recognition result transmitted by the recognition module, including the task group identifier and process stage keywords;

[0102] The retrieval module extracts the task group ID currently associated with the user and the task process stage number P currently in which the user is located. u ;

[0103] S2. Extract the nearest task subnode

[0104] The retrieval module obtains all task sub-node number sets from the task flow axis of the current task group {P i}, and calculate the number difference Δ for each task subnode i =|P u -P i |, and extract the nearest task subnode P * ;

[0105] If there is a node with the smallest difference, select the task subnode containing the most recent archive time file as P * ;

[0106] S3,archive time chain construction and retrieval value extraction;

[0107] At task node P * In [1], the set of belonging archives D is obtained, and the belonging archives are constructed into an archive time chain in ascending order of archiving time, and each archive includes:

[0108] Archive time T i ;

[0109] Version number V i ;

[0110] Frequency of use i ;

[0111] Retrieve value B i , which is generated by default during the collection;

[0112] The storage module dynamically adjusts the search value of each version of multiple files of the same type in the same task subnode and belonging to the same version chain, taking into account their update frequency and usage behavior, so that the score better meets the user's actual query needs. Specifically, the adjustment method includes:

[0113] Z1. Analyze each file number V i , and classify them by file type, build a version chain group G under the same type of files, and sort the files in the version chain group G in ascending time order;

[0114] Z2, calculate the update time interval between adjacent files in the version chain group G, and use the sliding window to count the update density;

[0115] If the update density is greater than the preset density threshold, the segment is determined to be a high-frequency update segment;

[0116] Z3. In each high-frequency update segment, the last file with the timestamp d n , set its search value upper limit B n =B max ;

[0117] And for the intermediate version files d in the high-frequency update segment k , set its search value B k for:

[0118]

[0119] Where B k is the final revised retrieval value of the intermediate version file in the kth high-frequency update segment;

[0120] B b is the initial search value of the corresponding file;

[0121] η is the inhibitory regulatory factor, which is used to control the down-regulation amplitude and its value is in the range [0, B b ] interval;

[0122] U k is the usage frequency of the intermediate version files in the kth high-frequency update segment;

[0123] max(u) is the maximum usage frequency value of all files in the current version chain group G;

[0124] Indicates file d kFor the "unpopularity" of the most frequently used archives, the closer it is to 1, the lower the usage rate.

[0125] S4. For file d i , calculate the following impact factors in turn and perform normalization processing uniformly:

[0126] Update interval factor

[0127]

[0128] Where, T i is the filing time of the i-th file;

[0129] T i-1 The filing time of the i-1th file;

[0130] ∈ is a small constant that prevents the denominator from being zero, used for numerical stability;

[0131] Update frequency factor W f :

[0132]

[0133] Where n is the total number of files under the task subnode;

[0134] T1 is the filing time of the first file;

[0135] T n The filing date of the last file;

[0136] File version sequence factor

[0137]

[0138] Where V i is the version number of the i-th file (the result of numerical processing, such as v1.0→1.0, v3.2→3.2);

[0139] max(V) is within the current task subnode and is consistent with the current file d i The maximum version number contained in a collection of archives with the same attribute identifier (such as archive type, archive tag or classification identifier), reflecting the highest number in the version chain of archives of this type;

[0140] Historical frequency factor

[0141]

[0142] Where U i is the number of times the i-th file is accessed, called, or downloaded;

[0143] ∑ j U j The usage frequency of all files in the task subnode is combined;

[0144] S5. Priority score calculation:

[0145] The retrieval module calculates the retrieval value of each file and each impact factor through normalized linear weighting to obtain the priority score S corresponding to the i-th file. i ;

[0146] Specifically, the formula used is:

[0147]

[0148] Where B i is the search value of the corresponding file;

[0149] ω, α, β, γ, and δ are preset weighting systems for the corresponding items, and can be adjusted based on user behavior. For example, long-term users pay more attention to the version, while temporary users pay more attention to the frequency of hotspot usage.

[0150] The storage module analyzes the update intervals of different types of files in the same task subnode, including:

[0151] A1. Categorize multiple version chain groups G belonging to different file types in the task subnode. Each version chain group G consists of files with the same file attribute identifier and a version evolution relationship.

[0152] Extract the first archive time of each type of file in each version chain group G and the update frequency of the files in the group;

[0153] A2. Compare the chain groups G of each version as follows:

[0154] a1. Determine whether there is a version chain group G in which the first update time of each file is earlier than that of other types of files;

[0155] a2. Determine whether the update frequency of each file in the corresponding version chain group G in a1 is higher than the preset frequency threshold;

[0156] A3. If a version chain group G satisfies all the judgment conditions in a1 and a2, the storage module marks the file type to which the version chain group G belongs as the dominant type in the task subnode;

[0157] A4. During the scoring and ranking process, the search module performs weighted improvement on the initial search value of each file in the dominant file type version chain group G.

[0158] For example, in a government process management system, when a user accesses the "Budget Approval" stage, a task subnode under the "Financial Appropriation Process," the system needs to intelligently sort and recommend multiple different types of archived materials within this stage.

[0159] The files under this task sub-node include:

[0160] Type A: Budget Approval Request Form (number prefixed with "BA-")

[0161] Type B: Financial Instruction Document (number prefixed with "FB-");

[0162] Type C: Attachment file (number prefixed with "AF-");

[0163] The system performs the following operations:

[0164] q1. Version chain group classification and attribute extraction:

[0165] The system automatically classifies the above files into three version chain groups based on the type field, and marks them as

[0166] G A , G B , G C , where the archives in each group form an evolution sequence according to the filing number and time;

[0167] The system extracts the first archiving time of each version chain group (such as ) and updating frequency (i.e. the number of archives of each type per unit time);

[0168] q2. Comparative judgment:

[0169] System judgment Explain that the budget approval form is the first file type to enter the process;

[0170] At the same time, the statistical results show that the budget approval form was updated 5 times within 3 days, while the other file types were updated 2 times and 1 time, respectively, meeting the update frequency threshold of 3;

[0171] q3. Mark the dominant file type:

[0172] System Confirmation G A It meets the dual conditions of “earliest filing” and “high update frequency”, so type A (budget approval application form) is marked as the dominant file type;

[0173] q4. In the subsequent scoring calculation, the system will A The weight boost factor λA = 1.2 is set for all files to calculate their final priority scores, namely:

[0174]

[0175] The storage module sets a structural weight for each task sub-node in the same task group, indicating its stage importance in the overall task.

[0176] The recommendation module also records and analyzes the behavior data of the search archive group that the current user is accessing, including:

[0177] Reading dwell time: refers to the time a user spends on the recommended profile of the current task subnode;

[0178] Browsing ratio: refers to the ratio of the number of users actually visited to the total number of recommended archive groups;

[0179] Action count: refers to the number of user interactions such as clicking, collecting, downloading, and marking in recommended profiles;

[0180] The recommendation module performs normalized linear weighted calculation based on reading dwell time, browsing ratio and number of operation behaviors to obtain the behavior recognition score R u ,The higher the behavior recognition score value is, the higher the user recognition degree is;

[0181] The recommendation module extracts the current task subnode number P based on the task flow axis of the current task group C , and find its subsequent task subnode set {P1,P2,...,P n}, each task sub-node corresponds to a subsequent task stage;

[0182] For each subsequent task subnode P j , through the behavior approval score R u 、P j The historical jump probability H j and subsequent task subnode P j The structural weight W in the entire task process axis j , get the subsequent task subnode P j The incentive score Q is:

[0183] Q=R u ·H j W j ;

[0184] The recommendation module compares the incentive score Q with the preset incentive threshold. When the incentive score Q of any task sub-node is not less than the incentive threshold, the recommendation module will recommend the subsequent task sub-node P j Join the background pre-generation queue;

[0185] And the system pre-generates the subsequent task sub-node P in the background queue j , do the following:

[0186] Extract subsequent task subnode P j The following archive collection;

[0187] Calculate the priority score based on the node's archive time chain, version chain, and historical behavior data;

[0188] Generate a recommended profile group sorted by ratings;

[0189] The generated results are cached in the pre-read pool for quick call when the user jumps to the task subnode.

[0190] The recommendation module optimizes the background generation of recommended profile groups based on the process-aware mechanism, including:

[0191] The system records the user's access behavior information in the current task sub-node, including the time spent in the node, the proportion of recommended archives accessed, and the average archive reading time, and constructs a "user browsing rhythm model" for the task sub-node;

[0192] Specifically, the "user browsing rhythm model" is constructed as follows:

[0193] The total stay time T of the current task subnode c , Total number of accessed files N v , average browsing time per file As the core variable, the behavior vector B is constructed after normalization c ;

[0194] Used for subsequent calculation of matching degree with structural rhythm;

[0195] The browsing time of the current node is R t and the node structure weight W c Compare the two. If the difference is lower than the threshold, the user behavior is determined to be consistent with the structural goal, and the system considers that the current user is in a stable process state.

[0196] Browsing time ratio R t = Total stay time T of the current task subnode c / Total time T that the user has used in the current task group g ;

[0197] At the same time, the node structure weight is pre-set by the system or automatically assigned based on historical task execution data (for example, by statistics on the average number of files and historical node processing time);

[0198] When | R t -W cWhen |≤the preset judgment threshold (e.g., judgment threshold = 10%), the current user is considered to be in a stable process state;

[0199] Based on the current task flow position, historical jump probability, and task flow axis, the system identifies the set of subsequent task subnodes that the user may enter;

[0200] The historical jump probability is the system's record of the current task subnode P in the historical user behavior. c To the subsequent task subnode P j The proportion of jumps;

[0201] The recommendation module sets a jump threshold based on the connection topology and jump probability of the task process axis, and retains the task subnodes whose historical jump probability is not less than the jump threshold as the set of potential subsequent nodes;

[0202] At the same time, based on the steady-state judgment results, low-relevance sub-nodes with high behavioral jumps and irrelevant structures are filtered out;

[0203] Behavior jumpiness is achieved by counting the probability of users switching between non-adjacent process nodes. If the jump behavior of a task subnode has discontinuous characteristics or no obvious sequential correspondence in the history (for example, skipping more than 1-2 nodes), it is judged as a jump node;

[0204] Structural irrelevance determines whether there is a direct path connection through the task process axis. If there is no reachability, the structure is considered irrelevant and filtered;

[0205] The system performs personalized scoring and sorting operations on the selected subsequent task sub-nodes, generates its recommended profile group, and stores it in the cache or preload module, which can be displayed immediately when the user switches the process stage;

[0206] For each selected task sub-node, the system calls the current recommendation system's scoring model, calculates the priority scores of all files in each task sub-node, and generates a sorted recommended file group;

[0207] The recommendation results are cached in the edge cache module or the client preload queue and bound to the task subnode identifier. When the user switches to the subnode, the system can load and display it immediately.

[0208] Example 2

[0209] An intelligent archive retrieval and personalized recommendation system includes: an identification module, a storage module, a retrieval module and a recommendation module;

[0210] The storage module organizes each file into several task groups according to the task type, and establishes a task flow axis for each task group, dividing the corresponding task group into multiple task sub-nodes;

[0211] And the storage module collects the corresponding files under each task sub-node, assigns a search value to each file, and establishes a collection time axis based on the filing time of the file to form a file time chain corresponding to the task sub-node;

[0212] Based on the search request, the retrieval module obtains the process stage of the user's current task and extracts the task subnode closest to the current process stage. At the same time, it combines the archive time chain within the task subnode and the impact factor of each archive to calculate the priority score of each archive. Specifically, the impact factors of archives include: update time interval, update frequency, archive version order and historical usage frequency;

[0213] The recommendation module generates a search archive group based on the priority score of each archive and displays it visually;

[0214] At the same time, the recommendation module combines the user's browsing behavior trajectory, task sub-node access rhythm and steady-state judgment to predict the task sub-node that the user is about to enter, generate the corresponding recommendation file group in advance and preload it.

[0215] Compared with Example 1, the storage module in Example 2 establishes an initial connection path between archives with similar update times in different task subnodes, and calculates the jump utilization rate based on the subsequent continuous jump behavior of users. When the jump utilization rate is not less than a preset utilization threshold, the path is retained; otherwise, the connection is cancelled, which is used to dynamically optimize the horizontal recommendation relationship between task subnodes, specifically including:

[0216] Y1. The storage module regularly analyzes the file update time of all task nodes in the task group;

[0217] For two or more task subnodes P i , P j If the difference in archiving time between the archives in ... is less than the time threshold, it is considered as "synchronous update behavior";

[0218] And the storage module establishes an "initial connection path" for the corresponding files of each task sub-node determined to be "synchronous update behavior", which is recorded as an associated pair (d i ,d j ), and record the connection timestamp T link ;

[0219] Y2, the storage module records Y1 as an associated pair (d i ,d j )’s files are included in the “temporary recommendation connection pool” for subsequent behavior judgment;

[0220] If the current recommendation module needs to recommend profile d to the user i , then the associated file d is also promotedj The recommendation probability of

[0221] The storage module converts the "initial connection path" into a "tentative path" for subsequent behavior judgment and confirmation of whether it can be maintained in the long term;

[0222] Y3. During the user's use, the system records the i To file d j The actual access jump situation;

[0223] For all file association pairs that are converted to "tentative paths", count the actual number of user jumps within the preset time N jump (d i →d j ) and the ratio of the number of recommendations to obtain the jump utilization rate;

[0224] Y4. When the jump utilization rate is not less than the preset utilization rate threshold, the storage module will mark the file association pair (d i ,d j ) is solidified as a recommended path for the corresponding task sub-node, and in subsequent calculations, the corresponding path weighting factor is added to the scoring model to improve the priority score of each file in the file association pair;

[0225] Otherwise, the storage module converts the "tentative path" into an invalid link and removes it from subsequent recommendations.

[0226] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An intelligent archive retrieval and personalized recommendation system, characterized by: include: The storage module divides each file into several task groups according to the task type, and establishes a task flow axis for each task group, which is divided into multiple task sub-nodes; And assign retrieval values ​​to the archives under each task sub-node, and establish a collection timeline based on the archiving time to form an archive time chain; An identification module, used to identify a search request input by a user; The retrieval module obtains the process stage of the user's current task based on the search request, extracts the task subnode closest to it, and calculates the priority score based on the archive time chain and the impact factor of each archive in the subnode; The recommendation module generates a search profile group based on the priority score of each profile.

2. The intelligent archive retrieval and personalized recommendation system according to claim 1, characterized in that: The calculation method of the file priority score includes: S1. Determine the user's current process stage: S2. Extract the nearest task subnode S3, build archive time chain and extract search value; S4. For file d i , calculate the update time interval factor, update frequency factor, archive version sequence factor and historical usage frequency factor as influencing factors in turn, and perform normalization processing uniformly: S5. The search module calculates the search value of each file and each impact factor by normalized linear weighting to obtain the priority score corresponding to the i-th file.

3. The intelligent archive retrieval and personalized recommendation system according to claim 1, characterized in that: The storage module dynamically adjusts the search value of each version of multiple files belonging to the same version chain in the same type of files in the same task sub-node, taking into account their update frequency and usage behavior.

4. The intelligent archive retrieval and personalized recommendation system according to claim 3, characterized in that: Methods for adjusting retrieved values ​​include: Z1. Analyze each file number V i , and classify them by file type, build a version chain group G under the same type of files, and sort the files in the version chain group G in ascending time order; Z2, calculate the update time interval between adjacent files in the version chain group G, and use the sliding window to count the update density; If the update density is greater than the preset density threshold, the segment is determined to be a high-frequency update segment; Z3. In each high-frequency update segment, the last file with the timestamp d n , set its search value upper limit B n =B max ; And for the intermediate version files d in the high-frequency update segment k , set its search value B k for: Where B b is the initial search value of the corresponding file; η is the inhibition adjustment factor; U k is the usage frequency of the intermediate version files in the kth high-frequency update segment; max(u) is the maximum usage frequency value of all files in the current version chain group G.

5. The intelligent archive retrieval and personalized recommendation system according to claim 1, characterized in that: The storage module analyzes the update time intervals of different types of files in the same task subnode, including: A1. Classify multiple version chain groups G belonging to different file types in the task subnode, where each version chain group G is composed of files with the same file attribute identifier and a version evolution relationship; Extract the first archive time of each type of file in each version chain group G and the update frequency of the files in the group; A2. Compare the chain groups G of each version as follows: a1. Determine whether there is a version chain group G in which the first update time of each file is earlier than that of other types of files; a2. Determine whether the update frequency of each file in the corresponding version chain group G in a1 is higher than the preset frequency threshold; A3. If a version chain group G satisfies all the judgment conditions in a1 and a2, the storage module marks the file type to which the version chain group G belongs as the dominant type in the task subnode; A4. During the scoring and ranking process, the search module performs weighted improvement on the initial search value of each file in the dominant file type version chain group G.

6. The intelligent archive retrieval and personalized recommendation system according to claim 1, characterized in that: The storage module sets a structural weight for each task sub-node in the same task group; The recommendation module records, analyzes, and performs normalized linear weighted calculations on the behavior data of the search archive group that the current user is accessing to obtain a behavior recognition score; Behavioral data includes: reading dwell time, browsing ratio and number of operation behaviors; The recommendation module extracts the current task subnode number based on the task flow axis of the current task group and searches for its subsequent task subnode set {P1, P2, ..., P n }, each task sub-node corresponds to a subsequent task stage; For each subsequent task subnode P j , through behavioral recognition score, P j The historical jump probability and the structural weight of the subsequent task sub-node in the entire task process axis are used to obtain the subsequent task sub-node P j Incentive rating.

7. The intelligent archive retrieval and personalized recommendation system according to claim 6, characterized in that: The recommendation module compares the incentive score with the preset incentive threshold. When the incentive score of any task sub-node is not less than the incentive threshold, the recommendation module will assign the subsequent task sub-node P to the task. j Join the background pre-generation queue; And the system pre-generates the subsequent task sub-node P in the background queue j , do the following: Extract subsequent task subnode P j The following archive collection; Calculate the priority score based on the node's archive time chain, version chain, and historical behavior data; Generate a recommended archive group sorted by score and cache the generated results in the pre-read pool.

8. The intelligent archive retrieval and personalized recommendation system according to claim 7, characterized in that: The recommendation module optimizes the background generation of recommended profile groups based on the process-aware mechanism, specifically including: The system records the user's access behavior information in the current task sub-node, including the time spent in the node, the proportion of recommended archives accessed, and the average archive reading time, and constructs a "user browsing rhythm model" for the task sub-node; The browsing time ratio of the current node is R t and node structure weight W c Make comparisons; If | R t -W c When |≤the preset judgment threshold, the current user is considered to be in a stable process state.

9. The intelligent archive retrieval and personalized recommendation system according to claim 8, characterized in that: Based on the current task flow position, historical jump probability, and task flow axis, the system identifies the set of subsequent task subnodes that the user may enter; The recommendation module sets a jump threshold based on the connection topology and jump probability of the task process axis, and retains the task subnodes whose historical jump probability is not less than the jump threshold as the potential subsequent node set; At the same time, based on the steady-state judgment results, low-relevance sub-nodes with high behavioral jumps and irrelevant structures are filtered out; For each selected task sub-node, the system calls the current recommendation system's scoring model, calculates the priority scores of all files in each task sub-node, and generates a sorted recommended file group.

10. The intelligent archive retrieval and personalized recommendation system according to claim 9, characterized in that: The storage module establishes an initial connection path between archives with similar update times in different task sub-nodes, and calculates the jump utilization rate based on the subsequent continuous jump behavior of users. When the jump utilization rate is not less than a preset utilization threshold, the path is retained; otherwise, the connection is cancelled.

Images