Sales lead mining methods that integrate multi-modal data
By identifying the lifecycle states and state sensitivity spectrum of sales leads, generating smooth weight vectors, detecting mutations, and performing incremental updates, the problem of unstable scoring under multimodal data is solved, achieving the stability and reliability of scoring, and making it suitable for high-concurrency, low-latency sales conversion scenarios.
Patent Information
- Application Number
- CN202611275717.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
AI Technical Summary
Under the dynamic changes of multimodal data, existing technologies are prone to drastic fluctuations in sales lead scores, leading to score instability, which affects sales strategies and model credibility. Furthermore, existing mechanisms cannot simultaneously guarantee the real-time sensitivity and cross-stage continuity of scores.
By acquiring multimodal sales lead data streams, identifying lifecycle states, generating state labels, querying state sensitivity genealogy libraries, calculating smooth weight vectors, detecting mutations and delaying releases, monitoring state transitions for incremental updates, generating stability scores, and optimizing scores through feedback closed-loop calibration.
It achieves temporal consistency and logical coherence of scoring results, reduces the risk of scoring drift, improves the stability and credibility of scoring, supports high-concurrency, low-latency industrial-grade deployment, and enhances user trust.
Smart Images

Figure CN123115238A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multimodal sales lead scoring and stability control technology, and in particular to a sales lead mining method that integrates multimodal data. Background Technology
[0002] Currently, with the widespread adoption of SaaS CRM systems in sales automation and precision marketing, enterprises are increasingly relying on the intelligent fusion analysis of multimodal data for lead acquisition, management, and value assessment. Mainstream technical solutions typically employ fusion modeling modules to integrate heterogeneous data from multiple sources within the CRM, such as structured fields, social media behavior, and website browsing records, through methods like feature concatenation, weighted summation, or deep learning. This enables automatic scoring of lead conversion probabilities or potential scores. Dynamic weight adjustment mechanisms are often considered a key path to improve the sensitivity and business adaptability of the scoring model. These mechanisms can leverage rule engines, attention mechanisms, and gating networks to adaptively allocate the importance of each modality feature based on the real-time data status and historical performance of different leads.
[0003] However, the existing technology system faces the following prominent technical challenges: Firstly, the typical paradigm of real-time weight recalculation and score output makes the model prone to drastic score fluctuations under dynamic conditions of multimodal data changes. When some modal data (such as an explosion in social media interaction or abnormal growth in website behavior) changes abruptly, the model weights will adjust dramatically in an instant, causing sales lead scores to fluctuate wildly in a short period of time, exhibiting serious instability. This score fluctuation is not only detrimental to the sales team's long-term follow-up and strategy optimization of leads, but also directly affects the business credibility of the model's output results.
[0004] On the other hand, existing mechanisms either rely on static scoring windows to forcefully smooth the output or simply adjust the weight normalization method, failing to simultaneously guarantee the immediate sensitivity and cross-stage continuity of the scoring. Especially when sales leads are at critical transition points in the business lifecycle (such as switching from "first contact" to "multi-channel interaction"), the importance of model weights for different modalities should undergo structural leaps, but rapid dynamic adjustments often introduce new risks of scoring jumps. While some studies have introduced temperature scaling and gating mechanisms to attempt to alleviate this problem, in practice, there are still technical shortcomings such as insufficient robustness of fusion models and difficulty in effectively suppressing outlier data.
[0005] In actual business applications, these score fluctuations can lead to the following typical problems: sales teams frequently adjust lead follow-up strategies due to drastic score changes, increasing resource consumption; management cannot accurately obtain the value trend of leads, resulting in distorted decision-making references; and the trustworthiness of the AI scoring system decreases due to inconsistent output, requiring frequent manual intervention for review, thus defeating the original purpose of intelligent recommendation. Summary of the Invention
[0006] This application provides a sales lead mining method that integrates multi-modal data, aiming to solve one of the problems or issues of the existing technology mentioned in the background section above.
[0007] The sales lead mining method that integrates multi-modal data provided in this application specifically includes:
[0008] S1: Acquire multimodal sales lead data streams and identify the lifecycle status of each lead, generating lead status tags with timestamps.
[0009] S2: Based on the clue state label, query the preset state sensitivity spectrum library, extract the state stability threshold and modal sensitivity spectrum bound to the state label, and generate a state constraint parameter set.
[0010] S3: Based on the multimodal sales lead feature vector at the current moment and the state constraint parameter set, calculate the weighted moving average of the most recent effective weights to generate a basic smooth weight vector.
[0011] S4: Perform mutation detection based on the instantaneous difference between the basic smooth weight vector and the current multimodal sales lead feature vector. If the single-modal increase exceeds the preset mutation buffer threshold, the instantaneous change is temporarily stored and released with a delay, generating a buffer weight vector that has been suppressed by mutation.
[0012] S5: Monitor whether the clue state label has undergone a cross-stage transition. If a state transition is determined to have occurred, trigger a progressive interpolation update operation to correct the buffer weight vector to a state weight vector.
[0013] S6: Generate basic cue scores based on the state weight vector, and simultaneously capture the amplitude and rhythm characteristics of each modal residual signal to generate fluctuation suppression coefficients.
[0014] S7: Perform nonlinear decay or enhancement adjustment operation on the basic clue score based on the fluctuation suppression coefficient, and output the final clue score with stability confidence meta-label.
[0015] S8: Collect the manually labeled clue conversion results and compare them with the final clue score. If the absolute value of the deviation exceeds the preset joint feedback threshold, construct a stability and effectiveness joint feedback set.
[0016] The sales lead mining method that integrates multi-modal data provided in this application has the following beneficial effects:
[0017] (1) This application effectively overcomes the problems of score fluctuation and business misjudgment caused by frequent data updates in traditional SaaS CRM systems by constructing a score generation path driven by lead status. Existing technologies generally adopt a single-path architecture of "immediate weight recalculation - instant score output", which is prone to causing drastic score fluctuations when faced with continuous injection of customer behavior flow, especially causing repeated adjustments to sales strategies and resource misallocation in key conversion stages. This solution innovatively introduces a lead status anchoring unit, which automatically matches the corresponding stability threshold and modal sensitivity spectrum according to the business stage of the lead, and realizes differentiated modeling response for leads in different life cycle stages; combined with the three-channel collaborative control of historical smoothing, mutation temporary storage and state transition calibration in the weight buffer mechanism, it significantly improves the time consistency and logical coherence of the scoring process, greatly reduces the risk of score drift caused by non-essential feature disturbances, and enables the system to output stable and reliable scoring results in dynamic data environment.
[0018] (2) By designing a feedback closed-loop calibration layer, this scheme achieves synergistic optimization of scoring accuracy and stability. The system collects samples with manual annotations and scoring deviations exceeding a threshold to construct a "stability-effectiveness" joint feedback set, which is used to periodically fine-tune buffer parameters and suppression rules, rather than retraining the entire model. This greatly reduces operation and maintenance costs and enhances the maintainability of the system. This mechanism enables the scoring system to have adaptive evolution capabilities, allowing it to continuously approach real business judgment standards without interrupting service.
[0019] (3) This solution fully supports the industrial-grade deployment requirements of high concurrency and low latency, and the output includes a "stability confidence" meta tag, providing interpretability support for sales decisions. All modules are designed based on an event flow architecture, achieving millisecond-level state recognition and hundreds of millisecond-level scoring response, meeting the requirements of large-scale real-time lead processing in SaaS scenarios; at the same time, it transforms the complex stability control problem into state management and engineering buffer control tasks based on business semantics, getting rid of the dependence on complex algorithms such as comparative learning and adversarial training, and avoiding widely covered technical paths such as temperature scaling and gating adjustment, forming a technical closed loop with a clear action chain and data interface. This design not only improves the system's computational efficiency and deployment flexibility, but also significantly enhances user trust due to its strong interpretability and low parameter tuning threshold. It is particularly suitable for CRM sales conversion scenarios that require high-frequency human-machine collaboration, effectively ensuring the reliable implementation and long-term stable operation of the intelligent scoring system in real business environments.
[0020] In summary, this solution starts from the business essence of sales leads, transforming the abstract challenge of scoring stability into a modelable, controllable, and feedback-based state management problem. Through multi-level collaboration of state anchoring, buffer control, and closed-loop calibration, it achieves a systematic improvement in the time continuity, logical rationality, and business relevance of scoring results, demonstrating engineering practicality and unpredictable technological innovation that surpasses existing technologies. Attached Figure Description
[0021] Figure 1 This is the main flowchart of a sales lead mining method that integrates multi-modal data.
[0022] Figure 2 This is a sub-flowchart of a sales lead mining method that integrates multi-modal data.
[0023] Figure 3 This is another sub-flowchart of the sales lead mining method that integrates multi-modal data. Detailed Implementation
[0024] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0025] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0026] like Figure 1 As shown, this application provides a sales lead mining method that integrates multi-modal data, specifically including:
[0027] S1: Acquire multimodal sales lead data streams and identify the lifecycle status of each lead, generating lead status tags with timestamps.
[0028] S2: Based on the clue state label, query the preset state sensitivity spectrum library, extract the state stability threshold and modal sensitivity spectrum bound to the state label, and generate a state constraint parameter set.
[0029] S3: Based on the multimodal sales lead feature vector at the current moment and the state constraint parameter set, calculate the weighted moving average of the most recent effective weights to generate a basic smooth weight vector.
[0030] S4: Perform mutation detection based on the instantaneous difference between the basic smooth weight vector and the current multimodal sales lead feature vector. If the single-modal increase exceeds the preset mutation buffer threshold, the instantaneous change is temporarily stored and released with a delay, generating a buffer weight vector that has been suppressed by mutation.
[0031] S5: Monitor whether the clue state label has undergone a cross-stage transition. If a state transition is determined to have occurred, trigger a progressive interpolation update operation to correct the buffer weight vector to a state weight vector.
[0032] S6: Generate basic cue scores based on the state weight vector, and simultaneously capture the amplitude and rhythm characteristics of each modal residual signal to generate fluctuation suppression coefficients.
[0033] S7: Perform nonlinear decay or enhancement adjustment operation on the basic clue score based on the fluctuation suppression coefficient, and output the final clue score with stability confidence meta-label.
[0034] S8: Collect the manually labeled clue conversion results and compare them with the final clue score. If the absolute value of the deviation exceeds the preset joint feedback threshold, construct a stability and effectiveness joint feedback set.
[0035] Step S1: Acquire multimodal sales lead data streams and identify the lifecycle status of each lead, generating a lead status tag with a timestamp. Specifically, this includes:
[0036] S1.1: Perform protocol parsing and field standardization on multi-source heterogeneous raw data streams from CRM systems, social media platforms, and website access logs to eliminate data format differences and generate a unified multimodal sales lead feature vector sequence.
[0037] Perform multi-protocol parsing processing on multi-source heterogeneous raw data streams from CRM systems, social media platforms, and website access logs. Use the corresponding transport protocol decoder for different data sources to parse the HTTP, WebSocket, or custom API return structures into a recognizable set of raw fields.
[0038] Based on the original set of fields obtained from parsing, a field mapping table is constructed. Fields with the same logical meaning but different naming rules are mapped to normalized field names through a unified identifier, and redundant or conflicting fields are eliminated according to a preset field priority strategy.
[0039] Type calibration and unit unification are performed on the field values corresponding to the normalized field names. Text-type time data is converted into millisecond-level timestamps, currency amounts are adjusted according to a unified currency benchmark, and Boolean status values are mapped to binary numerical identifiers to achieve data type consistency.
[0040] Normalization is performed on numeric fields, and the standard value is calculated using the minimum-maximum normalization method.
[0041] Normalized numerical fields, uniformly coded categorical fields, and standardized time fields are concatenated in a preset modal order. A unified multimodal sales lead feature vector sequence is generated using a fixed-length padding and missing value placeholder strategy. A unique lead identifier is added to each vector to support subsequent status recognition processing.
[0042] By using multi-protocol parsing, field mapping, type calibration, and normalization, the heterogeneous raw data stream from the previous step is transformed into a multimodal sales lead feature vector sequence that is structurally consistent, numerically stable, and semantically aligned, thus providing a high-precision input foundation for subsequent behavioral feature extraction and lifecycle status determination.
[0043] S1.2: Extract key behavioral event indicators based on the multimodal sales lead feature vector sequence, and use the sliding time window algorithm to calculate the behavioral frequency density and interaction depth entropy value to construct a dynamic behavioral feature map that reflects the activity of leads.
[0044] S1.3: Based on the dynamic behavior feature map, match the preset business stage division rule base, execute multi-dimensional threshold judgment logic and state transition detection operation to determine the specific business life cycle stage of each clue and generate preliminary clue state classification results.
[0045] The business stage segmentation rule base is a predefined expert knowledge base used to automatically determine the business lifecycle stage (e.g., "first contact," "initial interaction," "multi-channel interaction," "confirmation of intent," "closing negotiation," etc.) of sales leads based on their multi-dimensional behavioral characteristics (e.g., behavior frequency density, interaction depth entropy, consecutive active days). This rule base is stored as stage description vectors and threshold conditions, providing stage matching criteria for step S1.3. It maps continuous features to discrete stage labels through Euclidean distance calculation and threshold determination, thereby generating timestamped lead status labels. The rule base is built offline based on historical lead conversion data and domain expert experience and can be regularly updated offline to adapt to business changes.
[0046] The business phase segmentation rule base adopts a relational table structure, with each record corresponding to a business phase and containing the following fields:
[0047] Stage identifier: A unique code that identifies a business stage.
[0048] Phase Name: A business-readable phase description (e.g., "Initial Contact", "Multi-channel Interaction").
[0049] Stage description vector: A fixed-length numerical vector, each component of which corresponds to a key behavioral feature (such as behavioral frequency density, interaction depth entropy value, number of consecutive active days, etc.), representing the expected value or central value of these features under this stage.
[0050] Feature matching threshold: A threshold vector of the same length as the stage description vector, where each component represents the allowable deviation range for that dimension (e.g., the allowable deviation for frequency density is ±2 times / day), used to determine whether the current feature belongs to that stage.
[0051] Transition Conditions: Optional field, defines additional conditions required to transition from the previous stage to the current stage (such as the percentage increase in consecutive active days), used for state transition detection.
[0052] Default weight matrix: The basic weight allocation of each modality feature in this stage (which can be further adjusted in subsequent steps).
[0053] Construction method:
[0054] (a) Offline initialization phase:
[0055] Feature definition: Based on business analysis, identify key behavioral features for differentiating stages (such as behavior frequency density, interaction depth entropy, number of consecutive active days, email reply rate, etc.), and perform normalization preprocessing on each feature.
[0056] Phase segmentation and center point identification: Extract the mean feature value of each lead at each phase from the historically labeled lead conversion data, and use it as the descriptive vector for that phase. If there is no historical labeled data, business experts can be invited to set typical feature values for each phase based on their experience.
[0057] Threshold setting: Calculate the standard deviation of each feature within each stage, and take 1.5 times the standard deviation as the initial matching threshold; it can also be set directly by experts. For certain key features (such as the number of consecutive active days), an absolute lower limit or upper limit can be set (e.g., no less than 10 days).
[0058] Transition condition configuration: Based on business logic, define additional constraints that must be met to transition from the previous stage to the current stage (e.g., the "multi-channel interaction" stage requires the "initial interaction" stage to last at least 3 days).
[0059] Rule base storage: The above data is stored in the system configuration center in JSON or database table format, and a query interface is provided.
[0060] (II) Online Inference Stage: After inputting the dynamic behavioral feature map of the current clue, the rule base calculates the Euclidean distance between the input feature vector and the description vector of each stage. If the distance of all features in a certain stage is less than the matching threshold corresponding to that stage, then that stage is marked as a candidate. If multiple stages meet the criteria, the final stage is selected based on the minimum distance sum or business priority. State transition detection is performed by combining the historical state sequence and transition conditions (e.g., by comparing the transition coefficient with the stability coefficient), and the final state label is output.
[0061] (III) Rule Base Maintenance and Updates: Regularly (e.g., monthly) recalculate the description vectors and thresholds for each stage using newly accumulated labeled lead data, and update the rule base. When a new stage is introduced or the stage division is adjusted, rule base entries can be manually modified through the management interface.
[0062] Through the aforementioned business stage segmentation rule base, this invention achieves automatic mapping from continuous behavioral features to discrete business stage labels, providing a stable reference benchmark for subsequent state-related weight adjustment and score smoothing.
[0063] Based on the dynamic behavior feature map generated in the previous sub-step, the multi-dimensional feature indicators of this map are used as the judgment criteria and imported into the business stage segmentation rule matching module to achieve the corresponding mapping retrieval between features and the rule base. The Euclidean distance is calculated on the input features using the pre-set stage description vectors in the rule base to extract the numerical distance with each business stage description vector, forming a stage matching distance matrix. This stage matching distance matrix is input into the multi-dimensional threshold judgment logic, and stage-by-stage judgment is performed according to preset threshold boundary conditions. When the distances of all key features in a certain stage are less than the matching threshold of the corresponding stage, a judgment flag for that stage is generated. A state transition detection operation is performed on any candidate stage judgment flag. Based on the difference vector between the judgment results of two adjacent stages in the time series, the state change rate is calculated and the transition coefficient is generated using the following formula:
[0064]
[0065] Where D(t) is the magnitude of the distance vector determined in the current stage, D(t−1) is the magnitude of the distance vector determined in the previous stage, and K... trans This represents the state transition coefficient. The transition coefficient is compared with a preset state stability coefficient. If the transition trigger condition is met, the state change flag is recorded, and the original stage judgment result is retained as the preliminary state classification output. Through the above processing method, the dynamic behavioral feature map from the previous step is transformed into preliminary state classification data reflecting the current business lifecycle stage of the clue, achieving a mapping and recognition effect between multimodal behavioral features and stage rules.
[0066] For example, in the sales lead analysis scenario of a SaaS CRM system, the input dynamic behavioral feature map includes a behavior frequency density of 12.5 times / day, an interaction depth entropy value of 2.85, and 14 consecutive active days. The matching threshold for the "multi-channel interaction" stage in the business stage rule base is set to a frequency density of no less than 10 times / day, an interaction depth entropy value in the range of 2.5 to 3.5, and no less than 10 consecutive active days. The Euclidean distance between the input features and the rule thresholds is calculated, resulting in a matching distance matrix feature vector modulus D(t) of 0.32 and a historical value D(t−1) of 0.45. These values are then substituted into the formula to calculate the state transition coefficient.
[0067]
[0068] Get K trans The value is approximately 0.711, which is greater than the preset stability coefficient of 0.65. Therefore, a state transition is determined, and a change marker is added. The classification result indicates that the clue is in the "multi-channel interaction" stage, and preliminary state classification data is output. This output data will serve as a stability anchor in the subsequent timestamp embedding step, thereby significantly improving the scoring stability during dynamic weight adjustment.
[0069] S1.4: Associate the preliminary clue status classification result with a high-precision system clock signal, and perform timestamp embedding and timing alignment operations to generate clue status labels with precise timestamps as stable anchor points for subsequent weight allocation.
[0070] Step S2: Based on the cue state label, query a pre-set state sensitivity spectrum library, extract the state stability threshold and modal sensitivity spectrum bound to the state label, and generate a state constraint parameter set. Specifically, this includes:
[0071] S2.1: Perform hash index calculation on the input timestamped clue status tags to locate the corresponding status record entry in the preset status sensitivity spectrum library and generate a target status record object containing the original configuration data.
[0072] The State Sensitivity Spectrum Library is a pre-configured parameter database used to store the modal sensitivity spectrum and state stability threshold of sales leads at different stages of the business lifecycle. The library is indexed by lead state tags, with each state corresponding to a set of quantitative indicators: the modal sensitivity spectrum (i.e., the sensitivity coefficient of each data modality, such as CRM data, social media behavior, website access logs, etc., to score fluctuations) and the state stability threshold (i.e., the maximum allowable score drift boundary for that stage).
[0073] The state sensitivity genealogy library uses a key-value pair storage structure, with each record corresponding to a business stage and containing the following fields:
[0074] Status label: A unique code or name that identifies the business stage (such as "first contact", "multi-channel interaction", "confirmation of intent", etc.), which is consistent with the lead status label generated by S1.
[0075] Modal sensitivity spectrum: a vector whose length is equal to the total number of modalities involved in the scoring (such as CRM data, social media interactions, website access logs, etc.). Each component represents the sensitivity coefficient of the corresponding modality in the current state (the value range is usually 0 to 1, and the larger the value, the stronger the impact of the fluctuation of the modality on the final score).
[0076] High-frequency fluctuation mode identifier list: A list of modes with sensitivity coefficients higher than a preset threshold (e.g., 0.7) selected from the modal sensitivity spectrum, used for key analysis during S2.3 adaptive threshold derivation.
[0077] State stability threshold: a scalar representing the maximum score drift boundary allowed in the current business phase (e.g., 0.12 points), used for setting the length of the dynamic sliding window and the upper limit of tolerance for mutation detection.
[0078] Historical transformation variance reference value: Stores the transformation rate variance of each modality in historical data during this stage, for use in adaptive threshold extrapolation.
[0079] Construction method:
[0080] (a) Offline initialization phase:
[0081] Status definition: Determine the set of status labels based on the list of stages defined in the business stage division rule base.
[0082] Modal sensitivity coefficient calibration: Business experts were invited to score the actual impact of each modality on the stage conversion, and the scores were normalized to serve as the sensitivity coefficients; From historical lead conversion data, the correlation between the fluctuation of each modal feature at different stages and the change in conversion rate was analyzed (such as Pearson correlation coefficient), and the absolute value of the correlation coefficient was taken as the sensitivity coefficient.
[0083] High-frequency fluctuation mode identification: For each state, the mode index with a sensitivity coefficient greater than a preset threshold (e.g., 0.7) is stored in the high-frequency fluctuation mode identification list.
[0084] State stability threshold setting: Based on the distribution of score fluctuations in historical data during this period, take the 95th percentile as the initial threshold; or set it directly based on expert experience (e.g., 0.1 to 0.2 points).
[0085] Historical transformation variance calculation: For each mode in each state, calculate the variance of the historical transformation rate time series and store it in the optional database.
[0086] (II) Online Inquiry and Dynamic Fine-tuning:
[0087] Query interface: Based on the hash index of the clue status label, quickly locate the corresponding status record and extract modality sensitivity spectrum, state stability threshold, etc.
[0088] Periodic fine-tuning: Based on the feedback set (the deviation between the manually labeled transformation results and the final score), the modal sensitivity coefficient and state stability threshold are incrementally updated using gradient descent or regression methods, and written back to the genealogy library.
[0089] (III) Maintenance and Version Management: The genealogy database supports version control, with the version number incremented after each batch update to ensure traceability. When business phases change (addition, deletion, or merging phases), the structure of the genealogy database must be updated synchronously.
[0090] Through the aforementioned state sensitivity spectrum library, this invention achieves decoupled storage and fast retrieval of business stages and modal sensitivity and stability thresholds, enabling weight adjustment and scoring smoothing to adapt to the characteristics of different stages, thereby improving the system's configurability and generalization ability.
[0091] S2.2: Based on the target state record object, analyze the original data of the modal sensitivity spectrum, and use normalization mapping to convert the sensitivity index of each data mode into a standard weight redistribution vector to generate a standardized modal sensitivity spectrum.
[0092] S2.3: Based on the high-frequency fluctuation mode identifier in the standardized modal sensitivity spectrum, and combined with historical business conversion variance data, perform adaptive threshold deduction calculation to determine the maximum allowable score drift boundary for the current business stage and generate a state stability threshold.
[0093] The high-frequency wave modes identified in the input standardized modal sensitivity spectrum are analyzed for wave intensity. During the analysis, the sensitivity coefficients and historical wave frequencies of each high-frequency wave mode are extracted as the calculation benchmark.
[0094] The above sensitivity coefficients are compared with the time series data of conversion rates retrieved from historical business conversion records to calculate the variance contribution of each high-frequency modality to the score fluctuation. Standardization is then used to eliminate the differences in measurement scales between different modalities.
[0095] The variance contribution values of each high-frequency mode are weighted and accumulated according to the sensitivity coefficient to obtain the comprehensive fluctuation amplitude index of the business stage, and a nonlinear mapping function parameter set between the fluctuation amplitude and the scoring drift boundary is established.
[0096] By using adaptive threshold extrapolation calculation, the comprehensive fluctuation amplitude index is input into the mapping function, and the slope and intercept of the function curve are dynamically updated using historical stage score drift statistics, thereby obtaining the maximum score drift boundary value of the current stage.
[0097] The maximum score drift boundary is calculated using the following formula:
[0098]
[0099] in, , , The variance of modal history transformation, where m is the total number of high-frequency modes. This represents the maximum score drift boundary value.
[0100] The calculation results are output as the state stability threshold, providing drift tolerance boundary parameters for subsequent dynamic sliding window calculations of the weight buffer.
[0101] Through the above processing method, the standardized modal sensitivity spectrum of the previous step is transformed into a quantifiable state stability threshold, realizing the adaptive setting of the score drift boundary based on historical fluctuations and sensitivity characteristics.
[0102] For example, in an embodiment targeting an intelligent sales assistance SaaS platform, high-frequency fluctuation modal identifiers for CRM modality, social media modality, and website access modality are obtained from a standardized modal sensitivity spectrum, corresponding to sensitivity coefficients of 0.85, 0.65, and 0.92, respectively. Historical two-year conversion result time-series data are retrieved, and the historical conversion variances for the above modalities are calculated to be 2.1, 1.7, and 2.8, respectively. Multiplying the sensitivity coefficients by the variances yields weighted contribution values: 1.785 for CRM modality, 1.105 for social media modality, and 2.576 for website access modality. Summing these three values and dividing by the number of high-frequency modalities (3), the maximum score drift boundary value is calculated using a formula to be 1.822, which is set as the state stability threshold. At this threshold, the subsequent weight buffer effectively suppresses instantaneous score drift caused by sudden increases in website access modality during historical weight smoothing calculations. The system score exhibits significantly improved stability and robustness during cross-stage state changes, and the credibility of the business conversion evaluation results is greatly enhanced.
[0103] S2.4: The standardized modal sensitivity spectrum and the state stability threshold are encapsulated in a structured manner to construct a composite data structure containing modal weight constraint rules and fluctuation tolerance boundaries, thereby generating the final set of state constraint parameters.
[0104] A hierarchical structure mapping is performed on the input standardized modal sensitivity spectrum and state stability threshold, associating the original weight constraint information and fluctuation tolerance boundary to independent structure nodes respectively.
[0105] Each modality sensitivity component is arranged into a fixed-order index vector, and a one-to-one correspondence is established between the index label and the modality identifier. The weight constraint rules are then encapsulated in a key-value pair manner.
[0106] The state stability threshold is bound together with a threshold type identifier as a single numerical parameter to form a tolerance boundary node that can be directly invoked.
[0107] Input the weight constraint rule nodes and tolerance boundary nodes into the composite data structure builder and perform binary tree nested encapsulation so that the modal weight constraints and fluctuation tolerance parameters can be automatically routed according to the node type when accessed.
[0108] The encapsulated composite data structure is output as a set of state constraint parameters, ensuring that modal sensitivity and stability thresholds can be directly read in subsequent weight buffer calculations, thereby achieving precise weight control based on state labels.
[0109] By using a structured encapsulation process, the standardized modal sensitivity spectrum and state stability threshold from the previous step are transformed into a composite data structure with fast retrieval and parallel access capabilities, thereby realizing the formal output of the state constraint parameter set.
[0110] For example, in a SaaS CRM environment, the input standardized modal sensitivity spectrum includes CRM data sensitivity of 0.75, social media data sensitivity of 0.40, and website access data sensitivity of 0.85, with a state stability threshold set to 0.12. The sensitivity vector [0.75, 0.40, 0.85] is arranged sequentially by modal index, with index 0 corresponding to CRM, index 1 to social media, and index 2 to website access, and encapsulated as a weight constraint rule node. The stability threshold node is set to threshold = 0.12, and type identifier = "maximum drift boundary". The composite data structure builder nests the weight constraint rule node and the tolerance boundary node in a binary fashion, generating the structure fields: weight rule = {CRM: 0.75, Social: 0.40, Web: 0.85}, tolerance boundary = {threshold: 0.12, type: "maximum drift boundary"}. In subsequent calls to the weight buffer, querying the state constraint parameter set simultaneously obtains the three-modal weight constraints and threshold boundaries, ensuring that the weights of each modality do not exceed the set fluctuation tolerance range in the historical weight smoothing calculation. Performance verification shows that the access latency of this encapsulation structure is significantly lower than that of the ordinary list structure, and the score fluctuation suppression effect is stable with no long-term lag.
[0111] like Figure 2 As shown, step S3 involves: based on the multimodal sales lead feature vector at the current moment and the state constraint parameter set, calculating the weighted moving average of the most recent effective weights to generate a basic smoothed weight vector. Specifically, this includes:
[0112] S3.1: Based on the modal sensitivity spectrum in the state constraint parameter set, perform dimension alignment and validity verification on the multimodal sales lead feature vector at the current moment to remove missing or abnormal modal data and generate a standardized multimodal feature sequence to ensure the integrity and consistency of the input data.
[0113] The feature vector of the multimodal sales lead at the current moment is obtained by parsing the protocol and standardizing the fields of the multimodal sales lead data stream at the current moment to eliminate data format differences.
[0114] S3.2: Query the historical weight repository using the standardized multimodal feature sequence, extract the five most recent valid weight records bound to the clue identifier, and perform a time decay factor weighting operation on the five most recent valid weight records to generate a historical weight candidate set with time priority.
[0115] The historical weight repository is a database used to persistently store the weight vectors calculated for each sales lead at different historical moments. Indexed by the lead's unique identifier and status tag, this repository records the most recent (e.g., the last five) valid weight records and their corresponding timestamps, providing a historical data source for the time decay factor weighting calculation in step S3.2. By querying this repository, the system can obtain the recent weight evolution trajectory of the same lead, thereby calculating a weighted moving average with time-series priority to generate a smoother underlying weight vector. This repository supports incremental writing and on-demand retrieval, is invoked in step S3.2, and updated in step S3.5.
[0116] The historical weight repository uses a relational or key-value storage structure. Each record corresponds to one weight calculation event for a sales lead and includes the following fields:
[0117] Clue unique identifier: The clue ID is bound to the multimodal feature vector generated in S1.1.
[0118] Timestamp: The system time (millisecond precision) at the moment the weights were generated.
[0119] Status label: The current stage of the business lifecycle of the lead (e.g., "first contact", "multi-channel interaction").
[0120] Weight vector: A floating-point vector with a length equal to the total number of modalities involved in the scoring. Each component represents the weight coefficient of the corresponding modality at that moment (the value range is usually 0 to 1).
[0121] Validity flag: Indicates whether the weighted record is a valid record (e.g., excluding invalid weights caused by abnormal data), for filtering during S3.2 retrieval.
[0122] In addition, the repository can maintain an index for each thread in descending order of timestamps to quickly retrieve the most recent N records.
[0123] Construction method:
[0124] (a) Initialization: When the system starts for the first time, a blank repository is created and a hash index is built according to the clue ID.
[0125] (ii) Write operation: After each weight update (e.g., after generating the basic smooth weight vector in S3.5, or after generating the state weight vector in S5.5), the system calls the write interface to store the current clue ID, timestamp, status label, and weight vector into the storage repository, and marks the validity flag as "valid". To control storage capacity, each clue can be set to retain a maximum of 20 most recent records, and when this limit is exceeded, the oldest record will be discarded based on the timestamp.
[0126] (III) Search Operations:
[0127] Input: Clue unique identifier.
[0128] Output: The N most recent valid weight records for this clue (usually N=5), sorted in descending order of timestamp. Each record contains a weight vector and a timestamp.
[0129] Implementation: Find the corresponding record list based on the clue ID, sort by timestamp, take the first N records, and filter out records marked as "invalid".
[0130] (iv) Maintenance and cleanup: Regularly clean up old records that have exceeded the retention period (e.g., 90 days) to reduce storage usage. When a lead is marked as "lost" or "converted", the relevant records can be archived or retained but not used for online queries.
[0131] Through the aforementioned historical weight repository, this invention achieves a time-series memory function for the evolution of sales lead weights, providing reliable historical data support for weighted moving average calculation, thereby improving the smoothness of weights and the stability of scoring.
[0132] Based on the standardized multimodal feature sequence, a unique matching retrieval of the clue identifier is performed, and the index position of the weight record corresponding to the clue is located in the historical weight repository to ensure that the retrieval results are consistent with the business relevance of the currently input feature sequence.
[0133] After locating the index position, the weight record extraction module is called to read the five most recent valid weight records bound to the clue. The weight vector of each record is fully parsed, split into a set of computable weight values according to modal components, and the timestamp information of each record is retained for subsequent time-series weighting.
[0134] Arrange the five most recent valid weight records in timestamp order, convert the record interval into a time decay factor, and calculate the decay coefficient corresponding to each record according to the preset time-sensitive mapping rule. The formula for calculating the decay coefficient is as follows:
[0135]
[0136] Where, α i Let λ be the decay coefficient of the i-th weight record. i Let Δt be the time decay rate corresponding to the i-th weight record. i Let be the time interval between the i-th record and the current time.
[0137] The attenuation coefficient is multiplied by the corresponding weight vector component to generate a weight component matrix with attenuation correction, and then aggregated according to the modal dimension to form a weight candidate set containing temporal priority information.
[0138] A unified structured data interface is constructed for the aggregated weight candidate set, including modality identifiers, decayed weight values, and original record time information, for use by the dynamic sliding window calculation model in the next step.
[0139] By using time decay factor weighting and time priority injection, the standardized multimodal feature data from the previous step is transformed into a set of candidate weights with historical continuity and time sensitivity, thus achieving a stable response to changes in modality weights in the subsequent smooth fusion stage.
[0140] For example, for a three-modal lead consisting of CRM, social media, and website logs, the input standardized feature sequence includes three dimensions: CTR, interaction frequency, and page dwell time. The five most recent valid weight records for this lead are retrieved from the historical weight repository, with timestamps of 2 hours, 5 hours, 8 hours, 15 hours, and 24 hours prior to the current time. Preset time decay rates λ are 0.05, 0.08, and 0.02, respectively. After calculating the time intervals Δt to 2, 5, 8, 15, and 24, the decay coefficient is calculated using a formula. For example, for the second record:
[0141]
[0142] The attenuation coefficient values are obtained, corresponding to the original weight vector of [0.35, 0.40, 0.25]. Each component is multiplied by the attenuation coefficient to generate a corrected vector of [0.27, 0.31, 0.19]. This process is repeated for five records, and the results are aggregated by modality to form a historical weight candidate set. In performance verification, after this set is input into the smooth window model, the short-term fluctuations of the basic smoothed weight vector are significantly reduced, and the sensitivity response latency is controlled within an acceptable range for business applications.
[0143] S3.3: Based on the historical weight candidate set and the state stability threshold in the state constraint parameter set, a dynamic sliding window calculation model is constructed, and normalized sliding mean fusion processing is performed on each modal weight component in the historical weight candidate set to generate a preliminary smooth weight matrix.
[0144] The historical weight candidate set with temporal priority and the state stability threshold in the state constraint parameter set are input together into the dynamic sliding window model initialization logic. The sliding window length and window step size are set according to the threshold value, and a window structure index matrix corresponding to each modality is generated.
[0145] Time series normalization is performed on each modal weight component in the historical weight candidate set. Linear normalization based on the range of maximum and minimum values is used to make the different modal weight components have uniform dimensions and relative proportions. The normalized weight components are then mapped to the corresponding positions in the window index matrix.
[0146] Within each window, a moving average fusion operation is performed, which sums the normalized weight components in the window according to the time priority coefficients, and simultaneously performs cumulative normalization on the weighted coefficient sequence to form a smoothed mean result vector.
[0147] Arrange the smoothed mean result vectors of all windows in modal order and form a matrix structure. Use matrix normalization to ensure that the smoothed mean distribution between different modes meets the fluctuation tolerance boundary set by the state stability threshold.
[0148] Boundary condition checks are performed on the values in the smoothing matrix to detect weight components that exceed the state stability threshold and correct them according to the maximum allowable deviation of the threshold. The output is a preliminary smoothing weight matrix that satisfies the stability constraints.
[0149] By using the dynamic sliding window and normalization fusion processing method described above, the historical weight candidate set from the previous step is transformed into matrix-smoothed weight data that can balance sensitivity and stability, thereby achieving the expected technical effect of suppressing short-term fluctuations.
[0150] S3.4: Based on the real-time distribution characteristics of the standardized multimodal feature sequence, perform gain compensation adjustment on the preliminary smoothing weight matrix to correct the sensitivity lag problem caused by long-period smoothing and generate a corrected smoothing weight vector.
[0151] S3.5: Encapsulate the modified smoothing weight vector into a standard data interface format and output the basic smoothing weight vector.
[0152] The weight buffer is the core data processing module in this invention, used to smooth the modal weights of sales leads, suppress instantaneous fluctuations, and achieve smooth transitions between states. Located between weight calculation and score generation, this module receives the basic smoothed weight vector from the historical smoothing channel, identifies abnormal fluctuations through a mutation detection unit, temporarily stores and delays the release of mutation components, and supports progressive weight updates during state transitions. Finally, it outputs a buffered weight vector with mutation suppression, providing stable and continuous weight input for the subsequent main scoring path. This module undertakes the core function of three-channel coordinated control—historical smoothing, mutation storage, and state transition calibration—in steps S3 to S5. Furthermore, the module provides a standardized data interface protocol to ensure unified data exchange format, time alignment, and reliable transmission with pre- and post-processing units (such as the historical weight repository, state constraint parameter set, and main scoring path).
[0153] The weight buffer consists of the following sub-modules:
[0154] Historical Smoothing Channel: Receives a set of historical weight candidates from S3, calculates a weighted moving average through a dynamic sliding window, and generates a basic smoothed weight vector. This channel suppresses short-term random fluctuations while preserving long-term trends.
[0155] The mutation detection module monitors the modal-by-modal difference between the current multimodal feature vector and the basic smoothing weight vector in real time, calculates the instantaneous increase ratio, and compares it with a preset mutation buffer threshold. When the increase of any modality exceeds the threshold, it is marked as an abnormal mode, and its instantaneous change is extracted. This module outputs a mutation event label set and the over-amplitude value.
[0156] Delayed buffer: Employs a first-in, first-out (FIFO) queue caching mechanism to temporarily store instantaneous changes in abnormal modes, with a release countdown set for each temporary item. The countdown is calculated based on the state stability threshold and the modality sensitivity coefficient. Modal changes not marked as abnormal are directly sent to the instant change buffer.
[0157] Gradient release unit: Monitors the release countdown in the delay buffer. When the countdown ends, it decomposes the temporarily stored over-amplitude change into multiple small increment steps through linear interpolation smoothing, generates a progressive compensation weight sequence, and gradually injects it into the weight vector in subsequent time slices.
[0158] Fusion Update Unit: The instantaneous change quantity is vector-superimposed with the progressively compensated weight sequence, and a weighted update is performed on the basic smooth weight vector to output a buffered weight vector that has undergone mutation suppression.
[0159] State transition interface: Receives cross-state transition control parameters from S5, triggers progressive interpolation update during state transition, and smoothly adjusts the buffer weight vector to the state weight vector corresponding to the target state.
[0160] Data Interface Encapsulation Unit: Responsible for standardizing, encapsulating, and parsing the input and output data of the weight buffer according to a preset data interface protocol. This unit ensures that:
[0161] Input data (basic smooth weight vector, multimodal feature vector, state constraint parameter set, etc.) must be accompanied by metadata tags (modal identifier, state tag, timestamp, etc.) and adopt a unified byte alignment and precision specification (e.g., double precision for floating-point numbers and little-endian encoding for integers).
[0162] Output data (buffered weight vector, state weight vector, etc.) is encapsulated into standardized messages, including a length prefix, checksum (such as CRC16) and protocol version number, to ensure the integrity and resolvability of cross-module transmission.
[0163] Provides a data validity verification interface to intercept or complete data with missing fields, incorrect data types, or outdated timestamps.
[0164] Construction method:
[0165] (a) Offline parameter calibration:
[0166] Mutation buffer threshold: The initial threshold is set based on the 95th percentile of normal fluctuations of each modality in historical data, or it can be configured by experts according to business sensitivity.
[0167] Release countdown step function: Define the mapping relationship based on the state stability threshold θ and the modal sensitivity coefficient μ.
[0168] Sliding window length: Set according to the business cycle (e.g., one week) and sampling frequency, usually between 5 and 10 times.
[0169] Data interface protocol parameters: define the maximum message length, verification algorithm (such as CRC-16 / CCITT), byte order (little endian), floating-point precision (double precision), etc.
[0170] (II) Online Operation Process
[0171] The following process is executed for each scoring cycle:
[0172] Historical smoothing: Call S3.2-S3.4 to generate the basic smoothing weight vector.
[0173] Mutation detection: Calculate the instantaneous increase rate modally and generate a set of mutation event markers.
[0174] Temporary processing: Abnormal modal changes are enqueued into a delayed buffer and a release countdown is set; normal modal changes are entered into an immediate buffer.
[0175] Gradient release: Scan the delay buffer, perform linear interpolation decomposition on the mutations that have expired on countdown, and output an asymptotically compensated weight sequence.
[0176] Fusion Update: The instantaneous changes and the progressively compensated weight sequence are superimposed on the base smooth weight vector to output a buffered weight vector.
[0177] State transition processing: If a state transition occurs, the transition matrix parameters of S5 are received, and the buffer weight vector is progressively interpolated to update it to the target state weight vector.
[0178] Data encapsulation: The output weight vector is encapsulated into a standard message according to the data interface protocol, with additional metadata such as timestamp and checksum, for use by downstream modules.
[0179] (III) Memory and Performance Optimization:
[0180] The delay buffer is implemented using a circular queue with a fixed capacity (e.g., 10 units) to avoid unlimited memory growth.
[0181] Gradient release operations use a pre-allocated incremental step cache to reduce dynamic memory allocation.
[0182] The data interface protocol uses zero-copy technology to reduce serialization / deserialization overhead.
[0183] To ensure interoperability between the weight buffer and preceding and following modules (such as historical weight storage, state constraint parameter set, main scoring path, etc.), the following data interface protocol is defined:
[0184] Request message format (input):
[0185] Fixed header (8 bytes): Protocol version number (2 bytes), message type (2 bytes), message length (4 bytes);
[0186] Variable load: Includes fields such as basic smoothing weight vector (floating-point array), multimodal feature vector (floating-point array), state constraint parameter set (structured key-value pairs), timestamp (millisecond-level integer), and clue identifier;
[0187] Tail check (2 or 4 bytes): CRC16 / CRC32 checksum.
[0188] Response message format (output):
[0189] Similar request messages have a payload of a buffer weight vector or a state weight vector, with an additional processing status code (0 indicates success, and non-zero indicates an error).
[0190] Field encoding rules:
[0191] Floating-point numbers are uniformly set to IEEE 754 double precision (8 bytes);
[0192] Integers are in little-endian byte order;
[0193] The string is encoded in UTF-8 and prefixed with a length field.
[0194] Timing alignment requirement: The generation timestamp of all input messages should be synchronized with the processing cycle of the weight buffer. If the deviation exceeds a preset threshold (e.g., 50ms), retransmission or dropping should be triggered.
[0195] Error handling: When an illegal message is parsed, the data interface encapsulation unit returns an error code and logs it, and does not participate in the weight calculation.
[0196] Through the aforementioned weight buffer and its standardized data interface protocol, this invention achieves loosely coupled integration between weight calculation modules, ensuring the reliability of data transmission, timing consistency, and system scalability.
[0197] The corrected smoothed weight vector is used as the input object, and data structure reorganization is performed according to the data interface protocol definition of the weight buffer to ensure that each modal weight component has a unified data type, fixed field order and precise decimal control in the interface message.
[0198] Field label embedding logic is applied to the recombined weighted dataset to encapsulate modal identifiers, state labels, and timestamp signals into multi-level metadata labels, and binds them to the corresponding weight components using index vectors, so as to ensure that the subsequent mutation detection module can quickly retrieve and access them according to modal and state dimensions.
[0199] Based on the preset sequence encoding rules, the encapsulated weighted data is processed into binary serialization, and a length prefix and check bit embedding mechanism is adopted to achieve high-speed decoding and transmission integrity verification.
[0200] The serialized weight data is compared with the interface protocol standard field by field through the interface verification logic to generate a structure verification report. When the report passes, the data is sent to the weight buffer output queue.
[0201] The standardized weight data is released synchronously at the time slice boundary by using the output queue scheduler, which ensures the strict temporal alignment between the basic smooth weight vector received by the mutation detection module and the multimodal sales lead feature vector.
[0202] Through the above encapsulation and verification process, the corrected smooth weight vector from the previous step is transformed into a basic smooth weight vector with a unified structure, complete metadata labels, and time-series alignment characteristics, enabling efficient data access and lossless transmission when the mutation detection module is called.
[0203] For example, when executing this sub-step for the sales lead weight buffer of a SaaS CRM system, the corrected smoothed weight vector consists of 8 modal components, corresponding to features such as CRM structured fields, social media interaction counts, website visit depth, email reply latency, phone connection rate, offline meeting frequency, ad click-through rate, and historical transaction correlation. The values range from 0 to 1, and are rounded to 4 decimal places. During the data structure reorganization phase, each component is assigned a float32 data type and arranged in modal number order. Metadata tags include a modal ID (integer type), a status tag code (enumeration type), and a high-precision timestamp (millisecond-level integer), which are bound to the corresponding weight component via an index vector. The serialization phase uses little-endian encoding and inserts a length prefix (2 bytes) and a CRC16 checksum into the message header. The interface verification logic checks the data type, field order, and length prefix for correctness field by field. After confirming no anomalies, the data is written to an output queue containing a depth-first scheduling strategy. The scheduler releases data at 50-millisecond time slice boundaries, ensuring that the basic smoothed weight vector and the multimodal sales lead feature vector input are synchronized into the mutation detection module. In this embodiment, the weight vector received by the mutation detection module has been verified to be processed with an average call latency of less than 2 milliseconds, significantly improving the real-time performance and stability of the score generation process.
[0204] like Figure 3 As shown, step S4 involves performing mutation detection based on the instantaneous difference between the basic smoothed weight vector and the current multimodal sales lead feature vector. If a single-modal increase is detected to exceed a preset mutation buffer threshold, the instantaneous change is temporarily stored and its release is delayed, generating a buffered weight vector that has undergone mutation suppression. Specifically, this includes:
[0205] S4.1: Obtain the basic smoothing weight vector and the current multimodal sales lead feature vector, perform modal-by-modal difference operation on the two to generate an instantaneous difference vector, and input the instantaneous difference vector into the modal amplification ratio calculation logic. Calculate the relative change amplitude of each modal component using the normalized gain method, and output the modal dynamic change spectrum containing the instantaneous amplification ratio of each modality.
[0206] Obtain the basic smoothing weight vector and the current multimodal sales lead feature vector, perform modal operator difference operation, subtract the corresponding basic smoothing weight component from the current feature value of each modal component, and obtain the instantaneous difference vector.
[0207] The instantaneous difference vector is input into the modal amplification ratio calculation logic. For each modal component, the absolute value of its difference is extracted and divided by the absolute value of the basic smoothing weight component to achieve normalization of the relative change amplitude.
[0208] The normalized gain method is used to multiply the relative change amplitude of each mode according to the sensitivity coefficient in the state constraint parameter set, so as to amplify the weight ratio of high-sensitivity mode change and suppress the ratio of low-sensitivity mode change.
[0209] The adjusted relative change magnitude vectors are arranged in modal order and encapsulated as modal dynamic change spectra to express the instantaneous increase ratio distribution of each mode relative to the smoothing baseline weight at the current moment.
[0210] Through the above processing method, the basic smoothing weight vector of the previous step is transformed into a modal dynamic change spectrum that can be used for mutation detection, realizing the standardized quantification of the change amplitude of each modality, and laying a data foundation for subsequent mutation buffer threshold comparison.
[0211] S4.2: Receive the modal dynamic change spectrum, perform threshold comparison operation on the instantaneous amplification ratio of each mode in the spectrum based on the preset mutation buffer threshold, use logic threshold to identify abnormal mode identifiers whose amplification exceeds the threshold, and generate a mutation event marker set carrying abnormal mode location information and overamplitude value.
[0212] S4.3: Based on the mutation event tag set, perform time slice temporary storage processing on the instantaneous changes of the modal marked as abnormal, use the first-in-first-out queue caching mechanism to lock the over-amplitude value in the delay buffer and set a release countdown, while retaining the original changes of the unmarked abnormal modal, and generate a hybrid pending weight reorganization that includes the temporarily stored changes and the instantaneous changes.
[0213] The input mutation event tag set serves as the execution object for this step, containing the anomalous modality location index and corresponding overamplitude value obtained from the preceding threshold determination. The mutation event tag set is input into the time-slice temporary storage processing logic, establishing a delay buffer index table for anomalous modalities, and allocating storage units within this buffer based on the overamplitude value in the tag set. An enqueue operation is performed on the instantaneous changes of anomalous modalities, using a first-in-first-out queue caching mechanism to lock the overamplitude value at the head of the delay buffer, and a release countdown parameter set is created. For unmarked anomalous modal components, their original changes are retained and directly written to the instant change buffer to ensure that non-nominal modal weights are not affected by the delay mechanism. The temporary changes and instant changes are structurally assembled to form a hybrid weight reassembly, where the change channel corresponding to the anomalous modality points to the delay buffer, while the change channel for the normal modality points to the instant update path. The data interface protocol verifies the field integrity and type matching of the hybrid weight group, ensuring input consistency for subsequent mutation release and weight fusion. By using time-slice temporary storage and first-in-first-out queue locking mechanism, the mutation event marking information of the previous step is transformed into abnormal weight delay data that can be released in a controllable manner, thereby achieving the effect of temporary isolation and buffer stabilization of single-modal mutation on the overall weight system.
[0214] S4.4: Monitor the release countdown status in the delay buffer. When the countdown ends, perform gradient release operation on the temporarily stored overamplitude value. Use linear interpolation smoothing method to decompose the locked mutation value into multiple small increment steps to generate a progressive compensation weight sequence distributed in subsequent time periods.
[0215] The system retrieves the countdown signal for releasing the over-amplitude value locked for abnormal modes from the delay buffer. Based on the remaining time slice of this signal, a state scan is performed to determine if the set release trigger condition has been met. For the temporarily stored over-amplitude change that meets the release condition, the gradient release operation controller is invoked, decomposing the overall change into multiple small incremental steps according to a preset progressive compensation strategy. A linear interpolation smoothing method is used to construct the time series distribution of the incremental steps, which is then sequentially injected into the weight vector in subsequent time slices. The starting value in the interpolation process is set to the weight baseline value before the release of the temporarily stored change, and the ending value is the target weight value after the release of the change. The interpolation formula is used to...
[0216]
[0217] Where Δ is the step increment, V is the target weight value, U is the current weight value, and N is the number of decomposition steps, ensuring that the step increment is numerically continuous and increases or decreases smoothly. The generated incremental step sizes form a progressively compensated weight sequence, stored in a sequence cache unit for subsequent fusion. Through the above processing method based on release countdown scanning, gradient decomposition, and linear interpolation smoothing, the temporarily stored mutation values from the previous step are transformed into continuous compensated weight data in the time domain, achieving orderly release of fluctuation amplitude in the mutation buffer and significantly improving scoring stability.
[0218] For example, in a certain SaaS CRM system, the abnormal modality cached in the delayed buffer is the website deep access behavior characteristic, with an excessive weight change of 0.25 and a release countdown of 3 scoring cycles. The controller sets the decomposition steps N to 5, the current weight value U to 0.60, and the target weight value V to 0.85. The overall change is decomposed into Δ steps:
[0219]
[0220] The calculated value is Δ = 0.0625. A step size sequence [0.6625, 0.7250, 0.7875, 0.8500] is generated using linear interpolation to ensure a smooth increase in weight values over the subsequent four periods. In performance verification, the scoring curve incorporating this progressively compensated weight sequence shows a significantly reduced fluctuation amplitude compared to the uncompensated condition. Furthermore, the scoring maintains continuity and robustness consistent with business state characteristics during the release of abrupt mode changes.
[0221] S4.5: By integrating the instantaneous changes with the generated progressive compensation weight sequence, a weighted update operation is performed on the basic smooth weight vector. The smoothed abrupt change effect is incorporated into the weight system by using vector superposition, and finally, a buffered weight vector with abrupt change suppression that eliminates instantaneous violent fluctuations is output.
[0222] Step S5: Monitor whether the clue state tag has undergone a cross-stage transition. If a state transition is determined to have occurred, trigger a progressive interpolation update operation to correct the buffer weight vector to a state weight vector. Specifically, this includes:
[0223] S5.1: Obtain the current clue status label and the previous historical clue status label, and compare the two based on the preset state transition judgment logic to generate a state transition detection signal containing the transition flag and the source target status identifier.
[0224] It receives the buffer weight vector after mutation suppression processing and the current clue status label as input data, and retrieves the historical clue status label stored in the status management cache at the previous moment as the comparison object.
[0225] Perform a field-by-field matching operation on the current clue status label and the historical clue status label of the previous time. Use status field index matching to ensure that the comparison results are unique and accurate.
[0226] The matching result is input into the preset state transition judgment logic. The judgment logic includes conditional branch operation and state legality verification operation based on the difference of state stage number. The difference judgment adopts the absolute value calculation method to ensure the directionality and magnitude of the detection across stages.
[0227] By using the state stage difference and the legality verification result, a state change judgment vector is constructed. This vector consists of three parts: a transition detection flag, a source state stage identifier, and a target state stage identifier. The flag is a binary signal used to indicate whether a transition has occurred.
[0228] The state change determination vector is encapsulated as a state transition detection signal, which serves as the trigger condition for cross-state weight transition matrix invocation and progressive interpolation update.
[0229] By generating and processing state transition detection signals, the application context of the buffered weight vector from the previous step is transformed into a data trigger signal with the ability to identify stage differences, thereby achieving stage adaptability and continuity in the weight update process.
[0230] For example, in a SaaS CRM system, the buffered weight vector after mutation suppression corresponds to the multimodal weight configuration of a sales lead. The current lead status label is "Multi-channel Interaction," and the previous one was "First Contact," with status stage numbers 5 and 3 respectively. The difference between the status stage numbers is calculated using |5−3|, resulting in a difference of 2. This difference, along with the validity check result, is input into the judgment logic. The judgment logic sets a transition trigger threshold of 1. When the difference is greater than the threshold, it outputs a transition flag of 1, a source state identifier of 3, and a target state identifier of 5. After being encapsulated into a state transition detection signal, this signal is sent to the cross-state weight migration matrix module, driving the matrix to select the migration path from the source state to the target state, achieving a progressive interpolation update of the weights. In this scenario, the state transition detection signal can initiate a smooth weight transition targeting the characteristics of the multi-channel interaction stage, significantly improving the stability and matching degree of the scoring mechanism during business stage changes.
[0231] S5.2: Based on the source target state identifier in the state transition detection signal, query the preset cross-state weight transition matrix library, extract the corresponding initial transition coefficient matrix and interpolation step size parameter, and construct a set of state transition control parameters to guide the smooth transition of weights.
[0232] The Cross-State Weight Transition Matrix Library is a pre-built knowledge base used to store weight transition parameters for sales leads as they transition between different business lifecycle stages. The library is indexed by combinations of source and target states, with each entry corresponding to a cross-state weight transition matrix (i.e., a transition coefficient matrix). Matrix elements represent how the modal weights of the source state should proportionally transition to the modal weights of the target state during the state transition process.
[0233] The cross-state weighted transition matrix library uses a key-value pair storage structure. Each record corresponds to a state transition path and contains the following fields:
[0234] Transition identifier: A unique key formed by concatenating the source state identifier and the target state identifier (e.g., "first contact → multi-channel interaction").
[0235] Source status identifier: Business stage code before the transition.
[0236] Target state identifier: Code of the business stage after the transition.
[0237] The migration coefficient matrix is an M×M square matrix, where M is the total number of modalities involved in the scoring (e.g., CRM data, social media behavior, website visit logs, etc.). Matrix elements T ij This represents the proportion of the migration contribution of the j-th mode of the source state to the i-th mode of the target state. Typically, the sum of each row of the matrix is 1 to maintain the energy conservation of the weights.
[0238] Interpolation step size parameter: a scalar or vector that defines the magnitude of each weight update during the interpolation process (e.g., 0.05 or a scale that is dynamically adjusted according to the time slice length).
[0239] Smoothing decay factor (optional): Used to control the convergence speed of weights during interpolation and prevent overshoot.
[0240] Construction method:
[0241] (a) Offline initialization phase:
[0242] Modality definition: Identify all modalities involved in the scoring (such as CRM data, social media interactions, website visit logs, etc.), fix their order, and obtain the total number of modalities M.
[0243] Status Enumeration: List all business stages (such as "First Contact", "Initial Interaction", "Multi-channel Interaction", "Confirmation of Intent", etc.) and number them.
[0244] Transition Matrix Calibration: Business experts were invited to create a transition matrix for each pair of states based on the changing importance of each modality at different stages. For example, from "first contact" to "multi-channel interaction," the weight of the social media modality should increase significantly; therefore, the source social media column corresponding to the target social media row in the transition matrix has a higher weight. From the weight change trajectory of historical leads, the average percentage change in the actual weight of each modality when a state transition occurs was statistically analyzed and normalized to serve as the initial value of the transition matrix.
[0245] Interpolation step size parameter setting: Based on the business's timeliness requirements for smooth weight transition, set the interpolation step size for each transition path (such as 0.05, 0.1, etc.), or dynamically calculate it based on the score update frequency.
[0246] Matrix library storage: Stores the constructed records in the system configuration center in JSON or database table format, providing a fast query interface based on source status + target status.
[0247] (II) Online Inquiry and Dynamic Fine-tuning:
[0248] Query interface: Based on the source state identifier and target state identifier in the state transition detection signal, retrieve the corresponding transition coefficient matrix and interpolation step size parameters from the matrix library.
[0249] Dynamic update: If the feedback set in step S8 shows that the weight smoothing effect of a certain transition path is not good (such as overshoot or slow convergence), the transition matrix elements or interpolation step size of the path can be adjusted offline to update the matrix library.
[0250] (III) Maintenance and Version Management:
[0251] The matrix library supports version control, incrementing the version number after each batch update to ensure traceability.
[0252] When a business phase changes (addition, deletion, or merging), all related transition entries must be updated synchronously.
[0253] Through the aforementioned cross-state weight transition matrix library, this invention achieves standardization and configurability of weight transition during state transitions, ensuring the smoothness of weight adjustment between different stages and the rationality of business operations.
[0254] Based on the source and target state identifiers in the state transition detection signal, a matrix library index retrieval operation is performed to map the combination key of the source state and the target state to the unique entry position in the cross-state weighted transition matrix library in order to locate the corresponding state transition record.
[0255] The state transition record is subjected to structural parsing, and the initial transition coefficient matrix encapsulated in the record is extracted. This matrix defines the mapping ratio between the modal weights of the source state and the modal weights of the target state, and retains the modal index information corresponding to its rows and columns.
[0256] Numerical verification and parameter type conversion are performed based on the interpolation step size parameters in the records. The step size parameters are unified into floating-point format and quantized according to the preset time slice length and weight update frequency so that they can be directly called by subsequent interpolation calculation logic.
[0257] The initial migration coefficient matrix is normalized, and the elements of the matrix are proportionally adjusted according to the modal sensitivity coefficients under the target state so that the normalized migration coefficient matrix numerically satisfies the constraint condition of the state stability threshold.
[0258] The normalized transition coefficient matrix and the quantized interpolation step size parameters are structurally encapsulated to generate a set of state transition control parameters to guide the smooth transition of weights, and to provide complete matrix and step size inputs for subsequent progressive interpolation operations.
[0259] Through the matrix retrieval, structural analysis, numerical processing and parameter encapsulation described above, the state transition detection signal from the previous step is transformed into a set of state transition control parameters that can directly drive the smooth calculation of weights, thereby achieving controllability and continuity in the cross-state weight adjustment process.
[0260] For example, in a SaaS CRM scenario, when the source state is identified as "first contact" and the target state is identified as "multi-channel interaction," the corresponding entry in the cross-state weight migration matrix library is a 5×5 initial migration coefficient matrix. The matrix element values range from 0.2 to 0.8, corresponding to the weight mapping ratios of five modalities: CRM fields, social media behavior, website access logs, email feedback, and telephone interaction. The original value of the interpolation step size parameter is 0.05, which is converted to 0.025 after time-slice quantization to match the weight update frequency of 0.5Hz. During normalization, the migration matrix elements are proportionally adjusted based on the modal sensitivity coefficients of the target state {0.9, 0.7, 0.8, 0.6, 0.5}. For example, the original value of 0.8 becomes 0.8×0.8 after adjustment in the third modality, resulting in 0.64, ensuring that the maximum element of the adjusted matrix does not exceed the state stability threshold of 0.65. The final output state transition control parameter set includes the adjusted matrix and a step size of 0.025. The transition weight sequence generated during the interpolation stage exhibits numerical continuity, which greatly improves the smoothness of cross-state weight switching and reduces the impact of abrupt changes.
[0261] S5.3: Using the state transition control parameter set, perform multiple rounds of linear interpolation operations on the current buffer weight vector and the baseline weight vector in the target state to generate an intermediate state transition weight sequence that changes progressively with the time slice.
[0262] The system receives the initial transition coefficient matrix and interpolation step size parameters from the state transition control parameter set. These are then indexed and mapped to the buffer weight vector and the baseline weight vector in the target state to establish a correspondence between matrix elements. Element-wise interpolation is performed on the buffer weight vector and the target state baseline weight vector to generate a difference vector reflecting the weight differences across states. The difference vector is then multiplied element-wise with the initial transition coefficient matrix to obtain a transition magnitude vector, which represents the scaling factor of weight change at each interpolation step size. Piecewise linear interpolation is performed on the transition magnitude vector using the interpolation step size parameters to construct an intermediate state transition weight sequence that varies progressively with time slices. The transition weight for each time slice is generated by vector superposition of the buffer weight vector and the corresponding interpolation increment. Through the interpolation process, the buffer weight vector from the previous step is smoothly transitioned to the target state baseline weight vector, obtaining the intermediate state weight sequence and achieving a gradual control effect for cross-state weight transition.
[0263] For example, in a SaaS CRM environment, this sub-step is performed on leads that are in the "first contact" state and have transitioned to the "intent confirmation" state. The buffer weight vector is [0.35, 0.25, 0.40], the target state baseline weight vector is [0.50, 0.20, 0.30], the initial transition coefficient matrix is [[0.4, 0.6, 0.8], [0.3, 0.5, 0.7], [0.2, 0.4, 0.6]], and the interpolation step size parameter is set to 3. First, the difference is calculated to obtain the difference vector [0.15, -0.05, -0.10], which has the mathematical form [0.50−0.35, 0.20−0.25, 0.30−0.40]. Multiplying the difference vector by the corresponding elements of the migration coefficient matrix yields the migration magnitude vector. For example, the calculation formula for the first time slice is [0.15×0.4,−0.05×0.3,−0.10×0.2], resulting in [0.06,-0.015,-0.02]. The migration magnitude vector is then evenly divided according to the interpolation step size, and three progressively changing increment vectors are generated using linear interpolation. For example, the first group is [0.02,-0.005,-0.0067]. These increment vectors are then sequentially superimposed with the buffer weight vector to obtain the intermediate state transition weight sequence: [0.37,0.245,0.3933]→[0.39,0.24,0.3866]→[0.41,0.235,0.3799]. This sequence smoothly moves the weights towards the target state within the three time slices without abrupt changes, ultimately achieving the expected technical effect of progressive weight migration across states.
[0264] S5.4: Based on the convergence stability index of the intermediate transition weight sequence, the interpolation step size is dynamically fine-tuned to eliminate local oscillations caused by data noise and output a gradually updated weight vector that has been smoothed and calibrated.
[0265] S5.5: Replace the original buffered weight vector with the smoothed and calibrated progressively updated weight vector as the final output to generate a state weight vector that can adapt to the characteristics of the new business stage and maintain numerical continuity.
[0266] Step S6: Generate a basic cue score based on the state weight vector, and simultaneously capture the amplitude and rhythm characteristics of the residual signals of each mode to generate a fluctuation suppression coefficient. Specifically, this includes:
[0267] S6.1: Obtain the state weight vector and the multimodal sales lead feature vector at the current moment, and perform linear mapping processing based on the weighted summation method to generate a basic lead score that reflects the immediate conversion potential of the lead.
[0268] S6.2: Using the basic lead score and the multimodal sales lead feature vector at the current moment, perform deviation separation processing based on the residual back-calculation mechanism to extract the multimodal residual signal sequence that characterizes the instantaneous deviation of each modality data.
[0269] Obtain the basic lead score and the multimodal sales lead feature vector dataset at the current moment, and establish the residual calculation input pair to ensure that the residual back-inference mechanism has complete comparison samples after weight mapping.
[0270] For each modal component in the multimodal sales lead feature vector, a baseline prediction value is calculated, and the feature components are weighted using the weight coefficients corresponding to the state weight vector to form a sequence of modal prediction contribution values.
[0271] The modality prediction contribution value sequence is differentially mapped to the basic clue score, and an instantaneous deviation value is generated for each modality component. The instantaneous deviation value is defined as the difference between the actual feature contribution and the theoretical contribution of the modality in the total score.
[0272] The instantaneous deviation values are normalized using the residual back-calculation method, unifying the deviations of features at different scales to the same dimension for subsequent amplitude and rhythm feature extraction. The normalization formula is as follows:
[0273]
[0274] Where v is the actual feature contribution value, p is the predicted contribution value, and σ is the historical standard deviation of the modality feature.
[0275] The normalized modal residual values are aggregated according to the modal index to form a multimodal residual signal sequence, which contains the instantaneous deviation of each mode at the current moment.
[0276] By using a residual back-calculation mechanism, the basic lead scores and multimodal sales lead feature vectors are transformed into multimodal residual signal sequences that can be analyzed by a fluctuation sensor, thereby achieving the technical effect of explicitly quantifying the sources of deviation within the score.
[0277] S6.3: Receive the multimodal residual signal sequence, perform amplitude energy aggregation processing based on sliding window statistics, and generate residual amplitude characteristic parameters that characterize the intensity of abrupt changes in single-modal data.
[0278] S6.4: Based on the time series characterization of the multimodal residual signal sequence, perform rhythm frequency extraction processing based on autocorrelation analysis to generate residual rhythm characteristic parameters that characterize the frequency of data fluctuations.
[0279] S6.5: The residual amplitude characteristic parameter and the residual rhythm characteristic parameter are fused together, and coefficient synthesis processing is performed based on a nonlinear mapping function to generate a fluctuation suppression coefficient for suppressing drastic fluctuations in the score.
[0280] The residual amplitude feature parameters and residual rhythm feature parameters are used as input data pairs and connected to the coefficient synthesis operation control module. Based on a preset multidimensional nonlinear mapping function library, a mapping function model matching the current business state label is selected, and the input data is normalized according to the defined parameters of the mapping function model to eliminate the difference in amplitude dimensions between different modes. A squared nonlinear enhancement operation is performed on the normalized residual amplitude feature parameters to improve the detection sensitivity of instantaneous changes, while a logarithmic compression operation is performed on the normalized residual rhythm feature parameters to reduce the numerical amplitude of high-frequency fluctuations. The enhanced amplitude feature and the compressed rhythm feature are multiplied and coupled within the mapping function model, and a weighted smoothing coefficient is introduced to perform an exponential decay operation on the product result to form a stability adjustment core value. A nonlinear suppression operator is constructed using this stability adjustment core value and a state stability threshold. The fluctuation suppression coefficient is generated through the operator's action and encapsulated into a standard numerical interface format for output, to be called by subsequent scoring adjustment steps. By using nonlinear product coupling and exponential decay processing, the residual amplitude characteristics and residual rhythm characteristics from the previous step are transformed into fluctuation suppression coefficients to eliminate drastic fluctuations in scores, thereby achieving a synergistic suppression effect on fluctuation amplitude and frequency.
[0281] Step S7: Based on the fluctuation suppression coefficient, perform a nonlinear decay or enhancement adjustment operation on the basic clue score, and output a final clue score with a stability confidence meta-label. Specifically, this includes:
[0282] S7.1: Obtain the amplitude and rhythm features of the modal residual signal, and perform fusion calculation on the amplitude and rhythm features of the modal residual signal based on a preset nonlinear mapping function to generate a dynamic fluctuation suppression coefficient vector.
[0283] S7.2: Receive the basic clue score and the dynamic fluctuation suppression coefficient vector, and based on the nonlinear attenuation enhancement method, perform an adaptive adjustment operation on the basic clue score using the dynamic fluctuation suppression coefficient vector to generate a preliminary corrected clue score.
[0284] This sub-step receives the basic cue score and dynamic fluctuation suppression coefficient vector as input conditions. The basic cue score is converted into a computable numerical matrix, and the dynamic fluctuation suppression coefficient vector is parsed into adjustment factors for the corresponding modes to form a coefficient matrix. Dimension matching and matrix normalization are performed on the coefficient matrix to ensure it is completely consistent with the basic cue score matrix in row and column space, guaranteeing numerical compatibility for subsequent calculations. The core mapping function of the nonlinear attenuation enhancement method is constructed, setting the transition boundary between attenuation and enhancement as the median of the dynamic fluctuation suppression coefficient vector, and establishing a piecewise nonlinear mapping rule. The normalized coefficient matrix and the basic score matrix are input into the nonlinear adjustment module, and adaptive adjustment is performed using a composite operation of multiplicative fusion and exponential mapping, according to the following formula:
[0285]
[0286] Where S is the basic scoring matrix, C is the dynamic fluctuation suppression coefficient matrix, and k is the gain adjustment factor dynamically calculated based on the spectral characteristics of the scoring fluctuations. The output preliminary corrected scoring matrix is encapsulated in a data format, bound with clue identifiers and timestamp information, forming a preliminary corrected clue score that can be used for subsequent stability confidence assessment. Through adaptive adjustment of the nonlinear attenuation enhancement method, the basic scoring results of the previous step are transformed into scoring data corrected for fluctuation suppression, achieving the expected technical effects of smooth suppression of instantaneous fluctuations and trend enhancement.
[0287] S7.3: Extract the high-frequency oscillation component statistics from the dynamic fluctuation suppression coefficient vector, and perform probability distribution fitting on the high-frequency oscillation component statistics based on the confidence assessment model to generate a stability confidence quantification index.
[0288] Receive the full data frame of the dynamic fluctuation suppression coefficient vector, call the high-frequency component extraction operator to perform fast Fourier transform and bandpass filtering operations on the coefficient sequence, filter out components below the preset lower frequency limit and above the preset upper frequency limit, and retain the complex amplitude matrix of the corresponding high-frequency oscillation component in the frequency domain spectrum.
[0289] The complex amplitude matrix is converted into a real oscillation amplitude vector by modulus operation, and local extreme points in the oscillation amplitude vector are identified based on peak detection to establish a statistical sample set of high-frequency oscillation amplitudes.
[0290] The central tendency calculation is performed using the high-frequency oscillation amplitude statistical sample set to generate a high-frequency oscillation statistical vector containing statistics such as mean, variance, skewness, and kurtosis.
[0291] The high-frequency oscillation statistic vector is input into the probability distribution fitting module of the confidence assessment model. The maximum likelihood estimation method is used to verify the distribution family hypothesis and estimate the parameters of the statistic, and the confidence quantification index is calculated.
[0292] The probability distribution fitting results are output as a stability confidence quantification index, and the parameter source and statistical sample identifier are bound in the numerical interface to achieve the traceability of the confidence index in subsequent scoring encapsulation.
[0293] By extracting high-frequency oscillation components, calculating statistics, and fitting probability distributions, the preliminary corrected score from the previous step is correlated with a stability confidence quantification index, thereby achieving a quantitative assessment and credibility characterization of the score's volatility risk.
[0294] S7.4: Obtain the preliminary corrected clue score and the stability confidence metric, and bind the stability confidence metric as a metadata tag to the preliminary corrected clue score based on the data encapsulation protocol to generate the final clue score with a stability confidence meta tag.
[0295] The process involves obtaining preliminary corrected clue scores and stability confidence metrics. The meta-tag binding interface of the data encapsulation protocol is invoked to perform index matching and data classification processing on the two types of input objects according to the field alignment rules specified in the protocol, generating an intermediate encapsulated data unit containing a score value field and a confidence index field. The score value field in the intermediate encapsulated data unit undergoes encoding format checks and numerical precision verification, using double-precision floating-point encoding to ensure numerical stability and read / write consistency of the score data during transmission. The stability confidence metrics field undergoes statistical range verification and unit consistency confirmation. Combined with a preset quantitative index standard template, it is mapped to the meta-tag attribute set specified in the protocol, forming a three-element structure containing attribute name, attribute type, and attribute value. The verified and mapped confidence attribute set is structurally merged with the score value field. Through meta-tag binding operations, the confidence attribute set is appended as an embedded meta-tag to the score value field, achieving integrated encapsulation of score data and stability metrics. The packaging function of the encapsulation protocol is invoked to generate the final data object conforming to the system output interface specification, and output as the final clue score data stream with accompanying stability confidence meta-tags.
[0296] By using the data encapsulation protocol meta tag binding processing method, the preliminary correction clue score and stability confidence quantification index of the previous step are transformed into the final score data interface that can be directly called by the downstream feedback closed-loop calibration layer, so as to achieve synchronous consistency of score output in terms of numerical, semantic and stability assessment information.
[0297] Step S8: Collect the manually labeled clue conversion results and compare them with the final clue score. If the absolute value of the deviation exceeds a preset joint feedback threshold, construct a stability and effectiveness joint feedback set. Specifically, this includes:
[0298] S8.1: Obtain the manually labeled lead conversion result data stream and the final lead score sequence with accompanying stability confidence meta-labels, perform timestamp alignment and identifier matching on both to generate a pairwise bias analysis dataset containing actual conversion values and predicted score values.
[0299] S8.2: Perform absolute difference calculation based on the actual conversion value and predicted score value in the pairwise deviation analysis dataset, and logically compare the calculated absolute deviation value with a preset joint feedback threshold to filter out abnormal score sample pairs whose deviation exceeds the threshold.
[0300] S8.3: Construct a joint feedback set of stability and effectiveness using the abnormal rating sample pairs and their corresponding clue status labels, and perform cluster statistical analysis on the sample feature distribution in the set to extract the key modal residual feature patterns that cause rating fluctuations.
[0301] S8.4: Based on the key modal residual feature patterns and the current state constraint parameter set, perform gradient descent optimization to iteratively update the state stability threshold and the weight coefficients in the modal sensitivity spectrum to generate a fine-tuned state constraint parameter set.
[0302] S8.5: Perform nonlinear regression fitting based on the fluctuation amplitude characteristics of the fine-tuned state constraint parameter set and the joint feedback set of stability effectiveness, reconstruct the decay curve parameters in the fluctuation suppression coefficient mapping rule, and output the updated fluctuation suppression coefficient mapping rule.
[0303] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.
[0304] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.
[0305] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A sales lead mining method that integrates multi-modal data, specifically including: S1: Acquire multimodal sales lead data streams and identify the lifecycle status of each lead, generating lead status tags with timestamps; S2: Based on the cue state label, query the state sensitivity spectrum library, extract the state stability threshold and modal sensitivity spectrum bound to the state label, and generate a set of state constraint parameters; S3: Based on the multimodal sales lead feature vector at the current moment and the state constraint parameter set, calculate the weighted moving average of the most recent effective weights to generate a basic smooth weight vector; S4: Perform mutation detection based on the instantaneous difference between the basic smooth weight vector and the current multimodal sales lead feature vector. If the single-modal increase exceeds the preset mutation buffer threshold, the instantaneous change will be temporarily stored and released with a delay, generating a buffer weight vector that has been mutated and suppressed. S5: Monitor whether the status label of the clue has undergone a cross-stage transition. If a state transition is determined to have occurred, trigger the progressive interpolation update operation to correct the buffer weight vector to the state weight vector. S6: Generate basic cue scores based on state weight vectors, and capture the amplitude and rhythm characteristics of residual signals of each mode to generate fluctuation suppression coefficients; S7: Perform nonlinear decay or enhancement adjustment on the basic clue score based on the fluctuation suppression coefficient, and output the final clue score with stability confidence meta-label.
2. The sales lead mining method based on multi-modal data according to claim 1, characterized in that, The state sensitivity spectrum library includes state labels, modal sensitivity spectrum, a list of high-frequency fluctuation mode identifiers, state stability thresholds, and historical transformation variance reference values.
3. The sales lead mining method based on multi-modal data according to claim 1, characterized in that, The multimodal sales lead feature vector is obtained by parsing the multimodal sales lead data stream according to the protocol and standardizing the fields to eliminate data format differences.
4. The sales lead mining method based on multi-modal data fusion according to claim 1, characterized in that, Following S7, the following also includes: S8: Collect the conversion results of manually labeled clues and compare them with the final clue score. If the absolute value of the deviation exceeds the preset joint feedback threshold, construct a stability and validity joint feedback set.
5. The sales lead mining method based on multi-modal data according to claim 1, characterized in that, S3 specifically includes: Based on the modal sensitivity spectrum in the state constraint parameter set, the feature vector of multimodal sales leads at the current moment is subjected to dimension alignment and validity verification, missing or abnormal modal data is removed and a standardized multimodal feature sequence is generated. The historical weight repository is queried using standardized multimodal feature sequences to extract the five most recent valid weight records bound to the clue identifier, and a time decay factor weighting operation is performed on the five most recent valid weight records to generate a historical weight candidate set with time priority. Based on the historical weight candidate set and the state stability threshold in the state constraint parameter set, a dynamic sliding window calculation model is constructed. Normalized sliding mean fusion processing is performed on each modal weight component in the historical weight candidate set to generate a preliminary smooth weight matrix. Based on the real-time distribution characteristics of the standardized multimodal feature sequences, gain compensation adjustment is performed on the initial smoothing weight matrix to correct the sensitivity lag problem caused by long-period smoothing and generate the corrected smoothing weight vector. The corrected smoothed weight vector is encapsulated into a standard data interface format, and the basic smoothed weight vector is output.
6. The sales lead mining method based on multi-modal data according to claim 1, characterized in that, S4 specifically includes: Obtain the basic smoothing weight vector and the current multimodal sales lead feature vector, perform modal-by-modal difference operation on the two to generate an instantaneous difference vector, and input the instantaneous difference vector into the modal amplification ratio calculation logic to calculate the relative change amplitude of each modal component and output the modal dynamic change spectrum. The modal dynamic change spectrum is received, and a threshold comparison operation is performed on the instantaneous amplification ratio of each mode in the spectrum based on a preset mutation buffer threshold. Abnormal mode identifiers with amplification exceeding the threshold are identified, and a mutation event marker set carrying abnormal mode location information and overamplitude value is generated. Based on the mutation event label set, the instantaneous changes of the modalities marked as anomalous are temporarily stored in time slices. The first-in-first-out queue caching mechanism is used to lock the over-amplitude values in the delay buffer and set a release countdown. At the same time, the original changes of the unmarked anomalous modalities are retained, and a hybrid pending weight reorganization is generated. Monitor the release countdown status in the delay buffer. When the countdown ends, perform gradient release operation on the temporarily stored overamplitude value, decompose the locked mutation value into multiple small increment steps, and generate a progressive compensation weight sequence distributed in subsequent time periods. By integrating the instantaneous changes with the generated progressively compensated weight sequence, a weighted update operation is performed on the basic smoothed weight vector, incorporating the smoothed abrupt changes into the weight system, and finally outputting a buffered weight vector that has eliminated instantaneous drastic fluctuations and suppressed abrupt changes.
7. The sales lead mining method based on multi-modal data according to claim 6, characterized in that, The modal dynamic change spectrum includes the instantaneous amplification ratio of each mode.
8. The sales lead mining method based on multi-modal data according to claim 6, characterized in that, The hybrid pending rights reorganization includes temporary changes and immediate changes.
9. The sales lead mining method based on multi-modal data according to claim 1, characterized in that, S5 specifically includes: Obtain the current clue status label and the previous historical clue status label, and compare the two based on the state transition judgment logic to generate a state transition detection signal containing the source target status identifier. Based on the source target state identifier in the state transition detection signal, a pre-set cross-state weight transition matrix library is queried, and the corresponding initial transition coefficient matrix and interpolation step size parameter are extracted to construct a set of state transition control parameters for guiding the smooth transition of weights. The state transition control parameter set is used to perform multiple rounds of linear interpolation on the current buffer weight vector and the baseline weight vector in the target state to generate an intermediate state transition weight sequence that changes progressively with time slices. The interpolation step size is dynamically fine-tuned based on the convergence stability index of the intermediate transition weight sequence to eliminate local oscillations caused by data noise and output a gradually updated weight vector that has been smoothed and calibrated. The original buffered weight vector is replaced by the smoothed and calibrated progressively updated weight vector as the final output to generate a state weight vector that can adapt to the characteristics of the new business stage and maintain numerical continuity.
10. The sales lead mining method based on multi-modal data according to claim 9, characterized in that, The state transition detection signal also includes a transition flag.
