A big data-based customer relationship management system

Abstract

The application discloses a kind of customer relationship management systems based on big data, it is related to data processing and business management technical field, the system includes: data fusion processing module, for receiving and processing from transaction record inside enterprise, customer service work order, social media interaction data and market research report and macroeconomic index data outside enterprise;Customer dynamic portrait engine, for building and continuously updating customer portrait based on unified customer data view;Intelligent behavior prediction module, for predicting the future behavior tendency of customer based on updated customer portrait;Adaptive strategy generator, for generating personalized marketing or service strategy according to predicted customer behavior probability;Real-time decision execution center, for coordinating the data flow and task scheduling between modules.The application realizes the real-time perception and accurate decision of customer state by integrating multi-source data processing, dynamic portrait updating, behavior prediction and strategy adaptive optimization.

Description

Technical Field

[0001] This invention relates to the fields of data processing and business management technology, and in particular to a customer relationship management system based on big data. Background Technology

[0002] Customer relationship management (CRM) is a core component of modern enterprise operations, aiming to optimize sales processes and improve customer satisfaction through systematic analysis of customer interaction data. With increasing enterprise informatization, customer data is characterized by its massive volume, multiple sources, and dynamic changes. Traditional analysis methods relying on manual experience are no longer sufficient to meet the demands of precision marketing and personalized services.

[0003] Among them, big data-based customer relationship management systems provide data support for enterprise decision-making by collecting, storing, and analyzing customer behavior data. These systems typically involve key technologies such as customer profiling, purchase tendency prediction, and marketing strategy optimization, with the core objective of extracting valuable business insights from massive amounts of data.

[0004] Existing customer relationship management systems (CRM) suffer from significant limitations: insufficient integration capabilities for multi-source heterogeneous data, resulting in single-dimensional and outdated customer profiles; limited ability of traditional statistical analysis models to capture non-linear relationships, making it difficult to accurately predict dynamic changes in customer behavior; and a lack of adaptive adjustment mechanisms to real-time market feedback, as marketing strategy generation relies heavily on historical rules. Furthermore, inefficient data flow between system modules leads to response delays when handling high-concurrency query requests, impacting the timeliness of enterprise decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide a customer relationship management system based on big data to solve the problems in existing customer relationship management systems, such as insufficient integration capabilities of multi-source heterogeneous data, single and lagging customer profile dimensions, limited ability of traditional statistical models to capture nonlinear relationships, lack of real-time adaptive adjustment mechanism for marketing strategies, and low efficiency of data flow between system modules leading to response delays.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A big data-based customer relationship management system, comprising:

[0008] The system includes a data fusion and processing module, a customer dynamic profiling engine, an intelligent behavior prediction module, an adaptive strategy generator, and a real-time decision execution center.

[0009] The data fusion and processing module receives and processes transaction records, customer service work orders, and social media interaction data from within the enterprise, as well as market research reports and macroeconomic indicator data from outside the enterprise. This module first performs standardized preprocessing on the multi-source heterogeneous data, specifically including word segmentation and sentiment analysis of unstructured text data, and missing value imputation and outlier detection and correction for numerical data.

[0010] Furthermore, the data fusion processing module employs data linking technology based on entity parsing and association rules to match and merge the same customer information from different data sources, forming a unified customer data view. This module also includes a built-in time-series data alignment unit, which can unify all customer-related data to the same timestamp benchmark, providing a consistent data foundation for subsequent dynamic analysis.

[0011] The dynamic customer profiling engine connects to the data fusion processing module to build and continuously update customer profiles based on a unified customer data view. This engine comprises static attribute units and dynamic behavior trajectory units. Static attribute units extract and maintain basic customer identity information and historical purchase records. Dynamic behavior trajectory units track and record in real-time customer page browsing paths, content click sequences, service inquiry frequency, and their changing patterns over time. The core of the dynamic customer profiling engine lies in its event-driven update mechanism. Whenever new customer interaction data is input into the data fusion processing module, the engine triggers a profile update calculation, ensuring that the customer profile always reflects its latest state.

[0012] The intelligent behavior prediction module connects to the customer dynamic profile engine to predict future customer behavior trends based on updated customer profiles. This module employs an ensemble learning model architecture, its core consisting of a gradient boosting decision tree algorithm and a long short-term memory neural network algorithm. The gradient boosting decision tree algorithm handles the customer's structured attribute features, uncovering high-order nonlinear interactions between features. The long short-term memory neural network algorithm specifically processes customer behavior sequence data from dynamic behavior trajectory units, capturing their time-dependent patterns. The intelligent behavior prediction module generates a probability value, expressed as a percentage, by weightedly fusing the outputs of these two algorithms to predict whether a customer will make a purchase, churn, or show interest in a specific product within a future specific time window.

[0013] The adaptive strategy generator connects to the intelligent behavior prediction module to generate personalized marketing or service strategies based on predicted customer behavior probabilities. This generator has a pre-built strategy rule base and an effect feedback learning unit. The strategy rule base stores various basic strategy templates, whose trigger conditions are associated with customer behavior probability thresholds. The effect feedback learning unit continuously monitors performance metrics such as the actual conversion rate and customer satisfaction rating of executed strategies. By comparing real-time performance metrics with historical baselines, the adaptive strategy generator dynamically adjusts the priority weights and trigger thresholds of various strategies in the strategy rule base, achieving strategy self-optimization.

[0014] The Real-Time Decision Execution Center, serving as the system's scheduling hub, connects to the data fusion and processing module, the customer dynamic profiling engine, the intelligent behavior prediction module, and the adaptive strategy generator. This center coordinates data flow and task scheduling between these modules. Internally, it features a message queue and stream processing engine to ensure that data is processed in an orderly and low-latency manner under high-concurrency scenarios. The Real-Time Decision Execution Center also provides an application programming interface (API), enabling the generated personalized strategies to be directly pushed to the enterprise's marketing automation platform, customer service system, or sales personnel's mobile terminals, thus completing the closed loop from insight to action.

[0015] In a preferred embodiment of the present invention, the entity parsing technology in the data fusion processing module specifically adopts a two-stage matching strategy combining rule-based matching and fuzzy matching based on machine learning models. In the first stage, the system uses preset precise matching rules, such as identical ID numbers or identical mobile phone numbers, for rapid matching. In the second stage, for records that fail to match precisely, the system uses a pre-trained random forest model to comprehensively calculate the similarity of multiple fields such as name, address, and email. When the comprehensive similarity score exceeds a preset threshold of 85%, it is determined to be the same customer entity.

[0016] In a preferred embodiment of the present invention, the update mechanism of the customer dynamic profile engine is specifically implemented as an incremental update mode. This mode does not recalculate the profile entirely every time new data is available, but rather performs partial updates only on the data dimensions that have changed. The system maintains a version history of a customer profile; each incremental update generates a new profile version and records the version timestamp, thereby supporting retrospective analysis of customer status.

[0017] In a preferred embodiment of the present invention, the weighted fusion strategy of the ensemble learning model in the intelligent behavior prediction module is dynamically adjustable. Initially, the system sets the output weights of the gradient boosting decision tree and the long short-term memory neural network to 50% each. After model deployment, the system periodically, for example weekly, evaluates the predictive performance of the two algorithms on an independent validation dataset and automatically fine-tunes their output weights based on the evaluation results to ensure that the fusion model always maintains optimal performance.

[0018] In a preferred embodiment of the present invention, the effect feedback learning unit of the adaptive strategy generator employs a multi-armed gambling machine algorithm framework for strategy optimization. This unit treats each selectable strategy as a gambling machine arm, and the strategy's execution effect, such as conversion rate, as the reward. By continuously exploring the effects of different strategies on different customer groups and utilizing historical reward information, the algorithm can gradually learn the optimal strategy that yields the highest expected return in a specific context, thereby achieving automation and optimization of strategy allocation.

[0019] In a preferred embodiment of the present invention, the message queue of the real-time decision execution center adopts a publish-subscribe model. Each module within the system acts as a publisher or subscriber of messages, communicating decoupledly through a central message broker. For example, when the customer dynamic profiling engine completes an update, it publishes a message to a specific topic, and the intelligent behavior prediction module, having subscribed to that topic, can be immediately triggered to execute prediction tasks. This architecture significantly improves the system's scalability and the efficiency of inter-module collaboration.

[0020] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0021] This invention achieves deep integration and standardization of multi-source heterogeneous data through a data fusion processing module, constructing a unified and time-aligned customer data view, fundamentally solving the data silo problem, providing a high-quality and consistent data foundation for subsequent accurate analysis, and significantly improving the completeness and accuracy of customer profiles.

[0022] This invention achieves real-time perception of customer status and accurate prediction of future behavior through the collaborative work of a dynamic customer profiling engine and an intelligent behavior prediction module. The dynamic profiling engine ensures the timeliness of the profiles, while the integrated learning model effectively captures the nonlinear relationships and time-series patterns in customer behavior, enabling the system to gain insights into customer intent far exceeding traditional statistical analysis models, and substantially improving prediction accuracy.

[0023] This invention constructs an intelligent closed loop that continuously learns from market feedback and acts rapidly by combining an adaptive strategy generator with a real-time decision execution center. The strategy generator automatically optimizes strategies based on actual results, while the decision execution center ensures low-latency execution of strategies, enabling enterprises to respond promptly to changes in customer needs and significantly improving the return on investment of marketing activities and customer service experience.

[0024] The system architecture of this invention adopts a modular design and a message queue-based communication mechanism. Each functional module has a clear responsibility and low coupling, which not only facilitates system maintenance and functional expansion, but also effectively ensures system processing efficiency and stability under high concurrency data pressure, meeting the requirements of modern enterprises for high performance and high reliability of customer relationship management systems. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the customer relationship management system based on big data proposed in this invention;

[0026] Figure 2 This is a schematic diagram illustrating the core principle framework of the integrated learning model and adaptive policy generator in this invention. Detailed Implementation

[0027] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0028] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0029] In the embodiments of the present invention, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of the present invention shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on the present invention; the term "multiple" in the present invention refers to two or more (including two).

[0030] Example 1

[0031] Please refer to the attached document. Figure 1 This embodiment details the specific technical implementation of a customer relationship management system based on big data. The system is physically deployed on a distributed computing cluster consisting of multiple high-performance servers. This cluster receives continuous data streams from various internal and external data sources through a load balancer. The core software architecture adopts a microservice design pattern, with each major functional module encapsulated as an independent, horizontally scalable service unit, exchanging data and calling services through predefined application programming interfaces (APIs). The entire system runs in a containerized environment to ensure resource isolation and rapid deployment capabilities.

[0032] The data fusion and processing module is the core hub for system data entry and preprocessing. This module, through a set of dedicated data acquisition agents, continuously receives transaction records, customer service tickets, and social media interaction data from within the enterprise, as well as market research reports and macroeconomic indicators from outside, with a throughput capacity of thousands of records per second. When this data enters the module, its format and protocol vary. Transaction records are typically in tabular form from a Structured Query Language (SCL) database; customer service tickets may exist in Extensible Markup Language (XML) or JavaScript object representation format; social media interaction data contains a large amount of unstructured text and image information; and market research reports and macroeconomic indicators may originate from third-party data vendor application programming interfaces (APIs) and be transmitted in comma-separated values ​​or specific binary formats. The module contains a unified data access gateway responsible for preliminary protocol parsing and format identification of all input data, and converting it into a unified intermediate data format defined internally by the system. This format includes standard fields such as data source identifier, original data body, timestamp, and data quality identifier.

[0033] The data fusion processing module is further divided into a data standardization preprocessing submodule, an entity parsing and data linking submodule, and a time-series data alignment submodule. The data standardization preprocessing submodule first processes the input data stream. For unstructured text data, such as text descriptions in social media comments or customer service tickets, this submodule calls a natural language processing engine for deep processing. This engine performs word segmentation, dividing continuous text into independent lexical units, and then performs part-of-speech tagging and named entity recognition on each lexical unit to identify key information such as names of people, places, and organizations. Further, the engine performs sentiment analysis, calculating the sentiment polarity score of the entire text using a pre-trained sentiment analysis model. This score is a floating-point number between -1 and +1, where negative values ​​represent negative sentiment, positive values ​​represent positive sentiment, and zero represents neutral sentiment. For numerical data, such as transaction amounts or customer ages, this submodule performs a data cleaning process. The process first detects missing values. For randomly missing data, imputation is performed using the K-nearest neighbor algorithm. This involves finding the K nearest neighbors of the record based on its similarity to other features (K is typically 5), and then using the weighted average of these neighbors' values ​​for that feature as the imputation value. For non-random missing data, special values ​​are marked according to business rules, or the global mean is used for imputation. Outlier detection uses a method based on the 3-sigma principle. This involves calculating the mean and standard deviation of each numerical feature, marking data points that are more than three times the standard deviation from the mean as outliers, and correcting them according to preset strategies, such as truncating using the upper or lower quartiles of the feature, or directly replacing them with boundary values.

[0034] The Entity Resolution and Data Linking submodule is responsible for matching and merging pre-processed data records from different data sources that point to the same real-world customer. This submodule employs a two-stage matching strategy. The first stage is rule-based exact matching. The system pre-configures an exact matching rule base containing multiple deterministic rules. For example, if two records have identical ID card number fields, they are considered to be the same customer; or if two records have identical mobile phone number fields and the number has been verified as valid, they are also considered to be the same customer. This stage is extremely fast, designed to quickly process high-quality data records. The second stage is machine learning model-based fuzzy matching, used to handle records that failed to match in the first stage. The system pre-trains a random forest model using historical data. This model takes the similarity of multiple fields between two records as input features, including but not limited to customer name, mailing address, and email address. For each field, the system calculates its similarity score; for example, for name, it uses an edit distance algorithm to calculate similarity; for address, it uses geocoding-based similarity calculation. The random forest model combines these field similarity features and outputs a comprehensive similarity score between zero and one. The system sets a preset matching threshold of 85%. When the combined similarity score of two records is greater than or equal to 85%, the system determines that they belong to the same customer entity. Successfully matched records are assigned a globally unique customer identifier, and all their attributes are merged into a unified customer data view. For conflicting attribute values, the system uses a confidence-weighted approach or a latest timestamp-first strategy for resolution.

[0035] The time-series data alignment submodule ensures that all customer-related data has a consistent time base. This module internally maintains a high-precision time synchronization service. All data flowing into the system, regardless of its original timestamp format, is converted to a unified Coordinated Universal Time (UTC) timestamp with millisecond-level accuracy. For data lacking a definite timestamp, such as certain static attributes, the system assigns a reasonable default timestamp based on its first entry into the system or according to business logic. In this way, the system provides a solid data foundation for subsequent time-series-based analysis.

[0036] The customer dynamic profiling engine is tightly integrated with the data fusion and processing module. Its core responsibility is to build and maintain a dynamic profile that reflects the latest status of customers in real time, based on a unified view of customer data. Logically, the engine is divided into static attribute units and dynamic behavior trajectory units. The static attribute units are responsible for extracting and persistently storing relatively stable basic customer information, including but not limited to customer identification information, registration date, demographic group, historical cumulative purchase amount, and product preference categories. This data is typically stored in a column-oriented distributed database to support fast query and aggregation analysis.

[0037] The Dynamic Behavior Tracing Unit focuses on capturing customer behavior sequences that change over time. This unit processes customer interaction event streams from the data fusion processing module in real time. These events include page view events, content click events, in-app search events, and service inquiry initiation events. Each event includes an event type, event target, timestamp, and associated customer identifier. The Dynamic Behavior Tracing Unit sorts and organizes these events according to customer identifier and timestamp, forming a behavior sequence for each customer. This sequence is stored in a specially optimized time-series database that supports efficient range queries and pattern matching operations. For example, for a specific customer, their behavior sequence might be recorded as follows: browsing product A's details page at time T1, adding product A to their shopping cart at time T2, inquiring about product A's logistics information at time T3, and completing the purchase of product A at time T4. The Dynamic Behavior Tracing Unit not only records these discrete events but also uses stream processing technology to calculate a series of derived metrics in real time, such as page visit frequency over the past 30 days, average session duration, and the percentage of views for specific product categories.

[0038] The core mechanism of the customer dynamic profiling engine lies in its event-driven incremental update mechanism. This mechanism does not employ the traditional scheduled batch full calculation mode. Instead, it is triggered by data inflow. Whenever the data fusion processing module completes the processing of a new batch of data and updates the unified customer data view, it sends an event notification to the customer dynamic profiling engine through a high-throughput message middleware. This notification contains a list of customer identifiers for which data updates have occurred, along with the data dimensions involved. Upon receiving this event, the customer dynamic profiling engine does not recalculate the entire customer profile; instead, it initiates an incremental update process. This process first loads the latest version of the customer's profile from persistent storage, and then performs only partial recalculation or numerical updates on the dimensions that have changed as specified in the notification. For example, if the update involves only a new transaction record, the engine will only update transaction-related attributes such as the customer's cumulative spending and last purchase time, without affecting their basic demographics or earlier behavioral sequences. Each successful incremental update generates a new profile version, which is assigned a unique version number and recorded with a precise update timestamp. All historical versions are retained in the version control system, which enables the system to support the retrospective and comparative analysis of customer status at any historical point in time, making in-depth behavioral analysis possible.

[0039] The intelligent behavior prediction module is the core analysis unit of the system. Connected to the customer dynamic profiling engine, its function is to predict customer behavior tendencies within a specific future time window based on the latest, dynamically updated customer profiles. This module employs a carefully designed ensemble learning model architecture, the core of which consists of two parallel prediction algorithms: a gradient boosting decision tree algorithm and a long short-term memory neural network algorithm. Please refer to the appendix. Figure 2 This diagram illustrates how this integrated model works in conjunction with the subsequent policy generator.

[0040] The gradient boosting decision tree algorithm primarily handles the structured attribute features in customer profiles. These features include numerical and categorical features extracted from static attribute units, as well as statistical features aggregated from dynamic behavioral trajectory units, such as the number of visits in the past 7 days and the average order value in the past 30 days. The algorithm iteratively trains a series of decision trees, each dedicated to correcting the residuals predicted by the previous tree. In each iteration, the algorithm calculates the negative gradient between the current model prediction and the true value, and then fits a decision tree to approximate this negative gradient direction. Ultimately, the model's prediction is a weighted sum of the predictions from all decision trees. This process automatically captures complex nonlinear relationships and higher-order interactions between features without requiring complex manual feature cross-design. Early stopping is used during model training to prevent overfitting; the iteration stops when the model's performance on independent validation sets no longer improves.

[0041] Long Short-Term Memory (LSTM) neural network algorithms are specifically designed to process customer behavior sequence data from dynamic behavior trajectory units. Customer behavior, such as a series of page views or product clicks, is essentially a time series, with subsequent behaviors often depending on previous behavioral patterns. LTM neural networks, due to their internal gating mechanisms, can effectively learn these long-term dependencies. The algorithm's input is a chronological sequence of customer behavior events. Each event is represented as a high-dimensional feature vector containing information such as event type and target object attributes. This sequence is fed into the LTM neural network. Internally, the network uses three gating structures—input gate, forget gate, and output gate—to determine which information should be retained, which should be forgotten, and what information should be output at the current time step. After processing through multiple time steps, the hidden state at the last time step is considered to encode the contextual information of the entire sequence, and then, through a fully connected layer and activation function, outputs the probability corresponding to a specific behavior prediction.

[0042] The intelligent behavior prediction module does not simply juxtapose the results of the two algorithms; instead, it combines them using a weighted fusion strategy. During initial system deployment, the output weights of the gradient boosting decision tree algorithm and the long short-term memory neural network algorithm are set at 50%. For each customer requiring prediction, the two algorithms run independently, each outputting the predicted probability of a purchase within the next 14 days, denoted as follows: and Then the final fusion prediction probability. Calculated using the following formula:

[0043]

[0044] This fusion probability value, a floating-point number between zero and one, is typically presented to downstream modules as a percentage. It integrates information from both static attributes and dynamic behavioral sequences, theoretically providing more robust and accurate predictions than a single model. Furthermore, this weighted fusion strategy is designed to be dynamically adjustable. The system periodically, for example every seven days, evaluates the predictive performance of the gradient boosting decision tree and the long short-term memory neural network on a recently generated validation dataset that was not used in model training. Performance metrics are typically expressed as the area under the curve (AUC) or the area under the precision-recall curve. Based on the evaluation results, the system automatically fine-tunes the weight allocation of the two algorithms. For example, if the evaluation shows that the long short-term memory neural network significantly outperforms the gradient boosting decision tree on recent data, the system might increase its weight from 50% to 60%, while decreasing the gradient boosting decision tree's weight to 40%, ensuring that the fusion model adapts to changes in data distribution and consistently maintains optimal predictive performance.

[0045] The adaptive strategy generator connects to the intelligent behavior prediction module, and its core task is to transform the predicted customer behavior probabilities into specific, actionable strategies. Internally, the generator maintains a strategy rule base that stores a large number of predefined basic strategy templates. Each strategy template contains several key components: a strategy identifier, a strategy description, a target behavior type, triggering conditions, specific action content, and initial priority weights. Triggering conditions are typically associated with the probability values ​​output by the intelligent behavior prediction module; for example, "When the customer churn probability is greater than 30%, trigger the strategy of sending retention coupons"; or "When the customer's interest probability in a certain type of product is greater than 75% and their historical purchase value is greater than 1,000 yuan, trigger the strategy of a dedicated account manager making a follow-up call."

[0046] In addition to a static rule base, the adaptive strategy generator integrates a powerful performance feedback learning unit. This unit continuously monitors the actual effectiveness of all executed strategies. Performance data comes from the enterprise's business systems; for example, the effectiveness of a marketing SMS strategy might be measured by subsequent conversion rates, while the effectiveness of a customer service strategy might be evaluated through subsequent customer satisfaction survey scores. The performance feedback learning unit uses a multi-armed slot machine algorithm framework to automate the strategy optimization process. In this framework, each available strategy is treated as an independent slot machine arm. Assigning a strategy to a customer is equivalent to pulling the corresponding arm once. The performance metrics observed after strategy execution, such as conversion rates, are quantified as the reward for this pull.

[0047] The core of the multi-armed gambling machine algorithm lies in striking a balance between exploration and exploitation. Exploration involves trying strategies that haven't been fully validated to gather more information; exploitation involves selecting the strategy with the highest known return. The algorithm dynamically calculates the upper bound of the expected return for each strategy using a specific strategy, such as an upper confidence bound algorithm. Each time a strategy needs to be assigned to a customer, the algorithm selects the strategy with the highest upper confidence bound. Over time, strategies that are truly more effective will have their return estimates stabilize and their upper confidence bounds increased, leading to more frequent selection; while less effective strategies will be gradually phased out. In this way, the performance feedback learning unit can automatically and continuously optimize the priority weights of each strategy in the strategy rule base, and even adjust their trigger thresholds, enabling the system to adaptively learn the most effective intervention methods for specific types of customers in the current market environment.

[0048] The real-time decision execution center, serving as the command and control hub of the entire system, establishes bidirectional communication connections with the data fusion processing module, customer dynamic profiling engine, intelligent behavior prediction module, and adaptive strategy generator. The core components of this center include a high-performance message queue system and a stream processing engine. The message queue system employs a publish-subscribe architecture. Various functional modules within the system, such as the customer dynamic profiling engine and intelligent behavior prediction module, act as message publishers. After producing new data or completing a computational task, they publish a message to a specific topic. Other modules interested in such events subscribe to these topics as subscribers.

[0049] For example, when the customer dynamic profiling engine completes an incremental update to a customer's profile, it publishes a message to a topic named "Customer Profile Updated," containing the identifier of the updated customer and the update timestamp. The intelligent behavior prediction module, having subscribed to the "Customer Profile Updated" topic, receives this message almost in real-time. Upon receiving the message, the intelligent behavior prediction module is immediately triggered, initiating the prediction calculation process for that specific customer. This message-based asynchronous communication mechanism effectively decouples the direct dependencies between modules, greatly improving the system's scalability and robustness. Even if a module's processing speed temporarily lags behind, the message queue can act as a buffer, preventing system crashes.

[0050] The stream processing engine handles data streams in high-concurrency scenarios, ensuring that data processing tasks are scheduled and executed in an orderly and efficient manner. It monitors the resource utilization and task queue status of the entire system and dynamically allocates computing resources. The real-time decision execution center also exposes a set of well-defined application programming interfaces (APIs). These interfaces enable external systems, such as enterprise marketing automation platforms, customer service systems, or sales personnel's mobile applications, to interact with the system. When the adaptive strategy generator generates a personalized strategy, the strategy is immediately pushed to the corresponding external execution system through the APIs of the real-time decision execution center. For example, a coupon distribution strategy is pushed to the marketing automation platform for execution; a high-value customer alert strategy is pushed to customer relationship management software and a pop-up notification appears on the salesperson's interface. This achieves a complete automated closed loop from customer data perception, behavior prediction, strategy generation to final action execution, ensuring extremely low latency and efficient execution of decisions.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A customer relationship management system based on big data, characterized in that, include: The data fusion processing module is used to receive and process transaction records, customer service work orders, social media interaction data from within the enterprise, and market research reports and macroeconomic indicator data from outside the enterprise; the data fusion processing module includes a data standardization preprocessing unit, an entity parsing and data linking unit, and a time-series data alignment unit; A dynamic customer profiling engine is used to build and continuously update customer profiles based on a unified customer data view; the dynamic customer profiling engine includes static attribute units and dynamic behavior trajectory units. The intelligent behavior prediction module is used to predict customers' future behavioral tendencies based on updated customer profiles. The intelligent behavior prediction module adopts an integrated learning model architecture, the core of which is composed of a gradient boosting decision tree algorithm and a long short-term memory neural network algorithm. The gradient boosting decision tree algorithm is used to process the structured attribute features of customers. An adaptive strategy generator is used to generate personalized marketing or service strategies based on predicted customer behavior probabilities. The adaptive strategy generator includes a strategy rule base and an effect feedback learning unit. The strategy rule base stores a variety of basic strategy templates, whose triggering conditions are associated with customer behavior probability thresholds. The real-time decision execution center is used to coordinate data flow and task scheduling between various modules; the real-time decision execution center includes a message queue and a stream processing engine; the real-time decision execution center also provides an application programming interface, so that the generated personalized strategies can be directly pushed to the enterprise's marketing automation platform, customer service system or sales personnel's mobile terminals.

2. The customer relationship management system based on big data according to claim 1, characterized in that, The two-stage matching strategy of the entity parsing and data linking unit specifically includes: the first stage adopts fast matching based on preset precise matching rules; the second stage, for records that fail to match precisely, uses a pre-trained random forest model to calculate the comprehensive similarity score of multiple fields, and when the comprehensive similarity score exceeds the 85% threshold, it is determined to be the same customer entity.

3. A customer relationship management system based on big data according to claim 2, characterized in that, The exact matching rules include matching conditions where the ID number or mobile phone number is exactly the same; the fields calculated by the random forest model include the similarity of multiple fields such as name, address, and email.

4. A customer relationship management system based on big data according to claim 1, characterized in that, The incremental update mechanism of the customer dynamic profile engine is specifically implemented as follows: only the data dimensions that have changed are partially updated, the system maintains the version history of the customer profile, and each incremental update generates a new profile version and records the version timestamp.

5. A customer relationship management system based on big data according to claim 1, characterized in that, The weighted fusion strategy of the intelligent behavior prediction module is dynamically adjustable; the system initially sets the output weights of the gradient boosting decision tree and the long short-term memory neural network to 50% each; after the model is deployed, the prediction performance of the two algorithms is evaluated periodically on an independent validation dataset, and the output weights are automatically fine-tuned based on the evaluation results.

6. A customer relationship management system based on big data according to claim 5, characterized in that, The periodic evaluation is conducted weekly; the automatic fine-tuning process is achieved by comparing the area under the curve performance metrics of the two algorithms on the validation dataset.

7. A customer relationship management system based on big data according to claim 1, characterized in that, The effect feedback learning unit uses a multi-armed gambling machine algorithm framework for strategy optimization. The multi-armed gambling machine algorithm treats each selectable strategy as a gambling machine arm and the strategy execution effect as a reward. By continuously exploring the effects of different strategies on different customer groups and utilizing historical reward information, it learns the optimal strategy with the highest expected return in a specific situation.

8. A customer relationship management system based on big data according to claim 7, characterized in that, The multi-armed gambling machine algorithm uses an upper confidence bound algorithm to balance exploration and exploitation; the strategy execution effect is quantified by conversion rate; and the historical return information is accumulated by continuously monitoring the actual effect of the executed strategy.

9. A customer relationship management system based on big data according to claim 1, characterized in that, The message queue of the real-time decision execution center adopts a publish-subscribe model; each module within the system acts as a publisher or subscriber of messages, and communicates decoupledly through the message broker of the center.

10. A customer relationship management system based on big data according to claim 9, characterized in that, The specific implementation of the publish-subscribe mode is as follows: when the customer dynamic profile engine completes the profile update, it publishes a message to a specific topic. The intelligent behavior prediction module subscribes to the topic and immediately triggers prediction calculation upon receiving the message.

Images