Shapley value and deep learning-based sku-level bid decision path mapping method and system

Abstract

The application discloses a SKU-level delivery decision path mapping method and system based on Shapley value and deep learning, and relates to the technical field of data processing. The method collects multi-channel SKU-level user behavior data and pre-processes the same, quantizes contact contribution values based on an improved Shapley value algorithm and an LSTM modeling, stores the same to an optimization database to support tracing, constructs a visual engine to output a key path, and integrates data to establish quality monitoring, thereby solving the problems that SKU-level data deep analysis, contact contribution accurate attribution and decision path efficient visualization are difficult to realize in the prior art, and having the advantages that SKU-level data deep analysis, contact contribution accurate attribution and decision path efficient tracing and dynamic visualization can be realized, and marketing efficiency and effect are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for mapping SKU-level delivery decision paths based on Shapley value and deep learning. Background Technology

[0002] With the rapid development of the digital marketing industry, refined management of brand advertising decisions has become a core requirement for improving marketing efficiency and effectiveness. In e-commerce and other fields, SKU (Stock Keeping Unit) serves as the basic unit for product circulation and sales. Different SKUs have significantly different target customer groups, market performance, and conversion paths, thus requiring differentiated advertising strategies for each SKU.

[0003] Structured campaign decision data typically possesses a unified data model and attributes, such as SKU exposure touchpoints, user interaction behavior, and conversion results. This data is essentially a centralized representation of multi-channel marketing activities at the data level. However, the data formats and statistical standards differ across different marketing channels (such as Xiaohongshu, Douyin, and Taobao), and user decision paths are both temporal and complex, posing significant challenges to the integration and analysis of SKU-level campaign data.

[0004] Currently, mainstream campaign decision-making methods primarily rely on simple attribution models (such as first-touchpoint attribution and last-touchpoint attribution) or traditional multi-touchpoint attribution techniques. A touchpoint refers to any channel or device through which a user interacts with marketing content, such as social media advertising, search engine advertising, email, and mobile applications. These touchpoints constitute the connection points between users and marketing content and are an important part of the user's behavior path. These methods have significant limitations: firstly, they cannot delve into granular data at the SKU level, only obtaining overall brand campaign performance, making it impossible to accurately identify the key conversion touchpoints for individual SKUs; secondly, existing attribution techniques struggle to quantify the true contribution of different touchpoints to SKU conversion and lack effective data traceability and visualization tools, failing to intuitively present the user's decision-making path. Furthermore, existing solutions are insufficient in multi-channel data fusion and data quality monitoring, making it difficult to handle the processing needs of massive amounts of multi-format data.

[0005] As brands accumulate a large number of SKUs and diversify their marketing channels, traditional manual analysis or crude attribution methods become inefficient, waste marketing resources, and fail to meet the demands of refined brand operations. This necessitates an automated processing method and system capable of SKU-level data refinement, accurate attribution, path tracing, and visualization to address the limitations of existing technologies. Summary of the Invention

[0006] The purpose of this application is to provide a SKU-level campaign decision path mapping method and system based on Shapley value and deep learning. It has the advantages of enabling in-depth analysis of SKU-level data, accurate attribution of touchpoint contributions, efficient tracing and dynamic visualization of decision paths, thereby improving marketing efficiency and effectiveness.

[0007] This application provides a SKU-level delivery decision path mapping method based on Shapley value and deep learning, including the following steps:

[0008] S100. Collect behavioral data of SKU-level users from multiple channels, and preprocess the behavioral data using the random forest algorithm to remove a preset proportion of noise data and outliers, thereby obtaining standardized data including exposure time, touchpoint type, conversion result and basic SKU attributes.

[0009] S200, based on the improved Shapley value multi-touch point attribution algorithm, combines the Long Short-Term Memory (LSTM) network to model the touch point sequence and quantify the touch point contribution value of each touch point to the conversion of a single SKU;

[0010] S300: Store the standardized data and the touchpoint contribution value into the optimized distributed time-series database InfluxDB, and support the decision path tracing of SKU-level users through data sharding, replication mechanism and caching technology;

[0011] S400 constructs a path heatmap generation engine, using WebGL 3D visualization technology and Apache Flink real-time computing technology to visualize and output key path nodes of SKU transformation.

[0012] S500: Integrate multi-channel data through the data fusion module, use the Apriori association analysis algorithm to mine data association relationships, and establish a real-time data quality monitoring mechanism. The multi-channel data includes behavioral data and standardized data of multi-channel SKU-level users. The data used to mine association rules includes touchpoint sequence data, conversion result data, channel association data, and SKU attribute association data. The objects of data quality monitoring include multi-channel data, touchpoint contribution values, and integrated data after fusion.

[0013] Furthermore, the random forest algorithm in step S100 includes multiple decision trees, uses the Gini coefficient as the feature selection criterion, and determines outliers based on the three-standard-deviation rule.

[0014] Furthermore, the Long Short-Term Memory (LSTM) network in step S200 allows for custom settings of the number of hidden layer nodes, learning rate, number of iterations, and batch size, and employs the Adam optimizer and cross-entropy loss function to optimize model parameters.

[0015] Furthermore, the optimization of the distributed time-series database InfluxDB in step S300 includes: dividing the data into 4 shards according to the data access frequency, with each shard corresponding to an independent storage node, and using a consistent hashing algorithm to allocate the data; setting 3 replicas for each shard, and using an asynchronous replication mechanism for replica synchronization.

[0016] Furthermore, in step S300, Redis caching technology is introduced, adopting an LRU replacement strategy, with the cache key being SKU_ID-touchpoint type-timestamp; the query optimization of the distributed time-series database InfluxDB adopts a B+ tree index structure, and a composite index is established for the exposure time and SKU_ID fields.

[0017] Furthermore, the WebGL 3D visualization technology in step S400 implements path node rendering through vertex shaders and fragment shaders. The vertex coordinate calculation formula is as follows:

[0018] Among them, the touchpoint timing order value is the time sequence number of the touchpoint in the user's decision path, representing the temporal position of the touchpoint; the scaling factor is the coordinate scaling coefficient of the timing order value, with a value range of [0.1, 1.0], used to adjust the display scale of the temporal dimension in three-dimensional space; the touchpoint contribution value is the quantitative contribution value of the touchpoint to SKU conversion; the weight coefficient is the coordinate scaling coefficient of the touchpoint contribution value, used to map the touchpoint contribution value to the height range of three-dimensional space for visual differentiation; the channel identifier code value is the unique code of the channel to which the touchpoint belongs, representing the channel source of the touchpoint;

[0019] The interactive features of the path heatmap include: locating nodes using ray detection technology when the mouse hovers over them, and returning the node's touch contribution value, exposure, and conversion rate in real time; and triggering an asynchronous request when a node is clicked to load the user behavior sequence data associated with that node.

[0020] Apache Flink real-time computing technology employs event-time semantics and uses a watermark mechanism to process out-of-order data.

[0021] Furthermore, in step S500, the data fusion module uses ETL tools to clean and transform the data, verifies the consistency of the data format through JSONSchema, and standardizes the touch point type fields using regular expressions;

[0022] The Apriori association analysis algorithm sets the support threshold for association rules to 0.05 and the confidence threshold to 0.7. Redundant association rules are removed through pruning strategies to uncover the association between touchpoint combinations and SKU conversion.

[0023] The data quality monitoring mechanism includes: periodically collecting data quality indicators, using a sliding window algorithm to calculate the average value of the indicators, and triggering an alarm when the indicators are lower than a preset threshold.

[0024] This application also proposes a SKU-level delivery decision path mapping system based on Shapley value and deep learning, used to execute the aforementioned SKU-level delivery decision path mapping method based on Shapley value and deep learning. The system includes:

[0025] The data preprocessing module is used to collect SKU-level user behavior data from multiple channels, remove outliers using the random forest algorithm, and output standardized data.

[0026] The attribution algorithm module, which includes an LSTM modeling unit and an adaptive adjustment unit, is used to execute an improved Shapley value multi-touchpoint attribution algorithm to quantify the touchpoint contribution value of touchpoints to SKU conversion.

[0027] The data storage module includes a distributed time-series database optimization unit and a Redis caching unit, which are used to store standardized data and contact point contribution values, and support millisecond-level traceability;

[0028] The visualization module is used to build a path heatmap generation engine, which outputs the critical path for interactive SKU conversion through WebGL 3D rendering and Flink real-time calculation.

[0029] The data fusion module, which includes a correlation analysis unit and a quality monitoring unit, is used to integrate data from multiple channels and ensure data reliability.

[0030] Furthermore, the LSTM modeling unit of the attribution algorithm module adopts a bidirectional LSTM structure, and the outputs of the forward and backward hidden layers are fused through a splicing operation.

[0031] Furthermore, the distributed time-series database optimization unit of the data storage module also includes an automated backup tool, which adopts a combination of incremental backup and full backup.

[0032] This application has the following beneficial effects:

[0033] This application provides a method and system for mapping SKU-level delivery decision paths based on Shapley value and deep learning. It collects and preprocesses SKU-level user behavior data from multiple channels, quantifies touchpoint contribution values ​​based on an improved Shapley value algorithm and LSTM modeling, stores the data in an optimized database for traceability, constructs a visualization engine to output critical paths, and integrates data to establish quality monitoring. This solves the problems in existing technologies that make it difficult to achieve in-depth analysis of SKU-level data, accurate attribution of touchpoint contributions, and efficient visualization of decision paths. It has the advantages of enabling in-depth analysis of SKU-level data, accurate attribution of touchpoint contributions, efficient traceability of decision paths, and dynamic visualization, thereby improving marketing efficiency and effectiveness. Attached Figure Description

[0034] Figure 1 A flowchart illustrating the SKU-level delivery decision path mapping method based on Shapley value and deep learning provided for this application;

[0035] Figure 2 A schematic diagram of the structure of the SKU-level delivery decision path mapping system based on Shapley value and deep learning provided in this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. A process can be terminated when its operation is completed, but it may also have additional steps not included in the drawings. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0037] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0038] In the digital marketing field, existing campaign decision-making methods rely on simple attribution models or traditional multi-touchpoint attribution techniques, which cannot delve into granular data at the SKU level. Differences in multi-channel data formats and the temporal complexity of user decision paths lead to difficulties in data integration, inaccurate quantification of touchpoint contributions, and a lack of effective path tracing and visualization tools. This makes the system inefficient in processing massive amounts of multi-format data, unable to meet the needs of SKU-level granular operations, and consequently affecting the accuracy of marketing resource allocation and decision-making efficiency.

[0039] For example, in an e-commerce platform, a fast-moving consumer goods (FMCG) brand operates hundreds of SKUs, with user interactions spanning multiple channels such as Xiaohongshu (Little Red Book) recommendations, Douyin (TikTok) short video exposure, and Taobao search clicks. Once a user completes a conversion, the existing system can only attribute the outcome based on the last touchpoint, failing to identify the potential impact of Douyin exposure on specific SKUs. Specifically, for high-priced SKUs, the contribution of Xiaohongshu content touchpoints is underestimated, leading to an over-allocation of budget to Taobao search while neglecting the long-term value of the recommendation channel. Furthermore, during the multi-channel data fusion process, inconsistent statistical standards result in fragmented touchpoint sequence data, making it impossible to fully trace the user's decision-making path and distorting the attribution results.

[0040] In this regard, such as Figure 1 As shown, this application proposes a SKU-level delivery decision path mapping method based on Shapley value and deep learning, which includes at least the following steps:

[0041] S100: Collect behavioral data of SKU-level users from multiple channels, and preprocess the behavioral data using the random forest algorithm to remove noise data and outliers by a preset proportion, so as to obtain standardized data including exposure time, touch point type, conversion result and basic SKU attributes.

[0042] S200, based on the improved Shapley value multi-touch point attribution algorithm, combines the Long Short-Term Memory (LSTM) network to model the touch point sequence and quantify the touch point contribution value of each touch point to the conversion of a single SKU;

[0043] S300: Store the standardized data and the contribution value of the touchpoint into the optimized distributed time-series database InfluxDB, and support the decision path tracing of SKU-level users through data sharding, replication mechanism and caching technology;

[0044] S400 constructs a path heatmap generation engine, using WebGL 3D visualization technology and Apache Flink real-time computing technology to visualize and output key path nodes of SKU transformation.

[0045] S500 integrates multi-channel data through the data fusion module, uses the Apriori association analysis algorithm to mine data relationships, and establishes a real-time data quality monitoring mechanism. The multi-channel data includes behavioral data and standardized data of multi-channel SKU-level users. The data used to mine association rules includes touchpoint sequence data, conversion result data, channel association data, and SKU attribute association data. The objects of data quality monitoring include multi-channel data, touchpoint contribution values, and integrated data after fusion.

[0046] In step S100, behavioral data of SKU-level users from multiple channels is collected. This behavioral data is preprocessed using a random forest algorithm to remove a predetermined proportion of noise and outliers, resulting in standardized data including exposure time, touchpoint type, conversion result, and basic SKU attributes. Specifically, user behavior data can be obtained from various marketing channels (e.g., social media platforms, e-commerce websites, search engines, etc.) through a log collection system. For example, data interfaces can be manually configured to periodically export CSV files from different platforms and then integrate them. Various methods can be used for data preprocessing. For example, a fixed threshold can be set to perform simple truncation or filtering, marking data exceeding the threshold as outliers and removing them. Noisy data can be smoothed using simple moving average or median filtering. Standardized data can be achieved by manually defining data patterns, mapping data fields from different sources to a unified structure. For example, all "click" events from all channels can be uniformly named "Click," and "view" events can be uniformly named "View."

[0047] In step S200, based on the improved Shapley value multi-touchpoint attribution algorithm, and combined with a Long Short-Term Memory (LSTM) network, the touchpoint sequence is modeled to quantify the touchpoint contribution value of each touchpoint to the conversion of a single SKU. The quantification of touchpoint contribution values ​​can be based on the traditional Shapley value algorithm. For example, all touchpoint combinations can be enumerated, the marginal contribution of each touchpoint in different combinations can be calculated, and then the average value can be taken as its contribution. Touchpoint sequence modeling can employ simple statistical methods. For example, the frequency of each touchpoint appearing in the conversion path can be counted, or the transition probability between touchpoints can be calculated to reflect its role in the sequence. Contribution values ​​can be simply allocated using a linear weighted model, for example, by distributing the contribution values ​​of all touchpoints equally, or by assigning different fixed weights to the touchpoints based on their position in the conversion path (such as the first touchpoint or the last touchpoint).

[0048] In step S300, the standardized data and the touchpoint contribution value are stored in the optimized distributed time-series database InfluxDB. Data sharding, replication mechanisms, and caching technologies support SKU-level user decision path tracing. Data storage can utilize a non-distributed relational database, such as MySQL. The standardized data and touchpoint contribution value are stored as fields in a regular table. Data tracing can be achieved by directly querying the database tables. For example, when querying the user decision path for a specific SKU, an SQL query is executed directly in the database to retrieve the relevant data. Caching technology can employ simple memory caching, such as storing a small amount of recently queried data in the application server's memory.

[0049] In step S400, a path heatmap generation engine is constructed, employing WebGL 3D visualization technology and Apache Flink real-time computing technology to visualize and output key path nodes for SKU conversion. The path heatmap can be constructed using a 2D charting library (such as ECharts or D3.js). For example, touchpoints can be used as nodes, paths as edges, and the size or color of nodes can represent their contribution value, generating a static 2D path map. Visualization output can employ traditional server-side rendering methods. For example, image files can be generated on the server side and then sent to the front end for display. Real-time computing can utilize batch processing, such as running a batch processing task at regular intervals (e.g., hourly or daily) to calculate and update path data. Key path nodes can be identified through manual analysis or by pre-defined rules. For example, the touchpoints with the highest conversion rates or the most frequently occurring touchpoint combinations can be marked as key path nodes.

[0050] In step S500, multi-channel data is integrated through a data fusion module. The Apriori association analysis algorithm is used to mine data relationships, establishing a real-time data quality monitoring mechanism. This multi-channel data includes behavioral data and standardized data of users at the SKU level across multiple channels. Data used for mining association rules includes touchpoint sequence data, conversion result data, channel association data, and SKU attribute association data. The objects of data quality monitoring include multi-channel data, touchpoint contribution values, and integrated data after fusion. Data fusion can be achieved manually using scripts. For example, a Python script can be written to read CSV files from different channels, perform field matching and merging, and then output to a unified data file. Association analysis can employ simple statistical methods. For example, the co-occurrence frequency of different touchpoint combinations can be statistically analyzed, and support and confidence scores can be manually calculated to discover potential association rules. Data quality monitoring can be conducted through regular manual checks. For example, data can be manually sampled and checked daily or weekly to verify data format, completeness, and accuracy. When problems are found, notifications are sent via email or instant messaging tools.

[0051] For example, suppose a brand wants to optimize its campaign strategy for a specific SKU (e.g., a product called "Smartphone X") across multiple digital marketing channels (e.g., e-commerce platform A, social media platform B, and search engine C). The problem this brand faces is that it cannot accurately understand the complex decision-making process of users before purchasing "Smartphone X," nor can it quantify the actual contribution of each marketing touchpoint to the final conversion, resulting in inefficient allocation of its marketing budget.

[0052] In step S100, the system automatically collects user behavior data related to "Smartphone X" from multiple channels, including e-commerce platform A, social media platform B, and search engine C. This data may include the time users spend browsing "Smartphone X" product pages on e-commerce platform A, events of clicking "Smartphone X" ads on social media platform B, searches for "Smartphone X" keywords on search engine C, and the conversion results of whether or not they ultimately purchase "Smartphone X". The collected raw data may contain inconsistent formats, noise, outliers such as invalid clicks or bot traffic. To address this issue, the system uses a random forest algorithm to preprocess this raw behavioral data. For example, by analyzing user behavior patterns, it identifies and removes abnormal clicks or browsing records that significantly deviate from normal user behavior. Simultaneously, the data is cleaned and formatted to generate standardized data containing consistent exposure time, touchpoint type, conversion results, and basic attributes of "Smartphone X" (such as price and configuration). This standardization process ensures the data quality and consistency for subsequent analysis.

[0053] In step S200, to accurately quantify the contribution of each touchpoint to the conversion of "Smartphone X", the system models the user's touchpoint sequence based on an improved Shapley value multi-touchpoint attribution algorithm and combined with a Long Short-Term Memory (LSTM) network. Specifically, for a series of user A's interactions before purchasing "Smartphone X" (e.g., social media ad exposure -> search engine click -> e-commerce platform browsing -> final purchase), the LSTM network analyzes the temporal order and contextual relationships of each touchpoint in the sequence, generating a temporal weight for each touchpoint to reflect its importance in the entire decision-making path. Subsequently, the improved Shapley value algorithm comprehensively considers these temporal weights to calculate the specific contribution of each touchpoint (such as social media ad, search engine click, e-commerce platform browsing) to user A's conversion of purchasing "Smartphone X". For example, if social media ads play a key guiding role in the early stages of user decision-making, and e-commerce platform browsing is the final step, then their respective contributions will be accurately quantified, avoiding the drawback of simply attributing all credit to the first or last touchpoint in traditional attribution models.

[0054] In step S300, to support the rapid tracing and analysis of the decision-making path of "Smartphone X" users, the system stores the standardized data generated above, along with the contribution value of each touchpoint, into the optimized distributed time-series database InfluxDB. This database uses a data sharding mechanism to distribute massive amounts of data across different nodes and a replication mechanism to ensure high availability and fault tolerance. Simultaneously, the introduced caching technology (e.g., temporarily storing frequently queried "Smartphone X" related path data) significantly improves data query speed. Therefore, when marketers need to trace the complete decision-making path of user A's purchase of "Smartphone X," the system can quickly retrieve and present all touchpoints and their contribution values ​​from user A's first contact to final conversion within milliseconds, solving the problem of low query efficiency in traditional databases when processing massive amounts of time-series data.

[0055] In step S400, to visually represent the critical conversion path of "Smartphone X," the system constructs a path heatmap generation engine. This engine utilizes WebGL 3D visualization technology to render each touchpoint in the user's decision-making path as a 3D node. The size, color, or height of the node visually reflects its contribution, exposure, or conversion rate. Simultaneously, combined with Apache Flink real-time computing technology, the system can process new user behavior data in real time and dynamically update the path heatmap. For example, when a new user purchases "Smartphone X," their decision path is immediately calculated and incorporated into the heatmap, allowing marketers to observe the most active and convertible critical path nodes for "Smartphone X" in real time. For instance, "social media advertising -> brand website -> e-commerce platform order" might be a high-contribution path. This visualization method significantly improves the efficiency of understanding complex user decision paths.

[0056] In step S500, to comprehensively integrate multi-channel data and ensure data quality, the system cleans, transforms, and unifies "Smartphone X" related data from different channels through a data fusion module. Using the Apriori association analysis algorithm, the system uncovers deep correlations between touchpoint sequence data, conversion result data, channel-related data, and SKU attribute-related data. For example, a strong correlation can be found between "seeing a review video of 'Smartphone X' on social media platform B" and "searching for 'Smartphone X' on e-commerce platform A and ultimately purchasing it." Simultaneously, the system establishes a real-time data quality monitoring mechanism to continuously monitor the completeness, accuracy, and consistency of multi-channel data, touchpoint contribution values, and the integrated data after fusion. Once data anomalies are detected (e.g., a sudden and significant drop in data volume from a channel or unreasonable fluctuations in touchpoint contribution values), the system immediately triggers an alarm to ensure the reliability of subsequent analysis. Through these steps, this method enables refined analysis, precise attribution, path tracing, and visualization of "Smartphone X" campaign decisions, effectively addressing the inefficiencies and resource waste faced by brands in multi-SKU, multi-channel marketing.

[0057] In summary, the method provided in this embodiment integrates advanced attribution algorithms, deep learning models, distributed storage and real-time computing technologies, as well as powerful visualization and data governance capabilities. This enables refined analysis, accurate attribution, efficient tracing, and intuitive visualization of SKU-level campaign decision-making paths, effectively overcoming the limitations of existing technologies in terms of efficiency, accuracy, and operability. It provides brands with more scientific and efficient marketing decision support.

[0058] In some implementations, the random forest algorithm in step S100 includes multiple decision trees, uses the Gini coefficient as the feature selection criterion, and determines outliers based on the three-standard-deviation rule.

[0059] The Random Forest algorithm comprises multiple decision trees. It's an ensemble learning method that improves model accuracy and robustness by constructing multiple decision trees and combining their predictions. Each decision tree independently classifies or regresses the data, and the final result is determined by the votes or average of all decision trees. This ensemble approach effectively reduces the risk of overfitting from a single decision tree and enhances the model's resistance to noisy data. For example, a Random Forest can be configured to contain 100 decision trees, or the number of decision trees can be dynamically adjusted based on cross-validation results to achieve optimal performance.

[0060] The Gini coefficient is used as the feature selection criterion. The Gini coefficient (Gini Impurity) is a commonly used feature selection criterion in decision tree algorithms, used to measure the purity of a dataset. When constructing a decision tree, the algorithm selects features that minimize the Gini coefficient as splitting nodes, resulting in child nodes containing more pure data categories. Besides the Gini coefficient, information gain or gain ratio can also be used as feature selection criteria; these also aim to find features that effectively distinguish between different categories of data.

[0061] Outlier detection is based on the three-standard-deviation rule. The three-standard-deviation rule is a statistically based outlier detection method applicable when data approximately follows a normal distribution. This rule states that if a data point differs from the mean of the dataset by more than three standard deviations, it is considered an outlier. This method is simple, intuitive, and easy to implement, effectively identifying extreme values ​​that deviate from the main distribution of the data. Alternatively, outlier detection can also be performed using the IQR (interquartile range) rule of boxplots, the density-based DBSCAN algorithm, or the distance-based LOF (Local Outlier Factor) algorithm.

[0062] For example, when performing step S100, the random forest algorithm can be configured to include 150 decision trees, with each decision tree having a maximum depth limit of 10 layers to balance model complexity and generalization ability. When constructing these decision trees, the feature selection criterion uses the Gini coefficient. For instance, if the decrease in the Gini coefficient after a node splits is less than a preset threshold (e.g., 0.01), the splitting stops. For outlier identification, the mean and standard deviation can be calculated for key numerical features in the standardized data, such as "exposure time" or "conversion result." Any data point exceeding the range of the mean plus or minus three standard deviations—for example, the exposure time of a certain SKU far exceeding three standard deviations of the average exposure time—is marked as an outlier and removed. In this way, it can be ensured that noise and outliers in behavioral data can be effectively and accurately identified and removed during the data preprocessing stage, thereby obtaining high-quality standardized data.

[0063] In some implementations, an improved Shapley value multi-touchpoint attribution algorithm is used to quantify the contribution of each touchpoint to the conversion of a single SKU. However, in actual user decision-making paths, the temporal relationships and contextual dependencies between touchpoints have a significant impact on the final conversion result. Traditional attribution methods may struggle to fully capture these complex temporal characteristics, leading to inaccurate quantification of touchpoint contribution values. For example, traditional Shapley value calculations treat the Shapley value as a static parameter in time, failing to consider how the contribution calculation needs to be adjusted to follow touchpoint changes in time, or failing to consider the impact of the temporal characteristics of multiple touchpoints on changes in contribution values. To address this, this application proposes an improved Shapley value multi-touchpoint attribution algorithm, the steps of which include:

[0064] S210, Define the contact point set Constructing a SKU conversion value function : ,in For a subset of touchpoints, This indicates the number of touchpoints corresponding to subsequent user interactions after browsing. This indicates the number of touchpoints corresponding to user browsing and interaction behaviors, and the conversion value coefficient indicates the value weight of conversion volume relative to exposure volume.

[0065] S220. Calculate the traditional Shapley value:

[0066] ,in, For the first The contact contribution value (i.e., the traditional Shapley value) of each contact represents the average marginal contribution of that contact across all combinations. N This is a collection of all touchpoints, with each element corresponding to a traceable user interaction touchpoint, including ad impressions and link clicks. S For the set of contact points N A non-empty subset represents a combination of touchpoints participating in a certain transformation path. For subset S The conversion value function refers to the comprehensive value quantification of SKU conversion, reflecting the conversion efficiency of this touchpoint combination. n The total number of contacts. Indicates to non-empty subsets New One contact point Represents the computation of non-empty subsets Number of contacts, symbol This represents the factorial operation;

[0067] S230, Captures contact timing sequences using a Long Short-Term Memory (LSTM) network. Based on long-term dependencies and temporal characteristics, output temporal weights. , A sequence of touchpoints for a single user to interact with a target SKU. T The time step for touch interaction. For the first t The touch feature vectors at each time step, where the hidden layer state update formula of the Long Short-Term Memory (LSTM) network is:

[0068]

[0069]

[0070] in, For the Long Short-Term Memory (LSTM) network, the first... t The hidden layer state vector at each time step is fused with the temporal feature information of all previous touchpoints; For the first t The time-series weight coefficient of each touchpoint quantifies the importance of the touchpoint in the entire time sequence. The higher the weight, the greater the impact on subsequent conversions. This is the weight matrix for updating the hidden layer states of the Long Short-Term Memory (LSTM) network, used to adjust the fusion ratio between the previous hidden state and the current touch feature. The bias term for updating the hidden layer state of the Long Short-Term Memory (LSTM) network is used to correct the initial offset of the hidden layer state. The weight matrix output by the time-series weights is used to map the hidden layer state to the touch point weight coefficients; This is a bias term for the time-series weight output, used to correct the output range of the weight coefficients;

[0071] S240, Introducing adaptive adjustment parameters to the traditional Shapley value Calculate the improved contact contribution value. As shown in the following formula:

[0072] ;

[0073] In the above formula, Adaptively adjusted based on data distribution characteristics, and satisfying Data centrality is quantified by the coefficient of variation, which is the ratio of standard deviation to mean, reflecting the dispersion of SKU conversion data.

[0074] Specifically, a touchpoint set is defined, and an SKU conversion value function is constructed. The touchpoint set refers to all interaction points a user might encounter during their decision-making process, such as ad impressions, link clicks, page views, and adding items to the cart. The SKU conversion value function is a quantitative model used to evaluate the overall value of a specific combination of touchpoints for the final conversion of an SKU. The touchpoint set can be predefined according to business needs; for example, it can consider interactions across all advertising platforms, content platforms, and social media. The SKU conversion value function can be a simple binary function (1 for successful conversion, 0 for failure) or a more complex multi-valued function considering conversion amount or profit. Alternatively, the touchpoint set can be defined using a dynamic identification mechanism, automatically identifying key interaction points in user behavior data through machine learning models. The SKU conversion value function can be constructed based on models such as logistic regression and support vector machines to output a conversion probability value between 0 and 1, serving as a quantification of its conversion value.

[0075] Calculate the traditional Shapley value. The traditional Shapley value is an allocation method in cooperative game theory used to fairly distribute the contribution of each participant in a cooperation. Here, it is used to quantify the average marginal contribution of each touchpoint to the final conversion of the SKU across all possible combinations of touchpoints. Calculating the Shapley value typically involves traversing all possible subsets of touchpoints and their permutations, calculating the change in the conversion value function before and after each touchpoint is added to the subset, and then averaging the results. This can be achieved through Monte Carlo simulation or exact calculation (for a small number of touchpoints). As an alternative approach, approximate calculation methods can be used, such as estimating the Shapley value by randomly sampling permutations of touchpoints, to reduce computational complexity, especially when the number of touchpoints is large.

[0076] The touchpoint time-series sequence is modeled using a Long Short-Term Memory (LSTM) network, outputting temporal weights. LSTM is a special type of Recurrent Neural Network (RNN) adept at processing and predicting time-series data. Here, LSTM is used to capture long-term dependencies and temporal features in the user touchpoint sequence, assigning a temporal weight to each touchpoint to reflect its importance in the overall decision path. The LSTM model can receive touchpoint feature vectors as input, which may contain information such as touchpoint type, exposure duration, and click location. Through training, the LSTM learns the temporal patterns between touchpoints and outputs a temporal weight corresponding to the time step. As another implementation, besides the basic LSTM, a Bi-LSTM can be used to simultaneously consider forward and backward contextual information, or an LSTM with an attention mechanism can be used to more accurately capture the importance of key touchpoints in the sequence.

[0077] An adaptive adjustment parameter is introduced to calculate the improved touchpoint contribution value. This parameter dynamically adjusts the touchpoint contribution value based on data distribution characteristics to more accurately reflect the actual influence of touchpoints under different levels of data dispersion. The coefficient of variation (the ratio of standard deviation to mean) serves as a quantitative indicator of data centrality, guiding the adaptive adjustment of the parameter. The adaptive adjustment parameter can be designed as a function inversely or directly proportional to the coefficient of variation; for example, when data dispersion is high, the parameter can appropriately amplify the contribution of certain touchpoints, and vice versa. This function can be optimized using rules of thumb or based on historical data. Alternatively, the adaptive adjustment parameter can be determined by combining reinforcement learning or Bayesian optimization methods, iteratively optimizing it to automatically learn and adjust based on real-time data distribution characteristics, maximizing the accuracy of attribution or the effectiveness of subsequent deployment decisions.

[0078] For example, suppose a user experiences a series of touchpoints before purchasing a certain SKU, including "search engine ad exposure," "click on the product details page," "add to cart," and "coupon redemption." These interactions can be included when defining the touchpoint set. The SKU conversion value function can be set as follows: if the user ultimately completes the purchase, the value is 1; otherwise, it is 0. When calculating the traditional Shapley value, the system simulates the impact of these touchpoints appearing in different orders on the final conversion and calculates the average contribution of each touchpoint. For example, "click on the product details page" may significantly increase the conversion probability in various combinations. Subsequently, the user's touchpoint sequence (e.g., exposure -> click -> add to cart -> coupon redemption) is input into a pre-trained LSTM model. The LSTM model analyzes this sequence and outputs the temporal weights of each touchpoint. For example, before the "coupon redemption" touchpoint, "add to cart" may be given a higher temporal weight because it is often a key step before conversion. Finally, the traditional Shapley value is fused with the temporal weights output by the LSTM, and the contribution value is adjusted adaptively based on the historical conversion data distribution of the SKU (e.g., if the conversion data of the SKU fluctuates greatly and has a high coefficient of variation). For example, if "Claim Coupon" is very close to the conversion in time series, and the conversion data concentration of the SKU is low, its contribution value may be further amplified to reflect its key role in a specific time series and data context.

[0079] Through the above technical solutions, this method, when quantifying the contribution value of touchpoints, not only considers the fair marginal contribution of touchpoints to SKU conversion, but also models the temporal sequence of touchpoints using a Long Short-Term Memory (LSTM) network, effectively capturing the temporal dependencies and contextual information between touchpoints in the user's decision-making path. The introduction of adaptive adjustment parameters allows the calculation of touchpoint contribution values ​​to be dynamically optimized based on data distribution characteristics, improving the accuracy and robustness of attribution. This enables the system to more accurately identify key touchpoints and their importance at different stages, thereby providing more refined and intelligent data support for SKU-level delivery decisions, significantly improving the effectiveness and conversion efficiency of delivery strategies.

[0080] In some implementations, the parameters of the Long Short-Term Memory (LSTM) network are set as follows: 128 hidden layer nodes, a learning rate of 0.001, 200 iterations, and a batch size of 32. The Adam optimizer and cross-entropy loss function are used to optimize the model parameters. The 128 hidden layer nodes refer to the dimension of the hidden state vector in the LSTM network, which determines the model's ability to learn and remember sequence information. Theoretically, more nodes allow the model to capture more complex patterns, but this also increases computational cost and the risk of overfitting. Setting the number of hidden layer nodes to 128 aims to balance the model's expressive power and computational efficiency. For example, depending on the complexity and size of the dataset, the number of nodes can be adjusted to 256 or 512 to handle more complex sequence data, or reduced to 64 or 32 to reduce computational overhead. The learning rate of 0.001 is the step size at which the optimizer updates the model parameters in each iteration. A suitable learning rate ensures that the model converges stably to the optimal solution. An excessively high learning rate may cause the model to oscillate or even diverge, while an excessively low learning rate will slow down convergence. 0.001 is a commonly used initial value that performs well in many deep learning tasks. Furthermore, learning rate scheduling strategies, such as exponential decay or cosine annealing, can be used to dynamically adjust the learning rate during training. The number of iterations (200) refers to the number of times the model is fully trained on the entire training dataset. Sufficient iterations allow the model to learn patterns in the data sufficiently, but too many iterations can lead to overfitting. Setting it to 200 iterations aims to ensure the model has enough chance to converge while avoiding overtraining. For example, this can be combined with an early stopping strategy to terminate training early when validation set performance no longer improves. The batch size (32) refers to the number of samples used for each model parameter update. A larger batch size provides more stable gradient estimates but requires more memory; a smaller batch size introduces more noise but helps the model escape local optima. Setting it to 32 is a commonly used value that strikes a balance between computational efficiency and model generalization ability. For example, depending on available computational resources, the batch size can be adjusted to 64 or 128 to accelerate training, or reduced to 16 to handle memory-constrained situations. The Adam optimizer is an adaptive learning rate optimization algorithm that combines the advantages of Adagrad and RMSprop, calculating an independent adaptive learning rate for each parameter. The Adam optimizer excels at handling sparse gradients and non-stationary objectives, and typically converges faster and performs better than traditional optimizers such as SGD (Stochastic Gradient Descent). Besides Adam, other optimizers such as SGD (with momentum), RMSprop, or Adagrad can also be used. The cross-entropy loss function is commonly used in classification tasks to measure the difference between the probability distribution predicted by the model and the true label distribution.For binary or multi-class classification problems such as SKU conversion prediction, the cross-entropy loss function can effectively guide model learning, making its output probability distribution closer to the true situation. For regression tasks, the mean squared error (MSE) loss function can be used.

[0081] In some implementations, optimizations to the distributed time-series database InfluxDB aim to improve its performance, availability, and scalability when handling large-scale time-series data. This typically involves adjusting the database's architecture, storage strategies, indexing mechanisms, and data distribution to accommodate high write throughput and complex queries. For example, InfluxDB's processing power can be enhanced by adjusting storage engine parameters, optimizing data model design, or configuring cluster mode. One approach is to divide data into four shards based on access frequency. This means logically or physically partitioning data into different storage units based on data activity or query frequency. This sharding strategy separates frequently accessed "hot data" from less frequently accessed "cold data," allowing for different storage media or optimization strategies for different shards. For example, hot data can be stored on high-performance solid-state drives (SSDs), while cold data can be stored on lower-cost hard disk drives (HDDs). Each shard corresponds to an independent storage node, meaning each data shard is deployed on a separate physical server or virtual machine. This design effectively distributes I / O load, avoids single points of failure, and allows each shard to scale its resources independently, thereby improving the overall system's concurrency and stability. Consistent hashing is a strategy used in distributed systems for data distribution and load balancing. This algorithm ensures that when the number of nodes in the cluster changes (e.g., adding or removing storage nodes), only a small amount of data needs to be migrated, rather than all data being redistributed. This significantly reduces system maintenance costs and data migration overhead, ensuring stable system operation. Each shard has three replicas, meaning that each data shard's data is stored as three identical copies on different storage nodes. This redundancy mechanism greatly enhances data reliability and system fault tolerance. Even if a storage node fails, the system can still read data from other replicas, ensuring continuous service availability. Asynchronous replication is used for replica synchronization, meaning that after completing a data write operation, the primary replica can return a success response to the client without waiting for all secondary replicas to confirm receipt. Subsequently, the primary replica asynchronously synchronizes the data to the secondary replicas. This mechanism effectively reduces write operation latency and improves system write throughput, making it suitable for scenarios with high write performance requirements and where a small amount of data loss is acceptable in extreme cases.

[0082] In some implementations, Redis caching technology is introduced in step S300, using an LRU replacement strategy, with the cache key being SKU_ID-touchpoint type-timestamp.

[0083] Introducing Redis caching technology refers to using Redis as a high-speed data storage layer to handle high concurrency and low latency data access requirements. Redis is an in-memory database that supports various data structures, and its read and write speeds far exceed those of traditional disk databases. In actual deployments, a Redis server cluster can be configured, and data interaction can be performed through a client connection pool, or a managed Redis service provided by a cloud service provider can be used to simplify operation and maintenance and ensure high availability. The LRU (Least Recently Used) replacement strategy means that when cache space is insufficient, the system will prioritize evicting the least recently used data items. LRU is an efficient cache eviction algorithm whose core idea is that recently accessed data is more likely to be accessed in the future. This strategy can be implemented by maintaining a doubly linked list and a hash table. The linked list records the access order, and the hash table is used for fast lookup, ensuring the accuracy and efficiency of the eviction mechanism. The cache key is SKU_ID-touchpoint type-timestamp, which means combining the unique identifier of the SKU (SKU_ID), the touchpoint type of the user interaction, and the timestamp of the interaction into a unique string as an index for the cached data. This combination method can accurately locate specific touchpoint behavior data for a specific SKU at a specific point in time, thereby improving cache hit rate and data retrieval accuracy. In implementation, the values ​​of these three fields can be combined into a unique cache key using string concatenation or a hash function.

[0084] By introducing Redis caching technology before the distributed time-series database InfluxDB, and combining it with an LRU replacement strategy and a specific cache key design, a high-efficiency data access layer is constructed. When the system needs to trace the decision path of a user at the SKU level, it first generates a unique cache key according to a preset format based on the target SKU's SKU_ID, touchpoint type, and timestamp. Subsequently, the system attempts to retrieve data from the Redis cache using this cache key. Since Redis is an in-memory database, its data read speed is extremely fast. If the cache is hit, the required standardized data and touchpoint contribution value can be directly obtained from Redis, thus avoiding query operations on the backend InfluxDB. This significantly reduces data access latency and alleviates the query pressure on the distributed time-series database. When the cache is missed, the system queries the corresponding data from InfluxDB, constructs a cache key based on the SKU_ID-touchpoint type-timestamp format, and stores it in Redis. To effectively manage the limited cache space, this application adopts an LRU replacement strategy. This means that when the Redis cache space reaches its preset limit, the system automatically identifies and evicts data items that have been accessed least recently, ensuring that the cache always retains the most active and most likely to be accessed again data. This strategy aligns perfectly with the characteristics of SKU-level user decision path tracing, as users often repeatedly access or focus on similar SKUs or touchpoint sequences within a short period. Through this collaborative mechanism, Redis caching technology, the LRU replacement strategy, and precise cache key design work together to form an intelligent and responsive data access system. This effectively solves the efficiency bottleneck that may exist in general caching technologies under massive data and high concurrency scenarios, ensuring real-time and rapid tracing of SKU-level user decision paths.

[0085] In some implementations, in step S300, the query optimization of the distributed time-series database InfluxDB adopts a B+ tree index structure, and a composite index is established for the exposure time and SKU_ID fields.

[0086] Query optimization in the distributed time-series database InfluxDB aims to improve the database's efficiency and response speed when handling data retrieval requests. This can be achieved in various ways, such as creating indexes, optimizing query statements, using materialized views, or adjusting data storage structures. Among these, indexes are a common and effective means of improving query performance, allowing the database system to quickly locate the required data without scanning the entire dataset. The B+ tree index structure is a widely used balanced tree index in database systems. Its characteristic is that all leaf nodes are located at the same level and contain all keys and pointers to the corresponding records. Leaf nodes are connected by pointers, facilitating range queries. B+ tree index structures can be applied to a single field to form a single-column index, or to multiple fields to form a composite index. The exposure time field records the timestamp of user behavior events and is a crucial dimension in time-series data analysis, often used for time-range queries, data aggregation, and trend analysis. The SKU_ID field is a unique identifier for a product's inventory unit, used to accurately identify and distinguish different products. In SKU-level data analysis, it is often used to filter data for specific products or to perform product-level aggregation. A composite index is an index created on multiple columns of a table, sorted according to the column order in the index. When a query condition includes the first column of the composite index or a prefix column, the database can efficiently utilize the index for data retrieval. For example, a composite index can be created with exposure time as the primary sort key and SKU_ID as the secondary sort key, or vice versa.

[0087] In the aforementioned method, to support SKU-level user decision path tracing, standardized data and touchpoint contribution values ​​are stored in the distributed time-series database InfluxDB. To address the query efficiency issue for specific SKUs within a specific time range in the context of massive data, this application further proposes query optimization for the distributed time-series database InfluxDB. Specifically, by adopting a B+ tree index structure and establishing a composite index for the exposure time field and the SKU_ID field, when the system needs to query the behavioral data of a certain SKU within a specific time period, the database query optimizer can quickly locate the relevant data using this composite index. The characteristics of the B+ tree index ensure that both time-range queries and precise queries for specific SKUs can be completed through efficient index traversal, significantly reducing disk I / O and data scanning volume. This optimization enables the data storage module to respond to data retrieval requests at a faster speed, thereby ensuring the real-time performance and accuracy of SKU-level user decision path tracing, effectively supporting the subsequent generation of path heatmaps and data fusion analysis.

[0088] In some implementations, the WebGL 3D visualization technology in step S400 renders path nodes using vertex shaders and fragment shaders, and the vertex coordinate calculation formula is:

[0089] Among them, the touchpoint timing order value is the time sequence number of the touchpoint in the user's decision path, representing the temporal position of the touchpoint; the scaling factor is the coordinate scaling coefficient of the timing order value, with a value range of [0.1, 1.0], used to adjust the display scale of the temporal dimension in three-dimensional space; the touchpoint contribution value is the quantitative contribution value of the touchpoint to SKU conversion; the weight coefficient is the coordinate scaling coefficient of the touchpoint contribution value, used to map the touchpoint contribution value to the height range of three-dimensional space for visual differentiation; the channel identifier code value is the unique code of the channel to which the touchpoint belongs, representing the channel source of the touchpoint.

[0090] WebGL 3D visualization technology is a JavaScript API for rendering interactive 2D and 3D graphics in any compatible web browser without the need for plugins. It achieves high-performance graphics display by leveraging the graphics processing unit (GPU) of the user's device to accelerate the rendering process. In this application, WebGL 3D visualization technology is used to present a stereoscopic view of the SKU conversion path on the browser side, allowing users to observe and analyze path nodes from different angles. Besides WebGL, libraries based on WebGL, such as Thres500 and js, can also be used for 3D visualization. The vertex shader is a programmable stage in the graphics rendering pipeline, responsible for processing data for each vertex, such as calculating the vertex's position, color, and texture coordinates in 3D space. It converts the input 3D coordinates into 2D coordinates on the screen and can perform operations such as transformations and lighting calculations on the vertices. The fragment shader (also called the pixel shader) is another programmable stage in the graphics rendering pipeline, responsible for calculating the final color of each pixel. It receives interpolated data (such as color and texture coordinates) output by the vertex shader and calculates the final pixel color displayed on the screen based on this data and texture information. By working together, these two shaders allow for fine-grained control over the visual attributes of path nodes, such as geometry, color, and transparency, thus clearly showcasing the characteristics of different touchpoints. Path node rendering refers to the process of transforming abstract touchpoint data into visible geometric shapes (such as spheres, cubes, or custom models) in three-dimensional space. This process involves mapping touchpoint attributes (such as timing, contribution value, and channel) to the visual attributes of the geometry (such as position, size, and color). For example, touchpoint timing ranking values ​​can be mapped to the X-axis or Z-axis, touchpoint contribution values ​​can be mapped to the Y-axis (height), and channel identifier codes can be mapped to color or shape. Besides directly rendering as geometry, path nodes can also be represented using particle systems or volume rendering techniques. The touchpoint timing ranking value in the vertex coordinate calculation formula is the temporal sequence number of the touchpoint within the user's decision path; for example, the first interaction between the user and the SKU is 1, the second is 2, and so on. It characterizes the temporal position of the touchpoint throughout the transformation process and is the foundation for constructing the path timeline. The scaling factor is a coefficient used to adjust the display scale of the temporal dimension in 3D space, with a value range of [0.1, 1.0]. By adjusting the scaling factor, the density of path nodes on the temporal axis can be controlled, making the path visually more compact or more expansive to adapt to different display needs and data densities. The touchpoint contribution value is quantified using an improved Shapley multi-touchpoint attribution algorithm, representing the quantified contribution of the touchpoint to SKU conversion. This value directly reflects the importance of the touchpoint and is a core indicator for evaluating touchpoint effectiveness. The weighting coefficient is a coordinate scaling factor used to map the touchpoint contribution value to a 3D spatial height range.By adjusting the weighting coefficients, the contribution values ​​of touchpoints of different magnitudes can be uniformly mapped to an appropriate visual height, thus intuitively distinguishing the importance of touchpoints through height differences. The channel identifier code value is a unique code for the channel to which a touchpoint belongs; for example, search engine advertising is 1, social media is 2, and email is 3. It represents the channel origin of the touchpoint and is key information for distinguishing different marketing channels.

[0091] By employing the methods described above, complex SKU-level user touchpoint data, including timing, contribution value, and channel information, is intuitively mapped into a three-dimensional space using WebGL 3D visualization technology and its specific vertex coordinate calculation formula. This visualization method allows users to clearly identify the temporal position, importance, and channel source of different touchpoints in the conversion path, thus overcoming the limitations of traditional two-dimensional charts in displaying multi-dimensional data. Users can more quickly and accurately understand the structure and key nodes of the SKU conversion path, effectively improving insights into user decision-making behavior patterns and providing strong visual support for optimizing marketing strategies.

[0092] In some embodiments, this application further proposes an interactive function for the path heatmap, which includes: locating nodes using ray detection technology when the mouse hovers over them, and returning the node's touch contribution value, exposure, and conversion rate in real time; and triggering an asynchronous request when a node is clicked to load the user behavior sequence data associated with that node.

[0093] The interactive functionality of the path heatmap refers to allowing users to interact with various nodes on the heatmap via mouse operations on the visual interface to obtain more detailed information or trigger further data queries. This feature aims to transform a static visual presentation into a dynamic data exploration tool, enhancing users' understanding of complex data patterns. Implementation methods include integrating event listeners into the front-end visualization framework to capture user mouse events and execute corresponding callback functions based on the event type. Another approach is to build a separate interaction layer that handles all user input and passes the processing results to the visualization rendering engine and data management module.

[0094] Node location via raycasting on mouse hover refers to the system's ability to accurately identify specific path nodes when the user's mouse cursor moves to a certain area on a path heatmap. Raycasting is a commonly used picking method in 3D space. It involves casting a virtual ray from the camera position to the mouse cursor's 2D coordinates on the screen, and then detecting the intersections of this ray with all 3D objects (path nodes) in the scene to determine the selected object. In 2D or simplified 3D scenes, bounding box or collider-based detection methods can also be used, or objects at pixel level can be identified by rendering a hidden buffer with a unique ID.

[0095] Real-time return of node touchpoint contribution value, exposure, and conversion rate means that once a node is located using raycasting technology, the system immediately retrieves and displays key business metrics related to that node. These metrics are typically pre-calculated and stored in association with the data model of the visualized node for quick access. When a node is located, the system uses the node's unique identifier to query these values ​​from in-memory data structures, front-end cache, or low-latency data services, and immediately presents them to the user in the form of floating tooltips, information panels, or status bars.

[0096] Asynchronous requests triggered by clicking a node mean that when a user clicks on a node in a path heatmap, the system initiates a non-blocking data request. Asynchronous requests allow the client to continue performing other operations while waiting for the server's response, avoiding interface lag. These requests are typically sent to the backend service via HTTP / HTTPS protocols and include a unique identifier for the clicked node or other necessary parameters. Implementation methods can include using the JavaScript `fetch` API or `XMLHttpRequest` object to send AJAX requests, or leveraging technologies like WebSocket to establish persistent connections for data interaction.

[0097] Loading the user behavior sequence data associated with the node refers to the process where, after a successful asynchronous request, the system loads and displays detailed user behavior data related to the clicked node, retrieved from the backend. This data may include a series of user interactions before or after reaching the touchpoint, such as specific ad click times, page views, dwell time, and shopping cart actions. This sequence data is typically stored in the distributed time-series database InfluxDB and can be queried using the node's identifier. Once loaded, this data can be presented in a separate details panel, pop-up window, or new visualization view for users to perform more in-depth analysis.

[0098] In building the path heatmap generation engine, employing WebGL 3D visualization technology and Apache Flink real-time computing technology, and visually outputting the critical path nodes of SKU conversion, uncertainties such as network latency and system failures may occur during the collection and transmission of multi-channel SKU-level user behavior data, leading to out-of-order data arrival at the processing system. If the real-time computing system fails to effectively process this out-of-order data, it may cause deviations in the calculation results, thereby affecting the accuracy of the critical path of SKU conversion. Consequently, the visualized decision path may not accurately reflect the actual user behavior trajectory and conversion contribution, thus impacting the effectiveness of ad placement decisions.

[0099] In response, the Apache Flink real-time computing technology in step S400 employs event-time semantics and uses a watermark mechanism to process out-of-order data.

[0100] Apache Flink is a real-time computing technology that is a stream processing framework whose core capability lies in performing stateful real-time computation on unbounded data streams. It can handle high-throughput, low-latency data streams and provides exact-once state consistency guarantees. This technology can be used for real-time analysis and processing of SKU-level user behavior data, such as calculating touchpoint contribution values ​​and aggregating user behavior sequences, providing real-time data support for the generation of path heatmaps. Implementation can include deploying a Flink cluster, defining data processing logic through the DataStream API or Table API, and submitting it to the cluster for execution. Another implementation is to utilize Flink's SQL interface to directly query and transform the data stream using SQL statements, simplifying the development process. Event-time semantics refers to the data stream processing system processing data based on the timestamp of the event, rather than the time the data arrives at the processing system. In distributed systems, data may arrive out of order due to network latency or system load; event-time semantics ensures that even if the data is out of order, the processing result is still based on the actual order of events. This is crucial for accurately analyzing user behavior paths, as the sequence of user actions directly affects the construction and attribution of their decision paths. Implementing event-time semantics typically involves embedding timestamps in data records and performing logical processing based on these timestamps in stream processing operators. The watermark mechanism is a core mechanism in Apache Flink for handling out-of-order event data. It's a marker indicating that "all data with event times less than or equal to this watermark have arrived so far." As the watermark advances, the system assumes that all events earlier than this watermark have arrived, thus triggering time-based window computation. The watermark mechanism effectively balances the real-time nature of data processing with the accuracy of results, avoiding delays caused by waiting for indefinitely out-of-order data while minimizing computational errors due to late data arrival. Watermarks can be generated based on the data source's timestamps, for example, periodically generated by subtracting a preset delay time from the currently observed maximum event time, or triggered based on specific events in the data stream. Handling out-of-order data refers to the ability of a stream processing system to take measures to ensure the correctness of computation results when the arrival order of data records is inconsistent with their actual occurrence order. In SKU-level campaign decision-making path mapping, the out-of-order arrival of user behavior data (such as impressions, clicks, and conversions) can severely impact the accurate construction of the path and the correct attribution of touchpoint contributions. By using event temporal semantics and a watermark mechanism, the system can identify and buffer out-of-order data, waiting for the watermark to advance to a sufficiently high value. This ensures that as much relevant data as possible has arrived before aggregation or window calculations are performed, resulting in more accurate calculation results.Out-of-order data can also be handled by setting a policy that allows late data, which means that late data is still accepted and processed for a period of time after the watermark, in order to further improve the completeness of the results.

[0101] In some implementations, the data fusion module in step S500 uses ETL tools to clean and transform data, verifies data format consistency through JSONSchema, and standardizes touch point type fields using regular expressions.

[0102] ETL (Extract, Transform, Load) tools are software tools used to extract data from different sources, transform (clean, format, aggregate, etc.) the data, and load the data into target systems (such as data warehouses or databases). Their role is to handle heterogeneous data sources, unify data formats, remove redundant and erroneous data, ensure data quality and consistency, and provide a reliable data foundation for subsequent data analysis and applications. For example, tools such as Apache NiFi or Talend can be used to configure data flows through a graphical interface, enabling automated data extraction, cleaning, and transformation. Another approach is to use Python scripts combined with the Pandas library to write custom data processing logic to perform programmatic cleaning and transformation operations on the raw data.

[0103] JSONSchema is a JSON-based description language used to define the structure, content, and format of JSON data. Validating data format consistency using JSONSchema ensures that input or transmitted JSON data conforms to expected structural specifications, such as field names, data types, required fields, and enumeration value ranges. This helps identify and correct formatting errors before data enters the system, avoiding processing anomalies caused by inconsistent data formats. For example, a JSONSchema file can be defined, specifying that touchpoint type fields must be strings and within a predefined list of enumerated values, and then a library such as `jsonschema` (Python) or `ajv` (JavaScript) can be used to perform real-time validation on the received JSON data. Another approach is to integrate JSONSchema validation logic into the data interface layer to ensure that all incoming data meets predefined format requirements before processing.

[0104] Regular expressions are a powerful tool for describing string patterns. Standardizing touchpoint type fields using regular expressions means unifying touchpoint type fields from different sources or with different expressions by defining specific matching patterns, making them conform to a preset standardized format. This can solve the data analysis difficulties caused by inconsistent touchpoint type expressions (e.g., "ad impressions," "ad displays," and "exposure" all refer to the same concept). For example, a regular expression can be defined to uniformly convert multiple expressions such as "ad impressions," "ad displays," and "exposure" into "exposure." Another approach is to pre-define a set of regular expression rules for different channel sources. When data enters, the corresponding regular expression is applied for matching and replacement based on its source, thereby standardizing touchpoint type fields.

[0105] In some implementations, in step S500, the support threshold for association rules in the Apriori association analysis algorithm is set to 0.05, and the confidence threshold is set to 0.7. Redundant association rules are removed through a pruning strategy to uncover the association between touchpoint combinations and SKU conversion.

[0106] The association rule support threshold is set to 0.05, meaning that a touchpoint combination or association rule must appear at least 5% of the total dataset to be considered "frequent" or "meaningful." This threshold is designed to filter out accidental touchpoint combinations that appear too infrequently in the user decision-making path, ensuring that subsequent analysis focuses on more common and representative behavioral patterns. This threshold can be adjusted based on historical data analysis results, for example, by evaluating the number and quality of rules at different thresholds, or according to business needs, such as reducing the threshold to discover more long-tail, uncommon conversion paths.

[0107] A confidence threshold of 0.7 means that an association rule (e.g., "If a user experiences touchpoints A and B, then they will convert") must have a confidence level of 70% or higher to be adopted. Confidence measures the reliability of a rule, specifically, what percentage of transactions containing the rule's antecedent (touchpoint combination) also contain the rule's consequent (SKU conversion). This threshold is designed to ensure that discovered association rules have high predictive power, avoiding decisions based on weak associations. This threshold can also be optimized based on domain expert experience or through methods such as cross-validation. For example, in scenarios where extremely high conversion prediction accuracy is required, the threshold can be further increased.

[0108] Removing redundant association rules through pruning strategies refers to using specific algorithms or logic to identify and remove rules that have redundant information or can be derived from simpler rules after the Apriori algorithm generates an initial set of association rules. For example, if both the rule "A->C" and "A,B->C" have high support and confidence, but "B" has no significant additional predictive power for "C", then "A,B->C" may be considered a redundant rule and pruned. Common pruning strategies can include lift-based pruning, which removes rules with lift close to 1, as this indicates that the antecedent and consequent are relatively independent; or pruning based on maximally frequent itemsets or closed frequent itemsets, retaining only the rules that best represent the data pattern.

[0109] Discovering the correlation between touchpoint combinations and SKU conversion refers to using the Apriori algorithm, optimized with the aforementioned parameters and strategies, to identify strong correlation patterns between specific touchpoint sequences or combinations and the final SKU conversion during the user's decision-making process from multi-channel SKU-level user behavior data and standardized data. These correlations are presented in the form of "if...then...", such as "If a user clicks on the product details page after being exposed to an advertisement, then there is a high probability that they will purchase that SKU."

[0110] In some implementations, this application further proposes a data quality monitoring mechanism, which includes: periodically collecting data quality indicators, calculating the average value of the indicators using a sliding window algorithm, and triggering an alarm when the indicators are lower than a preset threshold.

[0111] The data quality monitoring mechanism refers to a system or process for assessing, tracking, and reporting data quality status, aiming to ensure the accuracy, integrity, consistency, timeliness, and effectiveness of data throughout its lifecycle. This mechanism can be achieved by defining a series of data quality rules (e.g., field non-nullability, data type matching, numerical range restrictions, etc.) and developing automated scripts or tools to periodically check whether data conforms to these rules; alternatively, it can integrate a third-party data quality management platform that provides data profiling, data cleaning, data validation, and can generate data quality reports. Periodically collecting data quality indicators aims to ensure that data quality status is continuously and regularly tracked, allowing for the timely detection of potential problems. This can be achieved by automatically executing data quality checks and recording results at fixed time intervals (e.g., hourly, daily, or weekly); or by using an event-triggered mechanism based on data inflow, such as immediately initiating data quality indicator collection whenever a new batch of data enters the system. Using a sliding window algorithm to calculate the indicator mean helps smooth short-term fluctuations in data quality indicators, more accurately reflecting long-term trends or stable states of data quality, and avoiding frequent alarms triggered by momentary anomalies. Specifically, a fixed-size time window can be defined (e.g., data from the past 10 minutes or the past hour). Within this window, samples of all data quality indicators are collected, and the average of these samples is calculated. As time progresses, the window slides forward, continuously incorporating the latest data and removing the oldest. Alternatively, algorithms such as Exponentially Weighted Moving Average (EWMA) can be used to assign different weights to the data within the window, making recent data have a greater impact on the mean, thus more sensitively reflecting the latest changes. A preset threshold is a pre-defined numerical limit used to determine whether a data quality indicator is within an acceptable range. This threshold can be manually set as a fixed value based on historical data quality performance, business needs, or industry standards; for example, specifying that the data integrity indicator (non-empty rate) must not be lower than 99%. Alternatively, statistical methods can be used, such as calculating a confidence interval based on historical data distribution, treating indicator values ​​exceeding this interval as anomalies, and dynamically adjusting the threshold. Triggering alarms serves to promptly notify relevant personnel when data quality problems occur, enabling them to quickly intervene and resolve the issue, preventing the data quality problem from escalating further. Alarms can be sent via email, SMS, instant messaging tools (such as WeChat Work and DingTalk), etc. The notification content includes the alarm type, occurrence time, affected data range, and specific indicator values; or, alarm information can be integrated into a unified monitoring platform or operation and maintenance system, displaying alarm status through a visual interface, and supporting advanced functions such as alarm escalation and alarm suppression.

[0112] This data quality monitoring mechanism operates by continuously observing the health of the data. First, it collects data quality indicators for key data objects, such as multi-channel data, touchpoint contribution values, and integrated data, at predetermined time intervals or event triggers. These indicators may include multiple dimensions such as data completeness, accuracy, consistency, and timeliness. To avoid misjudgments caused by instantaneous data fluctuations, the system does not directly judge based on single-collection indicator values. Instead, it uses a sliding window algorithm to smooth these periodically collected indicators, calculating the indicator average over a certain time range. This average more stably reflects the true trend of data quality. Subsequently, the system compares the calculated indicator average with pre-set thresholds. Once the average of an indicator is found to be below the threshold, indicating a potential data quality problem, the system immediately triggers an alarm. This mechanism ensures that the quality of the data is continuously monitored and guaranteed throughout the entire SKU-level delivery decision-making path mapping process, from raw behavioral data to the final integrated data. In this way, data quality issues can be identified and resolved in a timely manner, thereby ensuring that the data relied upon for subsequent steps such as Shapley value attribution, LSTM network modeling, path heatmap generation, and Apriori association analysis are reliable, thus improving the accuracy and effectiveness of the entire deployment decision.

[0113] For example, a data quality monitoring mechanism can be implemented as follows: First, define a series of data quality indicators, such as the completeness of "multi-channel SKU-level user behavior data" (e.g., the non-empty rate of exposure time and touchpoint type fields), the reasonableness of "touchpoint contribution values" (e.g., whether the value range is between 0 and 1), and the consistency of "integrated data after fusion" (e.g., whether the attributes of the same SKU match in different channels). The system can be configured to collect these indicators periodically every 5 minutes. For each indicator, a sliding window of size 1 hour is used to calculate the average value of the indicator within the window. For example, if the preset threshold for the "non-empty rate of the exposure time field" is 98%, when the average value calculated by the sliding window is lower than 98%, the system will immediately send an alarm notification to the data operation and maintenance team via email and WeChat Work. The notification includes the specific indicator name, the current average value, the threshold, and the time of occurrence, so that operation and maintenance personnel can quickly locate and handle problems in the data source or data processing process.

[0114] like Figure 2 As shown in the embodiments of this application, a SKU-level delivery decision path mapping system based on Shapley value and deep learning is also provided, including:

[0115] The data preprocessing module 100 is used to collect SKU-level user behavior data from multiple channels, remove outliers using the random forest algorithm, and output standardized data.

[0116] The attribution algorithm module 200 includes an LSTM modeling unit and an adaptive adjustment unit, which are used to execute an improved Shapley value multi-touchpoint attribution algorithm to quantify the touchpoint contribution value of touchpoints to SKU conversion.

[0117] The data storage module 300 includes a distributed time-series database optimization unit and a Redis cache unit, which are used to store standardized data and contact point contribution values, and support millisecond-level traceability.

[0118] The visualization module 400 is used to build a path heatmap generation engine, which outputs the critical path for interactive SKU conversion through WebGL 3D rendering and Flink real-time calculation.

[0119] The data fusion module 500 includes a correlation analysis unit and a quality monitoring unit, which are used to integrate data from multiple channels and ensure data reliability.

[0120] In some implementations, the LSTM modeling unit of the attribution algorithm module 200 adopts a bidirectional LSTM structure, and the outputs of the forward and backward hidden layers are fused through a splicing operation.

[0121] Bidirectional Long Short-Term Memory (Bi-LSTM) is a neural network architecture capable of processing sequential data. Its core lies in utilizing two independent LSTM layers simultaneously: one processes the forward flow of the input sequence, and the other processes the backward flow. Through this design, Bi-LSTM can capture the forward and backward context information at each time step of the sequence, thus achieving a more comprehensive understanding of the elements within the sequence. For example, two independent LSTM networks can be used, one processing from the beginning to the end of the sequence, and the other from the end to the beginning; alternatively, a single LSTM unit can be designed with its internal control mechanism extended to handle both forward and backward inputs simultaneously. In the Bi-LSTM structure, the output of the forward hidden layer is the hidden state vector generated by the forward LSTM at each time step after processing the current input and all previous inputs, encoding information from the beginning of the sequence to the current time step. Similarly, the output of the backward hidden layer is the hidden state vector generated by the backward LSTM at each time step after processing the current input and all subsequent inputs, encoding information from the end of the sequence to the current time step. The outputs of these hidden layers are condensed representations of the sequence information in their respective directions. The concatenation operation is a mathematical operation that joins two or more vectors or matrices along a specific dimension. In bidirectional LSTM, concatenating the outputs of the forward and backward hidden layers means linking these two vectors end-to-end to form a longer new vector containing bidirectional contextual information. This fusion method preserves all information in both directions, providing richer and more comprehensive feature representations for subsequent processing. Besides concatenation, element-wise summation or element-wise multiplication can also be used to combine the outputs of hidden layers in two directions.

[0122] In some implementations, the distributed time-series database optimization unit of the data storage module 300 further includes an automated backup tool that uses a combination of incremental backup and full backup.

[0123] Automated backup tools refer to software or system components that can automatically perform data backup operations according to preset strategies without manual intervention. Their main function is to ensure data security and prevent data loss due to hardware failures, software errors, human error, or disasters. These tools can be integrated into database management systems, for example, through the database's built-in backup commands and scheduling functions; or they can run as independent third-party applications, interacting with the database interface to execute backup tasks. For example, operating system scheduled tasks can be used to trigger backup scripts, or professional backup software can be used for centralized management and scheduling. A combination of incremental and full backups typically involves performing full backups periodically and incremental backups daily between full backups. A full backup involves completely copying all selected data, creating a complete copy containing all data. Its advantages are simple and fast data recovery, but its disadvantages include long backup times and large storage space requirements. Incremental backups only back up data that has changed since the last backup of any type. Its advantages are fast backup speed and small storage space requirements, but its disadvantage is that data recovery requires sequentially restoring both the full backup and all subsequent incremental backups, making the process relatively complex and time-consuming. This combined approach aims to balance backup efficiency, storage costs, and ease of data recovery. In addition to incremental backups, differential backups can also be used, which back up all data that has changed since the last full backup. The recovery process is somewhere between a full backup and an incremental backup, requiring only a full backup and the most recent differential backup.

[0124] It should be noted that in some embodiments, the functions or modules of the system provided in this application embodiment can be used to execute the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0125] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of this application. The scope of this application is determined by the scope of the claims.

Claims

1. A SKU-level delivery decision path mapping method based on Shapley value and deep learning, characterized in that, It should include at least the following steps: S100. Collect behavioral data of SKU-level users from multiple channels, and preprocess the behavioral data using the random forest algorithm to remove a preset proportion of noise data and outliers, thereby obtaining standardized data including exposure time, touchpoint type, conversion result and basic SKU attributes. S200, based on the improved Shapley value multi-touch point attribution algorithm, combines the Long Short-Term Memory (LSTM) network to model the touch point sequence and quantify the touch point contribution value of each touch point to the conversion of a single SKU; S300: Store the standardized data and the touchpoint contribution value into the optimized distributed time-series database InfluxDB, and support the decision path tracing of SKU-level users through data sharding, replication mechanism and caching technology; S400 constructs a path heatmap generation engine, using WebGL 3D visualization technology and Apache Flink real-time computing technology to visualize and output key path nodes of SKU transformation. S500: Integrate multi-channel data through the data fusion module, use the Apriori association analysis algorithm to mine data association relationships, and establish a real-time data quality monitoring mechanism. The multi-channel data includes behavioral data and standardized data of multi-channel SKU-level users. The data used to mine association rules includes touchpoint sequence data, conversion result data, channel association data, and SKU attribute association data. The objects of data quality monitoring include multi-channel data, touchpoint contribution values, and integrated data after fusion.

2. The SKU-level delivery decision path mapping method based on Shapley value and deep learning according to claim 1, characterized in that, The random forest algorithm in step S100 contains multiple decision trees, uses the Gini coefficient as the feature selection criterion, and determines outliers based on the three-standard-deviation rule.

3. The SKU-level delivery decision path mapping method based on Shapley value and deep learning according to claim 1, characterized in that, In step S200, the Long Short-Term Memory (LSTM) network allows for custom settings of the number of hidden layer nodes, learning rate, number of iterations, and batch size. The Adam optimizer and cross-entropy loss function are used to optimize the model parameters.

4. The SKU-level delivery decision path mapping method based on Shapley value and deep learning according to claim 1, characterized in that, The optimization of the distributed time-series database InfluxDB in step S300 includes: dividing the data into 4 shards according to the data access frequency, with each shard corresponding to an independent storage node, and using a consistent hashing algorithm to allocate data; setting 3 replicas for each shard, and using an asynchronous replication mechanism for replica synchronization.

5. The SKU-level delivery decision path mapping method based on Shapley value and deep learning according to claim 1, characterized in that, In step S300, Redis caching technology is introduced, and an LRU replacement strategy is adopted. The cache key is SKU_ID-touchpoint type-timestamp. The query optimization of the distributed time series database InfluxDB adopts a B+ tree index structure, and a composite index is established for the exposure time and SKU_ID fields.

6. The SKU-level delivery decision path mapping method based on Shapley value and deep learning according to claim 1, characterized in that, The WebGL 3D visualization technology in step S400 renders path nodes using vertex shaders and fragment shaders. The vertex coordinate calculation formula is as follows: Among them, the touchpoint timing order value is the time sequence number of the touchpoint in the user's decision path, representing the temporal position of the touchpoint; the scaling factor is the coordinate scaling coefficient of the timing order value, with a value range of [0.1, 1.0], used to adjust the display scale of the temporal dimension in three-dimensional space; the touchpoint contribution value is the quantitative contribution value of the touchpoint to SKU conversion; the weight coefficient is the coordinate scaling coefficient of the touchpoint contribution value, used to map the touchpoint contribution value to the height range of three-dimensional space for visual differentiation; the channel identifier code value is the unique code of the channel to which the touchpoint belongs, representing the channel source of the touchpoint; The interactive features of the path heatmap include: locating nodes using ray detection technology when the mouse hovers over them, and returning the node's touch contribution value, exposure, and conversion rate in real time; and triggering an asynchronous request when a node is clicked to load the user behavior sequence data associated with that node. Apache Flink real-time computing technology employs event-time semantics and uses a watermark mechanism to process out-of-order data.

7. The SKU-level delivery decision path mapping method based on Shapley value and deep learning according to claim 1, characterized in that, The data fusion module in step S500 uses ETL tools for data cleaning and transformation, verifies data format consistency through JSONSchema, and standardizes touch point type fields using regular expressions; The Apriori association analysis algorithm sets the support threshold for association rules to 0.05 and the confidence threshold to 0.

7. Redundant association rules are removed through pruning strategies to uncover the association between touchpoint combinations and SKU conversion. The data quality monitoring mechanism includes: periodically collecting data quality indicators, using a sliding window algorithm to calculate the average value of the indicators, and triggering an alarm when the indicators are lower than a preset threshold.

8. A SKU-level delivery decision path mapping system based on Shapley value and deep learning, used to execute the SKU-level delivery decision path mapping method based on Shapley value and deep learning as described in any one of claims 1-7, characterized in that, include: The data preprocessing module is used to collect SKU-level user behavior data from multiple channels, remove outliers using the random forest algorithm, and output standardized data. The attribution algorithm module, which includes an LSTM modeling unit and an adaptive adjustment unit, is used to execute an improved Shapley value multi-touchpoint attribution algorithm to quantify the touchpoint contribution value of touchpoints to SKU conversion. The data storage module includes a distributed time-series database optimization unit and a Redis caching unit, which are used to store standardized data and contact point contribution values, and support millisecond-level traceability; The visualization module is used to build a path heatmap generation engine, which outputs the critical path for interactive SKU conversion through WebGL 3D rendering and Flink real-time calculation. The data fusion module, which includes a correlation analysis unit and a quality monitoring unit, is used to integrate data from multiple channels and ensure data reliability.

9. The SKU-level delivery decision path mapping system based on Shapley value and deep learning according to claim 8, characterized in that, The LSTM modeling unit of the attribution algorithm module adopts a bidirectional LSTM structure, and the outputs of the forward and backward hidden layers are fused through a splicing operation.

10. The SKU-level delivery decision path mapping system based on Shapley value and deep learning according to claim 8, characterized in that, The distributed time-series database optimization unit of the data storage module also includes an automated backup tool, which uses a combination of incremental backup and full backup.

Images