A marketing management system based on big data analysis

Abstract

The application relates to the technical field of big data analysis, and particularly discloses a marketing management system based on big data analysis. The system comprises a data fusion hub, a real-time portrait engine, a multi-touch attribution analysis module, a strategy optimizer and an execution gateway. By integrating multi-source user data, dynamically updating a user interest model, quantifying touch contribution weights and intelligently generating a marketing strategy, the system realizes global user view construction, real-time interest tracking, accurate attribution analysis and automated marketing decision-making, and finally improves marketing efficiency and the input-output ratio.

Description

Technical Field

[0001] This invention belongs to the field of big data analytics technology, specifically relating to a marketing management system based on big data analytics. Background Technology

[0002] In the field of information technology, big data analytics has become a core driver of enterprise decision-making and operational optimization. Through the collection, storage, processing, and analysis of massive amounts of data, it extracts valuable information to guide business practices. Marketing management, as a crucial aspect of enterprise operations, is increasingly relying on data analytics to improve the accuracy and efficiency of marketing activities.

[0003] Among them, marketing management systems based on big data analytics aim to build accurate user profiles by integrating user behavior data, and then formulate and optimize marketing strategies accordingly. The basic principle of such systems is to use data analysis models to understand user needs and predict their behavioral trends, thereby achieving personalized marketing outreach.

[0004] Existing technologies typically rely on multiple independent data sources, resulting in user data being scattered across different systems such as e-commerce platforms, social media, and offline stores, making it difficult to form a unified user view. Current user profiles are mostly static models with long update cycles, failing to capture rapidly changing user interests and purchasing intentions in real time, leading to delays in marketing opportunities.

[0005] Furthermore, traditional attribution models are overly simplistic and fail to accurately measure the complex, non-linear impact of multiple marketing touchpoints on the user's decision-making process, leading to misallocation of marketing resources. These issues severely restrict the improvement of marketing effectiveness in application scenarios that pursue real-time, precise marketing. Therefore, there is a need for an optimized marketing management system based on big data analytics. Summary of the Invention

[0006] The purpose of this invention is to provide a marketing management system based on big data analysis to solve the problems in the existing technology, such as the fragmentation of user data making it difficult to form a unified view, the lag in updating user profiles making it impossible to capture changes in user interests in real time, and the inaccurate measurement of the impact of marketing touchpoints due to the simplification of traditional attribution models.

[0007] This invention provides a marketing management system based on big data analysis, comprising: The data fusion hub is used to connect and integrate multi-source heterogeneous user data streams from e-commerce transaction platforms, social media applications, offline store sales systems, and behavior tracking points within mobile applications. The data fusion hub has a built-in unified identity resolution engine that aggregates fragmented records of the same user in different data sources into a structured data sequence under a single global user identifier through cross-channel user identifier matching and fuzzy association algorithms. The real-time profiling engine, connected to the output of the data fusion hub, is used to dynamically extract features and track interest drift in structured user data sequences. Based on a time-decay weighted sliding window model, the real-time profiling engine assigns higher weight to recent user interactions and continuously updates user preference vectors through an implicit feedback learning algorithm. The real-time profiling engine performs a full user profile refresh every 1 minute to ensure the timeliness of the user interest model. The multi-touchpoint attribution analysis module is connected to the real-time profiling engine and data fusion hub to reconstruct the complete decision-making path of users from initial touchpoint to final conversion. This multi-touchpoint attribution analysis module adopts a deep sequence model based on the attention mechanism to perform non-linear quantitative evaluation of the contribution of each marketing touchpoint in the path. The module outputs the attribution weight value of each touchpoint, which represents the relative importance of the touchpoint in the process of facilitating user conversion. The strategy optimizer receives attribution weight values ​​from the multi-touchpoint attribution analysis module and real-time user preference vectors from the real-time profiling engine. The strategy optimizer has a built-in reinforcement learning decision agent, which aims to maximize the marketing input-output ratio and dynamically generates the optimal marketing action instruction for the next moment based on the current user state and the set of available marketing channels. The execution gateway connects to the strategy optimizer and is used to translate the optimal marketing action instructions into API call instructions that can be recognized by each marketing channel, and drive the corresponding channel execution system to complete personalized content push, coupon distribution or advertising material display operations.

[0008] Furthermore, the workflow of the unified identity resolution engine is as follows: First, the input multi-source user data is standardized and preprocessed, including field alignment, format unification, and missing value imputation; then, a community detection algorithm based on graph neural networks is used to identify strongly connected components in the user identifier association graph, and multiple identifier nodes belonging to the same real user are merged into the same user entity; finally, a unique global user identifier is assigned to each confirmed user entity, and all associated data records are re-indexed under this identifier.

[0009] Furthermore, the implicit feedback learning algorithm in the real-time portrait engine is specifically implemented as follows: The algorithm maintains a high-dimensional user embedding vector, which is initialized with random values; whenever a new user behavior event is received, the algorithm calculates the inner product of the item embedding vector corresponding to the behavior and the current user embedding vector to obtain the predicted interaction probability; by comparing the predicted interaction probability with the actual observed interaction intensity, the loss function is calculated and the error is backpropagated, and the user embedding vector is updated in a stochastic gradient descent manner; the dimension of the user embedding vector is set to 128 dimensions, and the learning rate is set to 0.001.

[0010] Furthermore, the structure of the attention-based deep sequence model is as follows: the model input is a sequence of touchpoints on the user's decision path, and each touchpoint is represented as an embedding vector containing channel type, contact time point, and contact content features; the main body of the model is a 6-layer Transformer encoder stack, whose self-attention mechanism can capture long-range dependencies between touchpoints; the model output layer is a Softmax classifier, which maps the hidden state corresponding to each touchpoint in the sequence to a normalized attribution weight value, and the sum of the attribution weight values ​​of all touchpoints is 1.

[0011] Furthermore, the training process of the reinforcement learning decision agent is as follows: the agent's state space is defined as the real-time profile vector of the current user and the state of available marketing channels; the action space is defined as the specific combination of marketing actions to be performed for the user; the reward function is defined as the user conversion value caused by the marketing action minus the marketing cost; the agent is trained using a proximal policy optimization algorithm, and through interaction with the environment, i.e., the real user group, it collects experience data and iteratively updates its policy network parameters, and finally learns a marketing decision strategy that maximizes long-term cumulative rewards in a complex dynamic environment.

[0012] Furthermore, the execution gateway also has a cross-channel conflict resolution function; when the strategy optimizer generates multiple marketing actions that may have resource conflicts for the same user at the same time, the execution gateway filters and schedules the action sequence according to the preset channel priority rules and user fatigue threshold, so as to ensure that the user does not receive too much or contradictory marketing information in the same time period.

[0013] Furthermore, the system operates on a hierarchical control architecture, which includes a data layer, a computing layer, and an application layer. The data layer is responsible for the access, cleaning, and storage of multi-source data. The computing layer deploys the core algorithms of the real-time profiling engine, the multi-touchpoint attribution analysis module, and the strategy optimizer. The application layer provides a strategy configuration interface and an effect monitoring dashboard for marketing managers.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By using a data fusion hub and a unified identity resolution engine, this invention completely solves the problem of user data silos, constructs an accurate and complete global user view, lays a solid data foundation for subsequent precise analysis, and avoids decision-making biases caused by incomplete data.

[0015] 2. The real-time profiling engine, through minute-level dynamic updates and interest drift tracking, enables the system to keenly capture instantaneous changes in user intent, ensuring the timeliness and relevance of marketing outreach, and improving user response rates and the ability to capture marketing opportunities.

[0016] 3. The multi-touchpoint attribution analysis module adopts an advanced attention mechanism model, which breaks through the limitations of traditional linear attribution. It can accurately quantify the real contribution of each marketing touchpoint in the complex user journey, providing reliable data support for the scientific allocation of marketing resources and the optimization of channel strategies.

[0017] 4. The strategy optimizer uses reinforcement learning to achieve automated and intelligent marketing decisions. It can dynamically adjust marketing strategies based on real-time user status and historical feedback, continuously optimize the return on investment in marketing, and reduce the cost of manual intervention and the risk of subjective decision-making.

[0018] 5. The conflict resolution mechanism of the execution gateway ensures the consistency of the cross-channel marketing experience, avoids information overload and user annoyance, and maintains the long-term health of brand image and user relationships. The entire system forms a closed loop from data integration, real-time insights, intelligent decision-making to precise execution, comprehensively improving the efficiency and effectiveness of marketing management. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical architecture of the marketing management system based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the unified identity resolution engine in this invention; Figure 3 This is a logical flow diagram of the real-time portrait engine in this invention; Figure 4 This is a schematic diagram of the core principle framework of the multi-touch attribution analysis module in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the policy optimizer and the execution gateway in this invention. Detailed Implementation

[0020] The marketing management system based on big data analysis provided by this invention has an overall technical architecture that can be found in the appendix. Figure 1 The system consists of five core components: a data fusion hub, a real-time profiling engine, a multi-touchpoint attribution analysis module, a strategy optimizer, and an execution gateway. These components work together through a hierarchical control architecture to achieve closed-loop management of the entire process, from multi-source data integration and dynamic user profiling to attribution analysis, intelligent decision-making, and precise execution.

[0021] As the data foundation of the system, the data fusion hub is responsible for connecting and integrating user behavior data streams from four major categories of external data sources.

[0022] Specifically, user order records, product browsing logs and payment information provided by e-commerce trading platforms, user likes, comments, shares and follower relationship data collected by social media applications, member consumption details and store visit trajectories recorded by offline store sales systems, and page dwell time, function click sequences and search keywords captured by behavior tracking points within mobile applications are all transmitted to the data fusion center through standardized application programming interfaces.

[0023] The data fusion hub is equipped with multi-level buffer queues and streaming data pipelines to perform real-time peak reduction and asynchronous decoupling on the input high-frequency data.

[0024] The core of the data fusion hub is the unified identity resolution engine; please refer to the appendix for its principle framework. Figure 2 .

[0025] The unified identity resolution engine first performs standardized preprocessing on the original multi-source user data: to address the differences in field naming from different data sources, a global field mapping table is established to uniformly map fields such as user identifiers to global user identifiers, and to uniformly convert timestamp fields to international standard time formats. To address the issue of inconsistent data formats, type casting and encoding standardization operations are performed, such as converting text-based gender descriptions into binary values ​​and parsing unstructured geographic location information into latitude and longitude coordinates. To address the issue of missing values, an imputation algorithm based on user behavior sequence patterns is employed, such as using browsing history of the same user in adjacent time periods to predict missing product category tags.

[0026] After preprocessing, the unified identity resolution engine constructs a user identifier association graph, where nodes represent user identifiers from various data sources and edges represent the association strength between identifiers.

[0027] The engine employs a community detection algorithm based on graph neural networks, which aggregates node features through multi-layer graph convolutional networks to identify strongly connected components in the graph.

[0028] Specifically, the algorithm defines a node similarity metric function, calculates the structural equivalence and attribute consistency between each pair of nodes, and uses the principle of maximizing modularity to merge highly interconnected node clusters into the same user entity.

[0029] Each confirmed user entity is assigned a unique global user identifier of 128 bits. All associated data records are reorganized under this identifier through distributed indexing technology to form a user-centric structured data sequence.

[0030] The output of the data fusion hub is a time-ordered stream of user behavior events. Each event includes a global user identifier, event type, timestamp, event attribute set, and data source identifier.

[0031] The real-time profiling engine connects directly to the output of the data fusion hub; please refer to the appendix for its logical flow framework. Figure 3 The real-time profiling engine receives the structured user data sequence output from the data fusion center and performs dynamic feature extraction and interest drift tracking on it.

[0032] The engine maintains a distributed key-value database that stores the real-time profile vector corresponding to each global user identifier.

[0033] The profile vector contains 128 dimensions, each representing the strength of a user's preference across different interest categories.

[0034] The real-time profiling engine uses a time-decay weighted sliding window model to process user behavior sequences: the window size is set to 30 days, and events within the window are weighted according to the difference between their occurrence time and the current time. The weight decay function is exponential, with recent events having a significantly higher weight than historical events.

[0035] For each user behavior event, the engine extracts its implicit semantic features. For example, for a product browsing event, it maps it into a 128-dimensional item embedding vector through a pre-trained product embedding model; for a social interaction event, it extracts the content topic distribution vector of the interactive object.

[0036] The core update mechanism of the real-time profile engine is based on an implicit feedback learning algorithm, which captures changes in user interests by optimizing the inner product relationship between user embedding vectors and item embedding vectors.

[0037] Specifically, the algorithm maintains a 128-dimensional embedding vector for each user, which is initialized with Gaussian random values ​​with a mean of 0 and a standard deviation of 0.01.

[0038] When a new user behavior event is received, the algorithm calculates the inner product of the item embedding vector corresponding to the event and the current user embedding vector to obtain the predicted interaction probability value.

[0039] A binary cross-entropy loss function is constructed by comparing the predicted interaction probabilities with the actual observed interaction strengths, and the user embedding vector is updated by backpropagating the error using stochastic gradient descent. The learning rate is set to 0.001, and the batch size is set to 256.

[0040] The real-time user profile engine triggers a full user profile refresh task every 1 minute: First, it pulls all new behavior events of all users in the past minute from the data fusion hub, groups them by user, and performs vector update calculations in parallel; after the update is completed, it persists the latest profile vector to the database and publishes change notifications to the subscription components.

[0041] In addition, the engine also integrates an interest drift detection module, which identifies significant shifts in user interests by calculating the cosine similarity change rate of the user profile vector within a sliding window. When the change rate exceeds a threshold of 0.15, the user is marked as being in an interest transition period, which is then tracked by subsequent strategy modules.

[0042] The multi-touchpoint attribution analysis module connects to both the real-time profiling engine and the data fusion hub. Please refer to the appendix for its core principle framework. Figure 4 .

[0043] This module is responsible for reconstructing the complete decision-making path of users from the first marketing touchpoint to the final conversion, and quantifying the contribution of each marketing touchpoint in the path.

[0044] The module first obtains historical touchpoint sequence data of users from the data fusion center, including interaction records such as ad exposure, email opening, push notification clicks, and search keyword input; at the same time, it obtains real-time profile vectors of users when conversion events occur from the real-time profile engine.

[0045] The multi-touch attribution analysis module uses a deep sequence model based on an attention mechanism to process touch sequence data.

[0046] The model takes a variable-length sequence of touch points as input, with each touch point represented as a 256-dimensional embedding vector. The vector contains channel type encoding, timestamp embedding of the contact time point, contact content feature vector, and touch point context information.

[0047] The main body of the model is a 6-layer Transformer encoder stack, with each layer containing 8 attention heads and a hidden layer dimension of 512.

[0048] The encoder's self-attention mechanism can capture long-range dependencies between touchpoints, such as identifying causal relationships between early brand ad exposure and subsequent search behavior.

[0049] The model's output layer is a Softmax classifier that maps the final hidden state corresponding to each touchpoint in the sequence to a normalized attribution weight value. The sum of the attribution weight values ​​for all touchpoints is strictly equal to 1. The formula for calculating the attribution weight value is:

[0050] in, Indicates the first Attribution weight values ​​for each touchpoint For the first The hidden state vector corresponding to each contact point and The weight matrix and bias terms of the output layer are as follows: This represents the length of the contact sequence.

[0051] This formula achieves non-linear quantification of the contribution of touchpoints, which can accurately reflect the relative importance of different touchpoints in the user's decision-making path.

[0052] The multi-touchpoint attribution analysis module processes user conversion path data from the past hour in batches every 5 minutes, generates attribution reports, and stores them in the analysis results database.

[0053] The strategy optimizer receives attribution weight values ​​from the multi-touch attribution analysis module and real-time user preference vectors from the real-time profiling engine.

[0054] The strategy optimizer incorporates a reinforcement learning decision agent. This agent aims to maximize the return on marketing investment and dynamically generates optimal marketing action instructions based on the current user status and the set of available marketing channels.

[0055] The state space of the intelligent agent is defined as a 256-dimensional vector, which is formed by concatenating the user's real-time profile vector and the marketing channel state vector.

[0056] The marketing channel status vector contains real-time operational data such as the daily budget consumption ratio, average response rate, and fatigue index for each channel.

[0057] The action space is defined as a discrete-continuous hybrid space: the discrete part represents the type of marketing channel to be executed, and the continuous part represents the action intensity parameters for that channel, such as coupon value or advertising bid coefficient.

[0058] The reward function is defined as the user conversion value generated by the marketing action minus the marketing cost. The conversion value is allocated to each touchpoint through attribution weight values, while the cost includes direct financial costs and channel resource occupation costs.

[0059] The agent is trained using a proximal policy optimization algorithm: the policy network is a 3-layer fully connected neural network with 512 neurons in each layer and ReLU activation function; the value network has the same structure but independent parameters.

[0060] During training, the agent interacts with the environment, i.e., the real user group, and collects state-action-reward sample data, which is stored in the experience replay buffer. A policy update is performed after every 10,000 samples are accumulated.

[0061] During updates, the probability ratio is pruned to limit the magnitude of policy changes, ensuring training stability. The policy optimizer scans the queue of users waiting to make decisions every 10 seconds. For each user, it loads their latest state vector, calculates the probability distribution of each action through forward propagation, and samples to generate the final marketing action instruction.

[0062] The instruction content includes the target user's global identifier, action type, action parameters, execution time window, and priority identifier.

[0063] The execution gateway connects to the policy optimizer; please refer to the appendix for details on its multi-level interactions and data flow. Figure 5 .

[0064] The execution gateway is responsible for translating the optimal marketing action instructions generated by the strategy optimizer into application programming interface (API) call instructions that can be recognized by each marketing channel.

[0065] The gateway internally maintains a registry of channel adapters, with each adapter encapsulating the specific channel's interface protocol, authentication mechanism, and data format conversion logic.

[0066] For example, for email marketing channels, the adapter translates instructions into SMTP protocol email sending requests, populating the subject, recipient, personalized content template, and tracking pixels; for social media advertising channels, the adapter generates bid requests and creative upload instructions that conform to the platform's advertising management application programming interface specifications.

[0067] The execution gateway also has cross-channel conflict resolution capabilities: when the strategy optimizer generates multiple marketing actions for the same user that may have resource conflicts, the gateway arbitrates based on preset channel priority rules and user fatigue thresholds.

[0068] The channel priority rule is a static configuration table that defines the weight score of each channel. When a conflict occurs, the action of the channel with the highest weight is selected and executed.

[0069] The user fatigue threshold is calculated in real time based on the number of marketing outreaches a user receives in the past 24 hours and the negative feedback rate. When the number of outreaches exceeds 5 or the negative feedback rate exceeds 20%, the gateway automatically filters all non-urgent marketing actions for that user.

[0070] The execution gateway polls the instruction queue every 2 seconds, and the status of completed actions is fed back to the policy optimizer through a callback interface, forming a closed-loop learning process.

[0071] The system operates on a hierarchical control architecture, which includes a data layer, a computing layer, and an application layer.

[0072] The data layer deploys a distributed file system and columnar database, responsible for the access, cleaning and storage of multi-source data. The data retention strategy is divided into three levels: hot data, warm data and cold data. Hot data is retained for 30 days for real-time analysis, warm data is retained for 1 year for batch mining, and cold data is archived to object storage for historical auditing.

[0073] The computing layer employs containerized deployment of the core algorithms of the real-time profiling engine, multi-touchpoint attribution analysis module, and strategy optimizer. The resource scheduler dynamically allocates central processing unit and memory resources according to the load, ensuring that 95% of requests are responded to within 100ms.

[0074] At the application level, a networked strategy configuration interface and performance monitoring dashboard are provided to marketing managers. Administrators can adjust attribution model parameters, set channel budget limits, and define user segmentation rules through the interface. The dashboard displays real-time trend curves of core indicators such as conversion rate, average order value, channel contribution, and return on investment.

[0075] Through the coordinated operation of the aforementioned components, the entire system achieves full-domain integration of user data, minute-level updates of interest models, precise quantification of attribution weights, adaptive optimization of marketing strategies, and intelligent scheduling of cross-channel outreach.

[0076] The data fusion hub ensures the integrity and consistency of the user view, the real-time profiling engine guarantees the timeliness of capturing user interests, the multi-touchpoint attribution analysis module provides a scientific basis for performance evaluation, the strategy optimizer drives the automation and intelligence of the decision-making process, and the execution gateway maintains the consistency and comfort of the user experience.

[0077] The system is based on data-driven decision-making and has built an efficient, accurate and adaptive marketing management closed loop.

[0078] This embodiment further optimizes the community discovery algorithm in the unified identity resolution engine and the interest drift tracking mechanism in the real-time profiling engine to improve the system's performance and stability in scenarios with a massive user base and high concurrency data.

[0079] The community detection algorithm of the unified identity resolution engine has been upgraded to a hierarchical graph neural network architecture. The original single-layer graph neural network faces memory bottlenecks and computational latency issues when processing user identifier association graphs with hundreds of millions of nodes.

[0080] In this embodiment, the global graph is divided into multiple subgraphs according to the data source type. Community discovery is performed independently in each subgraph, and the results are merged through cross-subgraph alignment.

[0081] Specifically, the first step is to construct e-commerce sub-graphs, social sub-graphs, offline sub-graphs, and application tracking points. Figure 4 Each subgraph has a subgraph structure, and each subgraph uses a two-layer graph convolutional network for node embedding learning. The graph convolution operation is defined as follows: ; in, To add a self-loop adjacency matrix, This is the corresponding degree matrix. For the first The node feature matrix of the layer, For trainable weight matrix, This is the ReLU activation function.

[0082] After the subgraph community is discovered, the engine performs entity alignment by calculating the similarity between nodes across subgraphs: it calculates the cosine similarity of the embedding vectors of nodes in different subgraphs, and when the similarity exceeds the threshold of 0.8, it determines that they are the same user and merges them under the same global user identifier.

[0083] This hierarchical processing strategy reduces the computational complexity of graphs from cubic to linear, enabling the system to support real-time parsing of 1 billion user identifiers.

[0084] The interest drift tracking mechanism of the real-time profile engine has been enhanced to a multi-scale sliding window model. The original single-window model struggles to simultaneously capture both short-term fluctuations in user interests and long-term preference evolution.

[0085] This embodiment designs three parallel sliding windows: a short-term window of 7 days to capture immediate changes in interest; a medium-term window of 30 days to stabilize interest representation; and a long-term window of 90 days to model basic preferences. Each window independently maintains a user profile vector, and the update algorithm is the same as described above.

[0086] The engine has added a window fusion module, which dynamically aggregates multi-window vectors through an attention mechanism: it calculates the importance weight of each window vector, and the weight depends on the freshness of the window and the intensity of the user's recent activities. The final user profile vector is the weighted sum of the window vectors.

[0087] In addition, the interest drift detection module has been upgraded to a state recognizer based on a hidden Markov model, which divides the user's interest state into three types: stable period, fluctuating period and transition period. Different update strategies are adopted for different states: the learning rate is reduced to 0.0001 to reduce noise during the stable period, the learning rate is increased to 0.01 to accelerate adaptation during the fluctuating period, and an external knowledge graph is introduced to assist semantic alignment during the transition period.

[0088] Multi-scale models enable the system to more finely depict the evolution trajectory of user interests, improving the long-term predictive accuracy of profiles while maintaining response speed.

[0089] The sequence model in the multi-touch attribution analysis module supports incremental training and online learning.

[0090] The original model requires periodic full retraining and cannot adapt to changes in marketing channels in real time.

[0091] This embodiment adds an incremental update interface to the model output layer. When a new channel is detected or the strategy of an old channel is adjusted, only the touchpoint sequence of the affected users is locally recalculated, and elastic weight consolidation technology is used to prevent catastrophic forgetting.

[0092] The reinforcement learning agent of the policy optimizer introduces a hierarchical reward design, separating short-term conversion rewards from long-term user lifetime value rewards, and balancing immediate effects and lasting relationships through value function decomposition techniques.

[0093] The conflict resolution mechanism of the execution gateway increases context awareness, dynamically adjusting channel priority based on the user's current device type, geographical location, and activity status. For example, mobile users are given priority for push notifications, while desktop users are given priority for displaying advertisements.

[0094] Through component-level optimization and cross-module collaboration, the system has achieved stable and efficient operation in an environment of continuously increasing data volume, number of users, and business complexity.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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 process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A marketing management system based on big data analytics, characterized in that, include: The data fusion hub is used to connect and integrate multi-source heterogeneous user data streams from e-commerce trading platforms, social media applications, offline store sales systems, and in-app behavior tracking points. The real-time profiling engine, connected to the output of the data fusion hub, is used to perform dynamic feature extraction and interest drift tracking on structured user data sequences. The real-time profiling engine is based on a time decay weighted sliding window model, which gives higher weight to the user's recent interactive behavior and continuously updates the user preference vector through an implicit feedback learning algorithm. The multi-touchpoint attribution analysis module, connected to the real-time profiling engine and data fusion hub, is used to reconstruct the complete decision-making path of a user from initial touchpoint to final conversion. This module employs a deep sequence model based on an attention mechanism to non-linearly quantify the contribution of each marketing touchpoint in the path. The input to the deep sequence model is a variable-length touchpoint sequence, with each touchpoint represented as an embedding vector. The main body of the model is a Transformer encoder stack, which captures long-range dependencies between touchpoints through a self-attention mechanism. The model output layer maps the hidden state corresponding to each touchpoint in the sequence to a normalized attribution weight value, with the sum of all attribution weight values ​​equal to 1. The multi-touchpoint attribution analysis module outputs the attribution weight value for each touchpoint. The strategy optimizer receives attribution weight values ​​from the multi-touchpoint attribution analysis module and real-time user preference vectors from the real-time profiling engine. The strategy optimizer has a built-in reinforcement learning decision agent, which aims to maximize the marketing input-output ratio and dynamically generates the optimal marketing action instruction for the next moment based on the current user state and the set of available marketing channels. The execution gateway connects to the strategy optimizer and is used to translate the optimal marketing action instructions into API call instructions that can be recognized by each marketing channel, and drive the corresponding channel execution system to complete personalized content push, coupon distribution or advertising material display operations.

2. The marketing management system based on big data analysis according to claim 1, characterized in that, The data fusion hub has a built-in unified identity resolution engine that aggregates fragmented records of the same user in different data sources into a structured data sequence under a single global user identifier through cross-channel user identifier matching and fuzzy association algorithms.

3. A marketing management system based on big data analysis according to claim 2, characterized in that, The workflow of the unified identity resolution engine includes: First, the input multi-source user data is standardized and preprocessed, including field alignment, format standardization, and missing value imputation; Next, a community detection algorithm based on graph neural networks is used to identify strongly connected components in the user identifier association graph and merge multiple identifier nodes belonging to the same real user into the same user entity. Finally, each confirmed user entity is assigned a unique global user identifier within the system, and all associated data records are reindexed to that identifier.

4. A marketing management system based on big data analysis according to claim 1, characterized in that, The implicit feedback learning algorithm in the real-time portrait engine is specifically implemented as follows: The algorithm maintains a high-dimensional user embedding vector, which is initialized with random values. Whenever a new user behavior event is received, the algorithm calculates the inner product of the item embedding vector corresponding to the behavior and the current user embedding vector to obtain the predicted interaction probability. By comparing the predicted interaction probabilities with the actual observed interaction strengths, the loss function is calculated and the error is backpropagated, and the user embedding vector is updated using stochastic gradient descent.

5. A marketing management system based on big data analysis according to claim 1, characterized in that, The structure of the attention-based deep sequence model includes: The model input is a sequence of touchpoints along the user's decision path, with each touchpoint represented as an embedding vector containing channel type, contact time, and contact content features; The main body of the model is a Transformer encoder stack, whose self-attention mechanism can capture long-range dependencies between touchpoints. The model's output layer is a Softmax classifier that maps the hidden state corresponding to each touchpoint in the sequence to a normalized attribution weight value.

6. A marketing management system based on big data analysis according to claim 1, characterized in that, The training process of the reinforcement learning decision agent includes: The state space of an intelligent agent is defined as the real-time profile vector of the current user and the state of available marketing channels; the action space is defined as the specific combination of marketing actions to be performed for that user. The reward function is defined as the user conversion value generated by the marketing action minus the marketing cost; The agent is trained using a proximal policy optimization algorithm. It collects experience data and iteratively updates its policy network parameters by interacting with the environment, i.e., the real user group.

7. A marketing management system based on big data analysis according to claim 1, characterized in that, The execution gateway also has a cross-channel conflict resolution function; when the strategy optimizer generates multiple marketing actions for the same user that may have resource conflicts, the execution gateway filters and schedules the action sequence according to preset channel priority rules and user fatigue threshold.

8. A marketing management system based on big data analysis according to claim 3, characterized in that, The standardized preprocessing also includes data type casting and encoding normalization operations, converting text descriptions into binary values ​​and parsing unstructured geographic location information into latitude and longitude coordinates.

9. A marketing management system based on big data analysis according to claim 4, characterized in that, The real-time profile engine also integrates an interest drift detection module, which identifies significant shifts in user interests by calculating the cosine similarity change rate of the user profile vector within a sliding window; when the change rate exceeds a threshold, the user is marked as being in an interest transition period.

10. A marketing management system based on big data analysis according to claim 1, characterized in that, The marketing management system based on big data analytics runs on a layered control architecture, which includes a data layer, a computing layer, and an application layer. The data layer is responsible for accessing, cleaning, and storing multi-source data; the computing layer deploys the core algorithms of the real-time profiling engine, multi-touchpoint attribution analysis module, and strategy optimizer; and the application layer provides marketing managers with a strategy configuration interface and performance monitoring dashboard.

Images