Precision marketing analysis method and system based on big data

Abstract

The invention relates to the technical field of precision marketing analysis based on big data, and discloses a precision marketing analysis method and system based on big data. Sequentially executing transaction log preprocessing, extracting a customer transaction sequence and constructing a commodity transfer frequency matrix; constructing a commodity transition probability matrix based on the matrix, and calculating a diffusion retention coefficient; the commodity popularity is initialized in combination with a time attenuation mechanism; constructing a directed purchase network and calculating the diameter of the directed purchase network as an upper limit of diffusion round times; executing heat diffusion iteration according to the transition probability and the retention coefficient, and judging whether convergence occurs or not; and finally normalizing the popularity vector and sorting, and outputting a precise putting target. According to the method, modeling behaviors are extracted from original data, a propagation network is constructed, and indirect propagation of user interests is simulated in combination with time sequence popularity, so that automatic modeling from behavior data to a delivery target is realized. According to the scheme, the accuracy and the service decision-making efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of precision marketing analysis methods based on big data, and specifically to a precision marketing analysis method and system based on big data. Background Art

[0002] With the rapid development of e-commerce and mobile internet, the accumulation of massive amounts of transaction data has provided a rich "big data" foundation for precision marketing. Traditional precision marketing analysis methods primarily rely on collaborative filtering, RFM models, and rule-based feature clustering techniques. Collaborative filtering calculates the similarity between users or products to recommend products that similar users have historically liked. RFM models focus on the time, frequency, and amount of a user's most recent purchase, categorizing users into several static levels. In recent years, some studies have also attempted to apply Markov chains to purchase sequence analysis, predicting the next purchase by constructing a product transition probability matrix. Other approaches have drawn on graph theory to model the relationship between products and purchases as a graph structure, employing concepts such as PageRank and heat diffusion to measure node influence.

[0003] First, most collaborative filtering-based recommendation methods focus on calculating static similarities between users or products, insufficiently considering the temporal characteristics of user purchase sequences and struggling to capture the decaying influence of recent purchases. Furthermore, similarity calculations often introduce global or local fixed weight parameters, requiring manual tuning and being highly sensitive to parameter selection, resulting in poorly stable recommendation results. Secondly, while the RFM model reflects timeliness, its stratification criteria are generally set empirically and lack unified data-driven metrics, making it difficult for stratification results to accurately reflect users' dynamic value. Furthermore, most such models cluster at the user level, neglecting the spread and influence of products. Furthermore, approaches that directly apply Markov chains to purchase sequences typically employ fixed damping coefficients or adjustable parameters for heat diffusion. However, these parameters are often determined by prior experience and lack objective quantitative basis. The number of iterations is also often set empirically, subject to over- or under-iteration issues, making it impossible to ensure that information is fully disseminated throughout the purchase network. In terms of graph theory methods, many studies directly map product purchase relationships into undirected graphs or dense directed graphs, and use algorithms such as PageRank to evaluate node popularity. However, they do not distinguish the impact of network structure differences and path lengths, making it difficult to determine the appropriate number of iterations based on actual topological characteristics. In addition, some methods rely on closed-loop monitoring or online feedback adjustments, which increases system complexity and computational overhead and is not conducive to large-scale offline batch processing. In addition, most existing technologies directly sort and output the heat vectors by numerical value after calculation, but ignore the importance of normalization for cross-category or cross-regional placement decisions, which may lead to marketing targets being biased towards products with high transaction volumes but not the most valuable. Overall, the industry lacks a precision marketing analysis method that can automatically drive parameter calculations, fully consider the topological characteristics of the purchase network and the impact of time attenuation, and complete full-graph information diffusion and normalized sorting without closed-loop feedback.

[0004] To this end, this case aims to propose a precision marketing analysis method and system based on big data. By systematically processing transaction logs, this system establishes purchase transfer relationships between products. Furthermore, through heat diffusion modeling and network communication mechanisms, it conducts in-depth customer behavior mining and predicts product popularity. Combined with normalization and sorting, it outputs a precise list of target products for marketing. By leveraging graph theory and probabilistic diffusion mechanisms, this system conducts structured analysis of massive amounts of user behavior data, addressing the pain points of traditional marketing strategies, such as "ambiguous targets," "delayed responses," and "inefficient delivery." Summary of the Invention

[0005] The present invention provides a precision marketing analysis method and system based on big data, which helps solve the problems mentioned in the above background technology.

[0006] The present invention provides the following technical solution: a precision marketing analysis method based on big data, comprising:

[0007] Preprocess transaction logs, extract customer transaction sequences, and construct a global product transfer frequency matrix;

[0008] Based on the transfer frequency matrix, construct a transfer probability matrix between commodities;

[0009] Calculate the diffusion retention coefficient based on the out-degree information of the product;

[0010] Initialize the product popularity vector based on transaction records and time decay kernel function;

[0011] Construct a directed purchasing network between products and calculate the diameter of the network;

[0012] Perform heat diffusion iterative calculation based on the transition probability matrix and diffusion retention coefficient;

[0013] Determine whether the heat diffusion process converges. If the convergence condition is met, complete the calculation and obtain the final heat vector;

[0014] Normalize the heat vector, sort it in descending order by heat score, and output the sorting result as the target of precision marketing delivery.

[0015] Optionally, preprocessing the transaction logs to extract customer transaction sequences and construct a global commodity transfer frequency matrix specifically includes:

[0016] Let the transaction log be an ordered set of tuples Where N is the set Total number of records in u i is the customer ID of the i-th transaction; p i ∈{1,2,...,P} is the product number of the i-th transaction, P is the total number of products; t i is the discretized timestamp of the i-th transaction;

[0017] Culling Collection Any one of u i ,p i ,t i missing records;

[0018] Sort the set by customer ID u in ascending order, and then by time t in ascending order. Aggregate by segments to get the time series of each customer: {(p u,j ,t u,j )|j=1,2,...,n u}, n u =#{i|u i =u}; where u is the current customer ID; n u is the number of transactions of customer u; p u,j is the product number corresponding to the j-th transaction of customer u; tu,j The timestamp corresponding to the j-th transaction of customer u;

[0019] For each customer sequence, take two adjacent transactions (p u,j →p u,j+1 ) and count, statistical global conversion frequency matrix:

[0020] A=[A p,q ] P×P , Among them, A p,q The total number of times product p subsequently purchases product q; It is an indicator function, which has a value of 1 if the logic in the brackets is true, otherwise it is 0;

[0021] If no (p,q) transition occurs, then A p,q =0.

[0022] Optionally, constructing a transition probability matrix between commodities based on the transition frequency matrix specifically includes:

[0023] For each product number p, calculate its total out-degree frequency

[0024] Set the transition probability matrix to:

[0025] M=[M p,q ] P×P ,

[0026] Among them, M p,q is the conditional probability that the next purchase will be transferred to product q given that the current product is p.

[0027] Optionally, calculating the diffusion retention coefficient based on the out-degree information of the product specifically includes:

[0028] Calculate average outdegree

[0029] Setting the Diffusion Retention Coefficient

[0030] Optionally, initializing the product popularity vector based on transaction records and a time decay kernel function specifically includes:

[0031] Set the global maximum timestamp

[0032] Set the discrete-time decay kernel to Where Δt is the transaction time difference;

[0033] Calculate the initial heat:

[0034]

[0035] Among them, n p =#{i|p i =p}, for product p in the set The number of transactions in t p,i The timestamp of the i-th transaction of commodity p; is the initial popularity of product p.

[0036] Optionally, constructing a directed purchasing network between products and calculating the diameter of the network specifically includes:

[0037] Constructing a directed graph Where V = {1, 2, ... P}, E = {(p, q) | A p,q >0}; where V is a node set, representing P kinds of goods; E is a directed edge set, if A p,q >0, then there is an edge (p→q);

[0038] For each source node p, use the breadth-first search method to find the shortest path length to each reachable node q:

[0039]

[0040] Where d(p,q) is the shortest path length from node p to q; l is the path length; if it is unreachable, then d(p,q) = ∞;

[0041] Set the network diameter to

[0042] Optionally, performing heat diffusion iterative calculation based on the transition probability matrix and the diffusion retention coefficient specifically includes:

[0043] Let iteration index k = {0, 1, ..., D-1};

[0044] Set the iterative update function to:

[0045]

[0046] in, is the popularity of product p after the kth round;

[0047] The vector form is h (k+1) =α·M T ·h (k) +(1-α)·h (0) Among them, M T is the transpose of the matrix M.

[0048] Optionally, the process of determining whether the heat diffusion process has converged, and if so, completing the calculation to obtain the final heat vector, specifically includes:

[0049] The iteration ends when k=D, and the final heat vector is obtained:

[0050]

[0051] Optionally, the heat vectors are normalized, sorted in descending order by heat score, and the sorting results are output as precision marketing delivery targets, specifically including:

[0052] Calculate the normalized heat score

[0053] Press c p Sort in descending order, generate a sequence

[0054] Take the first K items {p1,p2,...,p K}As a precision marketing delivery target, output the required list and ranking.

[0055] A system for implementing the big data-based precision marketing analysis method and system, comprising:

[0056] Computation module: performs pre-processing of transaction logs, extracts customer transaction sequences, constructs a global product transfer frequency matrix, performs matrix construction, heat initialization, and iterative calculations;

[0057] Sorting module: normalizes, sorts and outputs the final heat vector.

[0058] The present invention has the following beneficial effects:

[0059] 1. Through the dual aggregation strategy of "ascending customer + ascending timestamp," a user transaction sequence is formed, and then a global product transfer frequency matrix is ​​constructed. This processing method avoids the common drawback of "using users as static analysis units" and can preserve the sequential and dynamic nature of consumer behavior, making the subsequent migration matrix more consistent with the actual purchase logic. Missing value elimination and normalization processing are also introduced to enhance data robustness. Traditional collaborative filtering or recommendation methods often ignore the time dimension and transaction order. This solution overcomes the limitations of static models by constructing a dynamic evolution graph network based on transfer frequency. It can reflect the probabilistic relationship of "pre-purchase" between products, significantly improving the accuracy of sequence modeling and the usability of marketing deduction.

[0060] 2. By normalizing the out-degree of the frequency matrix to form a transition probability matrix, the purchasing relationship between products is transformed from a "discrete statistical" model to a "directed weighted graph," establishing a propagation path for subsequent popularity diffusion. This structure, with its probabilistic meaning, facilitates integration into Markov chains or propagation models, enhancing the method's theoretical rigor and adaptability. Previous recommendation systems mostly relied on rating matrices or content similarity, lacking directionality and dynamic migration logic. In contrast, this solution transforms actual transactions into quantifiable "conditional probabilities," making the model more relevant to commercial realities and easier to understand and scale.

[0061] 3. By automatically calculating the average out-degree of each item and setting the diffusion retention coefficient, we address the instability issues associated with manual parameter adjustment in traditional diffusion models. This retention coefficient balances the diffusion heat with the initial heat, thereby avoiding dilution caused by excessive diffusion in high-density networks and preventing diffusion stagnation in low-density networks. Existing graph diffusion recommendation algorithms, such as PageRank and HeatS, rely on fixed coefficients and struggle to adapt to the varying network sparsity in real-world business scenarios. This method, combined with the dynamic setting of item out-degree, enhances the model's adaptability and stability.

[0062] 4. Define a time-decay kernel function based on transaction time differences, and weight the product's purchase frequency and freshness to form an initial popularity vector. This strategy highlights the impact of recent purchases on market popularity, helping to dynamically capture shifts in consumer preferences and improve the timeliness of subsequent predictions and rankings. Existing models typically set initial popularity based on cumulative sales or total clicks, neglecting the critical factor of timeliness and leading to delayed marketing delivery. In contrast, this method introduces a time kernel function to address this key shortcoming.

[0063] 5. We abstract the relationship between products and their purchases into a directed graph model, using the graph's "diameter" as the theoretical upper limit for the number of diffusion rounds. This ensures that all nodes are properly covered during the heat diffusion process, ensuring both diffusion integrity and controlling computational costs. Traditional diffusion methods typically use a fixed number of iterations or rely on empirical judgment, which can lead to insufficient heat diffusion or overfitting. The "network diameter-limited round" strategy proposed in this solution is both theoretically sound and practically controllable, significantly improving system efficiency and convergence stability.

[0064] 6. Set an iterative update function with a retention term, incorporating both initial heat and transition probabilities into the heat update, achieving directed diffusion of heat within the network. Matrix multiplication is used to express the heat update process, supporting parallel optimization and acceleration. Compared to conventional label propagation or average diffusion models, this method enhances the influence of initial behavioral data on the results through the retention term, making the model more robust when dealing with new products (cold start) while preventing heat from being absorbed by ineffective products.

[0065] 7. Using the network diameter as the iteration termination criterion avoids setting artificial convergence thresholds while ensuring that heat is fully propagated to the network's end nodes. This "structural constraint-driven convergence" approach is theoretically sound and efficient. Traditional methods rely on numerical changes to determine convergence, which can lead to misjudgments or localized oscillations. This approach, through a deterministic termination mechanism driven by the graph structure, provides stronger convergence guarantees.

[0066] 8. The final heat vectors are normalized and sorted to ensure comparability and usability, efficiently outputting the top target products and forming an actionable precision marketing list. This sorted output directly aligns with marketing objectives, truly achieving a seamless "algorithm-to-decision" approach. Compared to traditional methods, which often produce output that is further removed from the business, this solution directly generates a list of available products based on the "target" of the campaign, lowering the operational understanding barrier and improving feasibility to meet commercial deployment requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0069] Example, see Figure 1 , precision marketing analysis methods based on big data, including:

[0070] Preprocess transaction logs, extract customer transaction sequences, and construct a global product transfer frequency matrix;

[0071] Based on the transfer frequency matrix, construct a transfer probability matrix between commodities;

[0072] Calculate the diffusion retention coefficient based on the out-degree information of the product;

[0073] Initialize the product popularity vector based on transaction records and time decay kernel function;

[0074] Construct a directed purchasing network between products and calculate the diameter of the network;

[0075] Perform heat diffusion iterative calculation based on the transition probability matrix and diffusion retention coefficient;

[0076] Determine whether the heat diffusion process converges. If the convergence condition is met, complete the calculation and obtain the final heat vector;

[0077] Normalize the heat vector, sort it in descending order by heat score, and output the sorting result as the target of precision marketing delivery.

[0078] By preprocessing the data to extract customer transaction sequences and construct a product transfer frequency matrix, modelable purchasing behaviors are extracted from the raw data. Next, by constructing a transition probability matrix and calculating the diffusion retention coefficient, the model is transformed into a graph-theoretic diffusion model, enabling the diffusion of interest between products. A time decay mechanism is further employed to initialize the interest, ensuring that the model reflects the influence of recent transaction behavior. A product purchase network is then constructed and its diameter is calculated, serving as an upper limit on the number of diffusion rounds, making the diffusion process controllable. By iterating the diffusion of interest based on the retention coefficient and transition probability, the indirect propagation of user interest is simulated. Finally, the interest vectors are normalized and sorted to accurately identify high-value products. This method achieves full-process modeling from behavioral data to marketing objectives, addressing the issues of unclear goals and delayed responses in traditional precision marketing, thereby improving decision-making efficiency and business conversion rates.

[0079] The pre-processing of transaction logs, extracting customer transaction sequences and constructing a global commodity transfer frequency matrix specifically includes:

[0080] Let the transaction log be an ordered set of tuples Where N is the set Total number of records in u i is the customer ID of the i-th transaction; p i ∈{1,2,...,P} is the product number of the i-th transaction, P is the total number of products; t i The discretized timestamp of the i-th transaction; normalize the transaction data structure and provide a unified format for subsequent processing;

[0081] Culling Collection Any one of u i ,p i ,t i Missing records; ensure data quality and prevent subsequent calculations from errors due to outliers or missing values;

[0082] Sort the set by customer ID u in ascending order, and then by time t in ascending order. Aggregate by segments to get the time series of each customer: {(p u,j ,t u,j )|j=1,2,...,n u}, n u =#{i|u i =u}; where u is the current customer ID; n u is the number of transactions of customer u; p u,j is the product number corresponding to the j-th transaction of customer u; tu,j The timestamp corresponding to the j-th transaction of customer u forms a complete purchase time series for each customer, laying the foundation for counting adjacent purchase pairs;

[0083] For each customer sequence, take two adjacent transactions (p u,j →p u,j+1 ) and count, statistical global conversion frequency matrix:

[0084] A=[A p,q ] P×P , Among them, A p,q The total number of times product p subsequently purchases product q; is an indicator function, which takes a value of 1 if the logic in the brackets is true, and 0 otherwise. It measures the absolute frequency of subsequent purchases of product q for product p and is used to construct a purchase migration network.

[0085] If no (p,q) transition occurs, then A p,q =0; ensure matrix integrity, no additional filling value is required.

[0086] By establishing a standard data structure for transaction logs, customer IDs, product numbers, and timestamps are systematically processed, making the raw data uniform and operational. After data cleaning, each customer's transaction time series is extracted by aggregating customer IDs and time in ascending order. Further, by analyzing adjacent transaction pairs, the "subsequent purchase frequency" between products is calculated to form a global transfer frequency matrix. This mechanism addresses the issues of traditional models that ignore user purchase order and lack behavioral chains, enhancing the ability to model dynamic user behavior. Furthermore, by traversing the entire customer sequence, a "behavioral graph between products" is formed, providing a solid foundation for subsequent transition probability calculations and graph structure establishment, helping to improve model accuracy and the credibility of behavior predictions.

[0087] The step of constructing a transition probability matrix between commodities based on the transition frequency matrix specifically includes:

[0088] For each product number p, calculate its total out-degree frequency Summarize the total out-degree frequency of product p for probability normalization;

[0089] Set the transition probability matrix to:

[0090] M=[M p,q ] P×P ,

[0091] Among them, M p,qis the conditional probability that the next purchase will be transferred to product q given that the current product is p; it represents the probability that the next purchase will be product q under the condition that the purchased product is p, realizing the mapping from frequency to probability.

[0092] The method focuses on "constructing a product transfer probability matrix" and normalizing the frequency matrix to form a probability graph with behavioral logic. By calculating the "total out-degree frequency" of each product—the total number of times it was subsequently purchased as a predecessor product—and using this as the denominator, the transfer frequencies of all subsequent products are normalized to form a "transition probability matrix." This method maps behavioral frequencies to probabilities, reflecting the user's probabilistic preference for purchasing other products after purchasing a certain product. This process resolves the issue of the original frequency metric being incomparable across categories and scales, making the transfer relationship between products more quantitative and establishing a mathematical foundation for the diffusion of popularity. As the weighted edge weights of a directed graph, this matrix serves as a key bridge in subsequent heat diffusion calculations, enhancing the scientific nature of the marketing model and the accuracy of promotion.

[0093] The calculation of the diffusion retention coefficient based on the out-degree information of the product specifically includes:

[0094] Calculate average outdegree Measuring the overall purchase migration density as the basis for coefficient calculation;

[0095] Setting the Diffusion Retention Coefficient The ratio of initial heat to newly diffused heat is retained in controlling heat diffusion, and it is completely data-driven, without the need for manual parameter adjustment.

[0096] A method for calculating a "diffusion retention coefficient" based on average out-degree is proposed. By calculating the average out-degree of all product nodes in the transfer graph, the connection density and activity level of the purchase network are indirectly measured. On this basis, a "diffusion retention coefficient" is set to control the weight ratio of "historical heat" and "newly diffused heat" in heat diffusion. This coefficient is generated entirely based on dynamic data, avoiding the problems of poor generalization and adaptability caused by traditional models that rely on manual experience to set parameters. On the one hand, it can effectively suppress heat dilution in high-density networks and prevent the loss of popularity of popular products; on the other hand, it can still maintain the initial heat in low-density networks, preventing information islands. Therefore, this step enhances the flexibility and stability of the heat propagation model and provides a controllable heat evolution mechanism for subsequent iterative calculations.

[0097] Initializing the product heat vector based on transaction records and the time decay kernel function specifically includes:

[0098] Set the global maximum timestamp Unified "current moment" benchmark for calculating time differences;

[0099] Set the discrete-time decay kernel to Where Δt is the transaction time difference; a smooth decay weight is given to the transaction time difference Δt to highlight the impact of recent purchases;

[0100] Calculate the initial heat:

[0101]

[0102] Among them, n p =#{i|p i =p}, for product p in the set The number of transactions in t p,i The timestamp of the i-th transaction of commodity p; It is the initial popularity of product p; it comprehensively reflects the purchase freshness and frequency of product p, and provides an initial value for diffusion.

[0103] The "initialization heat vector" has been refined, and by introducing a time-decay kernel function, product transaction records are converted into heat values ​​with time-weighted values. First, a global maximum timestamp is set to measure the recency of the transaction time. Then, a time-decay kernel function is set so that newer transactions assign higher initial heat to products. Finally, the initial heat vector is constructed by combining the number of product transactions and their time distribution. This solves the problems of "static heat" and "cold start difficulties" in existing recommendation systems, namely, the inability to reflect the potential of new products or recent hot trends. This step uses dynamic time series modeling to make subsequent diffusion and dissemination more closely aligned with actual market conditions, improving the real-time and precision of marketing responses.

[0104] The step of constructing a directed purchasing network between commodities and calculating the diameter of the network specifically includes:

[0105] Constructing a directed graph Where V = {1, 2, ... P}, E = {(p, q) | A p,q >0}; where V is a node set, representing P kinds of goods; E is a directed edge set, if A p,q >0, then there is an edge (p→q); the purchase migration relationship is abstracted into a graph structure, where nodes represent products and edges represent possible purchase paths;

[0106] For each source node p, use the breadth-first search method to find the shortest path length to each reachable node q:

[0107]

[0108] Where d(p,q) is the shortest path length from node p to q; l is the path length; if unreachable, d(p,q) = ∞; breadth-first search is an existing technique; it efficiently traverses the graph structure and calculates the shortest path length d(p,q) between each node pair;

[0109] Set the network diameter to Determine the upper limit of diffusion iteration rounds to ensure that all reachable paths are covered.

[0110] By abstracting the transfer relationships between products into a "directed purchase graph" and calculating the graph's maximum propagation path (diameter) using a breadth-first search algorithm, the system limits the number of subsequent diffusion iterations. By abstracting product nodes and their migration paths into graph nodes and edges, the system's visualization capabilities are enhanced while also providing a clear propagation space for heat diffusion. Introducing network diameter as a maximum round control criterion avoids resource waste and convergence oscillations caused by infinite iterations. Compared to traditional approaches that set fixed rounds or error thresholds, this graph-driven dynamic strategy is more theoretically sound and data-adaptable, ensuring algorithm convergence, efficiency, and complete coverage of heat diffusion.

[0111] The heat diffusion iterative calculation is performed based on the transition probability matrix and the diffusion retention coefficient, specifically including:

[0112] Assume that the iteration index k = {0, 1, ..., D-1}; define the range of iteration rounds to close the diffusion process;

[0113] Set the iterative update function to:

[0114]

[0115] is a diffusion term that transfers heat from all source nodes p to node q according to the transfer probability; It is a reserved item to ensure that the initial heat has a fixed proportion in each round;

[0116] in, is the popularity of product p after the kth round;

[0117] The vector form is h (k+1) =α·M T ·h (k) +(1-α)·h (0) Among them, M T It is the transpose of the matrix M, which facilitates the use of matrix operations to accelerate batch calculations.

[0118] A "heat diffusion iteration function" is defined, and by introducing a "transition probability matrix" and a "diffusion retention coefficient," the iterative propagation of heat in the purchasing network is achieved. Each iteration consists of two parts: a diffusion term that probabilistically diffuses heat to subsequent product nodes; and a retention term that partially retains the original heat to prevent complete dilution. This update mechanism is implemented through matrix multiplication, facilitating large-scale parallel computing and improving the computational efficiency of the system. Compared to recommendation methods based on static ratings or collaborative matrices, this method simulates the propagation of user preferences in the network, capturing the user's indirect interest chain, and effectively increasing the exposure opportunities of unpopular and long-tail products, thereby achieving higher user conversion rates and product diversity in actual marketing.

[0119] The process of determining whether the heat diffusion process has converged, and if so, completing the calculation to obtain the final heat vector, specifically includes:

[0120] The iteration ends when k=D, and the final heat vector is obtained:

[0121]

[0122] Since the network diameter covers all reachable paths, it ensures that the heat of the entire graph has been fully diffused and no additional convergence criteria are required.

[0123] The "convergence termination condition of heat diffusion" is clarified, that is, it is terminated after the iteration reaches the network diameter round, and the final heat vector is directly output. This strategy is based on the aforementioned graph structure reasoning and believes that within the maximum path length, all heat has been effectively propagated in the network. Therefore, there is no need for the complex error comparison and convergence detection steps in traditional methods. On the one hand, this method greatly simplifies program logic and computing resources, and on the other hand, it avoids the occurrence of "false convergence" or "unending oscillation" in high-noise or high-sparse networks, thereby improving the stability, certainty and predictability of the system. It is a more reliable and engineering-implementable termination strategy.

[0124] Normalizing the heat vectors, sorting them in descending order by heat score, and outputting the sorting results as precision marketing delivery targets specifically include:

[0125] Calculate the normalized heat score Convert popularity values ​​into relative scores to facilitate comparison between products of different sizes;

[0126] Press c p Sort in descending order, generate a sequence Clarify the marketing priority of each product;

[0127] Take the first K items {p1,p2,...,p K Output the required list and ranking as the target of precision marketing delivery; provide a directly available product list for subsequent delivery strategies.

[0128] A mechanism for normalizing, sorting, and outputting the final heat vector is proposed, allowing the model output to be directly converted into a precision marketing strategy execution list. First, by normalizing the heat, the problem of incomparable heat between different products is solved; then, by sorting in descending order, the marketing priority is clarified, and finally, the top N products and their ranking are output as the basis for decision-making. Through this process, this method realizes a closed-loop chain from raw behavioral data → behavioral graph → heat evolution → marketing decision, truly achieving "data-driven automated marketing". Compared with traditional methods that rely on labeling systems or manual experience judgment, this method is more objective, real-time, efficient, and highly automated, greatly improving marketing effectiveness and resource utilization.

[0129] This embodiment also provides a system for a precision marketing analysis method based on big data, including:

[0130] Computation module: performs pre-processing of transaction logs, extracts customer transaction sequences, constructs a global product transfer frequency matrix, performs matrix construction, heat initialization, and iterative calculations;

[0131] Sorting module: normalizes, sorts and outputs the final heat vector.

[0132] From a system implementation perspective, the entire method is divided into a computation module and a sorting module to facilitate engineering deployment and system integration. The computation module is responsible for all data processing and analysis logic, including data cleaning, frequency matrix generation, heat initialization, and iterative calculations; the sorting module focuses on normalization processing and result output. This modular division facilitates the system's distributed architecture design, allowing for independent expansion on big data platforms, improving the system's maintainability and computational concurrency. Furthermore, this modular design facilitates subsequent functional upgrades or algorithm replacements, ensuring the system's long-term evolution and commercial scalability. It is a highly practical engineering implementation strategy.

[0133] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0134] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. The precision marketing analysis method based on big data is characterized by: include: Preprocess transaction logs, extract customer transaction sequences, and construct a global product transfer frequency matrix; Based on the transfer frequency matrix, construct a transfer probability matrix between commodities; Calculate the diffusion retention coefficient based on the out-degree information of the product; Initialize the product popularity vector based on transaction records and time decay kernel function; Construct a directed purchasing network between products and calculate the diameter of the network; Based on the transition probability matrix and the diffusion retention coefficient, perform heat diffusion iterative calculation; Determine whether the heat diffusion process converges. If the convergence condition is met, complete the calculation and obtain the final heat vector; Normalize the heat vectors, sort them in descending order by heat score, and output the sorting results as the target for precision marketing.

2. The precision marketing analysis method based on big data according to claim 1, characterized in that: The pre-processing of transaction logs, extracting customer transaction sequences and constructing a global commodity transfer frequency matrix specifically includes: Let the transaction log be an ordered set of tuples Where N is the set Total number of records in u i is the customer ID of the i-th transaction; p i ∈{1,2,...,P} is the product number of the i-th transaction, P is the total number of products; t i is the discretized timestamp of the i-th transaction; Culling Collection Any one of u i ,p i ,t i missing records; Sort the set by customer ID u in ascending order, and then by time t in ascending order. Aggregate by segments to get the time series of each customer: {(p u,j ,t u,j )|j=1,2,...,n u }, n u =#{i|u i =u}; where u is the current customer ID; n u is the number of transactions of customer u; p u,j is the product number corresponding to the j-th transaction of customer u; t u, The timestamp corresponding to the j-th transaction of customer u; For each customer sequence, take two adjacent transactions (p u,j →p u,j+1 ) and count, statistical global conversion frequency matrix: A=[A p,q ] P×P , Among them, A p,q For business The total number of times product p subsequently purchases product q; It is an indicator function, which has a value of 1 if the logic in the brackets is true, otherwise it is 0; If no (p,q) transition occurs, then A p,q =0.

3. The precision marketing analysis method based on big data according to claim 2, characterized in that: The step of constructing a transition probability matrix between commodities based on the transition frequency matrix specifically includes: For each product number p, calculate its total out-degree frequency Set the transition probability matrix to: M=[M p,q ] P×P , Among them, M p,q is the conditional probability that the next purchase will be transferred to product q given that the current product is p.

4. The precision marketing analysis method based on big data according to claim 3 is characterized in that: The calculation of the diffusion retention coefficient based on the out-degree information of the product specifically includes: Calculate average outdegree Setting the Diffusion Retention Coefficient 5. The precision marketing analysis method based on big data according to claim 4 is characterized in that: Initializing the product heat vector based on transaction records and the time decay kernel function specifically includes: Set the global maximum timestamp Set the discrete-time decay kernel to Where Δt is the transaction time difference; Calculate the initial heat: Among them, n p =#{i|p i =p}, for product p in the set The number of transactions in t p,i The timestamp of the i-th transaction of commodity p; is the initial popularity of product p.

6. The precision marketing analysis method based on big data according to claim 5, characterized in that: The step of constructing a directed purchasing network between commodities and calculating the diameter of the network specifically includes: Constructing a directed graph Where V = {1, 2, ... P}, E = {(p, q) | A p,q >0}; where V is a node set, representing P kinds of goods; E is a directed edge set, if A p,q >0, then there is an edge (p→q); For each source node p, use the breadth-first search method to find the shortest path length to each reachable node q: Where d(p,q) is the shortest path length from node p to q; l is the path length; if it is unreachable, then d(p,q) = ∞; Set the network diameter to 7. The precision marketing analysis method based on big data according to claim 6, characterized in that: The heat diffusion iterative calculation is performed based on the transition probability matrix and the diffusion retention coefficient, specifically including: Let iteration index k = {0, 1, ..., D-1}; Set the iterative update function to: in, is the popularity of product p after the kth round; The vector form is h (k+1) =α·M T ·h (k) +(1-α)·h (0) Among them, M T is the transpose of the matrix M.

8. The precision marketing analysis method based on big data according to claim 7, characterized in that: The process of determining whether the heat diffusion process has converged, and if so, completing the calculation to obtain the final heat vector, specifically includes: The iteration ends when k=D, and the final heat vector is obtained:

9. The precision marketing analysis method based on big data according to claim 8, characterized in that: Normalizing the heat vectors, sorting them in descending order by heat score, and outputting the sorting results as precision marketing delivery targets specifically include: Calculate the normalized heat score Press c p Sort in descending order, generate a sequence Take the first K items {p1,p2,...,p K }As a precision marketing delivery target, output the required list and ranking.

10. A system using the big data-based precision marketing analysis method according to claim 9, characterized in that: include: Computation module: performs pre-processing of transaction logs, extracts customer transaction sequences, constructs a global product transfer frequency matrix, performs matrix construction, heat initialization, and iterative calculations; Sorting module: normalizes, sorts and outputs the final heat vector.

Images