A method and system for automatic processing of business leads based on SPM

Abstract

The present invention discloses a method and system for automatic processing of the entire process of business leads based on SPM, which belongs to the field of information processing technology. The method comprises: collecting multi-source lead trajectory data, building a cross-system federated feature pool and extracting basic feature combinations to form original lead units; performing feature enhancement on the original lead units to obtain lead feature vectors; obtaining a system endpoint load matrix, combining the lead feature vectors, generating an initial endpoint mapping strategy using a reinforcement learning algorithm, and outputting an optimal endpoint mapping strategy through risk balance algorithm game optimization; deploying a context responder based on the optimal strategy, generating and pushing customized dynamic interactive content streams, and collecting interactive behavior feedback data at the same time; establishing a lead life cycle tracking chain, and dynamically adjusting the transition probability matrix and termination state probability in a pre-built lead state probability model based on the collected feedback data; the present invention realizes the precise processing and effective push of lead data.

Description

Technical Field

[0001] The present invention belongs to the field of information processing technology, and in particular to a method and system for automatic processing of the entire process of business leads based on SPM. Background Art

[0002] In the logistics sales process, the processing of business leads by manufacturing companies, such as equipment procurement needs of potential customers, logistics solution cooperation intentions, etc., faces many technical problems that need to be solved urgently: First, there are barriers to multi-source data integration. Lead data is scattered across multiple channels such as official website consultation, industry exhibitions, and supply chain systems, covering structured data, semi-structured data, and unstructured data. In the cross-system integration process, it is difficult to form a unified lead view due to heterogeneous data formats and privacy leakage risks; second, the lead value assessment method is one-sided. Traditional methods rely on manual experience or a single dimension for assessment, ignoring the relationship between dynamic behavior time series characteristics and real-time context, resulting in high-potential leads being mismatched with resources, while low-value leads are occupied. Excessive manpower is used. Secondly, the endpoint allocation mechanism is rigid. The load of endpoints responsible for following up on leads, such as sales teams and customer service systems, is constantly changing. Static allocation strategies can easily lead to overloaded endpoints and idle endpoints. There is also a lack of risk assessment of the matching degree between endpoint processing capabilities and lead value. In addition, there is a problem of homogeneity in interactive content. Pushed product solutions, quotations, and other content are generated based on fixed templates and are not dynamically adjusted based on the lead subject's immediate response tendency and contextual fit, resulting in a low interaction conversion rate. Finally, there are gaps in lead lifecycle tracking. The transfer of lead status relies on manual recording. There is a lack of data-driven dynamic prediction models, and the entire process data cannot be traced, making it difficult to accurately optimize follow-up strategies. These problems affect the quality and efficiency of information processing and cannot meet the current demand for efficient information processing. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention proposes a full-process automatic processing method and system for business leads based on SPM, which collects multi-source lead trajectory data, builds a cross-system federal feature pool and extracts basic feature combinations to form original lead units; performs feature enhancement on the original lead units to obtain lead feature vectors; obtains the system endpoint load matrix, combines the lead feature vectors, uses a reinforcement learning algorithm to generate an initial endpoint mapping strategy, and outputs the optimal endpoint mapping strategy through risk balance algorithm game optimization; deploys a context responder based on the optimal strategy, generates and pushes customized dynamic interactive content streams, and collects interactive behavior feedback data at the same time; establishes a lead life cycle tracking chain, and dynamically adjusts the transition probability matrix and termination state probability in the pre-built lead state probability model based on the collected feedback data; the present invention realizes the accurate processing and effective push of lead data.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A method for automatically processing the entire process of business leads based on SPM, comprising:

[0006] S1: Collect multi-source clue trajectory data through a distributed heterogeneous data interface, build a cross-system federated feature pool, and extract the basic feature combination of a single independent clue entity from the cross-system federated feature pool to form an original clue unit; the original clue unit is the basic feature combination of a single independent clue entity;

[0007] S2: Use the dynamic evolution modeling engine to enhance the features of the original clue units, calculate the three-dimensional value evaluation vector of each original clue unit, and obtain the clue feature vector;

[0008] S3: Obtain the load matrix of each endpoint of the system, combine it with the clue feature vector, use the reinforcement learning algorithm to generate the initial endpoint mapping strategy, and use the risk balance algorithm to perform game optimization on the initial endpoint mapping strategy to output the optimal endpoint mapping strategy;

[0009] S4: Based on the optimal endpoint mapping strategy, a context responder is deployed for the endpoint or lead subject. Based on the lead feature vector, a customized dynamic interactive content stream is generated and pushed to the endpoint or lead subject. Meanwhile, feedback data on the lead subject's interactive behavior towards the pushed content is collected in real time.

[0010] S5: Establish a clue lifecycle tracking chain, and dynamically adjust the transition probability matrix and terminal state probability in the pre-built clue state probability model based on the interactive behavior feedback data collected by the pre-established clue lifecycle tracking chain; the clue state probability model is constructed based on the hidden Markov model; the clue state probability model includes a clue state space, a transition probability matrix and a terminal state probability distribution.

[0011] Specifically, the process of building the cross-system federated feature pool includes:

[0012] Differential privacy technology is used to desensitize the multi-source clue trajectory data. The desensitized multi-source clue trajectory data is mapped to a unified vector space through the feature hashing algorithm to obtain distributed feature vectors, and the federated averaging algorithm is used to aggregate the distributed feature vectors.

[0013] Specifically, the specific steps of S2 include:

[0014] S2.1: Preprocess the original clue unit to separate static features and dynamic time series features; the static features include user attributes and device information; the dynamic time series features include behavior sequences and interaction logs;

[0015] S2.2: Set the sliding window size and sliding step size, sample from the dynamic time series features with the set sliding window size and sliding step size to obtain a sampled input sequence, and input the sampled input sequence into the pre-built time series convolutional network model to obtain a dynamic pattern feature vector;

[0016] S2.3: Use the dot-product attention mechanism to calculate the similarity score between the dynamic pattern feature vector and the current context feature vector, and obtain the association weight after normalization.

[0017] S2.4: Use a multi-layer perceptron to perform weighted summation of static features and dynamic pattern features according to their associated weights to obtain an enhanced feature representation;

[0018] S2.5: Extracting immediate responsiveness features from the enhanced feature representation and inputting them into a pre-trained immediate responsiveness prediction model based on logistic regression to output the predicted probability of the immediate responsiveness of the cue;

[0019] S2.6: Obtain the similarity score between the dynamic pattern feature vector and the current context feature vector from S2.3 to obtain the context fit;

[0020] S2.7: Input the enhanced feature representation into the pre-trained potential prediction model based on the gradient boosting tree to obtain the conversion potential value of the lead into actual business results;

[0021] S2.8: Concatenate the evaluation results of the three dimensions of the clue's immediate responsiveness, predicted probability, situational fit, and conversion potential value into a vector and normalize them so that the values ​​of each dimension are within the same range to form a clue feature vector.

[0022] Specifically, the specific steps of S3 include:

[0023] S3.1: Using sensors to collect load indicators of each endpoint in the system in real time, the collected load indicators of each endpoint are stored according to the endpoint-load indicator correspondence to form an initial load data set; the load indicators include CPU utilization, memory utilization, and network bandwidth utilization;

[0024] S3.2: Clean the initial load data set and fill missing values ​​with the mean to form a load matrix; the elements in the load matrix are the values ​​of the corresponding endpoints on the corresponding load indicators;

[0025] S3.3: Obtain the clue feature vector, and concatenate and integrate the clue feature vector and the vectorized form of the load matrix to obtain the global decision vector;

[0026] S3.4: Define the state, action, and reward function for reinforcement learning; the state is the global decision vector; the action is the selection of any endpoint for mapping; the reward function is implemented based on load balancing and performance indicators;

[0027] S3.5: Train the reinforcement learning model based on the historical global decision vector to obtain a trained reinforcement learning model. Then, input the current global decision vector into the trained reinforcement learning model. The trained reinforcement learning model selects an endpoint for mapping based on the learned strategy. Through iterative repetition, an initial endpoint mapping strategy is obtained.

[0028] S3.6: Define two-dimensional risk factors and construct a risk objective function based on the weighted sum of the two-dimensional risk factors; the two-dimensional risk factors include a load risk factor and a value risk factor;

[0029] The load risk factor is a load risk value obtained by calculating the load change of each endpoint after the mapping request based on the load matrix and the initial endpoint mapping strategy;

[0030] The value-risk factor is a quantified value-risk value of each endpoint mapping decision based on the value of the request and the processing capability of the endpoint;

[0031] The initial endpoint mapping policy is a set of endpoint mapping decisions.

[0032] Specifically, the specific steps of S3 further include:

[0033] S3.7: Consider endpoint mapping decision-making as a game process, where endpoints serve as game players and clue features serve as factors influencing the game outcome.

[0034] S3.8: For each endpoint and each clue feature combination, calculate the utility value of selecting the endpoint for mapping under that combination. Based on the calculated utility values, construct the endpoint-clue game utility matrix;

[0035] S3.9: Find the Nash equilibrium solution based on the endpoint-clue game utility matrix and use the Nash equilibrium solution as a constraint;

[0036] S3.10: Load risk and value risk are combined to form a multi-objective space, where the horizontal axis represents load risk and the vertical axis represents value risk, and the initial endpoint mapping strategy corresponds to an initial solution in the multi-objective space;

[0037] S3.11: Use a non-dominated sorting genetic algorithm to find the Pareto front in a multi-objective space and select the optimal endpoint mapping strategy from the Pareto front based on the Nash negotiation solution;

[0038] S3.12: Generate an executable endpoint mapping instruction table based on the optimal endpoint mapping strategy.

[0039] The specific steps of S3.9 include:

[0040] S3.9.1: Obtain an endpoint-cue game utility matrix. Based on the endpoint-cue game utility matrix, determine the game participants and strategy sets, where the game participants are each endpoint and the strategy set is a combination of all mappable clue features. Each row in the endpoint-cue game utility matrix represents the strategy set of an endpoint.

[0041] S3.9.2: The endpoint-clue game utility matrix is ​​used as the core data of the game model. For each endpoint, its payment function is the corresponding utility value, that is, the payment when the i-th endpoint selects the j-th clue feature combination is ;

[0042] S3.9.3: For each row of the endpoint-cue game utility matrix, find the maximum utility value in that row;

[0043] S3.9.4: Repeat for each row and find their optimal choice under the strategy set;

[0044] S3.9.5: If there is a clue feature combination , so that for all endpoints i, the clue feature combination is selected The maximum utility can be obtained when , that is, for any clue feature combination j, ,but is a pure strategy Nash equilibrium solution;

[0045] S3.9.6: If there is no pure strategy Nash equilibrium solution, then find a mixed strategy Nash equilibrium, including:

[0046] Assume that the probability of endpoint i selecting clue feature combination j is ;

[0047] Set the hybrid policy combination for all endpoints to ,in, represents the probability distribution of the mixed strategy of the mth endpoint, and satisfies ;

[0048] For endpoint i, its expected utility function is for and The product of the clue feature combination The cumulative sum under , where Represents endpoint i selection clue feature combination The probability of Represents endpoint i selection clue feature combination The utility value of represents the clue feature combination selected by endpoint i;

[0049] For each endpoint i, solve the mixed strategy that maximizes its expected utility when the mixed strategy of non-self endpoints is given, that is, solve the optimal response function ,in, Indicates logical NOT, represents a non-endpoint i, represents the mixed strategy combination of non-endpoint i, represents the optimal response mixed strategy of endpoint i;

[0050] If satisfied , then we get the mixed strategy Nash equilibrium solution;

[0051] S3.9.7: Use pure strategy Nash equilibrium solutions or mixed strategy Nash equilibrium solutions as constraints;

[0052] If a pure strategy Nash equilibrium solution is found, the constraint is to select a clue feature combination for each endpoint ;

[0053] If a mixed strategy Nash equilibrium solution is found, the constraint condition is that each endpoint has its mixed strategy probability Select clue feature combination .

[0054] Specifically, generating a customized dynamic interactive content stream based on the clue feature vector described in S4 and pushing it to the endpoint or clue subject includes:

[0055] S4.1: Determine the target endpoint or lead subject based on the optimal endpoint mapping strategy and deploy a context responder for the endpoint or lead subject;

[0056] S4.2: Obtain a clue feature vector; the clue feature vector includes the predicted probability of immediate responsiveness, context fit, and conversion potential value;

[0057] S4.3: Extract contextual fit information from the clue feature vector, screen the content template library based on the extracted contextual fit information, and select the content template with the highest contextual fit;

[0058] S4.4: Use the conversion potential value in the clue feature vector to classify the content information density into different levels;

[0059] S4.5: Set the content push frequency based on the predicted probability of immediate responsiveness in the clue feature vector;

[0060] S4.6: Based on the divided content information density levels, the content information is filled into the selected content template with the highest contextual fit. At the same time, the content flow is dynamically adjusted based on the real-time feedback of the clue subject during the interaction process;

[0061] S4.7: Combine the filled and adjusted content in order and logic to form a customized dynamic interactive content flow, and push the generated customized dynamic interactive content flow to the target endpoint or lead subject according to the set push frequency.

[0062] Specifically, the distributed heterogeneous data interface supports multiple data protocols; the multi-source clue trajectory data includes structured data, semi-structured data and unstructured data; the cross-system federated feature pool uses differential privacy technology to desensitize feature vectors, and realizes cross-system feature dimension alignment through feature hashing algorithm; the basic feature combination includes clue source channel, clue generation time, clue subject attributes and historical interaction records; the clue life cycle tracking chain is built based on blockchain technology; the clue life cycle tracking chain includes full-process nodes of clue creation, feature enhancement, endpoint mapping, interaction push and status update, and each node records operation time, operation subject and data change log.

[0063] An SPM-based full-process automatic processing system for business leads, including: feature construction module, enhancement and evaluation module, endpoint mapping module, interaction and feedback module, and lead status adjustment module;

[0064] The feature construction module collects multi-source clue trajectory data through a distributed heterogeneous data interface, constructs a cross-system federated feature pool, and forms an original clue unit;

[0065] The enhancement and evaluation module uses a dynamic evolution modeling engine to perform feature enhancement on the original clue unit and calculates a three-dimensional value evaluation vector to obtain a clue feature vector;

[0066] The endpoint mapping module is used to obtain the endpoint load matrix, combine it with the clue feature vector, and generate the optimal endpoint mapping strategy through the reinforcement learning algorithm and the risk balance algorithm;

[0067] The interaction and feedback module deploys context responders according to the optimal endpoint mapping strategy, generates and pushes customized dynamic interactive content streams, and collects interactive behavior feedback data;

[0068] The lead state adjustment module is used to establish a lead life cycle tracking chain and dynamically adjust the transition probability matrix and terminal state probability in the pre-built lead state probability model based on the interaction behavior feedback data.

[0069] Specifically, the endpoint mapping module includes: a load data acquisition unit, an initial strategy generation unit, and a game optimization unit;

[0070] The load data acquisition unit is used to acquire the load matrix of each endpoint in real time;

[0071] The initial strategy generating unit generates an initial endpoint mapping strategy using a reinforcement learning algorithm;

[0072] The game optimization unit uses a risk balance algorithm to optimize the initial endpoint mapping strategy and outputs the optimal strategy in combination with an adaptive weight adjustment mechanism.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] 1. The present invention proposes a full-process automatic processing system for business leads based on SPM, and optimizes and improves the architecture, operation steps and processes. The system has the advantages of simple process, low investment and operation costs, and low production work costs.

[0075] 2. The present invention proposes a full-process automatic processing method for business leads based on SPM. By collecting multi-source lead trajectory data, a cross-system federal feature pool is constructed, and basic feature combinations are extracted to form original lead units. The dynamic evolution modeling engine is then used to enhance the features of the original lead units, and the three-dimensional value evaluation vector is calculated to obtain the lead feature vector. This series of operations helps to comprehensively and accurately explore the value of leads and improve the effectiveness and accuracy of lead processing.

[0076] 3. The present invention proposes a full-process automatic processing method for business leads based on SPM, which obtains the system load matrix and combines it with the lead feature vector to generate and optimize the endpoint mapping strategy, deploys the context responder according to the optimal strategy, generates and pushes customized dynamic interactive content streams, and collects interactive behavior feedback data in real time. The lead status probability model is then dynamically adjusted based on the feedback data. This process realizes the intelligent and personalized processing of leads, can effectively improve the lead conversion rate, optimize the allocation of system resources, and continuously adapt to changes in lead status, thereby improving the overall lead processing effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a schematic diagram of a method for automatically processing the entire process of business leads based on SPM according to the present invention;

[0078] Figure 2 This is a principle flow chart of a method for automatic processing of business leads based on SPM in the present invention;

[0079] Figure 3 This is an architecture diagram of a full-process automatic processing system for business leads based on SPM in the present invention. DETAILED DESCRIPTION

[0080] Example 1

[0081] See also Figure 1-Figure 2 The present invention provides an embodiment of a method for automatically processing business leads through the entire process based on SPM, wherein SPM (Specific Processing Model) is a specific processing model, which in this application refers to a lead status probability model based on logistics sales, and includes the following steps:

[0082] S1: Collect multi-source clue trajectory data through a distributed heterogeneous data interface, build a cross-system federated feature pool, and extract the basic feature combination of a single independent clue entity from the cross-system federated feature pool to form an original clue unit; the original clue unit is the basic feature combination of a single independent clue entity;

[0083] The collected multi-source clue trajectory data is set based on the preset state feature requirements of the pre-built clue state probability model.

[0084] The distributed heterogeneous data interface supports multiple data protocols; the multi-source clue trajectory data includes structured data, semi-structured data and unstructured data; the cross-system federated feature pool uses differential privacy technology to desensitize feature vectors and realizes cross-system feature dimension alignment through feature hashing algorithm; the basic feature combination includes clue source channel, clue generation time, clue subject attributes and historical interaction records.

[0085] Among them, the distributed heterogeneous data interface includes: adapters that support Kafka, gRPC and RESTful API protocols, which are used to asynchronously collect multi-source trajectory data from CRM systems, user behavior log libraries and third-party data platforms.

[0086] Furthermore, the specific steps of S1 include:

[0087] (1) Design interfaces for different data sources to ensure compatibility of data formats and communication protocols;

[0088] (2) Formulate data collection frequency, data scope, and collection methods. Data scope refers to the time window and user group, and collection methods include real-time streaming collection and batch collection;

[0089] (3) During the data collection process, the data is preliminarily cleaned to remove noise and duplicate data, and necessary format conversion is performed;

[0090] (4) Connect various data sources to the collection system through designed interfaces;

[0091] (5) Extracting data related to the clue trajectory from various data sources, such as user behavior logs, transaction records, and location information;

[0092] (6) Integrate the extracted data according to a unified data model to form multi-source clue trajectory data;

[0093] (7) Based on business needs, define the features to be extracted, such as user attributes, behavioral characteristics, and transaction characteristics;

[0094] (8) Extracting defined features from multi-source clue trajectory data to form an original feature set;

[0095] (9) The extracted original feature set is stored in a pre-built cross-system federated feature pool, and operations such as version management and access control are performed;

[0096] (10) Identify a single independent clue entity, such as a user or a transaction, from multi-source clue trajectory data;

[0097] (11) For each identified clue entity, extract the relevant basic feature combination from the cross-system federated feature pool;

[0098] (12) The extracted basic features are combined and encapsulated to form original clue units, which provide a basis for subsequent feature enhancement and value evaluation.

[0099] The process of building the cross-system federated feature pool includes:

[0100] Differential privacy technology is used to desensitize the multi-source clue trajectory data. The desensitized multi-source clue trajectory data is mapped to a unified vector space through the feature hashing algorithm to obtain distributed feature vectors, and the federated averaging algorithm is used to aggregate the distributed feature vectors.

[0101] Furthermore, differential privacy technology is used to desensitize multi-source clue trajectory data. The data from different systems is mapped to a unified vector space through a feature hashing algorithm, and a federated averaging algorithm is used to aggregate distributed features, including:

[0102] (1) Set a privacy budget and control the level of privacy protection based on data sensitivity and security requirements;

[0103] (2) For each data point in the multi-source clue trajectory data, noise is added according to the Laplace mechanism, so that the data can maintain a certain degree of usability while preventing individual information leakage. Through actual testing, it is ensured that the desensitized multi-source clue trajectory data meets the differential privacy requirements while minimizing the impact on data quality. Among them, Laplace is the existing technical content in this field and is not the inventive solution of this application, so it will not be described in detail here;

[0104] (3) Selecting the feature hash function Murmur Hash to map the desensitized multi-source clue trajectory data into a preset fixed-length vector space, applying the feature hash function to each data feature in the desensitized multi-source clue trajectory data to convert it into a point in the vector space, ensuring that data from different systems can be compared and aggregated in the same space. At the same time, an open addressing method is used to resolve the hash conflict problem and ensure the accuracy and uniqueness of the mapping. The feature hash function Murmur Hash is a prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0105] (4) Initialize the same model parameters on the data nodes of different systems. Each data node calculates the feature hash vector based on the local data and performs model training or feature extraction based on the feature hash vector.

[0106] (5) Each data node uploads the model parameters or feature vectors calculated locally to the central server. The server aggregates the uploaded parameters using a federated averaging algorithm to obtain global model parameters or feature representations. The federated averaging algorithm is an algorithm used in federated learning to aggregate multiple local models to obtain a global model. It is a prior art in this field and is not an inventive solution of this application. It will not be described in detail here.

[0107] (6) The server distributes the aggregated global model parameters or feature representations back to each participant. Each participant updates the local model and prepares for the next round of training or feature extraction.

[0108] S2: Use the dynamic evolution modeling engine to enhance the features of the original clue units, calculate the three-dimensional value evaluation vector of each original clue unit, and obtain the clue feature vector;

[0109] The three-dimensional value evaluation vector includes immediate responsiveness, situational fit, and conversion potential value.

[0110] S3: Obtain the load matrix of each endpoint of the system, combine it with the clue feature vector, use the reinforcement learning algorithm to generate the initial endpoint mapping strategy, and use the risk balance algorithm to perform game optimization on the initial endpoint mapping strategy to output the optimal endpoint mapping strategy;

[0111] S4: Based on the optimal endpoint mapping strategy, a context responder is deployed for the endpoint or lead subject. Based on the lead feature vector, a customized dynamic interactive content stream is generated and pushed to the endpoint or lead subject. Meanwhile, feedback data on the lead subject's interactive behavior towards the pushed content is collected in real time.

[0112] The interactive behavior feedback data includes: click-through rate, content dwell time, operation response delay, interaction depth, and error trigger rate, and is transmitted back to the tracking chain in real time through the message queue.

[0113] The context responder adopts a generative adversarial network architecture, and its generator network includes:

[0114] Feature encoding layer: Use Bi-LSTM to extract context features;

[0115] Content generation layer: Generates interactive content sequences through the Transformer decoder;

[0116] Optimizer: Uses Wasserstein distance to constrain content relevance.

[0117] S5: Establish a clue lifecycle tracking chain, and dynamically adjust the transition probability matrix and terminal state probability in the pre-built clue state probability model based on the interactive behavior feedback data collected by the pre-established clue lifecycle tracking chain; the clue state probability model is constructed based on the hidden Markov model; the clue state probability model includes a clue state space, a transition probability matrix and a terminal state probability distribution.

[0118] It should be noted that the clue state space, transition probability matrix, and terminal state probability distribution are the core components of the clue state probability model constructed based on the hidden Markov model. Together, these three components constitute the model's mathematical description of the evolution of states throughout the clue lifecycle. This enables the clue state probability model to dynamically infer the current state of a clue, predict its future evolution path, and evaluate its final conversion value based on feedback data from the clue's interactive behavior. The clue state space refers to the set of all possible hidden states a clue may be in throughout its lifecycle. These hidden states cannot be directly observed but can only be inferred indirectly through the interactive behavior of the clue subject. The transition probability matrix describes the probability of a clue transitioning from one state to another, reflecting the dynamic evolution between states. The terminal state probability distribution describes the probability distribution of a clue ultimately residing in any terminal state, characterizing the possible outcomes at the end of the clue lifecycle.

[0119] It should be emphasized that in this application, a full-process automated closed-loop system from lead acquisition to dynamic optimization is mainly constructed. The core goal is to improve the conversion efficiency of business leads, the rationality of resource allocation and the accuracy of interaction through data-driven intelligent processing, and ultimately achieve automated and intelligent management of the entire life cycle of leads. Among them, S1 solves the problem of where the lead data comes from and how to integrate it, S2 solves the problem of how to quantify the value of leads, S3 solves the problem of which endpoint the leads should be assigned to, such as the sales team and customer service system processing problem, S4 solves the problem of how to effectively interact with lead entities, such as users and enterprises, and S5 solves the problem of how to make the system continue to evolve. The essence of the optimization of the lead state probability model in the present invention is to adjust the parameters of the lead state probability model, such as the transition probability matrix, through interactive feedback data, so that the lead state probability model can more accurately characterize the evolution law of the lead state, such as the probability of high-intent leads being converted into transactions in the next step and the risk of loss of low-response leads. After the lead state probability model is optimized, it will feed back to the previous links of the entire process.

[0120] Among them, the hidden Markov model is the existing technical content in this field and is not an inventive solution of this application, so it will not be described here in detail.

[0121] The lead lifecycle tracking chain is built based on blockchain technology; the lead lifecycle tracking chain includes full-process nodes of lead creation, feature enhancement, endpoint mapping, interactive push and status update, and each node records operation time, operation subject and data change log.

[0122] In the present invention, by introducing the perspective of game theory, the competition and cooperation relationship between different endpoints when facing different clue characteristics can be considered more comprehensively.

[0123] For example, the input is assumed to be an initial clue of an e-commerce enterprise consulting on its official website about its purchasing needs for intelligent sorting robots, where the e-commerce enterprise is the subject of the clue. Specifically, the process includes: obtaining multi-source clue trajectory data, where the multi-source clue trajectory data includes a purchase demand form filled out on the official website, official website browsing logs for the past three days, and customer service online chat records; after differential privacy desensitization, feature hash mapping to a unified vector space, and federated average algorithm aggregation, a cross-system federated feature pool is constructed and its basic feature combination is extracted to output the original clue unit, where the original clue unit is a basic feature combination including clue source, generation time, subject attributes, and historical interactions; then, the static features and dynamic time series features of the original clue unit are separated, and an input sequence is obtained by sliding window sampling. The input sequence is input into a time series convolutional network to output a dynamic pattern feature vector, where the static feature refers to the company size, and the dynamic time series feature refers to the browsing history; the dot product attention is combined to calculate the similarity score between the dynamic pattern feature vector and the preset production expansion scenario, and after multi-layer perceptron fusion and three-dimensional evaluation, a clue feature vector containing immediate responsiveness, situational fit, and conversion potential value is output; then, endpoint A (sales team, load rate 60%) and The load matrix of endpoint B (customer service system, CPU usage 55%) is combined with the clue feature vector. Reinforcement learning uses the clue feature vector + load matrix vectorization as the state, selects the endpoint as the action, and generates an initial mapping strategy (mapped to endpoint A) based on the load balance. The risk balance algorithm then calculates the load risk factor (e.g., endpoint A has a load rate of 65% after mapping, with a risk value of 0.3) and the value risk factor (e.g., endpoint A has a historical conversion rate of 70%, with a risk value of 0.2). The Nash equilibrium solution is found through the game utility matrix, and the optimal endpoint mapping strategy is finally output, that is, the mapping to endpoint is confirmed. A; then deploy a situational responder for endpoint A, screen the production expansion plan template based on the situational fit, set high-density content according to the conversion potential value, including technical parameters and spot inventory, and set daily push + real-time pop-up reminder based on immediate responsiveness to generate content flow, such as solution summary-inventory table-consultation entrance push, collect interactive behavior feedback data such as clicking the inventory table twice and replying to the quotation request; finally, based on the feedback data recorded in the blockchain tracking chain, dynamically adjust the clue state probability model of the hidden Markov model, output the adjusted transition probability matrix and termination state probability, and complete the whole process processing.

[0124] For example, in the above embodiment, the current state probability is: based on the adjusted parameters, it is inferred that the probability that the e-commerce enterprise lead is currently in a clear intention state is 80%, and the probability of being in a hesitation period is 20%; the transition probability prediction is: the probability of the lead transferring from a clear intention state to a transaction state is predicted to be 60%, and the probability of transferring to a loss state is 10%; the termination state probability is: the probability of the lead being successfully converted in the end is predicted to be 70%, and the probability of stopping loss is 30%.

[0125] Specifically, a method for automatically processing the entire process of business leads based on SPM also includes:

[0126] S6: Use the clue status probability model dynamically adjusted in S5 as the core decision-making basis, output the current status evaluation results of the clue and the state evolution law, and reversely empower the dynamic optimization of S2 to S4 to form a closed-loop iteration of the entire process. Specifically, it includes: outputting the association rules between lead status and features to the dynamic evolution modeling engine of S2, such as the weight of the consultation frequency feature of the lead in the past 7 days in a high-intention state needs to be increased by 25%, so that the calculation of the three-dimensional value evaluation vector is more in line with the actual state of the lead, and the accuracy of the lead feature vector is enhanced; outputting the lead status-endpoint processing efficiency association data to the strategy generation link of S3, such as the conversion rate of the initial contact state leads processed by the intelligent customer service endpoint increased by 18%, which serves as a supplementary decision feature of the reinforcement learning algorithm, so that the optimal endpoint mapping strategy can balance the load while being more adapted to the processing resource requirements of the current state of the lead; outputting the key trigger conditions for the lead state transition to the context responder of S4, such as when the probability of the lead transitioning from the intention cultivation state to the decision-making period state exceeds 50%, the quotation plan push rule needs to be triggered to guide the real-time adjustment of the customized dynamic interactive content flow, and realize the upgrade from initial feature-based push to state evolution-based push.

[0127] The specific steps of S2 include:

[0128] S2.1: Preprocess the original clue unit to separate static features and dynamic time series features; the static features include user attributes and device information; the dynamic time series features include behavior sequences and interaction logs;

[0129] It's important to explain that static features are relatively stable over a period of time and don't change frequently. Common static features include user attributes and device information. User attributes include age, gender, region, occupation, and education level, while device information includes device type, operating system, device brand, and screen resolution. Based on the definition of static features, fields related to static features are filtered from the original lead units. For example, fields such as age, gender, and region are filtered from user registration information, and fields such as device type and operating system are filtered from device information collection data. Dynamic time series features are features that change over time and exhibit time series characteristics, reflecting user behavior and interactions at different points in time. Common dynamic time series features include behavior sequences and interaction logs. Behavior sequences include page browsing records, button click records, and keyword search records. Interaction logs include conversation records with customer service and feedback records of system prompts. Records related to dynamic time series features are filtered from the original lead units. For example, user page browsing records and button click records are filtered from user behavior tracking systems, and conversation records between users and customer service are filtered from interaction logging systems.

[0130] S2.2: Set the sliding window size and sliding step size, sample the dynamic time series features using the set sliding window size and sliding step size to obtain a sampled input sequence, and input the sampled input sequence into a pre-built time series convolutional network model to obtain a dynamic pattern feature vector. The time series convolutional network model is prior art in this field and does not constitute an inventive solution of the present application, and is not described in detail here.

[0131] S2.3: Use the dot-product attention mechanism to calculate the similarity score between the dynamic pattern feature vector and the current context feature vector, and obtain the association weight after normalization.

[0132] Furthermore, the specific steps of S2.3 include:

[0133] (1) Obtaining a dynamic pattern feature vector and a current context feature vector. The dynamic pattern feature vector is extracted from dynamic temporal features, for example, by processing user behavior sequences through a temporal convolutional network. The current context feature vector is a vector encoded with information related to the current context, such as the current time, the page the user is currently on, and current market trends.

[0134] (2) Using the dot product attention mechanism, calculate the similarity score between the dynamic pattern feature vector and the current context feature vector, including:

[0135] The dynamic pattern feature vector is used as the query vector and the current context feature vector is used as the key vector;

[0136] Calculate the dot product of the query vector and the key vector to get a similarity score. The dot product mainly measures the degree of similarity between the two vectors in direction. The larger the value, the more similar the two vectors are.

[0137] (3) Normalize the similarity scores and convert them into probability distribution to obtain the association weights.

[0138] S2.4: Use a multi-layer perceptron to perform a weighted summation of the static features and the dynamic pattern features according to the associated weights to obtain an enhanced feature representation. The multi-layer perceptron is prior art in this field and does not constitute an inventive solution of this application, and is not described in detail here.

[0139] S2.5: Extracting immediate responsiveness features from the enhanced feature representation and inputting them into a pre-trained immediate responsiveness prediction model based on logistic regression to output the predicted probability of the immediate responsiveness of the cue;

[0140] Furthermore, the specific steps of S2.5 include:

[0141] (1) Obtaining enhanced feature representations and extracting immediate responsiveness features from the enhanced feature representations. The immediate responsiveness features include operation behavior patterns, interaction behavior patterns, and context matching. For example, through an online learning platform, a user first browses the course catalog, then enters the course trial page, and then clicks the sign up now button. This operation sequence reflects the user's interest in the course and has a high immediate responsiveness.

[0142] (2) The extracted immediate responsiveness features are normalized, and the preprocessed immediate responsiveness feature vector is input into the pre-trained immediate responsiveness prediction model based on logistic regression. The logistic regression model calculates the predicted probability of the immediate responsiveness of the clue based on the input feature vector.

[0143] Furthermore, the construction and training process of the immediate responsiveness prediction model based on logistic regression includes:

[0144] (1) Collect historical immediate responsiveness features, use statistical methods to perform feature selection on the historical immediate responsiveness features, and obtain the filtered historical immediate responsiveness features;

[0145] (2) Divide the data of the filtered historical immediate responsiveness features to obtain a training set and a validation set;

[0146] (3) The loaded logistic regression model is trained using the training set and verified using the validation set to obtain a pre-trained immediate responsiveness prediction model based on logistic regression. The logistic regression model is a prior art in this field and is not an inventive solution of the present application, and is not described in detail here.

[0147] S2.6: Obtain the similarity score between the dynamic pattern feature vector and the current context feature vector from S2.3 to obtain the context fit;

[0148] S2.7: Input the enhanced feature representation into a pre-trained potential prediction model based on a gradient boosting tree to obtain a conversion potential value for converting the lead into actual business results. The potential prediction model is constructed based on a gradient boosting tree, which is prior art in this field and does not constitute an inventive solution of this application, and is not described in detail here.

[0149] S2.8: Concatenate the evaluation results of the three dimensions of the clue's immediate responsiveness, predicted probability, situational fit, and conversion potential value into a vector and normalize them so that the values ​​of each dimension are within the same range to form a clue feature vector.

[0150] The specific steps of S3 include:

[0151] S3.1: Use sensors to collect load indicators from each endpoint in the system in real time. These indicators are stored according to endpoint-load indicator correspondence to form an initial load data set. These load indicators include CPU usage, memory usage, and network bandwidth usage. These endpoints include sales teams, customer service systems, agents, and other core entities involved in lead processing for manufacturing enterprises.

[0152] S3.2: Clean the initial load data set and fill missing values ​​with the mean to form a load matrix. In the load matrix, rows represent endpoints, columns represent different load indicators, and elements of the load matrix are the values ​​of the corresponding endpoints on the corresponding load indicators.

[0153] S3.3: Obtain the clue feature vector, and concatenate and integrate the clue feature vector and the vectorized form of the load matrix to obtain the global decision vector;

[0154] Among them, the vectorized form of the load matrix is ​​to directly splice each row or an endpoint of the load matrix to obtain the load vectorized representation; the global decision vector obtained after integration integrates the system's load status and clue information, providing a basis for subsequent reinforcement learning decisions.

[0155] S3.4: Define the state, action, and reward function for reinforcement learning; the state is the global decision vector; the action is the selection of any endpoint for mapping; the reward function is implemented based on load balancing and performance indicators;

[0156] In the present invention, the reward is set to be proportional to the load balancing degree of the endpoint. When the clue is mapped to the endpoint with lower load, a higher reward is given. At the same time, considering performance indicators such as response time, if the clue can be responded to quickly on the mapped endpoint, an additional reward is also given. On the contrary, if the clue causes the endpoint load to be too high or the response time is too long, a negative reward is given.

[0157] S3.5: Train a reinforcement learning model based on the historical global decision vector to obtain a trained reinforcement learning model. Then, input the current global decision vector into the trained reinforcement learning model. The trained reinforcement learning model selects an endpoint for mapping based on the learned strategy. Through iterative repetition, an initial endpoint mapping strategy is obtained. The reinforcement learning model is prior art in this field and does not constitute the inventive solution of this application, and is not described in detail here.

[0158] It should be noted that during the training process, the reinforcement learning model gradually learns the strategy of selecting the optimal action under a given state by continuously trying different actions and observing the corresponding rewards. After multiple rounds of iterative training, a trained reinforcement learning model is obtained.

[0159] S3.6: Define two-dimensional risk factors and construct a risk objective function based on the weighted sum of the two-dimensional risk factors; the two-dimensional risk factors include load risk factors and value risk factors ;

[0160] Among them, the smaller the value of the risk objective function, the lower the risk of the endpoint mapping strategy in terms of load and value.

[0161] The load risk factor is a load risk value obtained by calculating the load change of each endpoint after the mapping request based on the load matrix and the initial endpoint mapping strategy;

[0162] For example, the increase in the endpoint's CPU usage, memory usage, and network bandwidth usage is calculated, and the impact of these increases on endpoint performance is comprehensively considered and quantified into a load risk value. The higher the load risk value, the greater the load pressure the endpoint faces after mapping, and the greater the probability of performance problems.

[0163] The value-risk factor is a quantified value-risk value of each endpoint mapping decision based on the value of the request and the processing capability of the endpoint;

[0164] For example, mapping high-value requests to endpoints with weaker processing capabilities can cause request processing to fail or degrade in quality, creating value risks.

[0165] The initial endpoint mapping policy is a set of endpoint mapping decisions;

[0166] S3.7: Consider endpoint mapping decision-making as a game process, where endpoints serve as game players and clue features serve as factors influencing the game outcome.

[0167] S3.8: For each endpoint and each clue feature combination, calculate the utility value of selecting the endpoint for mapping under that combination. Based on the calculated utility values, construct the endpoint-clue game utility matrix;

[0168] Among them, the endpoint-clue game utility matrix shows the mapping utility under different combinations of endpoint and clue features, providing a data basis for finding the optimal strategy.

[0169] Furthermore, the specific steps of S3.8 include:

[0170] (1) Identify the factors that influence utility value, including:

[0171] Endpoint resource factors: 1) Computing resources: The number of CPU cores and main frequency of the endpoint. The more computing resources, the stronger the clue processing ability, and the higher the utility value. 2) Storage resources: The storage capacity of the endpoint, which stores clue-related data to avoid processing interruptions due to insufficient storage, which affects the utility value. 3) Network bandwidth: The larger the network bandwidth, the faster the data transmission speed, the higher the data exchange efficiency during clue processing, and the corresponding increase in utility value.

[0172] Lead characteristic factors: 1) Lead value: The potential commercial value of the lead. For example, high-value leads, such as customers with large purchasing intentions, can bring higher returns when mapped to endpoints, resulting in a high utility value. 2) Lead complexity: The business logic and data processing difficulty involved in the lead. Complex leads require more powerful endpoint processing capabilities. If the endpoint can meet these requirements, the utility value will be higher.

[0173] (2) Factors affecting the quantitative utility value include:

[0174] Endpoint resource factor quantification: 1) Computing resources: The number of CPU cores and main frequency are weighted together. For example, the resource quantification value is calculated by weighted summation of the number of CPU cores and main frequency. 2) Storage resource quantification value: Storage capacity is directly used as the quantification value. 3) Network bandwidth: Network bandwidth is used as the quantification value.

[0175] Quantification of lead characteristic factors: 1) Lead value: Evaluated based on the potential revenue of the lead. For example, for e-commerce leads, the potential purchase amount is estimated based on the user's historical purchase amount and the price range of browsed products. 2) Lead complexity: Quantified by analyzing the number of business process steps involved in the lead or the amount of data processed. Lead complexity is calculated by taking the weighted sum of the number of business process steps involved and the amount of data processed.

[0176] Quantification of service quality factors: 1) Response time: The average response time T of an endpoint processing similar leads is obtained through historical data or simulation tests. The shorter the response time, the higher the quantified value, and the reciprocal of the response time can be used as part of the quantified value. 2) Reliability: Quantified based on the historical failure rate of the endpoint. The reliability quantified value is equal to the difference between 1 and the historical failure rate of the endpoint.

[0177] (3) Based on the influencing factors of the quantified utility value, a utility value calculation formula is constructed to obtain the utility value of selecting the i-th endpoint for mapping under the j-th clue feature combination. ;

[0178] In the present invention, the utility value calculation formula is the weighted sum of resource quantization value, storage capacity, network bandwidth, clue value, clue complexity, the inverse of average response time, and reliability quantization value. The utility value of selecting the i-th endpoint for mapping under the j-th clue feature combination is obtained by weighted summation. .

[0179] (4) The calculated utility value Fill in the matrix to construct the endpoint-clue game utility matrix .

[0180] S3.9: Find the Nash equilibrium solution based on the endpoint-clue game utility matrix and use the Nash equilibrium solution as a constraint;

[0181] Among them, the Nash equilibrium solution means that during the game process, no player can improve his or her utility by unilaterally changing his or her strategy. The Nash equilibrium solution is used as a constraint condition to ensure that the optimal endpoint mapping strategy obtained is stable during the game process.

[0182] S3.10: Load risk and value risk are combined to form a multi-objective space, where the horizontal axis represents load risk and the vertical axis represents value risk, and the initial endpoint mapping strategy corresponds to an initial solution in the multi-objective space;

[0183] S3.11: Use a non-dominated sorting genetic algorithm to find the Pareto front in the multi-objective space, and select the optimal endpoint mapping strategy from the Pareto front based on the Nash negotiation solution. The non-dominated sorting genetic algorithm is prior art in this field and does not constitute the inventive solution of this application, and is not described in detail here.

[0184] Among them, the solutions on the Pareto front are the set of solutions that cannot further improve any objective without compromising other objectives, that is, these solutions achieve a relative balance between load risk and value risk; and the optimal endpoint mapping strategy selected from the Pareto front can effectively reduce load risk and value risk while ensuring system load balance and performance, and is stable during the game process.

[0185] S3.12: Generate an executable endpoint mapping instruction table based on the optimal endpoint mapping strategy.

[0186] The specific steps of S3.9 include:

[0187] S3.9.1: Obtain the endpoint-cue game utility matrix. Based on the endpoint-cue game utility matrix, determine the game participants and strategy set, where the game participants are each endpoint and the strategy set is all mappable clue feature combinations;

[0188] It should be noted that each row in the endpoint-clue game utility matrix represents the strategy set of an endpoint.

[0189] S3.9.2: The endpoint-clue game utility matrix is ​​used as the core data of the game model. For each endpoint, its payment function is the corresponding utility value, that is, the payment when the i-th endpoint selects the j-th clue feature combination is ;

[0190] S3.9.3: For each row of the endpoint-cue game utility matrix, find the maximum utility value in that row;

[0191] S3.9.4: Repeat for each row and find their optimal choice under the strategy set;

[0192] S3.9.5: If there is a clue feature combination , so that for all endpoints i, the clue feature combination is selected The maximum utility can be obtained when , that is, for any clue feature combination j, ,but is a pure strategy Nash equilibrium solution;

[0193] S3.9.6: If there is no pure strategy Nash equilibrium solution, then find a mixed strategy Nash equilibrium, including:

[0194] Assume that the probability of endpoint i selecting clue feature combination j is ;

[0195] Set the hybrid policy combination for all endpoints to ,in, represents the probability distribution of the mixed strategy of the mth endpoint, and satisfies ;

[0196] For endpoint i, its expected utility function is for and The product of the clue feature combination The cumulative sum under , where Represents endpoint i selection clue feature combination The probability of Represents endpoint i selection clue feature combination The utility value of represents the clue feature combination selected by endpoint i;

[0197] For each endpoint i, solve the mixed strategy that maximizes its expected utility when the mixed strategy of non-self endpoints is given, that is, solve the optimal response function ,in, Indicates logical NOT, represents a non-endpoint i, represents the mixed strategy combination of non-endpoint i, represents the optimal response mixed strategy of endpoint i;

[0198] If satisfied , then we get the mixed strategy Nash equilibrium solution;

[0199] S3.9.7: Use pure strategy Nash equilibrium solutions or mixed strategy Nash equilibrium solutions as constraints;

[0200] If a pure strategy Nash equilibrium solution is found, the constraint is to select a clue feature combination for each endpoint ;

[0201] If a mixed strategy Nash equilibrium solution is found, the constraint condition is that each endpoint has its mixed strategy probability Select clue feature combination .

[0202] The customized dynamic interactive content stream is generated based on the clue feature vector as described in S4 and pushed to the endpoint or clue subject, including:

[0203] S4.1: Determine the target endpoint or lead subject based on the optimal endpoint mapping strategy and deploy a context responder for the endpoint or lead subject;

[0204] Among them, the context responder has the ability to process data, generate content and dynamically adjust.

[0205] S4.2: Obtain a clue feature vector; the clue feature vector includes the predicted probability of immediate responsiveness, context fit, and conversion potential value;

[0206] S4.3: Extract contextual fit information from the clue feature vector, screen the content template library based on the extracted contextual fit information, and select the content template with the highest contextual fit;

[0207] S4.4: Use the conversion potential value in the clue feature vector to classify the content information density into different levels;

[0208] It should be emphasized that the higher the conversion potential value, the greater the potential conversion possibility of the clue subject to the content. At this time, the content information density can be classified as a higher level to provide richer and more detailed information; conversely, if the conversion potential value is low, it should be classified as a lower level to avoid information overload.

[0209] S4.5: Set the content push frequency based on the predicted probability of immediate responsiveness in the clue feature vector;

[0210] It's important to emphasize that a higher predicted probability of immediate responsiveness indicates a greater likelihood that the lead will respond immediately to the content. This can lead to a higher push frequency to provide more relevant information to the lead. If the predicted probability is lower, the push frequency can be reduced to avoid excessive interruptions. The set push frequency will be used for subsequent content pushes.

[0211] S4.6: Based on the divided content information density levels, the content information is filled into the selected content template with the highest contextual fit. At the same time, the content flow is dynamically adjusted based on the real-time feedback of the clue subject during the interaction process;

[0212] S4.7: Combine the filled and adjusted content in order and logic to form a customized dynamic interactive content flow, and push the generated customized dynamic interactive content flow to the target endpoint or lead subject according to the set push frequency.

[0213] Among them, push methods include application push, SMS push, and email push.

[0214] Example 2

[0215] See also Figure 3 Another embodiment of the present invention provides a system for automatically processing business leads based on SPM, including:

[0216] Feature construction module, enhancement and evaluation module, endpoint mapping module, interaction and feedback module, and clue status adjustment module;

[0217] The feature construction module collects multi-source clue trajectory data through distributed heterogeneous data interfaces, builds a cross-system federated feature pool, and forms the original clue unit;

[0218] The enhancement and evaluation module uses a dynamic evolution modeling engine to enhance the features of the original clue unit and calculates the three-dimensional value evaluation vector to obtain the clue feature vector;

[0219] The endpoint mapping module is used to obtain the endpoint load matrix, combine it with the clue feature vector, and generate the optimal endpoint mapping strategy through the reinforcement learning algorithm and risk balance algorithm;

[0220] The interaction and feedback module deploys context responders based on the optimal endpoint mapping strategy, generates and pushes customized dynamic interactive content streams, and collects interactive behavior feedback data.

[0221] The lead state adjustment module is used to establish a lead lifecycle tracking chain and dynamically adjust the transition probability matrix and terminal state probability in the pre-built lead state probability model based on the interactive behavior feedback data to more accurately reflect the change of lead state.

[0222] The feature construction module includes: heterogeneous data access unit, federated feature processing unit, and clue generation unit;

[0223] Heterogeneous data access unit supports multiple data protocols and enables the collection of multi-source clue trajectory data;

[0224] The federated feature processing unit is used to perform data cleaning, feature normalization, and feature alignment, and build a cross-system federated feature pool;

[0225] The clue generation unit is used to extract basic feature combinations from the federated feature pool and generate original clue units with unique identifiers.

[0226] The enhancement and evaluation module includes: dynamic evolution modeling unit and evaluation calculation unit;

[0227] Dynamic evolution modeling unit, based on temporal feature extraction network and attention mechanism, performs feature enhancement on original clue unit;

[0228] The evaluation calculation unit is used to calculate the immediate responsiveness, situational fit and conversion potential values ​​respectively, and generate a clue feature vector.

[0229] The endpoint mapping module includes: load data acquisition unit, initial strategy generation unit, and game optimization unit;

[0230] A load data acquisition unit, used to obtain the load matrix of each endpoint in real time;

[0231] An initial strategy generation unit, which uses a reinforcement learning algorithm to generate an initial endpoint mapping strategy;

[0232] The game optimization unit uses the risk balance algorithm to optimize the initial endpoint mapping strategy and outputs the optimal strategy in combination with the adaptive weight adjustment mechanism.

[0233] The interaction and feedback module includes: responder deployment unit, interaction content generation unit, and feedback data collection unit;

[0234] A responder deployment unit, used to deploy a lightweight situational responder for a target endpoint or a clue subject;

[0235] An interactive content generation unit generates a customized dynamic interactive content flow based on the clue feature vector;

[0236] The feedback data collection unit is used to collect the interactive behavior feedback data of the clue subject in real time.

[0237] The clue status adjustment module includes: a tracking chain establishment unit and a model updating unit;

[0238] Tracking chain establishment unit, which establishes a lead life cycle tracking chain based on blockchain technology;

[0239] The model updating unit dynamically adjusts the transition probability matrix and terminal state probability in the pre-built clue state probability model based on the interactive behavior feedback data collected by the pre-established clue life cycle tracking chain.

[0240] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention. These are all protected by the present invention.

[0241] If the technical solution disclosed herein involves personal information, the product using the technical solution disclosed herein has clearly informed the individual of the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using the technical solution disclosed herein has obtained the individual's separate consent before processing the sensitive personal information and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, a clear and prominent sign is set to inform that the individual has entered the personal information collection scope and that personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that they agree to the collection of their personal information; or on the personal information processing device, when the personal information processing rules are notified by obvious signs / information, the individual's authorization is obtained through pop-up information or by asking the individual to upload their personal information. The personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A full-process automatic processing method for business leads based on SPM, characterized in that: include: S1: Collect multi-source clue trajectory data through a distributed heterogeneous data interface, build a cross-system federated feature pool, and extract the basic feature combination of a single independent clue entity from the cross-system federated feature pool to form an original clue unit; the original clue unit is the basic feature combination of a single independent clue entity; S2: Use the dynamic evolution modeling engine to enhance the features of the original clue units, calculate the three-dimensional value evaluation vector of each original clue unit, and obtain the clue feature vector; S3: Obtain the load matrix of each endpoint of the system, combine it with the clue feature vector, use the reinforcement learning algorithm to generate the initial endpoint mapping strategy, and use the risk balance algorithm to perform game optimization on the initial endpoint mapping strategy to output the optimal endpoint mapping strategy; S4: Based on the optimal endpoint mapping strategy, a context responder is deployed for the endpoint or lead subject. Based on the lead feature vector, a customized dynamic interactive content stream is generated and pushed to the endpoint or lead subject. Meanwhile, feedback data on the lead subject's interactive behavior towards the pushed content is collected in real time. S5: Establish a clue lifecycle tracking chain, and dynamically adjust the transition probability matrix and terminal state probability in the pre-built clue state probability model based on the interactive behavior feedback data collected by the pre-established clue lifecycle tracking chain; the clue state probability model is constructed based on the hidden Markov model; the clue state probability model includes a clue state space, a transition probability matrix and a terminal state probability distribution.

2. The method for automatically processing business leads based on SPM as claimed in claim 1, characterized in that: The process of building the cross-system federated feature pool includes: Differential privacy technology is used to desensitize the multi-source clue trajectory data. The desensitized multi-source clue trajectory data is mapped to a unified vector space through the feature hashing algorithm to obtain distributed feature vectors, and the federated averaging algorithm is used to aggregate the distributed feature vectors.

3. The method for automatically processing business leads based on SPM as claimed in claim 2, characterized in that: The specific steps of S2 include: S2.1: Preprocess the original clue unit to separate static features and dynamic time series features; the static features include user attributes and device information; the dynamic time series features include behavior sequences and interaction logs; S2.2: Set the sliding window size and sliding step size, sample from the dynamic time series features with the set sliding window size and sliding step size to obtain a sampled input sequence, and input the sampled input sequence into the pre-built time series convolutional network model to obtain a dynamic pattern feature vector; S2.3: Use the dot-product attention mechanism to calculate the similarity score between the dynamic pattern feature vector and the current context feature vector, and obtain the association weight after normalization. S2.4: Use a multi-layer perceptron to perform weighted summation of static features and dynamic pattern features according to their associated weights to obtain an enhanced feature representation; S2.5: Extracting immediate responsiveness features from the enhanced feature representation and inputting them into a pre-trained immediate responsiveness prediction model based on logistic regression to output the predicted probability of the immediate responsiveness of the cue; S2.6: Obtain the similarity score between the dynamic pattern feature vector and the current context feature vector from S2.3 to obtain the context fit; S2.7: Input the enhanced feature representation into the pre-trained potential prediction model based on the gradient boosting tree to obtain the conversion potential value of the lead into actual business results; S2.8: Concatenate the evaluation results of the three dimensions of the clue's immediate responsiveness, predicted probability, situational fit, and conversion potential value into a vector and normalize them so that the values ​​of each dimension are within the same range to form a clue feature vector.

4. The method for automatically processing business leads based on SPM as claimed in claim 3, characterized in that: The specific steps of S3 include: S3.1: Using sensors to collect load indicators of each endpoint in the system in real time, the collected load indicators of each endpoint are stored according to the endpoint-load indicator correspondence to form an initial load data set; the load indicators include CPU utilization, memory utilization, and network bandwidth utilization; S3.2: Clean the initial load data set and fill missing values ​​with the mean to form a load matrix; the elements in the load matrix are the values ​​of the corresponding endpoints on the corresponding load indicators; S3.3: Obtain the clue feature vector, and concatenate and integrate the clue feature vector and the vectorized form of the load matrix to obtain the global decision vector; S3.4: Define the state, action, and reward function for reinforcement learning; the state is the global decision vector; the action is the selection of any endpoint for mapping; the reward function is implemented based on load balancing and performance indicators; S3.5: Train the reinforcement learning model based on the historical global decision vector to obtain a trained reinforcement learning model. Then, input the current global decision vector into the trained reinforcement learning model. The trained reinforcement learning model selects an endpoint for mapping based on the learned strategy. Through iterative repetition, an initial endpoint mapping strategy is obtained. S3.6: Define two-dimensional risk factors and construct a risk objective function based on the weighted sum of the two-dimensional risk factors; the two-dimensional risk factors include a load risk factor and a value risk factor; The load risk factor is a load risk value obtained by calculating the load change of each endpoint after the mapping request based on the load matrix and the initial endpoint mapping strategy; The value-risk factor is a quantified value-risk value of each endpoint mapping decision based on the value of the request and the processing capability of the endpoint; The initial endpoint mapping policy is a set of endpoint mapping decisions.

5. The method for automatically processing the entire process of business leads based on SPM as claimed in claim 4, characterized in that: The specific steps of S3 also include: S3.7: Consider endpoint mapping decision-making as a game process, where endpoints serve as game players and clue features serve as factors influencing the game outcome. S3.8: For each endpoint and each clue feature combination, calculate the utility value of selecting the endpoint for mapping under that combination. Based on the calculated utility values, construct the endpoint-clue game utility matrix; S3.9: Find the Nash equilibrium solution based on the endpoint-clue game utility matrix and use the Nash equilibrium solution as a constraint; S3.10: Load risk and value risk are combined to form a multi-objective space, where the horizontal axis represents load risk and the vertical axis represents value risk, and the initial endpoint mapping strategy corresponds to an initial solution in the multi-objective space; S3.11: Use a non-dominated sorting genetic algorithm to find the Pareto front in a multi-objective space and select the optimal endpoint mapping strategy from the Pareto front based on the Nash negotiation solution; S3.12: Generate an executable endpoint mapping instruction table based on the optimal endpoint mapping strategy.

6. The method for automatically processing the entire process of business leads based on SPM as claimed in claim 5, characterized in that: The specific steps of S3.9 include: S3.9.1: Obtain an endpoint-cue game utility matrix. Based on the endpoint-cue game utility matrix, determine the game participants and strategy sets, where the game participants are each endpoint and the strategy set is a combination of all mappable clue features. Each row in the endpoint-cue game utility matrix represents the strategy set of an endpoint. S3.9.2: The endpoint-clue game utility matrix is ​​used as the core data of the game model. For each endpoint, its payment function is the corresponding utility value, that is, the payment when the i-th endpoint selects the j-th clue feature combination is ; S3.9.3: For each row of the endpoint-cue game utility matrix, find the maximum utility value in that row; S3.9.4: Repeat for each row and find their optimal choice under the strategy set; S3.9.5: If there is a clue feature combination , so that for all endpoints i, the clue feature combination is selected The maximum utility can be obtained when , that is, for any clue feature combination j, ,but is a pure strategy Nash equilibrium solution; S3.9.6: If there is no pure strategy Nash equilibrium solution, then find a mixed strategy Nash equilibrium, including: Assume that the probability of endpoint i selecting clue feature combination j is ; Set the hybrid policy combination for all endpoints to ,in, represents the probability distribution of the mixed strategy of the mth endpoint, and satisfies ; For endpoint i, its expected utility function is for and The product of the clue feature combination The cumulative sum under , where Represents endpoint i selection clue feature combination The probability of Represents endpoint i selection clue feature combination The utility value of represents the clue feature combination selected by endpoint i; For each endpoint i, solve the mixed strategy that maximizes its expected utility when the mixed strategy of non-self endpoints is given, that is, solve the optimal response function ,in, Indicates logical NOT, represents a non-endpoint i, represents the mixed strategy combination of non-endpoint i, represents the optimal response mixed strategy of endpoint i; If satisfied , then we get the mixed strategy Nash equilibrium solution; S3.9.7: Use pure strategy Nash equilibrium solutions or mixed strategy Nash equilibrium solutions as constraints; If a pure strategy Nash equilibrium solution is found, the constraint is to select a clue feature combination for each endpoint ; If a mixed strategy Nash equilibrium solution is found, the constraint condition is that each endpoint has its mixed strategy probability Select clue feature combination .

7. The method for automatically processing business leads based on SPM as claimed in claim 6, characterized in that: The customized dynamic interactive content stream is generated based on the clue feature vector as described in S4 and pushed to the endpoint or clue subject, including: S4.1: Determine the target endpoint or lead subject based on the optimal endpoint mapping strategy and deploy a context responder for the endpoint or lead subject; S4.2: Obtain a clue feature vector; the clue feature vector includes the predicted probability of immediate responsiveness, context fit, and conversion potential value; S4.3: Extract contextual fit information from the clue feature vector, screen the content template library based on the extracted contextual fit information, and select the content template with the highest contextual fit; S4.4: Use the conversion potential value in the clue feature vector to classify the content information density into different levels; S4.5: Set the content push frequency based on the predicted probability of immediate responsiveness in the clue feature vector; S4.6: Based on the divided content information density levels, the content information is filled into the selected content template with the highest contextual fit. At the same time, the content flow is dynamically adjusted based on the real-time feedback of the clue subject during the interaction process; S4.7: Combine the filled and adjusted content in order and logic to form a customized dynamic interactive content flow, and push the generated customized dynamic interactive content flow to the target endpoint or lead subject according to the set push frequency.

8. The method for automatically processing the entire process of business leads based on SPM as claimed in claim 7, characterized in that: The distributed heterogeneous data interface supports multiple data protocols; the multi-source clue trajectory data includes structured data, semi-structured data and unstructured data; the cross-system federated feature pool uses differential privacy technology to desensitize feature vectors, and realizes cross-system feature dimension alignment through feature hashing algorithm; the basic feature combination includes clue source channel, clue generation time, clue subject attributes and historical interaction records; the clue life cycle tracking chain is built based on blockchain technology; the clue life cycle tracking chain includes full-process nodes of clue creation, feature enhancement, endpoint mapping, interaction push and status update, and each node records operation time, operation subject and data change log.

9. A full-process automatic processing system for business leads based on SPM, which is used to implement the full-process automatic processing method for business leads based on SPM according to any one of claims 1 to 8, characterized in that: include: Feature construction module, enhancement and evaluation module, endpoint mapping module, interaction and feedback module, and clue status adjustment module; The feature construction module collects multi-source clue trajectory data through a distributed heterogeneous data interface, constructs a cross-system federated feature pool, and forms an original clue unit; The enhancement and evaluation module uses a dynamic evolution modeling engine to perform feature enhancement on the original clue unit and calculates a three-dimensional value evaluation vector to obtain a clue feature vector; The endpoint mapping module is used to obtain the endpoint load matrix, combine it with the clue feature vector, and generate the optimal endpoint mapping strategy through the reinforcement learning algorithm and the risk balance algorithm; The interaction and feedback module deploys context responders according to the optimal endpoint mapping strategy, generates and pushes customized dynamic interactive content streams, and collects interactive behavior feedback data; The lead state adjustment module is used to establish a lead life cycle tracking chain and dynamically adjust the transition probability matrix and terminal state probability in the pre-built lead state probability model based on the interaction behavior feedback data.

10. The SPM-based full-process automatic processing system for business leads according to claim 9, characterized in that: The endpoint mapping module includes: a load data acquisition unit, an initial strategy generation unit, and a game optimization unit; The load data acquisition unit is used to acquire the load matrix of each endpoint in real time; The initial strategy generating unit generates an initial endpoint mapping strategy using a reinforcement learning algorithm; The game optimization unit uses a risk balance algorithm to optimize the initial endpoint mapping strategy and outputs the optimal strategy in combination with an adaptive weight adjustment mechanism.

Images