Big data mining method and system based on digital enterprise management

Abstract

The application relates to the field of data processing, and specifically provides a big data mining method and system based on digital enterprise management, which comprises the following steps: obtaining multi-modal conversation data generated by a target enterprise in a digital service interaction process with a user, extracting a user demand feature set from the multi-modal conversation data, jointly training the user demand feature set based on a preset multi-task learning framework, generating a demand analysis model, calling the demand analysis model to analyze the user demand feature set, generating an optimization strategy set associated with a service process of the target enterprise, updating the digital service execution logic of the target enterprise according to the optimization strategy set, and incrementally training the demand analysis model based on user feedback data. The application can improve the response efficiency and functional adaptability of the digital service process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a big data mining method and system based on digital enterprise management. BACKGROUND

[0002] Currently, when a digital enterprise interacts with a user, it usually relies on conversational data (such as text tickets or voice recordings) to mine user needs and optimize service processes based on this. For example, some enterprises analyze the text feedback submitted by users to identify high-frequency keywords to infer service pain points; other enterprises use voice sentiment analysis technology to determine the level of user satisfaction during a call. In addition, there are also existing technologies that infer user preferences for specific functions by analyzing user click logs or page dwell time.

[0003] However, the above methods often process text, voice or behavior data in isolation, leading to one-sidedness in user demand analysis. In addition, existing service optimization strategy generation mechanisms are often limited to single-point adjustments (such as only modifying response scripts or local interface layouts), lacking the ability to coordinate the optimization of service response rules, function module architecture and interaction logic. This results in a "fragmented improvement" of the enterprise digital service system, making it difficult to form a global experience improvement. More importantly, once the traditional demand analysis model is deployed, it is usually not updated dynamically based on user feedback data, causing the model's decision-making accuracy to decline rapidly when business scenarios change or user behavior patterns migrate, ultimately affecting the sustained effectiveness of service optimization. SUMMARY

[0004] The purpose of the present application is to provide a big data mining method and system based on digital enterprise management. The embodiments of the present application are implemented as follows:

[0005] In a first aspect, an embodiment of the present application provides a big data mining method based on digital enterprise management, the method comprising: obtaining multi-modal conversation data generated by a target enterprise in a digital service interaction process with a user, the multi-modal conversation data comprising a text conversation segment, a voice conversation segment, and a user behavior sequence, wherein the user behavior sequence is used to record a time sequence and event attributes of operation events triggered by the user in an interaction interface; extracting a user demand feature set from the multi-modal conversation data, the user demand feature set comprising a text demand feature, a voice demand feature, and a behavior demand feature, wherein the behavior demand feature is used to describe operation preferences and functional demands of the user for the digital service; based on a preset multi-task learning framework, jointly training the user demand feature set to generate a demand analysis model, the demand analysis model being used to identify a user demand priority and predict a service optimization direction; calling the demand analysis model to analyze the user demand feature set, and generating an optimization strategy set associated with a service process of the target enterprise, the optimization strategy set comprising a service response rule adjustment strategy, a functional module iteration strategy, and an interaction interface dynamic configuration strategy; updating a digital service execution logic of the target enterprise according to the optimization strategy set, and incrementally training the demand analysis model based on user feedback data.

[0006] In a second aspect, an embodiment of the present application provides a computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.

[0007] The application can comprehensively capture the explicit demands and implicit behavior patterns of users in digital service interaction by integrating multi-modal data composed of text conversation segments, voice conversation segments and user behavior sequences. Among them, the text demand features can analyze the direct description of users on service functions, the voice demand features can mine the emotional tendency and urgency of users in voice interaction, and the behavior demand features can reveal the actual use preference of users on service functions through operation event sequences, thereby constructing a multi-dimensional user demand portrait. Moreover, the multi-task learning framework is adopted to jointly train the heterogeneous features, which can strengthen the accuracy of user demand priority recognition through feature complementation, and improve the robustness of service optimization direction prediction. Through generating an optimization strategy set covering service response rules, function modules and interaction interfaces, the response efficiency and function adaptability of the digital service process can be targetedly improved. In addition, the model is dynamically incrementally trained combined with user feedback data, which can continuously optimize the real-time decision-making ability of the demand analysis model, so that the enterprise digital service system forms a closed-loop optimization mechanism of "data collection-strategy generation-execution verification", and finally realizes the iterative upgrade based on user behavior feedback. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a flowchart of a big data mining method based on digital enterprise management provided by an embodiment of the application.

[0009] Figure 2 is a functional module architecture schematic diagram of a big data mining device provided by an embodiment of the application.

[0010] Figure 3 is a composition schematic diagram of a computer system provided by an embodiment of the application. DETAILED DESCRIPTION

[0011] The execution subject of the big data mining method based on digital enterprise management in the embodiment of the application is a computer system, including but not limited to a server, a personal computer, a notebook computer, a tablet computer, a smart phone and the like.

[0012] The embodiment of the application provides a big data mining method based on digital enterprise management as shown in Figure 1 , including:

[0013] Step 100: acquiring multi-modal conversation data generated by a target enterprise in a digital service interaction process with a user, the multi-modal conversation data including text conversation segments, voice conversation segments and user behavior sequences, wherein the user behavior sequences are used to record the time sequence and event attributes of operation events triggered by the user in an interaction interface.

[0014] Specifically, the multi-modal conversation data refers to a collection of various forms of data generated when an enterprise interacts with a user through a digital service platform. The text conversation segment is the dialogue content input and output by the user and the customer service system or intelligent assistant through text, such as the user sending “how to query the account balance” in the online consultation window and the corresponding system reply text. The voice conversation segment is the interaction content recorded in the form of voice call or voice message, such as the voice record of “I need to modify the bound mobile phone number” when the user calls the service hotline and the corresponding voice response. The user behavior sequence is the event stream of the user's operation in the digital service interface collected through the burying point technology, such as the user triggering the operation chain of “home page → transfer → input amount → confirm the account of the recipient → submit” in the mobile banking APP in turn, wherein each operation event contains a trigger timestamp, an interface element identifier and an operation type attribute. It should be noted that the time sequence of the user behavior sequence reflects the logical coherence of the user operation path, and the event attribute is used to distinguish the operation type and the associated function module, for example, the “transfer” operation event attribute contains the function code F103 and the interface area identifier Z05. Exemplarily, when the target enterprise is a commercial bank, during the interaction of a user with the mobile banking APP, the text conversation segment may include the intelligent customer service dialogue record “user: how to apply for a loan? System: please enter the ‘credit service’ module to submit an application”, the voice conversation segment may include the telephone banking recording “user: my credit card account billing failed, please help me process”, and the user behavior sequence may be recorded as the time sequence [2023-10-0109:00:00 click ‘account overview’ button; 2023-10-0109:00:05 slide to ‘transaction details’ page; 2023-10-0109:00:10 filter transaction type as ‘expenditure’]. In this scenario, the integrity of the multi-modal conversation data guarantees the comprehensiveness of the subsequent feature extraction, and the time sequence attribute of the user behavior sequence provides a basic data structure for analyzing the operation preference.

[0015] Step 200: extracting a user demand feature set from the multi-modal conversation data, the user demand feature set including text demand features, voice demand features and behavior demand features, wherein the behavior demand features are used to describe the operation preference and function demand of the user for the digital service.

[0016] Specifically, the user demand feature set is a structured feature set extracted from multi-modal data through natural language processing, speech recognition, and behavior pattern mining technology. The text demand feature is an intent classification and keyword vector obtained after semantic analysis of a text session segment. For example, the BERT model is used to identify the intent of the user query text "How to open the mobile payment function", and the feature vector [intent category: function opening; keyword weight: opening 0.92, mobile payment 0.87, function 0.79] is output. The speech demand feature is a demand urgency index and sentiment polarity score generated after speech-to-text conversion and sentiment analysis of a speech session segment. For example, after converting the user's voice "My transfer has not arrived, please handle it immediately!" into text, the sentiment analysis model outputs the feature [demand type: abnormal handling; urgency: 0.95; sentiment polarity: negative 0.88]. The behavior demand feature is an operation frequency statistic, path clustering label, and function usage depth index generated after pattern mining of user behavior sequences. For example, through sequence pattern analysis, it is found that a certain user group in the "financial supermarket" interface frequently performs the operation chain "sort by yield → filter low-risk products → view product details", thereby generating the feature [operation path pattern: P023; function usage depth: product comparison function call rate 82%]. It should be noted that the operation preference in the behavior demand feature reflects the repeated access tendency to a specific function module, for example, a certain user logs in three times in a row and directly accesses the "fund investment" page each time, and the function demand is mapped through the association between operation events and function codes, for example, the event cluster of frequently triggering "F205" code corresponds to the "cross-border remittance" function. Illustratively, in the securities trading APP scenario, the text demand feature may reflect the user's guidance demand for "how to set up a conditional order" through online consultation, the speech demand feature may capture the user's "I want to complain about order execution delay" in the phone, and the behavior demand feature may reveal the user's operation habit of repeatedly viewing the details page of a certain technology stock in the self-selected stock list. The three together constitute a complete user demand portrait.

[0017] Step 300: Based on the preset multi-task learning framework, the user demand feature set is jointly trained to generate a demand analysis model, which is used to identify user demand priority and predict service optimization direction.

[0018] Specifically, the preset multi-task learning framework can refer to a neural network architecture adopting a shared bottom network structure and independent task branches to jointly learn the representation of text, voice and behavior demand features. In this framework, the text demand features are converted into high-dimensional semantic vectors through a word embedding layer, the voice demand features are aligned with the text translation results after acoustic feature extraction, and the behavior demand features are extracted through a time convolution network to extract the periodicity and dependency patterns of the operation sequence. During the joint training process, the model simultaneously optimizes three task objectives: the first task is the demand priority classification based on the attention mechanism, such as classifying user demand into three priority levels of “emergency fault handling”, “routine function consultation” and “long-term experience optimization”; the second task is the service optimization direction prediction, i.e., outputting the probability distribution of optimization categories such as function module iteration, interface layout adjustment and response time improvement; and the third task is the feature reconstruction loss, which ensures that the shared layer can effectively fuse the complementary information of multi-modal features. It should be noted that the basis for demand priority recognition includes voice sentiment polarity, text intent urgency and behavior path abnormality, for example, when the user's voice sentiment score is negative 0.9 and there are three consecutive transaction failure events in the behavior sequence, the model will determine its priority as the highest level. The service optimization direction prediction is achieved by analyzing the association rules between features, for example, when the high-frequency access behavior of a certain function module and the complaint of “can't find the entry” in the text consultation appear at the same time, the model predicts that the entry visibility of the function needs to be optimized. Illustratively, in the e-commerce customer service scenario, the model may identify that the user's complaint about “order logistics stagnation” has the highest priority, and suggest optimizing the real-time push function of logistics status; at the same time, aiming at the user's behavior pattern of frequently using the collection but less triggering the purchase, it is predicted that the price reduction reminder function needs to be added to improve the conversion rate. Through multi-task joint training, the demand analysis model can establish the association mapping of cross-modal features, thereby improving the synergy of demand recognition and optimization suggestions.

[0019] Step 400: calling the demand analysis model to analyze the user demand feature set and generate an optimization strategy set associated with the target enterprise service process, the optimization strategy set including service response rule adjustment strategies, function module iteration strategies and interactive interface dynamic configuration strategies.

[0020] Specifically, the optimization strategy set is a set of operation instructions that convert abstract recommendations output by the model into executable service process improvement solutions. The service response rule adjustment strategy involves modifications to customer service ticket assignment logic, automatic response trigger conditions, and timeliness standards. For example, when the model identifies that the average response time for a certain type of consultation request exceeds a threshold, the strategy suggests automatically upgrading such requests to senior customer service handling and setting a response rule that must be resolved within two hours. The function module iteration strategy includes specific implementation schemes for function addition, deletion, performance optimization, and interface expansion. For example, to address the frequent demand of users to consult "how to export transaction records" through voice, the strategy suggests adding a "data export" sub-function to the "account management" module and providing CSV and PDF format options. The interactive interface dynamic configuration strategy refers to a rule library that adjusts interface element layout, visual style, and navigation path in real time based on user group behavior characteristics. For example, to address the operation preferences of the elderly user group, the strategy generates a heat map-driven layout scheme that enlarges the icons of frequently used functions by 20% and displays them at the top. It should be noted that the relevance of optimization strategies to the target enterprise service process is achieved through business process modeling. For example, in the bank account opening process, if the model detects that users generally abandon operations in the document upload link, the generated strategy will suggest integrating the identity card OCR recognition function into the mobile terminal to reduce manual input steps. Illustratively, in the online education platform scenario, the service response rule adjustment strategy may include upgrading the response priority of "course playback lag" complaints to P1 level and triggering immediate intervention by the technical team. The function module iteration strategy may propose adding a speed playback memory function to the video player. The interactive interface dynamic configuration strategy may dynamically arrange the difficulty order of exercises based on students' exercise habits. These strategies are connected to the API interface of the existing service process of the enterprise to ensure that optimization measures can be directly embedded into the existing digital service system.

[0021] Step 500: Update the digital service execution logic of the target enterprise according to the optimization strategy set, and incrementally train the demand analysis model based on user feedback data.

[0022] Specifically, the update of the digital service execution logic involves version iteration of the business process engine rule base, gray release of the function module, and parameter adjustment of the interface configuration management system. For example, for service response rule adjustment strategy, new ticket routing rules are deployed in the customer service system, so that requests that meet the conditions of "voice sentiment negative score > 0.8 and contain the keyword 'fault'" are automatically assigned to technical expert agents; for function module iteration strategy, AB testing method is used to batch online new data export function, and its usage rate and user satisfaction indicators are monitored; for interactive interface dynamic configuration strategy, condition display rules are configured in the content management system, and when the user device is identified as a tablet computer, the layout mode is automatically switched to the column layout mode. User feedback data includes explicit feedback (such as satisfaction score, complaint suggestion text) and implicit feedback (such as optimized function usage rate, operation path simplification degree), for example, after deploying a new interface layout, collect user click heat map data and positive evaluations such as "finding functions is more convenient" in conversation records. The incremental training process adopts an online learning mechanism, reweights the positive samples (such as 20% increase in optimized function usage rate) and negative samples (such as a button click rate decrease) in the feedback data and inputs them into the demand analysis model, and adjusts the network parameters to make the model continuously adapt to the user behavior patterns after service optimization. It should be noted that the execution logic update needs to follow the version rollback mechanism, and when the key indicators (such as task completion rate) decrease by more than the threshold value, it is automatically restored to the previous stable version. Illustratively, in the retail e-commerce scenario, the updated execution logic may include adding a virtual credit card payment option in the checkout process, and adjusting the recommendation weight of this function according to the proportion of complaints in user feedback "few payment methods"; at the same time, based on the user's usage duration data of the new version of the product filter, the model is incrementally trained to more accurately predict future interface optimization directions. Through the closed-loop optimization mechanism, the digital service system forms a virtuous cycle of continuous improvement.

[0023] In an implementation scheme, the step 200 of extracting a set of user demand features from the multi-modal conversation data comprises the following steps:

[0024] Step 210: performing semantic segmentation processing on the text conversation segment to generate a plurality of text sub-segments, wherein each text sub-segment contains at least one complete demand description sentence, and adjacent text sub-segments are connected by context association tags.

[0025] Specifically, the semantic segmentation processing refers to segmenting a continuous text conversation fragment according to semantic integrity by using natural language processing technology, to ensure that each text sub-fragment carries an independent demand semantic unit. The text conversation fragment is a multi-round dialogue text generated in the interaction process of a user and a digital service platform, for example, three messages continuously input by a user in an online customer service system: “How to reset the password?”→“I clicked the email link but the page reported an error”→“Please transfer to the human customer service”. By using a semantic segmentation model (such as a BERT-based sequence labeling model) to identify the boundaries of sentences and semantic independence, the above conversation is segmented into three text sub-fragments: [Sub-fragment 1: How to reset the password?], [Sub-fragment 2: I clicked the email link but the page reported an error], [Sub-fragment 3: Please transfer to the human customer service], and context association labels (such as “causal association: sub-fragment 2 is the operation result of sub-fragment 1”) are added to adjacent sub-fragments. It should be noted that the context association label is generated by analyzing the logical relationship between the sub-fragments, for example, when sub-fragment 1 is a function request and sub-fragment 2 is an operation exception feedback, the label is marked as a “problem-result” relationship. Exemplarily, in a bank APP consultation scenario, a user may continuously send “I want to apply for a credit card”→“What is the difference between gold card and platinum card”→“What materials do I need to prepare”, at this time the semantic segmentation processing will generate three independent text sub-fragments and add “process progression” association labels, to ensure that the context dependency relationship of the multi-round dialogue can be preserved during subsequent feature extraction.

[0026] Step 220: performing voiceprint feature separation and speech-to-text processing on the voice conversation fragment to generate a voice text set, and performing intent keyword extraction on the voice text set to generate a voice keyword sequence, wherein the intent keyword is associated with a function label in a preset enterprise service function library.

[0027] Specifically, the voiceprint feature separation refers to separating the voiceprint features of different speakers from a mixed voice stream, for example, in a call recording between a customer and a customer service personnel, distinguishing the customer voice segment from the customer service reply segment through a voiceprint clustering algorithm. The voice-to-text processing converts the separated customer voice segment into text data using automatic speech recognition (ASR) technology, forming a voice text set. For example, the customer voice content "I want to check the loan approval progress" is converted to generate the text "I want to check the loan approval progress". The intent keyword extraction extracts core keywords from the voice text by combining rule matching and deep learning models (such as the TextRank algorithm), for example, extracting the keywords "loan approval" and "progress" from the above text. The pre-set enterprise service function library is a pre-defined business function label system, for example, "loan approval" is mapped to function label F201, and "progress query" is mapped to operation label O305. The voice keyword sequence is generated by arranging the extracted keywords in the order of appearance and associating the function labels, for example, the sequence [loan approval-F201, progress query-O305]. Exemplarily, in an insurance claim scenario, the customer voice "My car insurance case number is XY123, what step is it handled now?" retains the customer voice segment after voiceprint separation, and the text obtained is "My car insurance case number is XY123, what step is it handled now?", the intent keyword extraction generates [car insurance-C101, case number-K205, progress-O401], where C101 corresponds to the car insurance business module, K205 is the case identification key field, and O401 is the progress query operation label. The sequence realizes the structured expression of the requirement semantics through the association of function labels.

[0028] Step 230: Analyzing the operation event attributes in the user behavior sequence, extracting core operation events with a trigger frequency higher than a pre-set threshold, and constructing a behavior pattern graph based on the time sequence of the core operation events, where the nodes in the behavior pattern graph represent operation event types, and the edges represent the time interval and the preceding and following dependency relationship triggered by the core operation events.

[0029] Specifically, the operation event attributes include event type code, trigger timestamp and associated interface element identification, for example, in an e-commerce APP, the user behavior sequence may include the event sequence [search "smartphone" -E101, filter "price <5000 yuan" -E205, view product detail page -E308]. The preset threshold is a statistical lower limit of frequency (such as the same event triggers in historical sessions ≥5 times), and the event trigger frequency is calculated by a sliding window statistics, for example, when the "filter price range" event occurs 60 times in 100 sessions, its frequency is 60%, and it is marked as a core operation event when it exceeds the threshold 30%. The behavior pattern graph models the time sequence relationship between the core operation events in a graph structure, the node is the event type code (such as E101, E205, E308), and the edge weight is determined by the average trigger time interval of adjacent events (such as E101→E205 average interval 3 seconds) and the conditional probability (such as the probability of triggering E101 after E205 is 85%). Exemplarily, in a securities trading APP, the user high-frequency trigger event [self-selected stock list -E501→individual stock time chart -E503→add warning condition -E507] forms a behavior pattern graph, and the edge weight between nodes is marked as "E501 to E503 average interval 2 seconds, trigger probability 92%". The graph reveals the user's operation inertia from viewing the list to in-depth analysis, providing a basis for identifying functional preferences.

[0030] Step 240: input the text sub-segment, voice keyword sequence and behavior pattern graph into a feature fusion network, calculate the correlation weight between text, voice and behavior features through the cross-modal attention layer in the feature fusion network, and according to the correlation weight, the heterogeneous features are weighted and spliced to generate the user demand feature set.

[0031] Specifically, the feature fusion network is a multi-modal neural network architecture, which determines the contribution weights of different modal features through calculating the similarity matrix between the semantic vectors of the text sub-fragments, the embedding vectors of the speech keywords sequence and the graph embedding vectors of the behavior pattern graph across the modal attention layer. For example, when the user text sub-fragment mentions "loan approval" and the speech keywords contain "urgent processing" and the behavior graph shows a high-frequency trigger "application status refresh" event, the attention layer may assign weights [text 0.6, speech 0.3, behavior 0.1], reflecting the dominance of text demand. Weighted concatenation linearly combines each modal feature vector according to the weight to generate a unified user demand feature vector, for example, the 768-dimensional BERT vector of the text sub-fragment, the 256-dimensional word embedding vector of the speech keywords and the 128-dimensional graph network vector of the behavior graph are fused into a 1024-dimensional feature vector after weighting. Exemplarily, in the medical appointment scenario, the user text sub-fragment "I want to make an appointment with a specialist in the digestive department" and the speech keywords [make an appointment-O01, digestive department-D12] and the high-frequency path [home page → department selection → doctor list → appointment period] in the behavior graph are calculated by the cross-modal attention layer, the text feature weight is 0.7 (strong semantic explicitness), the speech feature weight is 0.2 (supplement department information), and the behavior feature weight is 0.1 (verify operation habit), and the final fusion feature accurately represents the core demand of "quickly complete the appointment of the specified department specialist number". Through the complementary advantages of heterogeneous features, the completeness and explainability of the user demand feature set are significantly improved.

[0032] In an implementation scheme, the step 300 of jointly training the user demand feature set based on a preset multi-task learning framework to generate a demand analysis model includes the following steps:

[0033] Step 310: input the user demand feature set into a feature encoder, extract cross-modal time sequence features through a bidirectional long short-term memory network in the feature encoder, and generate a demand time sequence encoding vector.

[0034] Specifically, the feature encoder employs a bidirectional long short-term memory network (BiLSTM) to model the temporal dependency in the user demand feature set. The user demand feature set contains the fusion feature vectors of text sub-segments, speech keyword sequences, and behavior pattern graphs, such as the feature vector sequence [feature vector 1: password reset demand, feature vector 2: email link error feedback, feature vector 3: artificial customer service transfer request] generated in the user's continuous three conversations in the bank APP scenario. The bidirectional long short-term memory network captures the forward and reverse context associations through two layers of LSTM units in the forward and backward directions, respectively. For example, when analyzing feature vector 2, the forward unit remembers the "password reset" operation background of feature vector 1, and the backward unit predicts the "human intervention" demand tendency of feature vector 3. Cross-modal temporal features refer to the collaborative change rules of text, speech, and behavior features in the time dimension, such as when the user behavior pattern graph shows a high-frequency trigger "password modification" event and the speech keyword sequence contains an "urgent" intent, the temporal feature will strengthen the expression of the urgency of such demand. The demand temporal encoding vector is a high-dimensional dense vector output by the BiLSTM, such as encoding the three feature vector sequences into a 768-dimensional vector [0.23, -0.56,..., 1.02], which reflects the change of demand intensity at different time points. For example, in the e-commerce after-sales scenario, the user demand feature set may contain the feature sequence "return application → logistics stagnation complaint → prompt processing", and the bidirectional long short-term memory network generates an encoding vector to represent the evolution process of the user from functional request to emotional appeal by analyzing the causal relationship between events in the sequence.

[0035] Step 320: input the demand temporal encoding vector into the shared feature layer of the multi-task learning framework, and filter the key feature nodes across sessions through the sparse attention mechanism in the shared feature layer to generate a shared feature tensor.

[0036] Specifically, the shared feature layer is a feature abstraction module shared by the main task branch and the auxiliary task branch in the multi-task learning framework. The sparse attention mechanism calculates the importance scores of different time step feature nodes in the demand time sequence encoding vector, and only keeps the key nodes with scores higher than the threshold. For example, in the insurance claim settlement scenario, the user demand time sequence encoding vector contains five time step feature nodes, corresponding to "report submission → material supplement → progress query → request for follow-up → satisfaction feedback", respectively. The sparse attention mechanism identifies that the attention scores of "progress query" and "request for follow-up" nodes are 0.91 and 0.88 (threshold 0.7), respectively, while the scores of other nodes are lower than the threshold, so only these two nodes are kept for subsequent calculation. Cross-session key feature nodes refer to features that appear repeatedly or have a decisive influence on demand evolution in multiple rounds of interaction. For example, the "request for follow-up" node appears in three consecutive user sessions, indicating a high sensitivity to processing timeliness. The shared feature tensor is a three-dimensional matrix formed by arranging the screened key nodes in chronological order. For example, the above two nodes are encoded as the matrix [node 3: progress query, node 4: request for follow-up], which adapts to the input requirements of the main task and auxiliary task branches. For example, in the medical appointment scenario, the shared feature layer may filter out the "department selection → number source query → appointment failure feedback" key nodes, accurately capturing the user experience pain points caused by system failures.

[0037] Step 330: input the shared feature tensor into the priority prediction module of the main task branch, capture the demand intensity fluctuation pattern through the time convolution network in the priority prediction module, generate a demand intensity waveform diagram, and input the demand intensity waveform diagram into a sequence of gated recurrent units to output a user demand priority score.

[0038] Specifically, the priority prediction module employs a temporal convolution network (TCN) to extract multi-scale features from the temporal patterns in the shared feature tensor. The temporal convolution network captures dependencies at different time spans through dilated convolution kernels, e.g., using three layers of convolution kernels with dilation coefficients of 1, 2, and 4, respectively, to identify short-term fluctuations (e.g., demand changes within a single session), medium-term trends (e.g., demand evolution over three consecutive sessions), and long-term patterns (e.g., monthly periodic high-frequency demand types). The demand intensity fluctuation pattern refers to the periodic, sudden, or gradual variation of user demand in the time dimension, e.g., in the securities trading scenario, user demand intensity is significantly higher during the opening period than after the closing period. The demand intensity waveform is a two-dimensional feature map output by the temporal convolution network, with the horizontal axis representing time steps and the vertical axis representing intensity values. The peak of the waveform corresponds to the time of occurrence of the key demand node. The gated recurrent unit (GRU) sequence further models the waveform graph as a sequence, controlling information flow through update and reset gates, e.g., when analyzing sudden peaks in the waveform graph, the reset gate suppresses historical irrelevant information, and the update gate enhances the influence weight of the current peak. The user demand priority score is a scalar value output by the GRU sequence at the final time step, e.g., the score interval is 0-1, and a score of 0.8 indicates that it needs to be handled with priority. Illustratively, in the government service platform scenario, the temporal convolution network may identify that after the user submits a "residence certificate handling" request, "material audit status query" is initiated for three consecutive days, and the waveform graph shows that the intensity value reaches a peak on the third day. The GRU sequence outputs a priority score of 0.92 accordingly, triggering the urgent handling process.

[0039] Step 340: input the shared feature tensor into the feature decomposer of the auxiliary task branch, separate the functional demand feature vector and the service experience feature vector through the orthogonal projection matrix in the feature decomposer, wherein the functional demand feature vector is aligned with the function code dimension in the preset enterprise service function library.

[0040] Specifically, the feature decomposer adopts an orthogonal projection matrix to linearly transform the shared feature tensor, ensuring that the functional requirement feature vector and the service experience feature vector are irrelevant in the vector space. The column vectors of the orthogonal projection matrix are composed of the functional code basis vectors in the enterprise service function library, such as the “account management-F01” “loan service-F02” “investment and financial planning-F03” basis vectors in the bank service function library, and the matrix dimension is [d x k] (d is the shared feature dimension, and k is the number of functional categories). The functional requirement feature vector is the vector component aligned with the functional code basis vector after projection, for example, the user request “increase transfer limit” after projection, the projection length of the vector in the “account management-F01” basis vector direction accounts for 85%. The service experience feature vector is the remaining component, representing non-functional requirements such as operation fluency and interface friendliness, for example, the user feedback “transfer confirmation page loads slowly” is reflected in the service experience feature vector as high-dimensional numerical fluctuations. Exemplarily, in the aviation ticket booking scenario, the feature decomposer decomposes the shared feature tensor into a functional requirement feature vector (mapped to the “flight query-F11” “change and sign-F12” basis vector) and a service experience feature vector (including “payment lag” “seat map loading failure” experience problems), achieving precise isolation of demand types.

[0041] Step 350: input the functional requirement feature vector into the classifier, match the mapping relationship between the functional code and the user request through the multi-head self-attention mechanism in the classifier, and generate a user demand type label.

[0042] Specifically, the classifier adopts a multi-head self-attention mechanism to calculate the similarity weight of the functional requirement feature vector and each basis vector in the preset functional code library. The multi-head self-attention mechanism captures the matching relationship in different subspaces through parallel multiple attention heads, for example, one group of attention heads focuses on the semantic similarity of functional categories, and the other group pays attention to the operation process correlation. The matching process is realized by calculating the dot product similarity of the functional requirement feature vector and the functional code basis vector, for example, the dot product value of the vector and the “loan approval-F201” basis vector is 0.93, and the dot product value of the vector and the “account query-F101” basis vector is 0.15, then it is determined that the user demand type label is “loan approval-F201”. The user demand type label is the identifier of the basis vector with the highest similarity in the functional code library, for example, in the retail e-commerce scenario, the matching degree of the functional requirement feature vector and the “return application-F305” basis vector is 0.89, and the classifier outputs the label F305. Exemplarily, in the education platform scenario, the similarity of the user functional requirement feature vector and the “course playback-F501” “assignment submission-F502” basis vectors is 0.76 and 0.24 respectively, and the multi-head self-attention mechanism generates the label F501 accordingly, accurately reflecting that the user's core demand is the course review function.

[0043] Step 360: input the service experience feature vector into a generator, copy the high-frequency operation events in the user behavior sequence through the pointer network in the generator, and generate an optimized suggestion text in combination with the intent words in the voice keyword sequence.

[0044] Specifically, the generator adopts a pointer network to locate the key operation events to be copied from the service experience feature vector. The pointer network determines the reference position of the high-frequency events in the user behavior sequence through attention distribution, for example, in the government APP scenario, the service experience feature vector shows that the user frequently triggers the "material upload retry" event (frequency up to 78%), and the pointer network attention weight peaks at the 5th, 9th, and 13th time steps of the behavior sequence, corresponding to three retry operations. The intent words in the voice keyword sequence are converted into word vectors through an embedding layer, for example, the keywords "stuttering" and "unable to submit" after the user's voice is converted into text are encoded into 256-dimensional vectors. The optimized suggestion text generation adopts a sequence-to-sequence architecture to combine the operation event description selected by the pointer network with the intent words, for example, the output is "optimize the response speed of the material upload interface and reduce repeated submission operations". Exemplarily, in the social platform scenario, the service experience feature vector indicates that the user frequently triggers the "picture loading failure" event, and the voice keywords include "delay" and "blur", and the generator accordingly suggests "improve the stability of the picture compression algorithm and increase the loading progress prompt function".

[0045] Step 370: input the user demand priority score, user demand type label, and optimized suggestion text into a joint loss calculator to calculate the regression loss value of the main task branch, the classification loss value of the auxiliary task branch, and the text generation loss value.

[0046] Specifically, the joint loss calculator calculates three types of losses in parallel: the main task branch adopts mean square error (MSE) to measure the deviation between the predicted value and the true value of the priority score, for example, the true score is 0.9 and the predicted value is 0.85, the regression loss value is (0.9-0.85)2=0.0025; the auxiliary task branch adopts cross-entropy loss to evaluate the difference between the classification result and the true label, for example, the true label is F201 and the predicted probability distribution is [F201:0.88, F101:0.12], the cross-entropy loss is -log(0.88)=0.127; the text generation loss adopts negative log-likelihood loss (NLL) to measure the matching degree of the generated text and the standard suggestion, for example, the word overlap rate of the generated text and the reference text is 80%, the NLL loss is 1.2. Exemplarily, in the logistics tracking scenario, the main task predicts a score of 0.75 (true value 0.8), the classifier outputs the label F401 (true F401), and the edit distance between the generated text and the manually annotated suggestion is 2, the joint loss calculator outputs the regression loss 0.0025, the classification loss 0.09, and the text generation loss 1.5, respectively.

[0047] Step 380: input the regression loss value, classification loss value and text generation loss value into a dynamic weight allocator, adjust the weight coefficients of each loss value through a gradient variance normalization algorithm in the dynamic weight allocator, and generate a balanced total loss value.

[0048] Specifically, the dynamic weight allocator uses a gradient variance normalization algorithm to dynamically adjust the weights according to the influence of each task loss on parameter updating. The gradient variance normalization algorithm first calculates the gradient variance of each loss function, for example, the regression loss gradient variance is 0.003, the classification loss is 0.008, and the text generation loss is 0.012; then the weights are allocated in inverse proportion to the variance, and the larger the variance, the smaller the weight to suppress gradient shock. For example, the regression loss weight = 1 / 0.003 ≈ 333, the classification loss weight = 1 / 0.008 = 125, and the text generation weight = 1 / 0.012 ≈ 83, and the normalized weight coefficients are 0.62, 0.23 and 0.15 respectively. The balanced total loss value is the weighted sum of each loss value multiplied by its weight coefficient, for example, 0.0025 x 0.62 + 0.09 x 0.23 + 1.5 x 0.15 = 0.236. Exemplarily, in the hotel reservation scenario, the regression loss is 0.004, the classification loss is 0.15, and the text generation loss is 2.0, and after weight allocation, the total loss is 0.004 x 0.5 + 0.15 x 0.3 + 2.0 x 0.2 = 0.437, effectively balancing the influence of different orders of magnitude losses on the model.

[0049] Step 390: back-propagate the total loss value to the shared feature layer, main task branch and auxiliary task branch, and synchronously optimize the network weights of the feature encoder, priority prediction module, classifier and generator through the parameter updater.

[0050] Specifically, the back-propagation process drives the parameter updater (such as the Adam optimizer) to adjust the weight matrix by calculating the gradient of the total loss with respect to the network parameters. The parameter updater uses an adaptive learning rate mechanism, for example, the BiLSTM layer learning rate of the feature encoder is set to 1e-4, the multi-head self-attention layer learning rate of the classifier is 3e-5, and the pointer network layer learning rate of the generator is 5e-5. Synchronous optimization means updating the network parameters of all branches simultaneously in one iteration, ensuring that the representation ability of the shared feature layer is consistent with the optimization direction of the task-specific modules. For example, in the bank risk control scenario, the feature encoder enhances the extraction ability of the "abnormal login" feature through gradient descent, the classifier simultaneously improves the recognition accuracy of the "account lock" label, and the generator optimizes the generation quality of the "multi-device login reminder" suggestion. Exemplarily, after 300 iterations, the mean square error of the priority prediction module decreases from 0.25 to 0.08, the classification accuracy improves from 72% to 89%, and the BLEU score of the generated text increases from 0.65 to 0.82.

[0051] Step 3100: Repeat the steps of feature encoding to parameter updating until the prediction error of the user demand priority score in the validation set is lower than a preset threshold and the classification accuracy of the user demand type label tends to be stable, and generate the demand analysis model with training convergence.

[0052] Specifically, the preset threshold is set according to business needs, for example, the priority prediction error needs to be lower than 0.1 (mean square error), and the classification accuracy needs to be more than 85%. The validation set uses data samples independent of the training set, for example, randomly selecting 20% of the session records from the bank APP users as validation data. The determination conditions of training convergence include that the prediction error fluctuation range is less than ±0.01 for 5 consecutive epochs, and the classification accuracy changes by no more than 1% within 3 epochs. Illustratively, in the e-commerce customer service model training, after 1500 iterations, the priority prediction error of the validation set is stabilized at 0.07, the classification accuracy is 90.3%, and the consistency of the generated suggestion and the artificial audit result is 88%, so the model is determined to be converged and the training is terminated. The finally generated demand analysis model can be deployed to the production environment to process user requests in real time and output priority score, demand type label and optimization suggestion text, forming a closed-loop management process of demand recognition-classification-improvement.

[0053] In an implementation scheme, the step 400 calls the demand analysis model to analyze the user demand feature set and generates an optimization strategy set associated with the target enterprise service process, including the following steps:

[0054] Step 410: Input the user demand feature set into the main task branch of the demand analysis model to obtain the user demand priority score and the demand evolution trend curve, wherein the demand evolution trend curve is used to represent the change of user demand intensity in different time periods.

[0055] Specifically, the main task branch models the user demand feature set in time sequence through the pre-trained demand analysis model, and outputs the user demand priority score and demand evolution trend curve. The user demand priority score is a numerical indicator quantifying the urgency of user demand processing. For example, in the bank credit card application scenario, the priority score of the demand feature set generated in the user's three consecutive sessions gradually increases from 0.65 (initial application consultation) to 0.92 (material supplement prompt), reflecting the escalation of demand urgency. The demand evolution trend curve is a two-dimensional curve graph with time as the horizontal axis and demand intensity as the vertical axis. It is generated by fitting the intensity changes of the user demand feature set in the historical time period. For example, during the e-commerce promotion period, the user's demand intensity for "inventory query" increases from 0.3 to 0.9 within two hours after the start of the event, and the curve shows a steep upward trend. In the insurance claim settlement scenario, the user demand feature set includes the feature sequence "report submission → material review → claim payment inquiry", and the demand evolution trend curve shows that the demand intensity for "claim payment inquiry" reaches a peak of 0.87 on the fifth day after the report, and the model outputs a priority score of 0.85, prompting the priority handling of the service response at this node.

[0056] Step 420: input the user demand feature set into the auxiliary task branch of the demand analysis model to obtain the user demand type label and optimization suggestion text, wherein the user demand type label is bound to the service item code in the target enterprise service function library.

[0057] Specifically, the auxiliary task branch analyzes the user demand feature set through the classifier and generator module. The user demand type label is a unique identifier that maps to a pre-set service item in the enterprise service function library. For example, in the airline ticketing system, the "flight change - F102" label corresponds to the call interface of the change rules engine in the service function library. The optimization suggestion text is an improvement scheme description generated based on user behavior patterns and semantic features. For example, for the user's high-frequency "seat map loading failure" event and the voice keyword containing "stuttering", the suggestion text "optimize seat map rendering logic and preload flight seat data" is generated. The service item code binding is achieved by querying the label-interface mapping table in the function library. For example, when the user demand type label is "F102", it is automatically associated with the configuration parameter template of the change service API. Illustratively, in the government service platform, the user demand type label "G305" corresponds to the "residence permit handling progress query" service item, and the optimization suggestion text is "add short message notification trigger condition: send progress update automatically after material review is passed". The label is bound to the query interface code JZ-QUERY-003 of the "residence permit business module" in the government service database, ensuring the accuracy of the generated strategy.

[0058] Step 430: Sort the optimization suggestion texts according to the user demand priority score, filter optimization suggestions with priority higher than a preset threshold as candidate strategies, and perform correlation analysis on the candidate strategies and demand evolution trend curve to determine the time window for strategy execution.

[0059] Specifically, the preset threshold is the lower limit of priority set by the enterprise according to resource allocation capability, for example, suggestions with priority score ≥ 0.7 are included in the candidate strategies. The sorting process adopts descending arrangement, for example, in the online education scenario, the optimization suggestion list contains [suggestion A: optimize video cache mechanism (0.92), suggestion B: increase after-school exercise export function (0.68), suggestion C: repair chapter jump error (0.85)], and suggestions A and C are retained after filtering. The strategy execution time window is determined by analyzing the peak interval and periodicity of the demand evolution trend curve, for example, when the curve shows that the demand intensity of “video lag” is higher than 0.8 continuously from 20:00 to 22:00 in the evening of weekdays, the strategy generator sets the execution window of suggestion A to this time period. Exemplarily, in the retail e-commerce scenario, the candidate strategy “improve the concurrent processing capability of the page of the killing activity” is associated with the time period “demand intensity > 0.9 within 5 minutes after the start of the activity” in the demand trend curve, and the execution window is determined to be 30 minutes before the start of the activity to 10 minutes after the end of the activity, ensuring that resource scheduling matches the demand peak.

[0060] Step 440: Input the candidate strategies and time window into the strategy generator, and convert the optimization suggestion texts into executable configuration instructions through the rule engine in the strategy generator, wherein the rule engine performs semantic-to-instruction mapping based on a preset instruction template library.

[0061] Specifically, the rule engine parses the operation object and action in the optimization suggestion text by natural language understanding techniques, and matches the parameterized instructions in the instruction template library. The instruction template library is a set of pre-defined structured instructions, for example, the template "optimize {component} rendering logic" maps to the configuration instruction "SetRenderEngine(component={component}, preload=True)". In the medical appointment system, the optimization suggestion text "shorten the source refresh interval to 30 seconds" is parsed by the rule engine and matches the template "adjust {service} interval to {time}", generating the instruction "UpdateInterval(service='source refresh', interval=30)". The time window parameter is injected into the execution condition field of the instruction, for example, "ExecuteBetween('09:00', '18:00)' is added. Exemplarily, in the logistics tracking scenario, the candidate strategy "increase the logistics node push frequency" is converted into the instruction "IncreasePushFrequency(service='logistics status', frequency=2)", and the time window condition "ActiveWhen(demand_strength>0.8)" is attached, ensuring that the instruction is triggered only when the demand strength exceeds the threshold.

[0062] Step 450: According to the type of the configuration instruction, call the corresponding service configuration interface in the target enterprise database to generate the set of optimization strategies containing execution parameters and trigger conditions.

[0063] Specifically, the service configuration interface is a set of pre-packaged APIs in the target enterprise system, such as the "risk control rule configuration interface" of a bank core system or the "promotion activity management interface" of an e-commerce platform. The configuration instruction type is determined by analyzing the function name and parameter structure in the instruction, for example, the instruction "SetRenderEngine" calls the "RenderConfig" interface of the graphics service module. The execution parameters are extracted from the instruction and formatted into the JSON structure required by the interface, for example, "component='seatMap',preload=True" is converted to {"targetComponent":"seatMap","preloadPolicy":true}. The trigger condition is compiled into the effective rule of the interface, for example, the time window "20:00-22:00" is converted to the cron expression "020-22***". Exemplarily, in a government service platform, the configuration instruction "UpdateInterval(service='progress query',interval=30)" calls the "TimerConfig" interface of the "government business configuration center", generates the policy item {"interface name":"TimerConfig","parameters":{ "service":"progress query","interval":30},"trigger condition":"every day 8:00-20:00"}, and finally forms an optimized policy set that can be immediately deployed.

[0064] In an implementation scheme, the step 500 of updating the digital service execution logic of the target enterprise according to the optimized policy set comprises the following steps:

[0065] Step 510: Analyze the service response rule adjustment strategy in the optimized policy set, obtain the response rule identifier that needs to be modified and the corresponding parameter adjustment range, and lock the rule entry to be updated in the service rule library of the target enterprise according to the parameter adjustment range.

[0066] Specifically, the service response rule adjustment policy contains the business rule identifier and its parameter value range to be adjusted, for example, in the bank online customer service system, the rule identifier "RR-2023-005" corresponds to the "credit card application review response time limit rule", and the parameter adjustment range modifies the original rule "response time ≤ 24 hours" to "response time ≤ 12 hours". The service rule library is a database for storing business logic rules, and the specific entry is located by rule identifier index. In the parsing process, according to the rule identifier in the policy, an accurate matching query is performed in the service rule library, for example, the corresponding rule entry is retrieved with "RR-2023-005" as the key value, the current parameter value "max_response_time = 24" is extracted, and it is compared with the target value "max_response_time = 12" in the policy to generate a parameter difference report. After locking the rule entry to be updated, the system will add an exclusive lock to the entry to prevent concurrent modification conflicts. Exemplarily, in the insurance claim system, the service response rule adjustment policy requires that the case type matching threshold of the "case preliminary review rule R-INS-09" be lowered from 0.7 to 0.6, and the field "threshold = 0.7" of the corresponding entry in the service rule library is marked as to be updated.

[0067] Step 520: parse the function module iteration policy, obtain the code repository address of the current function module from the version management system of the target enterprise, and compare the function change requirements in the optimization policy with the differences of the current code based on the code difference analysis tool to generate a code patch file.

[0068] Specifically, the functional module iteration strategy contains a specific description of the functional code changes, such as the need to add inventory pre-occupation interface call logic to the "seckill activity module" of the e-commerce platform. The repository address where the module code is stored in the version management system (such as GitLab) is "http: / / gitlab.com / ecom / seckill-module". The code difference analysis tool (such as BeyondCompare) converts the requirement description in the strategy document into AST (Abstract Syntax Tree) nodes and compares the structure with the code in the latest branch of the repository to identify missing interface call statements and parameter passing paths. The code patch file records the added, deleted, and modified code lines in a unified difference format (Unified Diff), such as the patch file "seckill_v2.patch" containing the new code line "inventoryService.reserve(stockId, userId)". For example, in a medical appointment system, the functional module iteration strategy requires the "number source release module" to add a 30-minute automatic cancellation function for unpaid functions. The code difference analysis tool locates the timing task trigger point to be inserted in "ReleaseService.java" and generates a patch file containing "scheduler.addTask(newAutoCancelTask())".

[0069] Step 530: Parse the interactive interface dynamic configuration strategy, extract interface element layout parameters and user operation hot area adjustment information, and generate a configuration verification result by real-time preview rendering of the target enterprise's interactive interface prototype through the interface rendering engine.

[0070] Specifically, the interactive interface dynamic configuration strategy contains adjustment data for layout parameters (such as button position coordinates, font size) and operation hot areas (such as clickable area range). For example, in the government service platform APP, the strategy requires the "social security query" button on the "personal center" page to be adjusted from the lower right corner (coordinates X:320, Y:1200) to the top navigation bar (X:80, Y:200), and the hot area range to be expanded from 80x80 pixels to 120x120 pixels. The interface rendering engine (such as Figma Engine) loads the current interface prototype file, performs real-time rendering after injecting new layout parameters, and generates a visual preview interface. The configuration verification result includes rendering time, element overlap detection report, and hot area coverage rate indicators, such as detecting that the adjusted "social security query" button has a 10% overlap area with the "public accumulation fund entry", and the engine automatically generates a warning log and suggests coordinate fine-tuning to X:200, Y:200. For example, in the airline ticket booking APP, the interactive interface dynamic configuration strategy changes the "flight filter" folding panel to the default expanded state, and the rendering engine verifies whether the expanded panel height exceeds the screen visible area, and outputs a "fitting passed" verification result.

[0071] Step 540: Execute the service response rule adjustment strategy, function module iteration strategy, and interactive interface dynamic configuration strategy in sequence in the sandbox environment, monitor system logs and abnormal events generated during strategy execution, and dynamically schedule the execution order of the strategies according to the monitoring results.

[0072] Specifically, the sandbox environment is a simulation test platform isolated from the production environment, for example, a bank system sandbox contains a copy of the full business rule library and a mirror database. The execution order of the strategies is dynamically arranged according to the dependency relationship, for example, the interface configuration strategy is deployed first to ensure front-end compatibility, then the function module code is updated, and finally the service response rule is adjusted. The monitoring system collects log data such as process resource occupancy, API response delay, and error code distribution. When it is detected that the "inventory pre-occupies interface 500 error" occurs after the execution of the function module iteration strategy, the dynamic scheduler suspends the execution of the subsequent rule adjustment strategy, rolls back the code patch to the previous version, and re-triggers the test process. Illustratively, in a logistics tracking system, after the sandbox environment executes the interactive interface dynamic configuration strategy, it monitors that the map rendering component has memory leaks, the dynamic scheduler immediately terminates the strategy execution chain, generates an event report "map component version is incompatible", and suggests upgrading the map SDK version as a priority.

[0073] In one implementation scheme, the dynamic scheduling of the execution order of the strategies in step 540 can include the following steps:

[0074] Step 541: Build a strategy dependency graph, where the nodes in the strategy dependency graph represent the types of optimization strategies, and the edges represent the prerequisite relationships between the strategies.

[0075] Specifically, the strategy dependency graph is modeled using a directed acyclic graph (DAG) structure, for example, in an airline ticket booking system, strategy A ("increase luggage capacity calculation interface") must be executed before strategy B ("optimize luggage purchase process") because strategy B depends on the interface parameters provided by strategy A. The node attributes include strategy type labels (such as "interface extension" and "interface optimization"), and the edge weights represent the dependency strength (such as the weight of strategy A→strategy B is 0.9, indicating strong dependency). Illustratively, in a bank risk control system, strategy X ("upgrade identity verification SDK") and strategy Y ("adjust face recognition threshold") have a bidirectional dependency, and the execution order needs to be determined through topological sorting.

[0076] Step 542: Initialize a virtual service configuration environment in the sandbox environment, and load the service rule snapshot and database copy in the current production environment of the target enterprise.

[0077] Specifically, the virtual service configuration environment is built by Docker container technology, for example, creating an independent container cluster for bank policy testing, containing rule engine images, database images, and API gateway images. The service rule snapshot is a time point copy of the production environment rule library, for example, exporting the full data of the "risk control rule table" at 2023-11-05T00:00:00Z. The database replica is obtained in an incremental synchronization manner, for example, using MySQLBinlog log playback to generate test data consistent with the production environment. For example, in a medical appointment system, the sandbox environment loads the required number source database replica and appointment interface configuration snapshot of policy OPT-2023-MED-008.

[0078] Step 543: Execute the optimized policy set in the order of the topological sorting result of the policy dependency graph, while monitoring the resource occupancy rate and service interface response time during the execution of each policy.

[0079] Specifically, the topological sorting generates an acyclic execution sequence based on the Kahn algorithm, for example, the policy execution order is [A→C→B→D]. Resource occupancy rate monitoring includes CPU usage (such as policy A execution peak of 75%), memory occupancy (such as policy B causing JVM heap memory to grow to 4GB), and disk I / O (such as policy C writing 500MB of logs per second). Service interface response time statistics cover the entire link, for example, the average response time of policy D calling the "payment interface" increases from 50ms to 200ms. For example, in a logistics tracking system, the policy execution order is [update location refresh interval→ optimize map rendering logic→ increase push frequency], and an alarm event is generated when it is detected that the GPU occupancy rate of the map rendering logic policy exceeds 90%.

[0080] Step 544: When a certain policy execution timeout or system alarm event is detected, interrupt the current execution process, and start the rollback mechanism to undo the effects of the executed policy to the last checkpoint.

[0081] Specifically, the timeout threshold is dynamically set according to the policy type, for example, the interface class policy timeout time is 30 seconds, and the interface configuration class policy is 10 seconds. The rollback mechanism realizes state recovery through transaction logs (such as Undo Log), for example, when policy B fails, the database changes and API configuration are rolled back according to the log. The checkpoint is automatically created before each policy execution, for example, the database state and service configuration of the sandbox environment are snapshoted before policy C execution. For example, in an e-commerce promotion system, the policy "increase the capacity of the kill queue" execution timeout (more than 45 seconds without returning results), the rollback module undoes the database connection pool expansion operation, and restores to the resource quota before policy execution.

[0082] Step 545: According to the success rate of policy execution and resource consumption indicators recorded in the historical execution log, the reinforcement learning algorithm is used to dynamically adjust the policy execution order, and the policy combination with high success rate and low resource consumption is preferentially executed.

[0083] Specifically, the state space of the reinforcement learning algorithm is the combination of policy execution order, the action space is the order adjustment operation (such as swapping the positions of policy B and policy C), and the reward function R = success rate x 100 - resource consumption coefficient x 50. For example, the historical success rate of a certain policy combination is 95%, and the resource consumption coefficient is 0.3, then R = 95 x 100 - 0.3 x 50 = 9498.5. The model updates the policy priority through Q-Learning, for example, the policy sequence with high R value is preferentially selected. Exemplarily, in a bank system, policy X (success rate 98%, resource coefficient 0.2) and policy Y (success rate 85%, resource coefficient 0.6) are ranked by the reinforcement learning model as [X→Y] to maximize the overall reward.

[0084] Step 546: During dynamic scheduling, record the system stability indicators under different policy sequence combinations, and generate policy scheduling optimization suggestions as data input for subsequent incremental training.

[0085] For example, system stability indicators include service availability (such as 99.95%), error rate (such as 0.05%), and fault recovery time (such as an average of 2 minutes). The format of the optimization suggestion can be a JSON structure, including

[0086] {"strategy_sequence":["A","C","B"],"success_rate":97,"resource_cost":0.4,"recommendation":"Swap C and B to reduce GPU contention"}. Exemplarily, in an airline booking system, dynamic scheduling log analysis finds that the stability of policy [P→Q→R] (99.2%) is better than [R→Q→P] (98.1%), generating the suggestion "preferentially execute policy P to preload cache data". These data are input into the incremental training process of the demand analysis model, improving the accuracy of future scheduling decisions.

[0087] Step 550: Publish the verified optimized policy set to the production environment, record the policy effective timestamp and version number, and create a user feedback data collection channel associated with the optimized policy set.

[0088] Specifically, production environment deployments employ a blue-green deployment mechanism, such as deploying new rule entries to a backup server cluster and gradually replacing older versions through traffic switching. Policy version numbers are generated in the format of "year-serial number-Hash," such as "OPT-2023-019-8a3f2e." User feedback data collection channels utilize a combination of event tracking SDKs and APIs. For example, in an e-commerce app, a separate event "SeckillV2_Feedback" is created for the "Flash Sale Module Optimization V2" policy to collect user clickstream data and rating pop-up input. For instance, in a government service platform, policy version "OPT-2023-GOV-005" takes effect on 2023-11-05T00:00:00Z, and the feedback channel listens to the " / api / feedback / gov005" interface to receive user-submitted satisfaction ratings and text evaluations.

[0089] In one implementation, step 550, creating a user feedback data collection channel associated with the optimization strategy set, includes the following steps:

[0090] Step 551: Embed a feedback triggering component in the target enterprise's digital service interaction interface. The feedback triggering component automatically pops up a feedback form after detecting that the user has completed a service operation related to the optimization strategy.

[0091] Specifically, the feedback triggering component is an embedded code module based on listening to user behavior events. For example, in the transfer function page of a bank app, after a user completes the "cross-border remittance" operation (this function falls under the coverage of optimization strategy version OPT-2023-BANK-007), the triggering component activates the feedback form pop-up by listening to the "transaction_complete" event. The triggering logic of the feedback form is bound to the optimization strategy version number; for example, data collection is only triggered when the version number of the function module used by the user is ≥2.3.5. For instance, in an airline booking app, after a user completes the "baggage allowance purchase" operation that conforms to strategy version OPT-2023-AIR-012, a semi-transparent floating window pops up in the lower right corner of the interface, displaying a feedback request: "How was your service experience?"

[0092] Step 552: Configure a multi-dimensional feedback form template, which includes a service response speed rating slider, a multi-select list of function improvement effects, a star rating for interface layout adjustment satisfaction, and a free text input box.

[0093] Specifically, the service response speed rating slider is a horizontal slider control with a scale range from 1 (extremely slow) to 5 (extremely fast). For example, in the material upload function of the government service platform, if a user drags the slider to the 4.2 scale, it indicates that they are quite satisfied with the upload speed improvement effect of strategy OPT-2023-GOV-005. The function improvement effect multi-select list uses checkboxes to list the function items involved in the strategy. For example, in strategy OPT-2023-ECO-019 of the e-commerce APP, the list includes options such as "new product filter" and "intelligent recommendation algorithm". If a user checks "intelligent recommendation algorithm", it indicates that they approve of its improvement effect. The interface layout adjustment satisfaction rating is a five-pointed star click rating component. For example, strategy OPT-2023-MED-008 of the medical appointment APP adjusted the department navigation bar layout. If a user clicks 4 stars, it indicates that they are basically satisfied with the layout adjustment but still have room for improvement. The free text input box supports 200 characters of UTF-8 encoding. For example, if a user enters "The new filter lacks the price range memory function, and I have to reset it every time", it indicates that they are satisfied with the layout adjustment but still have room for improvement. For example, in an education platform, a multi-dimensional feedback form simultaneously collects student feedback on "course playback buffering speed" (slider rating 4.5), "automatic saving of chapter quizzes" (multiple selections), and "learning progress interface color scheme" (3 stars).

[0094] Step 553: Bind the user-submitted feedback form data to the version number of the currently effective optimization strategy, and add a timestamp and user anonymous identifier to form a structured feedback record.

[0095] Specifically, data binding is implemented by injecting the strategy version number into the HTTP request header. For example, when a user submits feedback, the front-end code automatically appends the header field "X-Strategy-Version:OPT-2023-ECO-019". The timestamp is accurate to the millisecond level (e.g., 2023-11-05T14:23:45.678Z), and the user's anonymous identifier is generated by encrypting the device ID using the SHA-256 hash algorithm, for example, "u_9f86d081894c7d55". The structured feedback record in JSON format contains {"strategy_version":"OPT-2023-AIR-012","timestamp":"2023-11-05T14:23:45.678Z","user_hash":"u_9f86d08","scores":{"speed":4.2,"function":["Intelligent Recommendation"],"ui_star":4},"text":"We hope to add sorting options for filtered results"}. For example, in a logistics tracking system, user feedback records are bound to strategy OPT-2023-LOG-003, and anonymous identifiers ensure data compliance.

[0096] Step 554: Perform real-time cleaning on the structured feedback records, remove abnormal data containing invalid characters or scores that exceed a reasonable range, and store the cleaned data in an independent partition of the feedback database.

[0097] Specifically, invalid character detection uses regular expression matching; for example, if "" appears in free text, then "" is detected. <script>alert()< / script> The cleaning rules are triggered when HTML tags such as "" are displayed. The rating range validation ensures that the slider rating ∈ [1,5] and the star rating ∈ {1,2,3,4,5}. For example, if a record has a "ui_star" field value of 6, it is automatically marked as abnormal and discarded. The feedback database is divided into independent storage partitions according to the policy version number. For example, the feedback data for policy OPT-2023-BANK-007 is stored in the HBase table under the path " / feedback / strategy / bank / 007". For instance, in an insurance claims system, if a user-submitted text review contains the special character "#ERROR#", the cleaning module will identify it and store the record in an isolated area for manual review.

[0098] Step 555: Establish an automatic synchronization mechanism between the feedback database and the demand analysis model training platform. When the newly added feedback data reaches the preset batch size, the incremental training process of the model will be triggered.

[0099] Specifically, the automatic synchronization mechanism is implemented through a Kafka message queue. For example, each partition of the feedback database registers as a Topic producer, and the training platform subscribes to Topics as a consumer. The preset batch size is typically set to 1000 records. When the feedback data for strategy OPT-2023-MED-008 accumulates to 1000 records, the Kafka producer sends a "training_trigger" event to the training platform to start the incremental training task. For instance, in a retail e-commerce scenario, if the feedback data for strategy OPT-2023-ECO-019 reaches the 1000-record threshold within 48 hours, the training platform automatically pulls the data and updates the model parameters, forming a closed loop of "data collection-cleaning-training".

[0100] In one implementation, step 500, which involves incrementally training the demand analysis model based on user feedback data, includes the following steps:

[0101] Step 560: Obtain user ratings and text evaluations of the optimization strategy's effectiveness in real time through the user feedback data collection channel, wherein the ratings include service response speed ratings, functionality usability ratings, and interface operation smoothness ratings.

[0102] In particular, user feedback data is collected through structured forms and open-ended text input boxes, for example, an insurance claim APP pops up a rating window after completing online claim, asking users to rate "material upload speed" "progress query convenience" and "interface loading smoothness" within a range of 1-5 stars. The text evaluation field allows input of up to 200 characters of improvement suggestions, for example, the user inputs "material upload still has lag, suggest compressing pictures while retaining clarity". The data collection channel encodes the rating data as a numerical vector [4, 3, 5], and the text evaluation is stored in its original UTF-8 format. Illustratively, in an education platform, the rating data [response speed: 4.2 / 5, function ease of use: 4.5 / 5, interface smoothness: 4.7 / 5] is collected for "course playback buffer strategy V3", and the text evaluation contains "the problem of audio and video out of sync when playing at double speed has not been solved".

[0103] Step 570: Perform sentiment analysis and keyword extraction on the text evaluation, generate a feedback feature vector, and store the feedback feature vector in association with the corresponding set of optimization strategies, forming a feedback training sample set.

[0104] In particular, sentiment analysis uses a RoBERTa-based model to calculate the text sentiment polarity score (-1 to +1), for example, the evaluation "operation process is obviously simplified" scores +0.93, while "the submit button is still difficult to find" scores -0.65. Keyword extraction uses the TF-IDF algorithm to select the top three words by weight, such as extracting "historical orders" "batch export" "function" from the evaluation "hope to add historical order batch export function". The feedback feature vector is formed by concatenating the sentiment score, keyword embedding vector and rating data, for example, merging [0.82, [0.34, -0.12,..., 0.78], 4.5] into a 1536-dimensional vector. When storing in association, the mapping relationship between the feedback feature vector and the set of optimization strategies is established through the strategy version number, for example, the feedback training sample set entry "OPT-2023-EDU-007" corresponds to the 1536-dimensional feature vector and label data. Illustratively, in an air ticket booking system, the text evaluation "change fee description is not prominent enough" is processed to generate a feedback feature vector, which is stored in association with the interface configuration strategy in the strategy version "OPT-2023-AIR-012".

[0105] Step 580: Extract a preset proportion of samples from the feedback training sample set as incremental training data, and perform data augmentation processing on the incremental training data, wherein the data augmentation processing includes synonym replacement of text evaluation, noise injection of rating data and boundary expansion of strategy execution parameters.

[0106] Specifically, the preset proportion is usually set to 70% of the total amount of feedback samples, for example, 700 samples are randomly selected from 1000 samples. Synonym replacement adopts back-translation technology, and the Chinese evaluation "Loading too slow" is translated into English "Loading too slow", and then translated back into "Loading speed is too slow". Score data noise injection adds Gaussian noise within the original value ±0.2, for example, 4.5 is disturbed to 4.3 or 4.7. Strategy execution parameter boundary expansion modifies the ±10% range of the policy effective time window, for example, the original execution parameter "effective period 09:00-18:00" is expanded to "08:06-18:54". Exemplarily, in the incremental training of the bank risk control model, the text evaluation "identity verification step is cumbersome" is enhanced to "identity verification process is complex", and the score data 4.2 is added to 4.15 after noise, and the strategy parameter "face recognition threshold 0.85" is expanded to the interval [0.765, 0.935].

[0107] In an implementation scheme, the process of the data enhancement processing in the above step 580 can include the following three strategies:

[0108] In strategy one, for synonym replacement of text evaluation, a word vector replacement strategy based on semantic similarity is adopted, which can specifically include:

[0109] Step A1: A special word vector library in the field of enterprise services is constructed, and the words in the special word vector library are associated with the evaluation terms frequently appearing in user feedback.

[0110] Specifically, the special word vector library is generated by analyzing high-frequency words in the historical user feedback text of the target enterprise, and is expanded in combination with the domain knowledge graph. For example, in a medical appointment system, the terms "stuttering", "slow loading", "delay" and the like frequently appear in user feedback, and the special word vector library maps such words and their synonyms (such as "delay" "stagnation") to a 300-dimensional word vector space, and ensures semantic correlation through Word2Vec model training. Each word in the word vector library is associated with its semantic weight in different contexts, for example, the cosine similarity of "stuttering" with "delay" in the interface operation scene is 0.92, and the similarity with "malfunction" in the system function scene is 0.85. Exemplarily, for the user evaluation "the interface is seriously stuttering when submitting an appointment", the synonym set of "stuttering" in the special word vector library includes "delay", "slow response", and "stuck", which provides a candidate word source for subsequent replacement.

[0111] Step A2: For each entity word in the text evaluation, calculate the cosine similarity of the word with other words in the special word vector library, and select candidate replacement words with a similarity higher than a preset threshold.

[0112] Specifically, entity words are extracted from the text evaluation by dependency syntax analysis, such as extracting the entity words "payment" and "slow loading" from the evaluation "the page loading is slow in the payment process". The cosine similarity calculation adopts the vector dot product and length normalization method, such as the vector dot product of "slow loading" and "delay" is 2.7, the length product is 3.0, and the cosine similarity is 0.9. The preset threshold is set according to business needs, such as words with a similarity of ≥0.8 are included in the candidate set. For example, in the user evaluation "video playback buffering time is too long" on the education platform, the similarity of the entity word "buffering time" with "loading time" and "waiting time" in the special word vector library is 0.88 and 0.79 respectively, and after screening, "loading time" is retained as the candidate replacement word.

[0113] Step A3: Based on context analysis, the word with the highest semantic consistency with the original text is selected from the candidate replacement words to replace the original word, so that the replaced text evaluation maintains logical coherence.

[0114] Specifically, the context analysis adopts a pre-trained BERT model to calculate the semantic coherence score of the replacement word and the original sentence. For example, in the original sentence "the page loading is slow in the payment process", after replacing "slow loading" with "delay", the BERT model outputs a coherence score of 0.93; when replaced with "stuck", the score is 0.65, because "stuck" implies an extreme situation of system crash, which is not consistent with the context of moderate performance problems in "payment process". Finally, the replacement word with the highest score is selected to generate the enhanced text "the page delay is obvious in the payment process". For example, in the aviation ticket booking scenario, the user evaluation "seat map rendering is too slow" is replaced by "seat map loading is slow", which not only preserves the original meaning but also introduces lexical diversity.

[0115] In strategy two, for noise injection of rating data, an algorithm combining Gaussian noise superposition and boundary constraint is adopted, which can include:

[0116] Step B1: Analyze the distribution of historical rating data and calculate the mean and variance of each rating dimension.

[0117] Specifically, the historical rating data is statistically analyzed by dimension (such as service response speed, functional ease of use). For example, the "interface operation smoothness" score of the education platform in the past 6 months of 1000 data has a mean of 4.2 and a variance of 0.3. The distribution is verified by kernel density estimation (KDE), for example, the smoothness score is concentrated between 4.0-4.5, showing a left-skewed distribution. For example, in the "transfer speed" score dimension of the bank APP, the mean is 4.5 and the variance is 0.2, indicating that users are generally satisfied with the transfer speed but have less fluctuation.

[0118] Step B2: Generate random noise according to the mean and variance, and superimpose the random noise into the original score data.

[0119] Specifically, the Gaussian noise generator generates random numbers based on the N(μ, σ2) distribution, where μ = 0 and σ = the square root of the original variance. For example, if the variance of the functional usability score is 0.3, a Gaussian noise of σ = √0.3 ≈ 0.55 is generated, and the noise value range is about ±1.1. The superposition process operates independently for each score dimension, for example, the original score 4.5 superimposes noise +0.3 to become 4.8, and superimposes -0.4 to become 4.1. Exemplarily, in the government service platform, the "material uploading convenience" score of 4.3 may be enhanced to 4.6 or 4.0 after injecting noise, increasing data diversity.

[0120] Step B3: Boundary check is performed on the score data after injecting noise, and if it exceeds the preset score range, it is truncated to make the data after noise injection conform to the actual business scenario.

[0121] Specifically, the preset score range is usually 1-5 points, and the boundary check is realized by comparing the noise value with the upper and lower limits. For example, the original score 4.8 superimposes noise +0.5 to get 5.3, which exceeds the upper limit and is truncated to 5.0; the score 2.5 superimposes noise -1.0 to get 1.5, which is lower than the lower limit and is truncated to 1.0. Exemplarily, in the e-commerce scenario, the "description accuracy" score of a certain commodity 4.9 after injecting noise +0.3 is modified to 5.0, avoiding the generation of abnormal data exceeding the actual range.

[0122] In strategy three, for the boundary expansion of the strategy execution parameter, the parameter space exploration strategy is adopted, which can specifically include:

[0123] Step C1: Identify the adjustable execution parameters in the optimization strategy set and determine their original value range.

[0124] Specifically, the adjustable parameters need to meet the continuous or ordered discrete attributes, such as the response time threshold of the service response rule, the transparency parameter of the interface element, etc. The original value range is obtained by analyzing the constraint conditions in the strategy file, for example, the "seat map preloading time" parameter of the airline ticketing system has an original range of [30 seconds, 60 seconds]. Exemplarily, in the logistics tracking system, the original value range of the "position refresh interval" parameter is [10 seconds, 30 seconds], which is determined by the "interval_min = 10, interval_max = 30" field in the strategy document.

[0125] Step C2: Extend the upper and lower limits of the parameter value range according to the preset expansion ratio to generate an expanded parameter space.

[0126] Specifically, the expansion ratio is usually set to 10%-20%, for example, expanding the "seat map preloading time" upper limit of 60 seconds by 20% to 72 seconds and the lower limit of 30 seconds to 24 seconds to form a new range [24 seconds, 72 seconds]. The parameter space dimension is consistent with the number of parameters, for example, containing 3 adjustable parameters, the expanded space is a three-dimensional hypercube. Illustratively, in a medical appointment system, the "number source refresh interval" parameter is expanded by 15% from [30 seconds, 90 seconds] to [25.5 seconds, 103.5 seconds], which expands the adaptability of the model to different configurations.

[0127] Step C3: Uniformly sample in the expanded parameter space to generate new policy execution parameter combinations, and combine them with the original parameter combinations to form enhanced training data.

[0128] Specifically, the uniform sampling uses Latin hypercube design to ensure the coverage of the parameter space, for example, generating 100 sampling points in a two-dimensional parameter space, and the value range of each dimension is divided into 10 intervals, and only one point is selected in each interval. After the new parameter combinations are combined with the original combinations, the amount of training data is expanded to 2-3 times the original size. Illustratively, in a bank risk control system, after the "transaction review threshold" is expanded from [0.7, 0.9] to [0.63, 0.97], new parameters such as threshold values 0.68, 0.85, and 0.93 are generated by sampling, which are input into the model training together with the original values 0.7, 0.8, and 0.9.

[0129] Step 590: Input the enhanced incremental training data into the demand analysis model, freeze the parameters of the auxiliary task branch in the demand analysis model, only fine-tune the hierarchical convolutional neural network and the time recurrent network in the main task branch, and update the demand priority prediction logic.

[0130] Specifically, parameter freezing is achieved by setting the gradient calculation flag of the auxiliary task branch to False, for example, the pointer network parameters in the generator remain fixed. The hierarchical convolutional neural network uses multi-scale convolution kernels (3x3, 5x5, 7x7) to extract local patterns of feedback features, and the time recurrent network (such as LSTM) models the time series change of demand intensity after the strategy takes effect. The fine-tuning process uses a lower learning rate (such as 1e-5) to only update the weight matrix of the main task branch. Illustratively, in a retail e-commerce demand analysis model, the incremental training data enhances the response weight of the hierarchical convolutional neural network to "promotion activity related" feedback features, and the time recurrent network adjusts the prediction sensitivity to "demand fluctuations during major promotions".

[0131] Step 5100: In the fine-tuning process, an elastic weight solidification algorithm is used to protect important parameters in the main task branch, wherein the elastic weight solidification algorithm calculates the protection strength according to the Fisher information amount of the parameters in the historical training.

[0132] Specifically, Fisher information is estimated by calculating the second derivative of the parameter on the historical training data, for example, the Fisher information of a certain convolution kernel in a hierarchical convolutional neural network is 8.3, indicating that it has a significant impact on the final prediction result. The elastic weight solidification algorithm sets a protection coefficient λ = 1 / (1+exp(-Fisher / 10)) for the parameter, when λ>0.9, the weight change amplitude is limited to ±5%. Illustratively, in the air ticket booking demand model, the LSTM unit parameter in the time recursive network that processes the "holiday demand peak" has a high Fisher information (12.7), and its weight adjustment is limited to the range of [-0.05, +0.05] during fine-tuning, avoiding catastrophic forgetting of historical learning achievements.

[0133] In an implementation scheme, the execution process of the elastic weight solidification algorithm in the above step 5100 includes the following steps:

[0134] Step 5110: After the completion of the historical training, calculate the Fisher information of each parameter in the main task branch of the demand analysis model, which is obtained by calculating the second derivative expectation value of the parameter on the training data.

[0135] Specifically, Fisher information reflects the sensitivity of the parameter to the impact on the model prediction result. The calculation process needs to traverse the historical training data set, calculate the second derivative of the loss function with respect to the parameter θ for each sample, and take the average as the final value. For example, the Fisher information of a certain convolution kernel parameter θ1 in a bank risk control model is 8.3, indicating that the parameter has high importance to risk score prediction. Illustratively, in the air demand prediction model, the Fisher information of the gate unit parameter θ2 in the time recursive network is 12.5, which is significantly higher than the 5.1 of the fully connected layer parameter θ3.

[0136] Step 5120: Construct a parameter importance matrix according to the Fisher information, wherein the element value in the parameter importance matrix represents the contribution of the corresponding parameter to the model prediction result.

[0137] Specifically, the parameter importance matrix is a tensor with the same dimension as the model parameter, and the element value is obtained by normalizing the Fisher information. For example, the Fisher information of parameter θ1 is 8.3, which is mapped to 0.83 in the matrix (assuming the maximum Fisher value is 10), and the Fisher information of parameter θ2 is 12.5, which is mapped to 1.0. The matrix is used for penalty term calculation in the subsequent loss function. Illustratively, in the demand model of the education platform, the parameter importance matrix value of the first convolution layer of the hierarchical convolutional neural network is [0.75, 0.92, 0.68], reflecting the importance difference of different convolution kernels.

[0138] Step 5130: In the incremental training process, define an elastic weight loss function that adds a parameter importance penalty term to the original loss function, which is proportional to the parameter's current weight relative to the historical weight and the Fisher information.

[0139] Specifically, the elastic weight loss function expression is L = L_original + λΣ(F_i*(θ_i-θ_i^0)^2), where λ is the penalty coefficient, F_i is the Fisher information of parameter θ_i, and θ_i^0 is the historical parameter value after training. For example, when the current value of θ1 is 0.85, the historical value is 0.82, and F_i = 8.3, the penalty term is 8.3*(0.85-0.82)^2 = 0.747. Exemplarily, in the medical appointment model, if the current offset of parameter θ2 is 0.1 and F_i = 12.5, its penalty term contribution is 12.5*(0.1)^2 = 0.125.

[0140] Step 5140: Minimize the elastic weight loss function by gradient descent algorithm, so that the update amplitude of key parameters is constrained, and non-key parameters allow large amplitude adjustment.

[0141] Specifically, in the gradient descent process, the gradient of high Fisher information parameters will be suppressed by the penalty term. For example, the gradient update amount of parameter θ1 is Where η is the learning rate. When λ = 0.1, F_i = 8.3, and θ1-θ1^0 = 0.03, the gradient correction caused by the penalty term is -0.498, which significantly reduces the parameter update amplitude. Exemplarily, in the logistics demand model, the update amount of key parameter θ3 is limited to ±0.02, while non-key parameter θ4 allows adjustment of ±0.15.

[0142] Step 5150: After each incremental training iteration, recalculate the Fisher information of the updated parameters, and dynamically adjust the parameter importance matrix to gradually expand the range of protected parameters.

[0143] Specifically, the Fisher information is recalculated on the new data after each iteration, and the newly added high importance parameters are added to the protection list. For example, after the fifth incremental iteration, the Fisher information of parameter θ5 in the education platform model increased from 4.2 to 9.1, exceeding the threshold of 7.0, and was included in the parameter importance matrix and subjected to the penalty term. The dynamic adjustment mechanism is realized by a sliding window algorithm, for example, only parameters with continuously increasing Fisher information in the last three iterations are retained. Exemplarily, in the e-commerce promotion model, after three incremental training, the number of protected parameters increases from the initial 56 to 89, covering more key feature extraction nodes.

[0144] Based on and Figure 1 Based on the same principle as the method shown, this embodiment of the invention also provides a big data mining device 10, such as... Figure 2 As shown, the device 10 includes:

[0145] The data acquisition module 11 is used to acquire multimodal conversation data generated by the target enterprise during the process of digital service interaction with users. The multimodal conversation data includes text conversation segments, voice conversation segments and user behavior sequences. The user behavior sequences are used to record the time sequence and event attributes of operation events triggered by users in the interactive interface.

[0146] Feature extraction module 12 is used to extract a set of user demand features from multimodal conversation data. The set of user demand features includes text demand features, voice demand features and behavioral demand features, among which behavioral demand features are used to describe users’ operational preferences and functional requirements for digital services.

[0147] Model training module 13 is used to jointly train the user demand feature set based on a preset multi-task learning framework to generate a demand analysis model. The demand analysis model is used to identify the priority of user demands and predict the direction of service optimization.

[0148] The strategy generation module 14 is used to call the demand analysis model to parse the user demand feature set and generate an optimization strategy set associated with the target enterprise service process. The optimization strategy set includes service response rule adjustment strategy, functional module iteration strategy and interactive interface dynamic configuration strategy.

[0149] The iterative optimization module 15 updates the digital service execution logic of the target enterprise according to the set of optimization strategies, and incrementally trains the demand analysis model based on user feedback data.

[0150] The above embodiments describe the big data mining device 10 from the perspective of a virtual module. The following describes a computer system from the perspective of a physical module, as detailed below:

[0151] The computer system provided in the embodiments of the present invention is as follows: Figure 3 As shown, the computer system 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one type, and the structure of this computer system 100 does not constitute a limitation on the embodiments of the present invention.

[0152] The embodiment of the present application provides a computer system, the computer system in the embodiment of the present application comprises: one or more processors;Memory;One or more computer programs, wherein the one or more computer programs are stored in the memory and are configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above-mentioned method is realized.

[0153] The above only describes some embodiments of the present application, and it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A big data mining method based on digitized enterprise management, characterized in that, The method comprises: obtaining multi-modal conversation data generated by a target enterprise in a digital service interaction with a user, the multi-modal conversation data comprising text conversation segments, voice conversation segments and a user behavior sequence, wherein the user behavior sequence is used to record the time sequence and event attributes of operation events triggered by the user in an interaction interface; extracting a user demand feature set from the multi-modal conversation data, the user demand feature set comprising text demand features, voice demand features and behavior demand features, wherein the behavior demand features are used to describe the operation preferences and functional demands of the user for the digital service; based on a preset multi-task learning framework, jointly training the user demand feature set to generate a demand analysis model, the demand analysis model being used to identify user demand priorities and predict service optimization directions; calling the demand analysis model to analyze the user demand feature set and generate an optimization strategy set associated with the service process of the target enterprise, the optimization strategy set comprising service response rule adjustment strategies, functional module iteration strategies and interaction interface dynamic configuration strategies; updating the digital service execution logic of the target enterprise according to the optimization strategy set, and incrementally training the demand analysis model based on user feedback data; based on the preset multi-task learning framework, jointly training the user demand feature set to generate a demand analysis model, comprising: inputting the user demand feature set into a feature encoder, extracting cross-modal time sequence features through a bidirectional long short-term memory network in the feature encoder to generate demand time sequence encoding vectors; inputting the demand time sequence encoding vectors into a shared feature layer of the multi-task learning framework, filtering cross-conversation key feature nodes through a sparse attention mechanism in the shared feature layer to generate a shared feature tensor; inputting the shared feature tensor into a priority prediction module of a main task branch, capturing demand intensity fluctuation patterns through a time convolution network in the priority prediction module to generate a demand intensity waveform graph, and inputting the demand intensity waveform graph into a sequence of gated recurrent units to output user demand priority scores; inputting the shared feature tensor into a feature decomposer of an auxiliary task branch, separating functional demand feature vectors and service experience feature vectors through an orthogonal projection matrix in the feature decomposer, wherein the functional demand feature vectors are aligned with the functional encoding dimensions in a preset enterprise service function library; inputting the functional demand feature vectors into a classifier, matching the mapping relationship between the functional encoding and the user request through a multi-head self-attention mechanism in the classifier to generate user demand type labels; inputting the service experience feature vectors into a generator, copying high-frequency operation events in the user behavior sequence through a pointer network in the generator, and generating optimization suggestion texts in combination with the intent words in the voice keyword sequence; inputting the user demand priority scores, the user demand type labels and the optimization suggestion texts into a joint loss calculator to calculate regression loss values of the main task branch, classification loss values of the auxiliary task branch and text generation loss values, respectively. The regression loss value, the classification loss value and the text generation loss value are input into a dynamic weight allocator, the weight coefficients of the loss values are adjusted by a gradient variance normalization algorithm in the dynamic weight allocator, and a balanced total loss value is generated; The total loss value is back-propagated to the shared feature layer, the main task branch and the auxiliary task branch, and the network weights of the feature encoder, the priority prediction module, the classifier and the generator are synchronously optimized by a parameter updater; The steps of feature encoding to parameter updating are repeatedly performed until the prediction error of the user demand priority score in the validation set is lower than a preset threshold and the classification accuracy of the user demand type label tends to be stable, and a training-converged demand analysis model is generated.

2. The method of claim 1, wherein, The user demand feature set is extracted from the multi-modal conversation data, including: The text conversation segment is subjected to semantic segmentation processing to generate a plurality of text sub-segments, wherein each text sub-segment contains at least one complete demand description sentence, and adjacent text sub-segments are connected by context association labels; The voice conversation segment is subjected to voiceprint feature separation and speech-to-text processing to generate a voice text set, and the voice text set is subjected to intent keyword extraction to generate a voice keyword sequence, wherein the intent keyword is associated with a function label in a preset enterprise service function library; The operation event attributes in the user behavior sequence are analyzed, core operation events with a trigger frequency higher than a preset threshold are extracted, and a behavior pattern graph is constructed based on the time sequence of the core operation events, wherein the nodes in the behavior pattern graph represent operation event types, and the edges represent time intervals and front-back dependency relationships triggered by the core operation events; The text sub-segments, the voice keyword sequence and the behavior pattern graph are input into a feature fusion network, the association weights between text, voice and behavior features are calculated by a cross-modal attention layer in the feature fusion network, and the heterogeneous features are weighted and spliced according to the association weights to generate the user demand feature set.

3. The method of claim 2, wherein, The demand analysis model is called to analyze the user demand feature set to generate an optimization strategy set associated with a target enterprise service process, including: The user demand feature set is input into the main task branch of the demand analysis model to obtain a user demand priority score and a demand evolution trend curve, wherein the demand evolution trend curve is used to represent the change of user demand intensity in different time periods; The user demand feature set is input into the auxiliary task branch of the demand analysis model to obtain a user demand type label and an optimization suggestion text, wherein the user demand type label is bound to a service item code in a target enterprise service function library; The optimization suggestion text is sorted according to the user demand priority score, candidate strategies with a priority higher than a preset threshold are screened, and the candidate strategies are associated with the demand evolution trend curve for analysis to determine a time window for strategy execution; The regression loss value, the classification loss value and the text generation loss value are input into a dynamic weight allocator, the weight coefficients of the loss values are adjusted by a gradient variance normalization algorithm in the dynamic weight allocator, and a balanced total loss value is generated; The total loss value is back-propagated to the shared feature layer, the main task branch and the auxiliary task branch, and the network weights of the feature encoder, the priority prediction module, the classifier and the generator are synchronously optimized by a parameter updater; The steps of feature encoding to parameter updating are repeatedly performed until the prediction error of the user demand priority score in the validation set is lower than a preset threshold and the classification accuracy of the user demand type label tends to be stable, and a training-converged demand analysis model is generated. The user demand feature set is extracted from the multi-modal conversation data, including: The text conversation segment is subjected to semantic segmentation processing to generate a plurality of text sub-segments, wherein each text sub-segment contains at least one complete demand description sentence, and adjacent text sub-segments are connected by context association labels; The voice conversation segment is subjected to voiceprint feature separation and speech-to-text processing to generate a voice text set, and the voice text set is subjected to intent keyword extraction to generate a voice keyword sequence, wherein the intent keyword is associated with a function label in a preset enterprise service function library; The operation event attributes in the user behavior sequence are analyzed, core operation events with a trigger frequency higher than a preset threshold are extracted, and a behavior pattern graph is constructed based on the time sequence of the core operation events, wherein the nodes in the behavior pattern graph represent operation event types, and the edges represent time intervals and front-back dependency relationships triggered by the core operation events; The text sub-segments, the voice keyword sequence and the behavior pattern graph are input into a feature fusion network, the association weights between text, voice and behavior features are calculated by a cross-modal attention layer in the feature fusion network, and the heterogeneous features are weighted and spliced according to the association weights to generate the user demand feature set. The demand analysis model is called to analyze the user demand feature set to generate an optimization strategy set associated with a target enterprise service process, including: The user demand feature set is input into the main task branch of the demand analysis model to obtain a user demand priority score and a demand evolution trend curve, wherein the demand evolution trend curve is used to represent the change of user demand intensity in different time periods; The user demand feature set is input into the auxiliary task branch of the demand analysis model to obtain a user demand type label and an optimization suggestion text, wherein the user demand type label is bound to a service item code in a target enterprise service function library; The optimization suggestion text is sorted according to the user demand priority score, candidate strategies with a priority higher than a preset threshold are screened, and the candidate strategies are associated with the demand evolution trend curve for analysis to determine a time window for strategy execution; inputting the candidate strategy and the time window into a strategy generator, and converting optimization suggestion text into executable configuration instructions through a rule engine in the strategy generator, wherein the rule engine performs semantic-to-instruction mapping based on a preset instruction template library; calling a corresponding service configuration interface in a target enterprise database according to the type of the configuration instructions, and generating the optimization strategy set containing execution parameters and trigger conditions.

4. The method of claim 3, wherein, updating the digital service execution logic of the target enterprise according to the optimization strategy set, including: parsing a service response rule adjustment strategy in the optimization strategy set, obtaining a response rule identifier that needs to be modified and a corresponding parameter adjustment range, and locking a to-be-updated rule item in a service rule library of the target enterprise according to the parameter adjustment range; parsing the function module iteration strategy, obtaining a code repository address of the current function module from a version management system of the target enterprise, and comparing differences between function change requirements in the optimization strategy and the current code based on a code difference analysis tool to generate a code patch file; parsing the interactive interface dynamic configuration strategy, extracting interface element layout parameters and user operation hot area adjustment information, and generating a configuration verification result by real-time preview rendering of an interactive interface prototype of the target enterprise through an interface rendering engine; sequentially executing the service response rule adjustment strategy, the function module iteration strategy and the interactive interface dynamic configuration strategy in a sandbox environment, monitoring system logs and abnormal events generated in the strategy execution process, and dynamically scheduling the strategy execution order according to the monitoring result; publishing the verified optimization strategy set to a production environment, recording a strategy effective timestamp and a version number, and creating a user feedback data collection channel associated with the optimization strategy set.

5. The method of claim 1, wherein, incrementally training the demand analysis model based on the user feedback data, including: real-time acquisition of user score data and text evaluation of optimization strategy execution effect through the user feedback data collection channel, wherein the score data includes service response speed score, function ease of use score and interface operation fluency score; performing sentiment analysis and keyword extraction on the text evaluation to generate a feedback feature vector, and associating and storing the feedback feature vector with the corresponding optimization strategy set to form a feedback training sample set; extracting a preset proportion of samples from the feedback training sample set as incremental training data, and performing data enhancement processing on the incremental training data, wherein the data enhancement processing includes synonym replacement of text evaluation, noise injection of score data and boundary expansion of strategy execution parameters; inputting the enhanced incremental training data into the demand analysis model, freezing parameters of auxiliary task branches in the demand analysis model, only fine-tuning a hierarchical convolutional neural network and a time recurrent network in a main task branch, and updating demand priority prediction logic; in the fine-tuning process, an elastic weight solidification algorithm is used to protect important parameters in the main task branch, wherein the elastic weight solidification algorithm calculates protection strength according to Fisher information quantity of parameters in historical training.

6. The method of claim 5, wherein, the data enhancement processing includes the following strategies: For synonym replacement of text evaluation, a word vector replacement strategy based on semantic similarity is adopted, including: A special word vector library for enterprise service fields is constructed, and the words in the special word vector library are associated with evaluation terms frequently appearing in user feedback; For each entity word in the text evaluation, the cosine similarity between the entity word and other words in the special word vector library is calculated, and candidate replacement words with a similarity higher than a preset threshold are screened; Based on context analysis, the word with the highest semantic consistency with the original text is selected from the candidate replacement words for replacement, so that the replaced text evaluation maintains logical coherence; For noise injection of scoring data, a Gaussian noise superposition combined with boundary constraint algorithm is adopted, including: The distribution law of historical scoring data is analyzed, and the mean and variance of each scoring dimension are calculated; Random noise conforming to Gaussian distribution is generated according to the mean and variance, and the random noise is superimposed on the original scoring data; Boundary check is performed on the scoring data after noise injection, and if it exceeds the preset scoring range, it is truncated to make the noise-injected data conform to the actual business scenario; For boundary expansion of strategy execution parameters, a parameter space exploration strategy is adopted, including: Identify the adjustable execution parameters in the optimization strategy set and determine their original value range; According to the preset expansion ratio, the upper and lower limits of the parameter value range are extended to generate an expanded parameter space; New strategy execution parameter combinations are uniformly sampled in the expanded parameter space, and they are combined with the original parameter combinations to form enhanced training data.

7. The method of claim 6, wherein, The execution process of the elastic weight solidification algorithm includes the following steps: After the completion of historical training, the Fisher information of each parameter in the main task branch of the demand analysis model is calculated, and the Fisher information is obtained by calculating the second derivative expectation value of the parameter on the training data; According to the Fisher information, a parameter importance matrix is constructed, wherein the element value in the parameter importance matrix represents the contribution of the corresponding parameter to the model prediction result; In the incremental training process, an elastic weight loss function is defined, which adds a parameter importance penalty term to the original loss function, and the parameter importance penalty term is proportional to the Fisher information and the deviation of the current parameter weight from the historical weight; The elastic weight loss function is minimized by the gradient descent algorithm, so that the update amplitude of the key parameters is constrained, and the non-key parameters are allowed to be adjusted by a large amplitude; After each incremental training iteration, the Fisher information of the updated parameters is recalculated, and the parameter importance matrix is dynamically adjusted to gradually expand the range of protected parameters.

8. The method of claim 4, wherein, The creation of the user feedback data collection channel associated with the optimization strategy set includes: Embed a feedback trigger component in the digital service interaction interface of the target enterprise, which automatically pops up a feedback form after detecting that the user has completed a service operation related to the optimization strategy; Configure a multi-dimensional feedback form template, which includes: a service response speed score slider for collecting users' perception scores on service delay time after policy execution; a function improvement effect multi-choice list for collecting users' recognition of new or optimized functions; an interface layout adjustment satisfaction star rating for quantifying users' acceptance of changes in the interactive interface; a free text input box for receiving users' unstructured descriptions of the policy execution effect; binding the feedback form data submitted by the user with the current effective optimization strategy version number, adding a timestamp and a user anonymous identifier to form a structured feedback record; real-time cleaning of the structured feedback records, removing abnormal data containing invalid characters or scores beyond a reasonable range, and storing the cleaned data in a separate partition of the feedback database; establishing an automatic synchronization mechanism between the feedback database and the demand analysis model training platform, triggering the model incremental training process when the new feedback data reaches a preset batch size.

9. A computer system, characterized by comprise: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processor, the method as claimed in any one of claims 1-8 is implemented.

Images