Statistical analysis method for data called by knowledge base

Abstract

The invention discloses a knowledge base calling data statistical analysis method, and relates to the technical field of knowledge base data processing. The method comprises the steps of collecting behavior data, processing the data, constructing a characteristic spectrum, analyzing the characteristic spectrum, outputting an evaluation index, carrying out weighted calibration on the evaluation index, generating a user portrait and health degree comprehensive score, generating a decision report, and carrying out self-adaptive updating according to the decision report. According to the method, through multi-dimensional behavior analysis, feature map construction and a closed-loop optimization mechanism, deep insight and self-adaptive adjustment and optimization of the use condition of the knowledge base are realized, and the personalized service level, the content health degree and the overall operation efficiency of the knowledge base are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge base data processing, and particularly relates to a knowledge base calling data statistical analysis method. BACKGROUND

[0002] The knowledge base is a structured digital asset library integrating explicit and implicit knowledge in a professional field and supporting search and sharing. The core value of the calling data statistics of the knowledge base lies in optimizing knowledge management driven by data. By analyzing data such as high-frequency / low-calling entries, search behaviors, and user preferences, the structure of the knowledge base can be optimized, knowledge gaps can be filled, and search efficiency and correlation can be improved.

[0003] At present, in the use data analysis and optimization decision of the knowledge base system, simple statistical indicators are generally relied on for isolated accounting and artificial experience judgment. This method cannot perform deep fusion and context association analysis on the massive, multi-source, and time-series behavior data generated in the interaction process between the user and the knowledge base.

[0004] Therefore, the present application provides a knowledge base calling data statistical analysis method to solve the above problems. SUMMARY

[0005] The main purpose of the present application is to provide a knowledge base calling data statistical analysis method to solve the problems raised in the above background.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is as follows: a knowledge base calling data statistical analysis method, the method comprising the following steps:

[0007] S1, collecting multi-dimensional behavior data of a user in a knowledge base system in real time to generate original behavior data;

[0008] S2, cleaning, fusing, and structuring the original behavior data stream to generate a standardized behavior data sequence;

[0009] S3, based on the knowledge base topology structure and the semantic association network, performing feature enhancement and context association analysis on the standardized behavior data sequence to construct a user-knowledge interaction feature map;

[0010] S4, inputting the user-knowledge interaction feature map into a multi-task evaluation model of fusion graph embedding and lightweight time-series model for analysis, and outputting multi-dimensional original evaluation indexes including demand preference intensity, search path efficiency, and content correlation degree;

[0011] S5, integrating a historical behavior mode library and an expert rule library, dynamically weighting and adaptively calibrating the multi-dimensional original evaluation indexes to generate a user portrait and a knowledge base health degree comprehensive score;

[0012] S6. Based on the user profile and the comprehensive score of the knowledge base health, generate an intelligent decision report that includes personalized ranking strategies, knowledge gap warnings, and content optimization suggestions;

[0013] S7. Based on the intelligent decision report, the content weight, sorting algorithm and display logic of the knowledge base are dynamically adjusted through an adaptive update engine driven by heuristic rules and lightweight optimization algorithms, and the evaluation model parameters are updated synchronously.

[0014] Preferably, the real-time acquisition of multi-dimensional behavioral data in S1 includes the following steps:

[0015] S11. Capture explicit behavioral data of users on the knowledge base interface through tracking technology, including search keyword sequences, item click flow, page dwell time, page turning frequency, and content download or collection records.

[0016] S12. Collect implicit user behavior data through the log analysis system, including the modification history of query statements, mouse hover trajectory, scroll speed changes, and operation interruption information;

[0017] S13. Associate user identity attributes with contextual environment data, including user role, department, access time period, device type and task context label, and synchronize and align all data according to global timestamp to generate raw behavioral data stream with multi-dimensional labels.

[0018] Preferably, the processing of the raw behavioral data stream in step S2 includes the following steps:

[0019] S21. A combined method based on rules and statistical outlier detection is used to detect and clean the original behavioral data stream, removing noise data generated by invalid clicks, machine crawlers, and misoperations.

[0020] S22. Lightweight natural language processing techniques are used to segment unstructured search keywords and query statements, remove stop words, extract stems, and perform word frequency-inverse document frequency vectorization representation.

[0021] S23. By using time window sliding and session segmentation techniques, the continuous behavioral data stream is divided into session sequences based on independent retrieval tasks, and the behavioral data within each session is serialized, encoded, and normalized to generate standardized behavioral data sequences.

[0022] Preferably, the construction of the user-knowledge interaction feature map in step S3 includes the following steps:

[0023] S31. Construct a knowledge base association graph, where nodes represent knowledge items, edges represent co-occurrence relationships, tag similarity, and jump frequency between items, and assign basic popularity weights to nodes.

[0024] S32, mapping the user's standardized behavior data sequence to the knowledge base association graph based on the graph embedding method of random walk, dynamically updating the access heat of the nodes and the relationship strength of the edges in the graph, and adding a virtual node representing the user's implicit demand;

[0025] S33, extracting high-order features from the generated dynamic interaction graph, including the subgraph diffusion range triggered by user behavior, the change of the aggregation coefficient of core knowledge nodes, and the deviation degree of the retrieval path from the optimal path, and encoding as a graph structure feature vector.

[0026] Preferably, the analysis in S4 using a multi-task evaluation model includes the following steps:

[0027] S41, constructing a multi-task evaluation model architecture, wherein the graph embedding branch processes the user-knowledge interaction feature graph, the bidirectional LSTM branch based on the attention mechanism processes the time sequence dependence of the user behavior sequence, and the lightweight semantic branch based on BERT fine-tuning processes the deep semantics of the query intent;

[0028] S42, inputting the graph structure feature vector and the behavior sequence vector into the multi-task evaluation model, respectively calculating the preference intensity score of the user to a specific knowledge field, the efficiency score of its retrieval path relative to the group average efficiency, and the correlation degree score of its query content and the coverage range of the knowledge base;

[0029] S43, dynamically integrating the outputs of the three branches through a weighted fusion strategy, and using a Softmax function with a smoothing factor to generate a normalized multi-dimensional original evaluation index.

[0030] Preferably, the construction of the multi-task evaluation model architecture in S41 further includes the following steps:

[0031] S411, using a pre-training strategy based on transfer learning to initialize the model, so that it can quickly adapt to new users or new knowledge fields, and realize small sample adaptation by fine-tuning a small number of parameters;

[0032] S412, introducing a consistency training mechanism based on clustering, constructing positive and negative sample pairs, and training the model to distinguish between efficient retrieval behavior and inefficient / ineffective behavior, in order to improve the model's ability to recognize differences in behavior patterns;

[0033] S413, deploying a sliding window incremental update module, using the recent generated behavior data stream to incrementally update the model, and using an elastic weight consolidation strategy to prevent catastrophic forgetting of the model during incremental learning.

[0034] Preferably, the dynamic weighting and adaptive calibration in S5 includes the following steps:

[0035] S51, establish an expert rule base to store the knowledge base optimization targets and priority weights set for different business scenarios and user roles;

[0036] S52, analyze the user's historical behavior sequence to extract the learning curve trend, interest migration pattern, and retrieval ability change characteristics as personalized weighting factors;

[0037] S53, use a method combining weighted fusion and Bayesian inference to fuse the multi-dimensional original evaluation indicators, expert rules, and personalized weighting factors to calculate the user portrait score reflecting individual characteristics and the health degree comprehensive score reflecting the overall state of the knowledge base.

[0038] Preferably, the method combining weighted fusion and Bayesian inference in S53 includes the following steps:

[0039] S531, design a time decay function and a confidence accumulation function to make the recent and high-frequency behavior data have a higher influence weight on the user portrait, and to improve the evaluation confidence with the accumulation of data;

[0040] S532, use a rule quantization mapping table to handle the fuzziness and uncertainty in the expert rules to convert qualitative rules into quantifiable influence factors;

[0041] S533, establish a feedback correction loop to use the user's subsequent behavior on the adjusted knowledge base as a feedback signal to dynamically adjust the weight parameters in the weighted fusion and the Bayesian prior distribution.

[0042] Preferably, the generation of the intelligent decision report in S6 includes the following steps:

[0043] S61, based on the user portrait score, use a hybrid algorithm of collaborative filtering and content recommendation to generate a personalized knowledge item ranking list and potential interest knowledge push suggestions for the user;

[0044] S62, based on the knowledge base health degree comprehensive score, identify knowledge coverage blind spots, hot overload areas, and knowledge aging items to generate optimization suggestions for content supplement, reconstruction, or archiving;

[0045] S63, use automated report generation technology to integrate the ranking strategy, push suggestions, optimization suggestions, and key indicator visual charts into a structured report, and push it to the knowledge base management platform through an interface.

[0046] Preferably, the formation of the closed-loop iterative system in S7 includes the following steps:

[0047] S71, the adaptive updating engine takes the knowledge base health degree comprehensive score and the user search efficiency improvement rate as optimization targets, adopts heuristic rules and gradient descent optimization algorithm to dynamically adjust the weight distribution strategy and interface display rule of the knowledge base;

[0048] S72, a model and threshold joint debugging mechanism is set, when the user group behavior distribution drifts or the knowledge base content is updated on a large scale, the retraining and calibration process of the multi-task evaluation model and related evaluation threshold is automatically triggered;

[0049] S73, the optimized knowledge base configuration and the updated model parameters are deployed synchronously, and a new round of behavior data is continuously collected.

[0050] The application has the following beneficial effects:

[0051] 1. In the application, the multi-task evaluation model is constructed by fusing graph embedding and lightweight time sequence model, and the user-knowledge interaction feature graph is analyzed, so that demand preference intensity, search path efficiency and content correlation degree and other multi-dimensional original evaluation indexes can be accurately extracted from massive and multi-source behavior data, the problems of simple statistics, single evaluation dimension and lack of deep semantic association in traditional methods are solved, and the accuracy and comprehensiveness of quantitative insight into user behavior and knowledge base usage state are improved.

[0052] 2. In the application, by integrating historical behavior patterns and expert rules, and using the method of weighted fusion and Bayesian inference combination to dynamically weight and adaptively calibrate multi-dimensional indexes, highly personalized user portraits and knowledge base health degree comprehensive scores reflecting overall state can be generated, the problems of user portrait static and one-sided, knowledge base state evaluation lag and lack of personalized consideration are solved, and a reliable data foundation is provided for realizing accurate personalized service and scientific content management.

[0053] 3. In the application, the adaptive updating engine is driven based on intelligent decision report, and the model and threshold joint debugging mechanism is set, so that the knowledge base configuration can be dynamically adjusted, the sorting algorithm can be optimized, and the evaluation model parameters can be updated synchronously, forming an iterative system of analysis-decision-optimization, solving the problems of traditional knowledge base system optimization relying on artificial experience, slow response and unable to continuously evolve, and significantly improving the automation and intelligent operation level and long-term adaptability of the knowledge base system. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The method flow chart of the knowledge base calling data statistical analysis method of the application;

[0055] Figure 2 The flow chart of collecting multi-dimensional behavior data in S1 of the knowledge base calling data statistical analysis method of the application;

[0056] Figure 3 A flow chart for processing the original behavior data stream in S2 of the knowledge base calling data statistical analysis method of the present application;

[0057] Figure 4 A flow chart for constructing the user-knowledge interaction feature map in S3 of the knowledge base calling data statistical analysis method of the present application;

[0058] Figure 5 A flow chart for analysis in S4 of the knowledge base calling data statistical analysis method of the present application using a multi-task evaluation model;

[0059] Figure 6 A flow chart for dynamic weighting and adaptive calibration in S5 of the knowledge base calling data statistical analysis method of the present application;

[0060] Figure 7 A flow chart for generating an intelligent decision report in S6 of the knowledge base calling data statistical analysis method of the present application;

[0061] Figure 8 A flow chart for forming a closed-loop iterative system in S7 of the knowledge base calling data statistical analysis method of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0063] Please refer to Figures 1-8 The present application provides a technical solution: a knowledge base calling data statistical analysis method, which comprises the following steps:

[0064] S1, collecting multi-dimensional behavior data of users in a knowledge base system in real time to generate original behavior data;

[0065] S2, cleaning, fusing and structuring the original behavior data stream to generate a standardized behavior data sequence;

[0066] S3, based on the knowledge base topology structure and semantic association network, performing feature enhancement and context association analysis on the standardized behavior data sequence to construct a user-knowledge interaction feature map;

[0067] S4, input the user-knowledge interaction feature map into a multi-task evaluation model combining graph embedding and lightweight time series model for analysis, and output multi-dimensional original evaluation indexes including demand preference intensity, search path efficiency and content correlation degree;

[0068] S5, integrate the historical behavior pattern library and the expert rule library, dynamically weight and adaptively calibrate the multi-dimensional original evaluation indexes, and generate a user portrait and a knowledge base health degree comprehensive score;

[0069] S6, according to the user portrait and the knowledge base health degree comprehensive score, generate an intelligent decision report containing personalized sorting strategy, knowledge gap warning and content optimization suggestion;

[0070] S7, based on the intelligent decision report, through the adaptive update engine driven by heuristic rules and lightweight optimization algorithm, dynamically adjust the content weight, sorting algorithm and display logic of the knowledge base, and synchronously update the evaluation model parameters.

[0071] S1, real-time collection of multi-dimensional behavior data includes the following steps:

[0072] S11, capture user explicit behavior data in the knowledge base interface through the point embedding technology, including search keyword sequence, entry click stream, page dwell time, page turning frequency, content download or collection record, including the following steps:

[0073] S111, implant standardized point embedding code in the key interactive elements (such as search box, knowledge entry link, page turning control, download button) of the knowledge base system Web interface or mobile terminal application, and generate standardized log messages containing event type, event target, timestamp, session ID when the user triggers the corresponding event;

[0074] S112, to avoid blocking the main thread and data loss, the generated log messages are temporarily stored in local IndexedDB or memory queue; a sliding time window and capacity double threshold trigger mechanism is adopted for batch asynchronous reporting; the time window , the upper limit of the queue capacity When any of the conditions or queue length is met, the buffered data is packaged and uploaded to the server-side log collection interface;

[0075] S113, the server-side interface receives the reported data packet, first performs format compliance verification and basic cleaning; the verification formula includes checking the monotonicity of the timestamp (for the same session ID) and the rationality of , where is the current server time, For the allowed maximum clock skew, usually set to 300 seconds), discard obviously invalid data packets;

[0076] S12, collect the implicit behavior data of the user through the log analysis system, including the modification history of the query statement, the mouse hovering track, the scrolling speed change and the operation interruption information, including the following steps:

[0077] S121, record the user interaction details by listening to the input event, the mouse movement event, the scrolling event, the focus loss event and other browser native events; for the query statement modification, record the input box content snapshot and the timestamp when each input event is triggered to form a modification history sequence, and record the mouse hovering track, the cursor moving in and out time and coordinates on the knowledge item; the scrolling speed is calculated by the time difference and the scrolling distance ; ;

[0078] S122, define a maximum allowed event interval within an operation session , by listening to the user's focus switching or long time without operation event, when detecting the eventless time , determine it as an "operation interruption", record the interruption time point and the last valid event before the interruption;

[0079] S123, associate the collected fine-grained implicit behavior log with the explicit behavior event stream reported in step S11 through the user ID and the session ID, align and aggregate them by using the timestamp to form a complete mixed event sequence within a single session of the user, including explicit and implicit behaviors;

[0080] S13, associate the user identity attributes and the context environment data, including the user role, the department to which the user belongs, the access time period, the device type and the task context label, and synchronize and align all the data according to the global timestamp to generate a raw behavior data stream with multi-dimensional labels, including the following steps:

[0081] S131, obtain the role and department attributes corresponding to the user ID from the enterprise unified identity authentication system. Parse the User-Agent string from the HTTP request header, identify the device type through the device fingerprint library, the access time period is directly converted from the event timestamp, and the task context label can be extracted from the business system module where the user stays before accessing the knowledge base or the task ID passed through the URL parameter;

[0082] S132, to solve the time synchronization problem between the client and the server, adopt the network time protocol (NTP) idea for clock compensation. In each data reporting, the client attaches its local timestamp and the time to receive server response The server records the time when the data packet is received. Estimated one-way network latency and clock offset It can be approximated as:

[0083] ;

[0084] Using the calculated The client event timestamps are calibrated to obtain the aligned global timestamps. ;

[0085] S133: Using the user ID and the calibrated global timestamp as the association key, the real-time behavioral events from S11 and S12 are fused with the user attributes and context data from S131. A complete set of multi-dimensional labels is attached to each behavioral event record. Finally, a structured raw behavioral data stream with multi-dimensional labels sorted by timestamp is output for subsequent processing steps.

[0086] The processing of the raw behavioral data stream in S2 includes the following steps:

[0087] S21. A combined method based on rules and statistical outlier detection is used to detect and clean the original behavioral data stream, removing noise data generated by invalid clicks, machine crawlers, and misoperations.

[0088] S211. Set hard rules based on business logic: when the time interval between consecutive clicks on the same knowledge item within a single session is less than 100 milliseconds, it is judged as an invalid click or machine behavior and is removed; when the page dwell time is less than a preset threshold, it is judged as a misoperation or invalid browsing and is marked.

[0089] S212. For key numerical indicators in the user behavior sequence (such as page dwell time, total session duration), adopt a standard deviation-based approach. Outlier identification methods for a given behavioral indicator dataset First, calculate its mean. and standard deviation :

[0090] ;

[0091] Set threshold (generally Choosing 2 or 3 will result in the value falling within the threshold range. Data points outside Identify them as outliers and perform cleaning or smoothing processes;

[0092] S213, for the cleaned data, check whether each session has complete start and end behavior, repair or reject the incomplete session;

[0093] S22, using lightweight natural language processing technology to segment, remove stop words, extract word stems and word frequency-inverse document frequency vectorization of unstructured search keywords and query sentences, including the following steps:

[0094] S221, using open source segmentation tools to segment Chinese search words; load the stop word table, remove the stop words without actual semantics, perform Porter stem extraction on English words, and unify the word form;

[0095] S222, based on the knowledge base of all documents, build a word table, and represent the preprocessed query sentence as a term frequency-inverse document frequency (TF-IDF) vector, the calculation formula is as follows:

[0096] Term frequency (TF): , wherein is the word ; The number of times it appears in the document (query sentence)

[0097] Inverse document frequency (IDF): , wherein is the total number of documents, and the denominator is the number of documents containing the word ;

[0098] TF-IDF weight: ;

[0099] Finally, each query sentence is represented as a TF-IDF feature vector based on the entire word table;

[0100] S223, to reduce the computational complexity, only the vector dimensions corresponding to the top K (such as K=200) feature words of the TF-IDF weight are retained, and a lightweight semantic vector is generated;

[0101] S23, by time window sliding and session segmentation technology, the continuous behavior data stream is cut into session sequences in units of independent retrieval tasks, and the behavior data in each session is serialized and normalized to generate standardized behavior data sequences, including the following steps:

[0102] S231, define the session timeout threshold , when the time interval between two adjacent behavior events exceeds , it is considered as the end of an independent retrieval session and the beginning of a new session, according to this rule, the continuous data stream is cut into multiple sessions ;

[0103] S232, for a single session S22, for each session, the behavior events within the session are arranged in chronological order of their timestamps, forming a behavior sequence, each event is encoded as a feature tuple, including: behavior type, associated knowledge item ID, duration, and the query semantic vector generated in step S22;

[0104] S233, for numerical features in the sequence (such as duration, page flipping frequency), min-max normalization is performed to scale them to the [0, 1] interval, for a feature , its normalized value The calculation formula is:

[0105] ;

[0106] where, and are the minimum and maximum values of the feature in the entire training data set, finally, the standardized behavior data sequence that can be used for model input is generated.

[0107] The construction of the user-knowledge interaction feature map in S3 includes the following steps:

[0108] S31, construct a knowledge base association graph, where nodes represent knowledge items, edges represent co-occurrence relationships between items, label similarity and jump frequency, and assign basic heat weights to nodes;

[0109] S32, use a graph embedding method based on random walk to map the user's standardized behavior data sequence to the knowledge base association graph, dynamically update the access heat of the nodes in the graph and the relationship strength of the edges, and add virtual nodes representing the user's implicit demand, including the following steps:

[0110] S321, for each access node in the user behavior sequence , start random walk on the knowledge base association graph, the calculation formula is:

[0111] ;

[0112] where is the transition probability, is the comprehensive edge weight, is the edge weight from node to its neighbor node , and is the neighbor set of node , and represents each neighbor node of node ;

[0113] S322, dynamic graph update: update the graph structure in real time according to user behavior:

[0114] Node access hotness update: When a user accesses a node , its hotness weight is updated in an exponential decay manner:

[0115]

[0116] where is the decay coefficient, is the time interval since the last access;

[0117] Edge relationship strength update: If a user accesses nodes and in a single session, the transition frequency component of the corresponding edge is updated as follows:

[0118]

[0119] where is the learning rate, and then the comprehensive edge weight is recalculated;

[0120] Virtual node creation: When a user performs a search but does not click on any results, a virtual node is inserted in the graph based on the TF-IDF vector of the search query, connected to the top K real nodes with the highest semantic similarity, and the edge weight is calculated based on semantic similarity;

[0121] S323, on the updated graph, learn node embedding vectors using the Node2Vec algorithm, control the breadth-first or depth-first tendency of the walk by adjusting the hyperparameters p and q, and use the Skip-gram model to optimize the objective function:

[0122]

[0123] where is the embedding vector of node , is the neighbor node obtained by the sampling strategy, is the normalization constant; S33, extract high-order features from the generated dynamic interaction graph, including the diffusion range of subgraphs triggered by user behavior, the change of the clustering coefficient of core knowledge nodes, the deviation degree of retrieval paths from optimal paths, and encode them into a graph structure feature vector.

[0124] The analysis in S4 using a multi-task evaluation model includes the following steps:

[0125]

[0126] ​​​​S41, construct a multi-task evaluation model architecture, wherein the graph embedding branch processes the user-knowledge interaction feature graph, the bidirectional LSTM branch based on the attention mechanism processes the time sequence dependence of the user behavior sequence, and the lightweight semantic branch based on BERT fine-tuning processes the deep semantics of the query intent;

[0127] S42, input the graph structure feature vector and the behavior sequence vector into the multi-task evaluation model, respectively calculate the preference intensity score of the user to a specific knowledge field, the efficiency score of the search path of the user relative to the group average efficiency, and the relevance score of the query content of the user to the coverage range of the knowledge base, including the following steps:

[0128] S421, splice the global representation vector of the graph and the representation vector of the behavior sequence , input a full connection layer, and calculate the original score of the preference intensity of the user to the pre-defined knowledge field:

[0129] ;

[0130] wherein, . The preference intensity score of the field is calculated by the Softmax function:

[0131] ;

[0132] S422, the efficiency score is based on the comparison between the actual search path length of the user and the group average optimal path length. Let the search path length of the user be , the group average optimal path length under this task type be . The calculation formula of the efficiency score is:

[0133] ;

[0134] This score is predicted and calibrated by a regression head (single neuron full connection layer) from the representation vector of the time sequence dependence branch;

[0135] S423, query content relevance score calculation: the relevance score measures the matching degree between the query intent and the overall content coverage of the knowledge base (represented by the TF-IDF vector center of all entries of the knowledge base). The cosine similarity is used for calculation:

[0136] ;

[0137] This score is from the representation vector of the semantic understanding branch​ After a fully connected layer, map to scalar, and weighted fusion with the cosine similarity described above as the final output;

[0138] S43, dynamically integrate the outputs of the three branches through a weighted fusion strategy, and use a Softmax function with a smoothing factor to generate a normalized multi-dimensional original evaluation index, including the following steps:

[0139] S431, dynamic weight learning: design a weight generation network to take the intermediate representations of the three branches as input, and output the dynamic fusion weights corresponding to the three tasks

[0140]

[0141] S432, index fusion and smoothing: multiply the original scores of the three tasks respectively with the corresponding dynamic weights, and normalize them. For multi-class outputs such as preference intensity, use a Softmax function with a smoothing factor to ensure numerical stability:

[0142]

[0143] where T is the temperature parameter, used to control the smoothing degree of the output distribution.

[0144] S433, combine the normalized indexes into a multi-dimensional original evaluation index vector:

[0145]

[0146] This vector is used as the input of the subsequent dynamic weighting and adaptive calibration module.

[0147] The construction of the multi-task evaluation model architecture in S41 also includes the following steps:

[0148] S411, use a pre-training strategy based on transfer learning to initialize the model, so that it can quickly adapt to new users or new knowledge fields, and achieve small sample adaptation by fine-tuning a small number of parameters;

[0149] S412, introduce a consistency training mechanism based on clustering, construct positive and negative sample pairs, and train the model to distinguish between efficient retrieval behavior and inefficient / inactive behavior, in order to improve the model's ability to recognize differences in behavior patterns, including the following steps:

[0150] S4121, in the feature space, use the K-Means algorithm to cluster the behavior sequence representation vectors of all users to discover different behavior patterns, and the number of clusters​​​​​ Determined by elbow rule, clustering centers are calculated by iterative update:

[0151] ;

[0152] where is the sample set assigned to the th cluster;

[0153] S4122, positive and negative sample pair construction: according to the search path efficiency score and clustering result, construct sample pairs:

[0154] Positive sample pair: randomly select two behavior samples with efficiency scores higher than the threshold value from the same cluster ; ;

[0155] Negative sample pair: first, select a high efficiency ( ) sample and a low efficiency ( ) sample from the same cluster ; second, randomly select two samples from different clusters ;

[0156] S4123, consistency contrast loss: use the constructed positive and negative sample pairs to add contrast learning loss in the representation space of the time sequence dependence branch , adopt normalized temperature scale cross entropy loss (NT-Xent):

[0157] ;

[0158] where, is the positive sample pair set, is the sample corresponding negative sample set, is the cosine similarity, is the temperature parameter, this loss encourages the model to map high efficiency behavior to similar representations and distinguish it from low efficiency behavior;

[0159] S413, deploy a sliding window incremental update module, use the recently generated behavior data stream to perform incremental update on the model, and use the elastic weight consolidation strategy to prevent catastrophic forgetting of the model in the incremental learning process, including the following steps:

[0160] S4131, set a sliding window with a fixed time length , only keep the user behavior data within the window for model update, the window slides forward with time, and the old data is removed from the window;

[0161] S4132、In incremental learning, to avoid overfitting on new data and forgetting old knowledge, an EWC regularizer is added to the loss function. First, the old data The importance of model parameters is evaluated by the diagonal elements of the Fisher information matrix :

[0162] ;

[0163] S4133、When new behavior data arrives, the total loss function of the model is :

[0164] ;

[0165] where, is the standard task loss, is the important parameter value of the model trained on old data, is a hyperparameter that controls the strength of regularization. By minimizing , the model is constrained to not deviate too far from its optimal value on the old task while adapting to new data , thus mitigating catastrophic forgetting. The update is performed using mini-batch stochastic gradient descent.

[0166] Dynamic weighting and adaptive calibration in S5 includes the following steps:

[0167] S51、Establish an expert rule base to store knowledge base optimization goals and priority weights set for different business scenarios and user roles;

[0168] S52、Analyze the user's historical behavior sequence to extract the learning curve trend, interest transfer pattern, and retrieval ability change characteristics as personalized weighting factors, including the following steps:

[0169] S521、Learning curve trend modeling: For the user's historical retrieval efficiency score sequence , an exponential decay model is used to fit the learning trend, the model is , where is the cumulative training number (or time ordinal number), is the potential maximum efficiency, is the learning rate; the parameters are fitted using nonlinear least squares method , the steepness factor of the learning curve is defined as the normalized value of , the larger the value, the faster the learning progress;

[0170] S522, Interest migration pattern quantification: Take the sequence of user's historical preference intensity vectors as time series, calculate the cosine similarity between adjacent time period preference vectors:

[0171]

[0172] Stability factor of interest migration defined as the mean of the similarity sequence ; meanwhile, calculate the concentration factor of interest defined as the inverse of the entropy of all user's historical preference vectors in each knowledge domain:

[0173]

[0174] where is a very small constant;

[0175] S523, Retrieval ability change feature extraction: In the sliding time window, calculate the moving average of user's retrieval efficiency score and the standard deviation , the stability factor of retrieval ability defined as , the ability trend factor obtained by calculating the linear regression slope of efficiency score in the window, positive value indicates that the ability is improving;

[0176] S524, Personalized weighting factor synthesis: Combine the above features into a personalized weighting factor vector , and map it to the personalized adjustment weight with the same dimension as the expert weight vector through a trainable fully connected layer:

[0177]

[0178] S53, Use the method of combining weighted fusion and Bayesian inference to fuse multi-dimensional original evaluation indicators, expert rules and personalized weighting factors, and calculate the user portrait score reflecting individual characteristics and the health degree comprehensive score reflecting the overall state of the knowledge base.

[0179] The method of combining weighted fusion and Bayesian inference in S53 includes the following steps:

[0180] S531, Design time decay function and confidence accumulation function, so that recent and high-frequency behavior data have higher influence weight on user portrait, and improve evaluation confidence with data accumulation, including the following steps:

[0181] ​​​​S5311, Time Decay Function Design: Assign a time-based decay weight to each historical behavior data point, assuming the current time is... The timestamp of a certain historical behavior data is The time difference is The time weights are calculated using an exponential decay function. :

[0182] ;

[0183] in, The decay time constant controls the rate at which the weights decay. The smaller the value (the newer the data), the closer the weight is to 1; conversely, the larger the value, the smaller the weight.

[0184] S5312, Frequency Reinforcement Factor: Reinforces high-frequency behavioral patterns. Statistically analyzes user activity within the most recent time window. Number of actions within the same knowledge domain or the same search pattern Frequency enhancement factor Smoothing enhancement using a logarithmic function:

[0185] ;

[0186] Then to Normalize it so that its maximum value is 1;

[0187] S5313, Comprehensive Weight Calculation: Combining the time decay weight and frequency enhancement factor to obtain the comprehensive weight of each historical data point. :

[0188] ;

[0189] When calculating historical statistics (such as the mean of preference and the mean of efficiency), a weighted average is used instead of a simple average.

[0190] S5314, Confidence Accumulation Function: The confidence level increases as data accumulates. Let the total amount of valid historical user behavior data currently used for evaluation be... Confidence level Defined as:

[0191]

[0192] in, For the growth rate parameter, when When smaller, A value close to 0 indicates a low confidence level; as... Increase Asymptotic to 1, indicating high confidence; this confidence will be used to adjust the relative weight of the prior distribution and the likelihood distribution in Bayesian inference;

[0193] S532、Using the rule quantization mapping table to handle the fuzziness and uncertainty in the expert rules, converting qualitative rules into quantifiable impact factors, including the following steps:

[0194] S5321、For the qualitative description in the expert rule base, convert it into a membership function of fuzzy logic;

[0195] S5322、Rule quantization mapping table construction: create a mapping table that maps the combination of fuzzy rules to a specific adjustment amount for the evaluation index;

[0196] S5323、Considering the uncertainty of the fuzzy rule itself, add an uncertainty measure to each adjustment amount , which is based on the clarity of the membership function and the credibility factor of the rule ;

[0197] ;

[0198] This uncertainty will be integrated into the subsequent Bayesian inference process, represented as an increase in the variance of the likelihood distribution;

[0199] S533、Establish a feedback correction loop, use the user's subsequent behavior on the adjusted knowledge base as a feedback signal to dynamically adjust the weight parameter in the weighted fusion and the Bayesian prior distribution, including the following steps:

[0200] S5331、Online adjustment of weight parameters: treat the weight parameter in the weighted fusion formula as adjustable, use an online learning strategy based on gradient descent, and use the overall improvement of the feedback signal (defined as a reward function such as ) as the target to fine-tune the weight parameter and update the formula as:

[0201] ;

[0202] where represents the weight parameter to be adjusted, is the online learning rate, is the gradient of the reward function to the parameter, obtained by policy gradient method or approximate gradient estimation;

[0203] S5332、Bayesian prior distribution update: update the prior distribution parameters and ​Considered as dynamic update, each time after evaluation and receiving feedback, the prior is updated according to the consistency of the posterior distribution and the feedback; if the feedback is good, the current posterior distribution is considered to be accurate, and it is taken as the prior distribution of the next evaluation, that is:

[0204]

[0205] Wherein, is the posterior mean, is the posterior variance, is the artificially added process noise variance, indicating that the user portrait itself may change slowly over time; if the feedback is not good, the prior variance is appropriately increased to reduce the weight of the prior, and make the subsequent evaluation more dependent on new data (likelihood).

[0206] The intelligent decision report generated in S6 includes the following steps:

[0207] S61, based on the user portrait score, using a collaborative filtering and content recommendation hybrid algorithm, generating a personalized knowledge item ranking list and potential interest knowledge push suggestions for the user;

[0208] S62, based on the knowledge base health score, identifying knowledge coverage blind area, hot spot overload area and knowledge aging items, generating optimization suggestions for content supplement, reconstruction or archiving, including the following steps:

[0209] S621, knowledge coverage blind area identification: calculate the demand-supply imbalance degree of each knowledge field kk. Let the time period be , the number of knowledge items in the field , the demand degree , the supply degree , and the coverage blind area index :

[0210]

[0211] If , it is determined that the field is a coverage blind area and needs to supplement the content;

[0212] S622, hot spot overload area identification: calculate the access heat of each knowledge item, and normalize it. For each field , calculate the heat concentration , using the HHI formula:

[0213] ​​​​ ;

[0214] Simultaneously calculate the total heat of the field , if , it is determined that it is a hotspot overload area, and reconstruction is suggested;

[0215] S623, knowledge aging item identification: for knowledge items , let the last update time be , the current time be , the content half-life be , and the timeliness score be :

[0216] ;

[0217] Simultaneously calculate its recent access decay rate , if , it is determined that it is a knowledge aging item, and archiving or updating is suggested;

[0218] S63, using an automated report generation technique, the ranking strategy, push suggestions, optimization suggestions, and key indicator visual charts are integrated into a structured report, and are pushed to the knowledge base management platform through an interface.

[0219] The closed-loop iterative system formed in S7 includes the following steps:

[0220] S71, the adaptive update engine takes the knowledge base health score and the user search efficiency improvement rate as optimization objectives, and dynamically adjusts the weight distribution strategy and interface display rules of the knowledge base using heuristic rules and gradient descent optimization algorithms, including the following steps:

[0221] S711, initial exploration guided by heuristic rules: based on the optimization suggestions in the intelligent decision report generated in S6, heuristic rules are formulated to initialize or adjust parameters:

[0222] If a coverage blind area is identified, the initial weight of the related field item is increased: , where is the adjustment coefficient;

[0223] If a hotspot overload is identified, the ranking weight decay factor of the popular item is reduced: ;

[0224] These rules provide an initial search direction and reduce the cost of random exploration;

[0225] S712, gradient descent optimization execution: a lightweight optimization algorithm based on policy gradient is used, and in each iteration, a small random disturbance is added to the parameters to obtain new parameters Run new configuration for a short time in online A / B testing environment, observe the change of reward function Parameter update adopts approximate gradient ascent:

[0226] ;

[0227] wherein, is the learning rate;

[0228] S72, set up model and threshold joint debugging mechanism, when user group behavior distribution drifts or knowledge base content is updated in large scale, automatically trigger the retraining and calibration process of the multi-task evaluation model and related evaluation threshold, including the following steps:

[0229] S721, distribution drift detection: calculate the current user behavior feature distribution periodically and the difference between the historical baseline distribution Adopt drift detection based on Wasserstein distance:

[0230] ;

[0231] wherein is the set of all joint distributions, if , it is determined that significant distribution drift occurs;

[0232] S722, content update monitoring: calculate the knowledge base content update rate , let the total number of knowledge entries be , the number of newly added, deleted or substantially modified entries in the current period be , then , if , it is determined that there is large-scale content update;

[0233] S723, automatic triggering condition judgment: when any of the following conditions is met , the system automatically generates a retraining task and adds the task to the priority queue;

[0234] S724, retraining and calibration:

[0235] Data preparation: collect the behavior data of all users in the recent time window as the new training set ;

[0236] Model retraining: use to retrain the multi-task evaluation model, and the training process is as described in S4, but incremental learning or fine-tuning from the latest checkpoint can be used to speed up;

[0237] Threshold recalibration: Recompute the statistical distribution of the evaluation metric based on the new validation set, and update the decision threshold (e.g. the threshold for efficient retrieval ) accordingly, e.g. set the threshold to the value at a certain percentile on the validation set:

[0238] ;

[0239] Performance verification: Verify the performance of the updated model on a new test set, ensuring that key metrics are not worse than the original model , otherwise roll back to the previous version;

[0240] S73, synchronize and deploy the optimized knowledge base configuration and the updated model parameters, continuously collect a new round of behavior data.

[0241] It should be noted that the relationship terms such as first and second, and the like, are used only to distinguish one entity or action from another, and do not necessarily require or imply that these entities or actions occur in any particular temporal or spatial sequence. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0242] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the following claims and their equivalents.

Claims

1. A method for statistical analysis of knowledge base retrieval data, characterized in that, The method includes the following steps: S1. Collect multi-dimensional behavioral data of users in the knowledge base system in real time and generate raw behavioral data; S2. The original behavioral data stream is cleaned, fused, and structured to generate a standardized behavioral data sequence; S3. Based on the knowledge base topology and semantic association network, feature enhancement and context association analysis are performed on the standardized behavioral data sequence to construct a user-knowledge interaction feature graph. S4. Input the user-knowledge interaction feature graph into the multi-task evaluation model of fusion graph embedding and lightweight time series model for analysis, and output multi-dimensional original evaluation indicators including demand preference intensity, retrieval path efficiency and content relevance. S5. Integrate the historical behavior pattern library and expert rule library, dynamically weight and adaptively calibrate the multi-dimensional original evaluation indicators, and generate a user profile and a comprehensive health score of the knowledge base. S6. Based on the user profile and the comprehensive score of the knowledge base health, generate an intelligent decision report that includes personalized ranking strategies, knowledge gap warnings, and content optimization suggestions; S7. Based on the intelligent decision report, the content weight, sorting algorithm and display logic of the knowledge base are dynamically adjusted through an adaptive update engine driven by heuristic rules and lightweight optimization algorithms, and the evaluation model parameters are updated synchronously.

2. The method for statistical analysis of knowledge base retrieval data according to claim 1, characterized in that: The real-time collection of multi-dimensional behavioral data in S1 includes the following steps: S11. Capture explicit behavioral data of users on the knowledge base interface through tracking technology, including search keyword sequences, item click flow, page dwell time, page turning frequency, and content download or collection records. S12. Collect implicit user behavior data through the log analysis system, including the modification history of query statements, mouse hover trajectory, scroll speed changes, and operation interruption information; S13. Associate user identity attributes with contextual environment data, including user role, department, access time period, device type and task context label, and synchronize and align all data according to global timestamp to generate raw behavioral data stream with multi-dimensional labels.

3. The method for statistical analysis of knowledge base retrieval data according to claim 1, characterized in that: The processing of the raw behavioral data stream in S2 includes the following steps: S21. A combined method based on rules and statistical outlier detection is used to detect and clean the original behavioral data stream, removing noise data generated by invalid clicks, machine crawlers, and misoperations. S22. Lightweight natural language processing techniques are used to segment unstructured search keywords and query statements, remove stop words, extract stems, and perform word frequency-inverse document frequency vectorization representation. S23. By using time window sliding and session segmentation techniques, the continuous behavioral data stream is divided into session sequences based on independent retrieval tasks, and the behavioral data within each session is serialized, encoded, and normalized to generate standardized behavioral data sequences.

4. The method for statistical analysis of knowledge base retrieval data according to claim 1, characterized in that: The construction of the user-knowledge interaction feature map in S3 includes the following steps: S31. Construct a knowledge base association graph, where nodes represent knowledge items, edges represent co-occurrence relationships, tag similarity, and jump frequency between items, and assign basic popularity weights to nodes. S32. Using a graph embedding method based on random walks, the standardized behavioral data sequence of users is mapped onto the knowledge base association graph, the access popularity of nodes and the relationship strength of edges in the graph are dynamically updated, and virtual nodes representing implicit user needs are added. S33. Extract high-order features from the generated dynamic interaction graph, including the subgraph diffusion range triggered by user behavior, the change in the clustering coefficient of core knowledge nodes, and the deviation between the retrieval path and the optimal path, and encode them into graph structure feature vectors.

5. The method for statistical analysis of knowledge base retrieval data according to claim 1, characterized in that: The analysis using the multi-task evaluation model in S4 includes the following steps: S41. Construct a multi-task evaluation model architecture, in which the graph embedding branch processes the user-knowledge interaction feature graph, the bidirectional LSTM branch based on the attention mechanism processes the temporal dependency of user behavior sequences, and the lightweight semantic branch based on BERT fine-tuning processes the deep semantics of query intent. S42. Input the graph structure feature vector and behavior sequence vector into the multi-task evaluation model to calculate the user's preference intensity score for a specific knowledge domain, the efficiency score of its retrieval path relative to the group average efficiency, and the correlation score between its query content and the knowledge base coverage. S43. The outputs of the three branches are dynamically integrated through a weighted fusion strategy, and a normalized multi-dimensional original evaluation index is generated using the Softmax function with a smoothing factor.

6. The method for statistical analysis of knowledge base retrieval data according to claim 5, characterized in that: The construction of the multi-task evaluation model architecture in S41 also includes the following steps: S411. The model is initialized using a pre-training strategy based on transfer learning, enabling it to quickly adapt to new users or new knowledge domains, and to achieve small-sample adaptation by fine-tuning a small number of parameters. S412. Introduce a cluster-based consistency training mechanism to construct positive and negative sample pairs and train the model to distinguish between efficient retrieval behavior and inefficient / ineffective behavior, so as to improve the model's ability to identify differences in behavior patterns. S413. Deploy the sliding window incremental update module to incrementally update the model using recently generated behavioral data streams, and adopt an elastic weight consolidation strategy to prevent catastrophic forgetting of the model during incremental learning.

7. The method for statistical analysis of knowledge base retrieval data according to claim 1, characterized in that: The dynamic weighting and adaptive calibration in S5 includes the following steps: S51. Establish an expert rule base to store the knowledge base optimization goals and priority weights set for different business scenarios and user roles; S52. Analyze users' historical behavior sequences, extract their learning curve trends, interest migration patterns, and search ability change characteristics, and use them as personalized weighting factors. S53. By combining weighted fusion and Bayesian inference, multi-dimensional original evaluation indicators, expert rules, and personalized weighted factors are integrated to calculate the user profile score reflecting individual characteristics and the comprehensive health score reflecting the overall state of the knowledge base.

8. A method for statistical analysis of knowledge base retrieval data according to claim 7, characterized in that: The method combining weighted fusion and Bayesian inference in S53 includes the following steps: S531. Design a time decay function and a confidence accumulation function to give higher weight to recent and high-frequency behavioral data in user profiles, and increase the evaluation confidence as data accumulates. S532. Use rule quantification mapping tables to handle the ambiguity and uncertainty in expert rules and transform qualitative rules into quantifiable influence factors. S533. Establish a feedback correction loop, using the user's subsequent behavior on the adjusted knowledge base as a feedback signal to dynamically adjust the weight parameters and Bayesian prior distribution in the weighted fusion.

9. A method for statistical analysis of knowledge base retrieval data according to claim 1, characterized in that: The process of generating an intelligent decision-making report in S6 includes the following steps: S61. Based on the user profile score, use a hybrid algorithm of collaborative filtering and content recommendation to generate a personalized knowledge item ranking list and potential interest knowledge push suggestions for the user. S62. Based on the comprehensive health score of the knowledge base, identify knowledge coverage blind spots, hotspot overload areas and knowledge aging items, and generate optimization suggestions for content supplementation, reconstruction or archiving; S63. Utilize automated report generation technology to integrate sorting strategies, push suggestions, optimization suggestions, and key indicator visualization charts into a structured report, and push it to the knowledge base management platform via an interface.

10. A method for statistical analysis of knowledge base retrieval data according to claim 1, characterized in that: The formation of the closed-loop iterative system in S7 includes the following steps: S71. The adaptive update engine aims to optimize the comprehensive health score of the knowledge base and the improvement rate of user retrieval efficiency. It uses heuristic rules and gradient descent optimization algorithms to dynamically adjust the weight allocation strategy and interface display rules of the knowledge base. S72. Set up a model and threshold joint debugging mechanism. When the distribution of user group behavior changes or the knowledge base content is updated on a large scale, the retraining and calibration process of the multi-task evaluation model and related evaluation thresholds will be automatically triggered. S73. Simultaneously deploy the optimized knowledge base configuration and updated model parameters to continuously collect a new round of behavioral data.

Images