Customer relationship management system based on big data

Abstract

The invention belongs to the field of customer relationship management based on big data, and particularly relates to a customer relationship management system based on big data, which comprises a multi-source sensing and acquisition module; a scene modeling module; a scene prediction and decision module; a strategy execution and closed loop optimization module; according to the method, a closed loop of acquisition-modeling-prediction decision-execution-attribution retraining is constructed, multi-source data is opened to form global customer portraits and scene representation, and a strategy list of human, time and place is generated according to individual probability evaluation and risk score and is executed in real time, so that personalized marketing is realized, resource waste is reduced, and the marketing efficiency is improved. And furthermore, customer loyalty is remarkably improved, a wastage rate is reduced, supply and demand matching is improved, inventory turnover is accelerated, promotion homogenization is avoided, and stable increase of revenue is promoted.

Description

Technical Field

[0001] This invention belongs to the field of big data customer relationship management, and specifically relates to a big data-based customer relationship management system. Background Technology

[0002] The Chinese patent application number is CN202411939071.8, which discloses a big data-based customer relationship management system. This invention provides a big data-based customer relationship management system, comprising: a customer segmentation module for classifying customers into different customer types based on the enterprise's industry characteristics, market positioning, and business objectives; a customer maintenance analysis module for acquiring and analyzing customer maintenance-related information to obtain customer analysis results; and a follow-up personnel configuration module for receiving customer analysis results, configuring personnel to obtain corresponding follow-up personnel, and sending the corresponding customer maintenance list to the follow-up personnel's smart terminals.

[0003] While this invention, in the field of information technology, utilizes big data technology to comprehensively collect customer information, accurately classifies customer types through complex algorithms, and precisely identifies expected and potential customers, providing enterprises with a basis for targeted marketing strategies, effectively improving market expansion efficiency, and enhancing enterprise competitiveness in the market, the following problems still exist in the retail industry: Due to the diversity of customers in the retail industry, customer behavior is complex, and existing customer relationship management systems cannot support personalized marketing, leading to wasted resources, decreased conversion and repurchase rates, lower loyalty, and increased customer churn. Furthermore, existing big data management systems, due to data silos, difficulty in data integration, and insufficient analytical capabilities, cannot form a comprehensive customer profile and provide real-time recommendations. This leads to an imbalance between supply and demand, resulting in slow inventory turnover and homogenized promotions, ultimately hindering revenue growth.

[0004] To address the aforementioned issues, this invention proposes a big data-based customer relationship management system. This system constructs a closed loop encompassing data collection, modeling, predictive decision-making, execution, and attribution retraining. It integrates multi-source data to create a comprehensive customer profile and contextual representation, and generates personalized strategy lists based on individual probability assessments and risk scores, executing them in real time. Summary of the Invention

[0005] This invention provides a big data-based customer relationship management system to address the following challenges: In the retail industry, the diversity of customers leads to the complexity of customer behavior. Existing customer relationship management systems cannot support personalized marketing, resulting in wasted resources, decreased conversion and repurchase rates, lower loyalty, and increased customer churn. Existing big data management systems, due to data silos, difficulty in data integration, and insufficient analytical capabilities, cannot form a comprehensive customer profile and provide real-time recommendations. This leads to an imbalance between supply and demand, resulting in slow inventory turnover and homogenized promotions, ultimately hindering revenue growth.

[0006] The technical solution adopted by the present invention to solve the above problems is as follows: a customer relationship management system based on big data, including: a multi-source perception and acquisition module: accessing external data and performing identity mapping and security authentication, and adopting an access governance mechanism for external data calls to obtain standardized raw event streams and unified timeline data sequences and inputting them into the scenario modeling module;

[0007] Context modeling module: Performs spatiotemporal matching and geofencing judgment on location trajectory data and geographic object database, and performs conversationalization, semantic completion and relation graph construction on event flow to obtain context representation and object representation vectors, which are then input into the context prediction and decision module;

[0008] Scenario prediction and decision-making module: Perform multi-task learning and uncertainty calibration on scenario representation and object representation vectors to obtain probabilistic assessments of behavior and business results and corresponding risk adjustment scores. Based on the probabilistic assessment and risk score, construct constraint functions, generate a list of strategies tailored to individuals, time and place, and distribute them to the strategy execution and closed-loop optimization modules.

[0009] Strategy execution and closed-loop optimization module: The strategy list is orchestrated by resources and the execution control and controlled evaluation methods are used to obtain the effect data and evaluation results that can be used to predict the model and strategy iteration. The data is then fed back to the preceding module and the model and strategy iteration is triggered.

[0010] As a preferred embodiment, the specific steps of the multi-source sensing and acquisition module are as follows:

[0011] The identity mapping unit obtains a unique, cross-platform verifiable customer ID by performing deterministic rule matching and probabilistic verification on device identifiers and account identifiers.

[0012] By collecting device and account identifiers, common-origin features based on time, location, network fingerprints, and login credentials are constructed. Deterministic matching is performed based on the same mobile phone number, login credentials, and device fingerprint. The matching probability is calculated using logistic regression, and the matching probability G of candidate ID pairs is compared with set primary and secondary thresholds P and PP. Generate a unified customer ID and update the mapping table in real time. Manual review is conducted at that time. If necessary, the merged IDs are discarded and written to the audit log, resulting in a traceable and rollbackable merged ID.

[0013] As a preferred embodiment, the specific steps of the scenario modeling module are as follows:

[0014] The spatiotemporal matching and geofencing determination unit collects location point sequences via GPS, performs noise reduction and smoothing, and matches customer movement paths and dwell behaviors with the map to obtain customer dwell points. It constructs a geofencing index based on the geometric boundaries and buffer radii of the POI dictionary, performs point-polygon determination point by point, and sets dwell thresholds to filter dwell points on the map, removing dwell points below the dwell threshold to obtain the session sequence.

[0015] As a preferred implementation, the sequence organization and context aggregation unit aggregates customer visit events and behavioral events along a unified time axis according to GPS trajectories and constructs touchpoint windows and path contexts based on customer dwell points to obtain sequentially recoverable session sequences and context features.

[0016] As a preferred embodiment, the specific steps of the scenario modeling module are as follows:

[0017] The semantic enhancement and consistency correction unit performs missing value imputation, standardizes definitions and eliminates overlapping customer conflicts and alias mapping for product types, purchasing stores and purchasing channels by performing missing value imputation on session sequences and context features, and injects external data to obtain a consistent set of elements and a standardized attribute dictionary of people-time-place-event-environment.

[0018] As a preferred embodiment, the specific steps of the scenario modeling module are as follows:

[0019] Spatiotemporal attention weighting is introduced into the relation representation construction and topological indexing units. By jointly weighting the neighborhood of candidate events spatiotemporally and relationally, a context representation that can be used for online retrieval and inference is obtained. The spatiotemporal attention weighting formula is as follows:

[0020] ,

[0021] Indicates an event The context representation vector, S is the representation vector, and a is the target event being represented. Let b be the set of candidate neighborhoods, and let b be a neighboring event in the neighborhood of the target event. For attention weights, Let D be the value vector of neighbor event b, and let D be the value vector. This is the summation symbol.

[0022] As a preferred implementation, the specific steps of the scenario prediction and decision-making module are as follows:

[0023] The object representation vector generation and interpretation unit performs numerical standardization, feature masking, and multi-view embedding on the elements and hierarchical labels of people, time, place, event, and environment, and uses learnable weight fusion and vector normalization to obtain the object representation vector. It then uses integral gradient to decompose the object representation into contribution values, establishes a mapping between the object representation vector and upstream features, and outputs a feature contribution list sorted by score and the impact of contribution values ​​on customers.

[0024] As a preferred implementation, the specific steps of the scenario prediction and decision-making module are as follows:

[0025] After receiving data from the uncertain calibration unit, the incremental revenue and risk-adjusted scoring unit estimates the individual probability assessment of customers and the risk score of shops by constructing an incremental revenue scoring algorithm formula, whereby the incremental revenue scoring algorithm formula is:

[0026] ,

[0027] For individual probability assessment and risk scoring, This represents the incremental revenue of sample i in channel j. For incremental revenue, i represents the sample, and j represents the channel. Let i be a measure of the uncertainty of sample i in channel j. As a measure of uncertainty, This is the cost weighting coefficient. Let F be the unit cost of sample i in channel j, and let F be the unit cost. This is the dot product symbol.

[0028] As a preferred implementation, the specific steps of the scenario prediction and decision-making module are as follows:

[0029] The constraint function construction unit constructs a Lagrange relaxation multi-constraint deployment selection method by incorporating budget, inventory, and channel capacity into the constraint function and using incremental revenue scores as constraint terms. The objective function of the constraint is obtained, and its algorithmic formula is as follows:

[0030] ,

[0031] Let ∑ represent the constrained objective function value, ∑ be the summation symbol, i be the sample, and k be the channel. This represents the incremental revenue of sample i in channel j. For incremental revenue, Let r represent the decision variable for sample i in channel j. This is the penalty coefficient for budget constraints. Let q be the unit cost of sample i in sample j, q be the unit cost, and K be the budget ceiling. Let u be the Lagrange multiplier for channel j, and V be the upper limit of the channel. Let S be the Lagrange multiplier for sample i, S be the Lagrange multiplier, and E be the upper limit of individual frequency control.

[0032] As a preferred implementation, the specific steps of the strategy execution and closed-loop optimization module are as follows:

[0033] By conducting gray-scale and A / B controlled experiments and recording individual-level operation receipts, structured execution logs and link feedback are obtained. By calculating the contribution of touchpoints on the receipt path and allocating them in combination with the path sequence, the marginal contribution of channels and touchpoints is obtained. By performing causal estimation on the treatment group and control group and outputting effect size and confidence interval, the net effect evaluation results of the strategy on business indicators are obtained. By writing behavioral and effect data back to the feature library and triggering retraining and parameter updates, a predictive model consistent with the data distribution is obtained. By accumulating failure examples and blacklist rules and performing version management, a reusable strategy knowledge base is obtained.

[0034] The beneficial effects of this invention are as follows:

[0035] 1. By introducing spatiotemporal attention weighting into relational representation construction and topological indexing units, and coordinating with sequence organization and context aggregation, semantic enhancement and consistency correction, a stable and interpretable context representation and an online searchable neighborhood representation are obtained. This enables the integration of heterogeneous data, accurate reconstruction of human-time-place-event-environment scenarios, and significant suppression of noise and cold start bias.

[0036] 2. By employing an incremental benefit evaluation algorithm with uncertainty calibration in the scenario prediction and decision-making modules, combined with multi-constraint optimization using Lagrange relaxation, a list of personalized, time-based, and location-based strategies that can be directly issued under constraints of budget, capacity, and frequency control is obtained. This can improve conversion efficiency and ROI while reducing ineffective outreach and resource waste under the same resources, thereby achieving traceable and auditable optimal decision-making. Attached Figure Description

[0037] Figure 1 This is a system module diagram of the present invention.

[0038] Figure 2 These are technical effect diagrams of the present invention, wherein the black bar chart represents the present invention, and the gray bar chart represents the prior art. Detailed Implementation

[0039] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0040] Example 1 Figure 1 A system module diagram of a big data-based customer relationship management system includes: a multi-source sensing and acquisition module; a scenario modeling module; a scenario prediction and decision-making module; and a strategy execution and closed-loop optimization module. The specific implementation steps are as follows:

[0041] Multi-source perception and acquisition module: By accessing external data and performing identity mapping and security authentication, and by adopting an access governance mechanism for external data calls, a standardized raw event stream and a unified timeline data sequence are obtained and input into the scenario modeling module;

[0042] Context modeling module: By performing spatiotemporal matching and geofencing judgment on location trajectory data and geographic object database, and by performing conversationalization, semantic completion and relation graph construction on event flow, context representation and object representation vectors are obtained and input into the context prediction and decision module;

[0043] Scenario prediction and decision-making module: By performing multi-task learning and uncertainty calibration on scenario representation and object representation vectors, it obtains probabilistic assessments of behavior and business results and corresponding risk adjustment scores. Based on individual probability assessments and risk scores, it constructs constraint functions, generates a list of strategies tailored to individuals, times, and locations, and distributes it to the strategy execution and closed-loop optimization modules.

[0044] Strategy execution and closed-loop optimization module: By using resource orchestration and execution control and controlled evaluation methods on the strategy list, it obtains effect data and evaluation results that can be used to predict the model and strategy iteration, which are then fed back to the preceding module and trigger the model and strategy iteration.

[0045] Example 2, this example is an explanation of Example 1, the multi-source perception module includes: an external data access unit; an identity mapping unit; a security authentication unit; an access governance unit; and an audit and link tracing unit;

[0046] The external data access unit connects the mobile terminal and IoT device SDK and gateway to obtain a standardized raw event stream, thereby unifying the entry point of multiple collection terminals and stably reporting the raw event stream; the standardized raw event stream refers to a parsable data sequence that is encoded according to unified field specifications, timestamps, customer IDs and data types, and outputs unprocessed events from each collection terminal in the order of occurrence.

[0047] The identity mapping unit performs deterministic rule matching and probability verification on device identifiers and account identifiers to obtain a unique customer ID that can be verified across terminals, thereby eliminating identity fragmentation and improving the accuracy of downstream associations.

[0048] Specifically, the steps for performing deterministic rule matching and probability verification on device identifiers and account identifiers are as follows: By collecting device identifiers and account identifiers, constructing common-source features such as time, location, network fingerprints, and login credentials, performing deterministic matching based on the same mobile phone number, login credentials, and device fingerprint, and combining this with logistic regression to calculate the matching probability, merging IDs according to a threshold, and recording audits, a cross-platform consistent unique customer ID and traceable matching credentials are obtained; the steps for merging IDs according to a threshold and recording audits are as follows: by comparing the matching probability G of candidate ID pairs with a set primary threshold P and a secondary threshold PP, ... Generate a unified customer ID and update the mapping table in real time. Manual review is conducted at that time. If necessary, the merged IDs are discarded and written to the audit log, thus obtaining a traceable and rollbackable ID merging result;

[0049] When the upstream of the security authentication unit initiates a call request, it performs digital signatures on the message and payload, and verifies the identity and resource access permissions of the calling entity, thereby obtaining a traceable data zipper containing the request ID, timestamp, signature and verification result, which can ensure compliance of collection and prevent tampering and unauthorized access.

[0050] The access governance unit performs traffic control, time domain alignment, and consistency assurance on the securely authenticated event records, and performs field-level desensitization and differential privacy perturbation on the privacy fields. It also performs access control, transmission encryption, storage encryption, and auditing and auditing for compliance control, thereby obtaining a standardized original event stream and a unified timeline data sequence, which is then sent to the scenario modeling module.

[0051] The audit and traceability unit records the request identifier, timestamp, signature and verification results of data calls to obtain data chains and audit evidence sets, which can support compliance verification, problem location and rollback recovery.

[0052] Example 3 is an explanation of Example 1. The scenario modeling module includes: a spatiotemporal matching and geofencing determination unit; a sequence organization and context aggregation unit; a semantic enhancement and consistency correction unit; a relation representation construction and topology indexing unit; and an object representation vector generation and interpretation unit.

[0053] The aforementioned spatiotemporal matching and geofencing determination unit performs matching and geofencing determination using the GPS trajectory and Point of Interest (POI) dictionary to obtain "human-land" relationship pairs, thereby enabling accurate location of touchpoint scenarios;

[0054] Specifically, the steps for matching GPS trajectory with the POI dictionary and performing geofencing are as follows: GPS location point sequences are collected, denoised, and smoothed. Customer stop points are obtained by matching their movement paths and dwelling behaviors with the map. A geofencing index is constructed using the geometric boundaries and buffer radii of the POI dictionary. Point-polygon determination is performed point by point, and a dwelling threshold is set to filter dwell points on the map, removing those below the threshold to obtain a session sequence. This sequence transforms the person-place-time relationship into a computable entity. The location point sequence includes fields such as timestamp, latitude and longitude, speed, and positioning accuracy to describe the customer's movement paths and dwelling behaviors. The session sequence includes visit events such as user identifier, point of interest identifier, entry time, exit time, and dwell duration.

[0055] The sequence organization and context aggregation unit aggregates customer visit events and behavioral events along a unified time axis according to GPS trajectories and constructs touchpoint windows and path contexts based on customer dwell points, thereby obtaining sequentially recoverable session sequences and context features, which can support semantic enhancement and relation representation.

[0056] The semantic enhancement and consistency correction unit performs missing value imputation, standardizes and eliminates overlapping customer conflicts and alias mapping for product types, purchasing stores and purchasing channels by performing missing value imputation on the conversation sequence and context features, and injects external data, thereby obtaining a consistent set of elements and a standardized attribute dictionary of people-time-place-event-environment, which can improve the context restoration and provide high-quality input for relation representation construction.

[0057] The relation representation construction and topology indexing unit constructs a heterogeneous relation graph using people-time-location-event-loop as nodes, generates weighted directed edges for customer visit events and repeat purchases in the session sequence, and solidifies event time windows and version tags. It also constructs a multi-level topology query index for the heterogeneous relation graph according to the session sequence, enabling relation retrieval and path enumeration capabilities that can be combined by entity, time, and spatial conditions, thereby improving the scene restoration accuracy and noise resistance. Furthermore, it introduces spatiotemporal attention weighting into the relation representation construction and topology indexing unit.

[0058] Specifically, the steps for introducing spatiotemporal attention weighting in relation representation construction and topology indexing units are as follows: By jointly weighting the neighborhood of candidate events spatiotemporally and relationally, a context representation that can be used for online retrieval and inference is obtained, thereby improving the context restoration accuracy and noise resistance. The spatiotemporal attention weighting formula is as follows:

[0059] ,

[0060] Indicates an event The context representation vector, S is the representation vector, and a is the target event being represented. Let b be the set of candidate neighborhoods, and let b be a neighboring event in the neighborhood of the target event. For attention weights, Let D be the value vector of neighbor event b, and let D be the value vector. The summation symbol;

[0061] After constructing the spatiotemporal attention weighting formula, higher weights can be assigned to neighborhoods that are temporally close, spatially close, and have stronger relationships, thereby obtaining a stable context representation, which can be used together with the object representation vector for subsequent learning and decision-making.

[0062] The object representation vector generation and interpretation unit performs numerical standardization, feature masking, and multi-view embedding on the elements and hierarchical labels of person-time-place-event-environment, and adopts learnable weight fusion and vector normalization to obtain the object representation vector. It then uses integral gradient to decompose the object representation into contribution values, establishes a mapping between the object representation vector and upstream features, and outputs a feature contribution list sorted by score and the impact of contribution values ​​on customers. For example, if the season is winter, the POI is a shopping mall, the shopping mall stay duration is 30 minutes, and it is a weekday evening, this indicates that the customer is more likely to go to the shopping mall to buy items on winter evenings. The context representation and object representation vectors are then sent to the context prediction and decision-making module.

[0063] The aforementioned element and hierarchical labels for people-time-location-event-environment are normalized using execution features and multi-view embedding. By applying numerical standardization, category coding, and time decay weighting to the original features of the elements and hierarchical labels for people-time-location-event-environment, a standardized feature set with consistent scale and uniform caliber is obtained. Then, attribute embedding, behavior sequence embedding, and relational neighborhood embedding are applied to the standardized feature set to obtain a multi-view vector representation covering static attributes and dynamic behaviors.

[0064] The aforementioned learnable weight fusion method applies attention weighting to the vector representations of multiple views and performs vector normalization to obtain object representation vectors.

[0065] Example 4 is an explanation of Example 1. The scenario prediction and decision-making module includes: a multi-task learning unit; an uncertainty calibration unit; an incremental revenue and risk adjustment scoring unit; a constraint function construction unit; and a multi-objective optimization and strategy list generation unit.

[0066] The multi-task learning unit concatenates context representation and object representation vectors in a fixed order, such as context representation + object representation, and performs numerical standardization and feature masking to obtain a unified input vector. This unified input vector is then input to a shared representation encoder to map a shared latent representation, thus providing a common feature base for multi-task learning. The extracted shared latent representation is then subjected to task adaptive decomposition via a task weighter. Gating coefficients are calculated for each task, and the outputs of the task weighter are weighted and fused to obtain task-specific representations. These task-specific representations are then normalized using a task head to obtain raw probabilities, which provide explicit probabilities for downstream evaluation and ranking. These raw probabilities are then input to an uncertainty calibration unit.

[0067] The aforementioned task-specific representation, through task-gated weighted fusion of shared latent vectors in a multi-task learning unit, enables dense vectors representing specific tasks, such as customer intent, needs, conversions, and churn, to be used for probability prediction, threshold management, and uncertainty calibration.

[0068] The uncertainty calibration unit receives the original probability, performs calibration using temperature calibration and entropy uncertainty measurement, and generates a stable trigger signal with threshold management to obtain a prediction result with consistent confidence.

[0069] Specifically, the steps for calibration using temperature calibration and entropy-based uncertainty measurement, and for generating stable trigger signals through threshold management, are as follows: The original probability is received and inversely calculated to obtain the original output score. A temperature parameter TEP is introduced, and the original score is scaled to obtain the calibrated original probability P. Normalization is then performed to calculate the sample uncertainty AP. A primary threshold T, a secondary threshold TT, and a sample uncertainty threshold AUP are set for judgment. When… When it is, a trigger signal is output; when If the signal is pending review, it is output; otherwise, it is discarded, and the temperature parameter TEP and threshold version are recorded and sent to the incremental return and risk adjustment scoring unit.

[0070] After receiving data from the uncertain calibration unit, the incremental revenue and risk adjustment scoring unit estimates the individual probability assessment of customers and the risk score of shops by constructing an incremental revenue scoring algorithm formula. This allows for increased actual incremental revenue and reduced false triggering rates within the same budget. The incremental revenue scoring algorithm formula is as follows:

[0071] ,

[0072] For individual probability assessment and risk scoring, This represents the incremental revenue of sample i in channel j. For incremental revenue, i represents the sample, such as a customer session, and j represents the channel, such as SMS and email. Let i be a measure of the uncertainty of sample i in channel j. As a measure of uncertainty, This is the cost weighting coefficient. Let F be the unit cost of sample i in channel j, and let F be the unit cost. The dot product symbol;

[0073] The calculated incremental revenue score is sent to the constraint function construction unit. By using the above formula, a ranking basis that takes into account both incremental revenue and uncertainty is obtained, which improves the stability of prediction and the controllability of business.

[0074] The constraint function construction unit constructs a Lagrange relaxation multi-constraint deployment selection method by writing budget, inventory, and channel capacity into the constraint function and using incremental revenue scores as constraint terms. This yields the objective function of the constraints, providing an executable solution space for the solver. The algorithmic formula for the objective function of the constraints is as follows:

[0075] ,

[0076] Let represent the constrained objective function value, ∑ represent the summation symbol, i represent the sample (e.g., customer conversation), and k represent the channel (e.g., SMS and email). This represents the incremental revenue of sample i in channel j. For incremental revenue, Let r represent the decision variable for sample i in channel j. This is the penalty coefficient for budget constraints. Let q be the unit cost of sample i in sample j, q be the unit cost, and K be the budget ceiling. Let u be the Lagrange multiplier for channel j, and V be the upper limit of the channel. Let S be the Lagrange multiplier for sample i, S be the Lagrange multiplier, and E be the upper limit of individual frequency control.

[0077] The multi-objective optimization and strategy list generation unit solves the problem jointly using an incremental benefit scoring algorithm and an objective function algorithm formula, and outputs selection results that satisfy the constraints. This yields a strategy list tailored to individuals, time, and location, enabling executable optimal decision-making under resource-constrained conditions.

[0078] Example 5, which explains Example 1, describes a strategy execution and closed-loop optimization module that includes: a resource orchestration and execution control unit; a load generation and delivery triggering unit; a controlled evaluation and receipt collection unit; a multi-touchpoint and path attribution unit; a causal effect identification unit; a feature backflow and model retraining unit; and a strategy knowledge accumulation and version management unit.

[0079] The resource orchestration and execution control unit maps the strategy list to channels such as SMS, App push, official account, WeChat Work, email and offline POS and performs concurrent scheduling to obtain a consistent execution sequence across channels, thereby ensuring that the implementation strategy can be implemented on time.

[0080] The load generation and delivery triggering unit fills in the implemented content and drives delivery with timed and event triggers, thereby obtaining a delivery load that meets the constraints of people-time-place, which can improve the reach and relevance.

[0081] The controlled evaluation and feedback collection unit, through grayscale and A / B comparison experiments and recording individual-level operation feedback, obtains structured execution logs and link feedback, which can accumulate evidence chains and support effect measurement; the individual-level operation feedback includes: delivery, opening, clicking, order placement and repeat purchase feedback.

[0082] The multi-touchpoint and path attribution unit calculates the touchpoint contribution on the receipt path and allocates it in combination with the path sequence to obtain the marginal contribution of the channel and the touchpoint, which can provide a quantifiable basis for resource reallocation.

[0083] The causal effect identification unit performs causal estimation on the treatment group and the control group and outputs the effect size and confidence interval, thereby obtaining the net effect evaluation result of the strategy on the operating indicators, which can improve the reliability and robustness of strategy selection.

[0084] The feature backflow and model retraining unit writes customer behavior and strategy implementation effect data back to the feature library and triggers retraining to update parameters, thereby obtaining a prediction model consistent with the data distribution. This can suppress model aging and improve prediction accuracy. The prediction model takes scenario representation and object representation vectors as input and adopts an online inference model with multi-task learning and uncertainty calibration to output probabilistic assessments of individual-level behavior and business results and corresponding risk adjustment scores.

[0085] The strategy knowledge accumulation and version management unit accumulates failure examples and blacklist rules and manages them in a versioned manner to obtain a reusable strategy knowledge base, thereby shortening the deployment cycle of predictive models. The failure examples refer to strategies that fail to reach the preset threshold and have negative effects during execution. If a strategy with negative effects is adopted, customers will still be lost, and the store's efficiency conversion rate will be very low. The blacklist rules are a restricted list set for specific subjects, such as users, devices, and accounts, based on compliance, unsubscription, risk, and quality control conditions. The blacklist rules include the scope of effectiveness and the time limit and are used to forcibly block access during the decision-making and execution stages.

[0086] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A big data-based customer relationship management system, characterized in that: a multi-source perception and acquisition module: accesses external data and performs identity mapping and security authentication, and adopts an access governance mechanism for external data calling, obtains standardized original event stream and unified time axis data sequence, and inputs into a scenario modeling module; a scenario modeling module: performs space-time matching and geographic fence judgment on positioning trajectory data and geographic object library, and performs sessionization, semantic completion and relationship graph construction on event stream, obtains scenario representation and object representation vector, and inputs into a scenario prediction and decision module; a scenario prediction and decision module: performs multi-task learning and uncertainty calibration on scenario representation and object representation vector, obtains probabilistic evaluation of behavior and business results and corresponding risk adjustment score, constructs a constraint function based on probability evaluation and risk score, generates a person-time-place-specific strategy list and issues it to a strategy execution and closed-loop optimization module; a strategy execution and closed-loop optimization module: adopts resource orchestration and performs control and controlled evaluation method on the strategy list, obtains effect data and evaluation results that can be used for prediction model and strategy iteration, and returns to the previous module and triggers model and strategy iteration.

2. The big data based customer relationship management system as claimed in claim 1, wherein: The specific steps of the multi-source perception and acquisition module are: an identity mapping unit obtains a cross-end verifiable customer unique ID by performing deterministic rule matching and probability checking on device identification and account identification; By collecting the device identifier and the account identifier, the homologous features of time, location, network fingerprint and login credential are constructed, the deterministic matching of the same mobile number, the same login credential and the same device fingerprint is performed, the matching probability is calculated by combining the logistic regression, and the matching probability G of the candidate ID pair is compared with the set first threshold P and second threshold PP, the unified customer ID is generated and the mapping table is updated when , manual review is performed when , and discarded when , and the ID merging result which is traceable and rollbackable is obtained.

3. The big data based customer relationship management system as claimed in claim 1, wherein: The specific steps of the scenario modeling module are: a space-time matching and geographic fence judgment unit performs denoising and smoothing on the positioning point sequence collected by GPS, matches the customer's moving path and stay behavior with the map to obtain the customer's stay points, constructs a geographic fence index from the geometric boundary and buffer radius of the POI dictionary, performs point-polygon judgment point by point, and sets a stay threshold to filter out stay points below the stay threshold, obtaining a session sequence.

4. The big data based customer relationship management system as claimed in claim 1, wherein: The specific steps of the scenario modeling module are: a sequence organization and context aggregation unit aggregates the customer's visit events and behavior events along the GPS trajectory according to the customer's stay points and constructs touch point windows and path context, obtaining a sequentially recoverable session sequence and context features.

5. The big data based customer relationship management system as claimed in claim 1, wherein: The specific steps of the scenario modeling module are: a semantic enhancement and consistency correction unit performs missing value interpolation, unified scale, and eliminates overlapping customer conflicts and alias mapping of product categories, purchase stores and purchase channels on the session sequence and context features, and injects external data to obtain a consistent person-time-place-thing-circle element set and standardized attribute dictionary.

6. The big data based customer relationship management system as claimed in claim 1, wherein: In the relationship representation construction and topology index unit, spatio-temporal attention weighting is introduced, which performs spatio-temporal and relationship joint weighting on the candidate event neighborhood to obtain scenario representation that can be used for online retrieval and inference, wherein the spatio-temporal attention weighting formula is: The specific steps of the scenario prediction and decision module are: ， representing an event scenario representation vector, S is a representation vector, a is a target event being represented, is a set of candidate domains, b is a certain neighbor event in the neighborhood of the target event, is an attention weight, is a value vector of the neighbor event b, D is a value vector, is a summation symbol.

7. The big data based customer relationship management system as claimed in claim 1, wherein: ​ The object representation vector generation and interpretation unit obtains the object representation vector by performing numerical standardization and feature mask and multi-view embedding on the elements and hierarchical labels of people-time-place-thing-circle, and adopting learnable weight fusion and vector normalization; and decomposes the contribution degree of the object representation by integral gradient, and maps the object representation vector and the upstream features to output the feature contribution list sorted by score and the impact of the contribution degree on the customer.

8. The big data based customer relationship management system as claimed in claim 1, wherein: The specific steps of the scenario prediction and decision module are: The incremental return and risk adjustment scoring unit estimates the individual probability evaluation of the customer and the risk score of the store by constructing an incremental return scoring algorithm formula after receiving the data of the uncertainty calibration unit, wherein the incremental return scoring algorithm formula is: ， for individual probability assessment and risk score, represents the incremental benefit of sample i in channel j, is the incremental benefit, i is the sample, and j is the channel, is the uncertainty measure of sample i in channel j, is the uncertainty measure, is the cost weight coefficient, is the unit cost of sample i in channel j, and F is the unit cost, is the dot product symbol.

9. The big data based customer relationship management system as claimed in claim 1, wherein: The specific steps of the scenario prediction and decision module are: The constraint function construction unit writes the budget, inventory and channel capacity into the constraint function, and constructs a multi-constraint release selection method of Lagrange relaxation by taking the incremental return score as a constraint term to obtain a constrained objective function, wherein the algorithm formula of the constrained objective function is: ， for the constraint objective function value, ∑ is a summation symbol, i is a sample, k is a channel, represents the incremental revenue of sample i in channel j, is the incremental revenue, represents the decision variable of sample i in channel j, r is the decision variable, is a budget constraint penalty coefficient, is the unit cost of sample i in sample j, q is the unit cost, K is the upper limit of the budget, is the Lagrange multiplier of channel j, u is the Lagrange multiplier, V is the upper limit of the channel, is the Lagrange multiplier of sample i, S is the Lagrange multiplier, E is the upper limit of the individual frequency control.

10. The big data based customer relationship management system as claimed in claim 1, wherein: The specific steps of the strategy execution and closed-loop optimization module are: Through gray-scale and A / B control experiments and recording of individual-level operation receipts, structured execution logs and link feedback are obtained; Through calculating the contact contribution on the feedback path and combining the path sequence for distribution, the marginal contribution of the channel and the contact is obtained; through causal estimation of the treatment group and the control group and output of the effect size and the confidence interval, the net effect evaluation result of the strategy on the business indicators is obtained; through writing the behavior and effect data back to the feature library and triggering retraining and parameter updating, a prediction model consistent with the data distribution is obtained; through sedimentation of failed examples and black list rules and version management, a reusable strategy knowledge base is obtained.

Images