An advertisement effect attribution analysis method and system based on multi-source data fusion

Abstract

The application discloses an advertisement effect attribution analysis method and system based on multi-source data fusion, relates to the technical field of data acquisition and processing, and comprises the following steps: collecting multi-source advertisement data of an advertisement activity to be evaluated and preprocessing, to generate a standardized attribution data set; extracting an effect evaluation index and locating an index correlation factor; identifying statistical characteristics of the index correlation factor based on historical interaction data and the standardized attribution data set, to construct a factor level of the index correlation factor with respect to the effect evaluation index; performing screening, filtering and dimension reduction processing on the index correlation factor according to the factor level, to obtain a dimension-reduced correlation factor; calculating factor state data of the dimension-reduced correlation factor corresponding to the historical interaction data, and determining an attribution analysis model of the advertisement activity to be evaluated according to the factor state data; and determining a real-time attribution result of the advertisement activity to be evaluated by using the attribution analysis model according to real-time conversion data. The application meets the actual needs of fine and intelligent advertisement operation.

Description

Technical Field

[0001] This application relates to the field of data acquisition and processing technology, and in particular to a method and system for attribution analysis of advertising effectiveness based on multi-source data fusion. Background Technology

[0002] In the field of digital marketing, ad performance attribution analysis aims to evaluate the contribution of each ad touchpoint in the user conversion path to the final conversion behavior, serving as a crucial basis for optimizing ad placement strategies. Currently, the industry widely adopts rule-based attribution models, such as first-click attribution, last-click attribution, or linear attribution. These models typically rely on limited interaction log data, such as impression records, click events, or transaction results, and quantify the value of touchpoints through pre-defined weighting logic.

[0003] However, the above methods face significant limitations in practical applications. On the one hand, their input data dimensions are relatively singular, mainly focusing on explicit behavioral indicators from the advertising side (such as click count and display frequency), rarely incorporating the user's full-link behavior in the conversion path (such as cross-channel browsing and multiple return visits), and failing to systematically integrate potential influencing factors such as channel attributes. This limitation in the data perspective makes it difficult for the attribution results to fully reflect the true driving factors in the complex user decision-making process.

[0004] On the other hand, existing attribution models generally employ static weight allocation strategies, with fixed attribution logic that cannot be dynamically adjusted based on contextual information such as the advertiser's industry characteristics (e.g., high-priced services versus fast-moving consumer goods) and marketing objectives (e.g., brand awareness enhancement and immediate sales conversion). Therefore, when faced with diverse advertising scenarios, the model's adaptability is limited, and the business guidance value of its attribution conclusions is consequently reduced.

[0005] In summary, existing advertising attribution technologies still have significant shortcomings in terms of data breadth, model flexibility, and adaptability, making it difficult to meet the actual needs of refined and intelligent advertising operations. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies and meet the actual needs of refined and intelligent advertising operations, this application provides a method and system for advertising effectiveness attribution analysis based on multi-source data fusion.

[0007] Firstly, the objective of this invention is achieved through the following technical solution: An advertising effectiveness attribution analysis method based on multi-source data fusion includes: Collect and preprocess multi-source advertising data from advertising campaigns to be evaluated, and build a deep correlation graph based on the relationship between touchpoints, elements, and conversions to generate a standardized attribution dataset; Extract the performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-constructed marketing performance evaluation system, and locate the correlation factors of the indicators; Obtain historical interaction data of the advertising campaign to be evaluated; based on the historical interaction data and the standardized attribution dataset, identify the statistical characteristics of the index-related factors; and construct the factor levels of the index-related factors with respect to the performance evaluation index based on the statistical characteristics. Based on the factor level, the index correlation factors are screened to obtain initial correlation factors. After filtering the initial correlation factors, dimensionality reduction processing is performed to obtain dimensionality-reduced correlation factors. Calculate the factor state data corresponding to the dimensionality-reduced correlation factor of the historical interaction data, and determine the attribution analysis model of the advertising campaign to be evaluated based on the factor state data; Obtain real-time conversion data of the advertising campaign to be evaluated, and determine the real-time attribution result of the advertising campaign to be evaluated based on the real-time conversion data and the attribution analysis model.

[0008] By adopting the above technical solutions, multi-source advertising data includes placement data, user behavior data, channel data, creative data, and conversion data. The standardized attribution dataset is generated by constructing a deep correlation graph based on the touchpoint-element-conversion relationship, effectively solving the problem of missing attribution information caused by the single data dimension and lack of integration of full-link behavior and channel / creative attributes in existing technologies. This is beneficial to improving the breadth and reliability of attribution analysis data coverage. By introducing an adaptive attribution model and combining it with dimensionality-reduced correlation factors obtained through factor level screening, filtering, and dimensionality reduction, the model construction overcomes the shortcomings of traditional rule-based models (such as last-click attribution) in terms of static weight allocation and inability to quantify the collaborative contribution of multiple touchpoints and the differentiated value of placement elements. The adaptive attribution model can reflect the dynamics of the path and analyze the influence of elements, thereby achieving a refined and quantitative evaluation of the contribution of each advertising touchpoint and placement element.

[0009] Then, by linking the performance evaluation indicators and related factors in the marketing performance evaluation system, and constructing factor levels based on historical interaction data, the optimization process of the attribution analysis model is driven. This allows the attribution logic to implicitly adapt to core evaluation objectives (such as brand exposure or sales conversion) in different business scenarios. This makes attribution analysis no longer dependent on fixed rules, but rather dynamically focuses on key influencing factors based on actual marketing goals, enhancing the model's generalization ability and adaptability across diverse industries and advertising objectives. This application achieves breakthroughs in three dimensions—data foundation, model architecture, and goal alignment—significantly improving the comprehensiveness, accuracy, and scenario adaptability of advertising performance attribution, meeting the actual needs of refined and intelligent advertising operations.

[0010] In a preferred embodiment of this application, the extraction of performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-constructed marketing performance evaluation system specifically includes: Identify the marketing objective category to which the advertising campaign to be evaluated belongs, including brand awareness, lead generation, or transaction conversion; Based on the marketing objective category, retrieve a set of candidate indicators associated with the category from a pre-built marketing performance evaluation system. The set of candidate indicators includes at least two of the following: exposure completion rate, click conversion rate, form submission rate, or transaction ROI. Obtain the key performance indicator (KPI) achievement data of the advertising campaign to be evaluated within the historical campaign period; Calculate the Pearson correlation coefficient between each candidate indicator in the candidate indicator set and the KPI achievement data, and use it as the correlation score of each candidate indicator; The relevance score is compared with a preset relevance threshold, and candidate indicators with relevance scores higher than the preset relevance threshold are selected as the performance evaluation indicators for the advertising campaign to be evaluated.

[0011] By employing the aforementioned technical solution, advertising campaigns are automatically categorized into three major marketing objectives: brand awareness, lead generation, and transaction conversion. A set of matching candidate metrics is dynamically retrieved, avoiding the subjectivity and lag inherent in traditional attribution analysis that relies on manual experience to set fixed metrics. The Pearson correlation coefficient is introduced as a quantitative tool, objectively assessing the linear correlation strength between each candidate metric and core business results based on real KPI achievement data from historical campaign periods. This effectively identifies key metrics that truly have explanatory and predictive power within the current campaign context, eliminating interference from noise or weakly correlated metrics.

[0012] In a preferred embodiment, this application further includes: extracting the performance evaluation indicators corresponding to the advertising campaign to be evaluated within a pre-constructed marketing performance evaluation system, and locating the correlation factors of the indicators; Based on the standardized attribution dataset, a user-level conversion path sequence is constructed, which records multiple ad touchpoints that the user encounters before conversion in chronological order. For each ad touchpoint, the corresponding delivery element feature vector is extracted. The delivery element feature vector includes channel identifier, material type code, delivery time segment tag and bid tier. Using the observed values ​​of the aforementioned performance evaluation indicators as the target variable, and the concatenated result of the feature vectors of all deployed elements as the input feature matrix, a gradient boosting tree model is trained. The average information gain value of each dimension feature in predicting the performance evaluation index is calculated using the built-in feature importance evaluation rule of the gradient boosting tree model. The advertising touchpoints or delivery elements corresponding to feature dimensions whose average information gain value is greater than a preset importance threshold are identified as correlation factors related to the performance evaluation index.

[0013] By adopting the above technical solution, the problem of "how to identify truly relevant factors from a massive amount of campaign elements" is solved. Structured campaign element features (such as channels, creatives, time periods, and bids) are extracted from the user-level conversion path sequence, refining the attribution object from "touchpoints" to "element dimensions." The built-in feature importance assessment mechanism (such as information gain) of the gradient boosting tree model is used to objectively quantify the predictive contribution of each element to the performance indicators. Key factors are automatically screened by setting importance thresholds, avoiding human experience bias while retaining the ability to capture nonlinear and interaction effects. This application overcomes the limitation of traditional attribution methods that only focus on touchpoint order and ignore element differences.

[0014] In a preferred embodiment of this application, the filtering of the initial correlation factors includes: Obtain a historical advertising campaign sample set corresponding to the initial association factor from a preset advertising database. The historical advertising campaign sample set includes the path coverage and average interaction depth of each initial association factor in the historical conversion path. A target sample is randomly selected from the historical advertising campaign sample set, and based on the path coverage and average interaction depth, samples of the same class with a similarity higher than a preset high threshold and samples of different classes with a similarity lower than a preset low threshold are identified. Configure the initial factor weights of the initial correlation factors; Extract a predetermined number of nearest neighbor samples of the same type and a predetermined number of nearest neighbor samples of different types from the samples of the same type and the samples of different types, respectively. The association factor weights of the initial association factor are calculated based on the distance relationships in the feature space between the initial factor weights, the target sample, the similar nearest neighbor samples, and the different near neighbor samples; When the weight of the correlation factor is not less than a preset weight threshold, the corresponding initial correlation factor is determined as the filtered correlation factor.

[0015] By adopting the above technical solutions, the problem of "initial association factors may contain low-discrimination or noisy factors" is solved. By introducing a historical advertising campaign sample set as a contextual reference, factor filtering is transformed into a discrimination task based on sample similarity. Sample similarity is defined by path coverage and average interaction depth, so that the division of samples of the same / different classes has a clear basis for advertising behavior. The weight of association factors is calculated by using the distance relationship between neighboring samples in the feature space, giving higher confidence to high-frequency and high-discrimination factors. Automatic filtering is achieved through weight thresholds, effectively eliminating factors that are unstable or have no discriminatory power in historical scenarios, thereby improving the quality of association factors input into the attribution analysis model.

[0016] In a preferred embodiment of this application, the step of calculating the association factor weight of the initial association factor based on the distance relationships in the feature space of the initial factor weight, the target sample, the similar nearest neighbor samples, and the different types of nearest neighbor samples includes: Calculate the first distance difference between the target sample and each of its nearest neighbors of the same class in the feature space, and calculate the second distance difference between the target sample and each of its nearest neighbors of different classes in the feature space; Obtain the distribution ratio of samples of the same type and samples of different types in the historical advertising campaign sample set, as the category prior probability; Based on the first distance difference and the prior probability of the same category, a positive update term is obtained; based on the second distance difference and the prior probability of different categories, a negative update term is obtained. The difference between the positive update term and the negative update term is calculated to obtain the single-round weight update amount; The single-round weight update amounts obtained from multiple rounds of random sampling are accumulated, and the correlation factor weight of the initial correlation factor is calculated based on the accumulation result and the initial factor weight.

[0017] By adopting the above technical solution, the distance difference between neighbors of the same class and neighbors of different classes is weighted separately (combined with the prior probability of the class) to construct an update mechanism that combines positive incentives and negative inhibitions; through multiple rounds of random sampling and weight accumulation, the randomness of a single sampling is smoothed and the stability of weight estimation is improved; the final output of the association factor weight comprehensively reflects the comprehensive performance of the factor in intra-class clustering and inter-class separation.

[0018] In a preferred embodiment of this application, the step of filtering and dimensionality reduction of the initial correlation factor to obtain a dimensionality-reduced correlation factor further includes: Calculate the variance inflation factor (VIF) for the feature columns corresponding to the initial correlation factors, and remove the top K feature columns with larger VIF values ​​to obtain the set of decollinearity correlation factors, where K is an integer; The path coverage and average interaction depth of the decollinearized factors in historical conversion paths are statistically analyzed. The path coverage is the proportion of the number of conversion paths containing the decollinearized factors to the total number of conversion paths, and the average interaction depth is the average position index of the factor in the conversion path. Based on the path coverage and average interaction depth, the path significance score S of each decollinearity correlation factor is calculated using a weighted scoring formula, which is: S=α×Coverage+β×(1 / Depth), Where Coverage is the path coverage rate, Depth is the average interaction depth, and α and β are preset weight coefficients with α+β=1; The path significance scores are sorted in descending order, and the top M decollinearity correlation factors are selected as the dimension reduction correlation factors, where M is a positive integer.

[0019] By adopting the above technical solutions, this application addresses the problem that "factors may still exhibit collinearity or low path representativeness after filtering." It eliminates highly collinear features through the VIF test, avoiding parameter instability caused by multicollinearity. It introduces path coverage (breadth) and average interaction depth (depth) to construct path saliency scores, prioritizing key factors that frequently appear and are positioned high in the user journey. A weighted scoring formula (S=α×Coverage+β×(1 / Depth)) flexibly balances coverage and depth to adapt to different marketing scenario preferences. Then, Top-M screening achieves controllable dimensionality compression, reducing model complexity while retaining information, significantly improving the interpretability and generalization ability of the attribution model.

[0020] In a preferred embodiment of this application, determining the attribution analysis model for the advertising campaign to be evaluated based on the factor state data includes: Based on the factor state data, a state transition graph of the user conversion path is constructed. The nodes of the state transition graph are advertising touchpoint types, and the edge weights are the transition frequencies of adjacent touchpoints in the historical path. Using the state transition diagram and the preset transition absorption state, the steady-state probability distribution of the Markov chain is solved to obtain the long-term access probability of each advertising touchpoint. Based on the deviation between the long-term access probability and the actual conversion rate, counterfactual inference is used to verify: estimate the expected change in the overall conversion rate after removing any ad touchpoint, and determine the expected change as the path-level contribution benchmark value of the ad touchpoint; Invoke a pre-built general XGBoost attribution model, which takes the characteristics of the delivery elements as input and the attribution contribution value as output; Using the path-level contribution benchmark value as a supervision signal, the prediction output of the general XGBoost attribution model in the dimensionality-reduced correlation factor feature space of the current advertising campaign is biased to obtain the calibrated real-time attribution prediction function. Determining the real-time attribution results of the advertising campaign to be evaluated includes: When the real-time conversion data of the advertising campaign to be evaluated is received, the real-time feature vector of the dimension-reduced correlation factor in the corresponding user conversion path is extracted and input into the calibrated real-time attribution prediction function. The system receives the predicted contribution scores of each dimension-reduced correlation factor output by the calibrated real-time attribution prediction function, and aggregates them according to the dimensions of advertising touchpoints and delivery elements to generate real-time attribution results that include the contribution ratio of channel, creative, and time period dimensions.

[0021] By adopting the above technical solution, Markov chains are used to model user path transition patterns, and path-level contribution benchmark values ​​are generated through counterfactual inference. These benchmark values ​​reflect the true marginal value of touchpoints in the sequence, overcoming the biases of rule-based models such as last clicks. A pre-built general XGBoost model is used as an efficient predictor, avoiding repeated training for each activity and improving system scalability. The path contribution benchmark value is used as a supervisory signal to calibrate the XGBoost output, achieving lightweight intelligence with "general model and scenario adaptation". In the real-time stage, the contribution scores of each element are quickly output based on the calibrated prediction function and aggregated by channel, creative, and time period to generate real-time attribution results that can directly guide the optimization of ad placement.

[0022] Secondly, the objective of this invention is achieved through the following technical solution: An advertising effectiveness attribution analysis system based on multi-source data fusion, the system comprising: The data acquisition and graph construction module is used to collect multi-source advertising data of the advertising campaign to be evaluated and preprocess it, and build a deep correlation graph based on the relationship between touchpoints, elements and conversions to generate a standardized attribution dataset. The feature recognition module is used to extract the performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-built marketing performance evaluation system, and to locate the factors related to the indicators. The filtering and dimensionality reduction module is used to acquire historical interaction data of the advertising campaign to be evaluated, identify the statistical characteristics of the indicator correlation factors based on the historical interaction data and the standardized attribution dataset, construct the factor levels of the indicator correlation factors with respect to the effect evaluation indicator according to the statistical characteristics, filter the indicator correlation factors according to the factor levels to obtain initial correlation factors, and perform dimensionality reduction processing on the initial correlation factors to obtain dimensionality-reduced correlation factors. The model optimization module is used to calculate the factor state data of the dimensionality-reduced correlation factors corresponding to the historical interaction data, and determine the attribution analysis model of the advertising campaign to be evaluated based on the factor state data. The attribution result output module is used to obtain the real-time conversion data of the advertising campaign to be evaluated, and to determine the real-time attribution result of the advertising campaign to be evaluated based on the real-time conversion data and the attribution analysis model.

[0023] Thirdly, the objective of this invention is achieved through the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described advertising effectiveness attribution analysis method based on multi-source data fusion.

[0024] Fourthly, the objective of this invention is achieved through the following technical solution: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described advertising effectiveness attribution analysis method based on multi-source data fusion.

[0025] In summary, this application includes at least one of the following beneficial technical effects: 1. This application achieves standardized fusion and full-link tracking of cross-channel data by collecting multi-source heterogeneous advertising data and constructing a deep correlation map of "touchpoint-element-conversion," effectively solving the problems of data silos and broken attribution paths in traditional methods. Based on this, it innovatively adopts a fusion architecture of Markov chains and XGBoost to construct an adaptive attribution model: using Markov chains to accurately characterize the state transition probabilities of user touchpoint sequences and scientifically quantify the contribution of each touchpoint to the conversion path; simultaneously, combining XGBoost's powerful nonlinear fitting and feature interaction capabilities, it dynamically corrects the attribution weights in complex scenarios. 2. It not only significantly improves the attribution accuracy for long-term, multi-touchpoint marketing campaigns, but also enhances the model's ability to adapt to real-time traffic fluctuations, ensuring the accuracy and timeliness of attribution results in dynamic market environments. Attached Figure Description

[0026] Figure 1 This is a flowchart of an advertising effectiveness attribution analysis method based on multi-source data fusion in one embodiment of this application; Figure 2 This is a flowchart of step S4 in an advertising effectiveness attribution analysis method based on multi-source data fusion in one embodiment of this application; Figure 3This is a schematic diagram of a device according to one embodiment of this application. Detailed Implementation

[0027] The present application will be further described in detail below with reference to the accompanying drawings.

[0028] In one embodiment, such as Figure 1 As shown, this application discloses an advertising effectiveness attribution analysis method based on multi-source data fusion, which specifically includes the following steps: S1: Collect multi-source advertising data of the advertising campaign to be evaluated and preprocess it, and build a deep correlation graph based on the relationship between touchpoints, elements and conversions to generate a standardized attribution dataset.

[0029] In this embodiment, multi-source advertising data consists of five types of core data that are structurally heterogeneous but semantically related and collected from the entire advertising delivery chain, including: (1) Delivery data: log data generated by the advertising delivery platform, recording the execution parameters of each ad exposure or click, with fields including exposure ID, click ID, advertising plan ID, bid amount, budget consumption, delivery start / end time, targeted audience tags, and bidding type; (2) User behavior data: user interaction events collected by the front-end tracking system on the ad landing page and subsequent paths, including page browsing, button clicks, form filling, product addition to cart, page dwell time, and bounce behavior, with each event carrying user device ID, session ID, timestamp, and page URL; (3) Channel data: metadata describing the quality and attributes of traffic sources, provided by channel partners or generated by internal evaluation systems, including channel name, channel level (first-level / second-level agent), new customer acquisition cost, historical conversion rate benchmark, device type distribution, and geographic coverage; (4) Creative data: structured description of advertising creative content, stored in the Creative Management library. Platform), including material ID, material type (short video, text and images, live slice, carousel), video duration, completion rate (25% / 50% / 75%), main visual color, copy keywords, A / B test group; (5) conversion data: final performance data confirmed by the business transaction system, including conversion event type (order placed, payment successful, registration completed), conversion time, order amount, user unique identifier (such as mobile phone number hash value), and affiliated advertising touchpoint ID.

[0030] A deep association graph is a directed attribute graph. Node types include users, touchpoints, features, and conversion events, while edge types include user-touchpoint (representing contact), touchpoint-feature (representing configuration), and touchpoint-conversion (representing facilitation). The graph is aligned across sources using unique IDs (such as user ID and exposure ID) and stored in the Neo4j graph database.

[0031] Specifically, the five types of data streams mentioned above are accessed in real time via a Kafka message queue, and the raw logs are preprocessed using standardization: For campaign data, the time format is standardized to ISO 8601, the bid amount is converted to RMB cents, and missing targeting tags are filled with "unknown"; for user behavior data, device ID and IP address are used for cross-session concatenation to construct a complete behavior sequence for a single user within 7 days, and invalid exposures with a dwell time of less than 1 second are removed; cross-session concatenation refers to using device ID as the primary concatenation rule and IP address as the secondary concatenation rule, and the same device ID is determined to be the same user regardless of whether the IP changes; when there is no device ID, the same IP + the same browser fingerprint + 7 days of sessions are determined to be the same user. For channel data, a mapping table is established based on the channel name, and internal codes such as "Douyin_Feed_V2" are unified as "Platform A Information Flow"; for material data, the version with the highest video completion rate is extracted as the representative feature of the material ID; for conversion data, a 7-day click attribution window rule is used to trace the conversion event back to the most recent valid click and verify the consistency of the user ID.

[0032] Subsequently, the system constructs a deep relational graph based on the three-element relationship of "touchpoint-element-conversion": with the advertising touchpoint (uniquely identified by exposure ID or click ID) as the core node; the bidding, time period, and targeting tags in the campaign data, the channel name and CAC in the channel data, and the material type and completion rate in the material data are used as the element attributes of the touchpoint and connected by the "touchpoint → element" edge; key events in the user behavior data (such as adding to cart and form submission) are used as intermediate state nodes and connected by the "touchpoint → behavior → conversion" chain structure; finally, all structured entities and relationships are imported into the graph database.

[0033] S2: Extract the performance evaluation indicators corresponding to the advertising campaigns to be evaluated within the pre-built marketing performance evaluation system, and locate the related factors of the indicators.

[0034] In this embodiment, a pre-built marketing effectiveness evaluation system is configured in a structured knowledge base within the configuration center of the attribution analysis system. Its core is a multi-dimensional mapping table that defines the static relationships between different marketing objective categories and candidate indicator sets, and provides standardized calculation formulas and business explanations for each indicator category, stored in the configuration center. The marketing effectiveness evaluation system categorizes marketing objectives into three types: brand awareness, lead acquisition, and transaction conversion. Each objective category is associated with a set of candidate indicators and calculation logic. Indicator association factors refer to the dimensions of the delivery elements that have significant predictive power for the selected effectiveness evaluation indicators, such as channel type and material format.

[0035] Brand awareness campaigns are suitable for activities aimed at expanding brand awareness, such as new product launches and holiday marketing. They focus on users' initial contact and interaction with the brand, with typical scenarios including splash screen ads and news feed exposure. Lead generation campaigns are suitable for activities requiring the collection of user intent, such as B2B services and high-priced items. They focus on potential customers' lead generation behavior, with typical scenarios including form submissions and phone appointments. Transaction conversion campaigns are suitable for activities aimed at direct sales, such as e-commerce promotions and FMCG promotions. They focus on the final transaction and return on investment efficiency, with typical scenarios including product purchase and payment completion.

[0036] Each candidate metric associated with a target category has undergone historical A / B testing to ensure a strong correlation with business results. For example, the candidate metric set associated with transaction conversion includes: Click-through rate (CVR = number of conversions / number of clicks), ROI (total GMV / ad spend), average order value (total GMV / number of orders), add-to-cart rate (number of users adding to cart / number of users clicking), etc. The calculation logic, time window (e.g., 7-day attribution), and filtering rules (e.g., excluding returned orders) for all metrics are clearly defined in the system.

[0037] Specifically, step S2 includes: S21: Identify the marketing objective category to which the advertising campaign to be evaluated belongs. Marketing objective categories include brand awareness, lead generation, or transaction conversion.

[0038] For example, after receiving an analysis request for an advertising campaign titled "Smart Home Appliance Special Promotion," the system first calls the advertising management API to obtain the campaign's creation configuration information. The "campaign_objective" field is valued as "SALES," which the system's built-in mapping rules interpret as "transaction conversion." If this field is missing (e.g., in historical campaigns), an alternative strategy is employed: analyzing the campaign's landing page URL path—if it contains the keywords " / order" or " / pay," it's determined to be a transaction conversion campaign; if it contains " / form" or " / contact," it's determined to be a lead generation campaign; and if it's a brand website homepage or video playback page, it's determined to be a brand awareness campaign.

[0039] S22: Based on the marketing objective category, retrieve a set of candidate indicators associated with the category from the pre-built marketing performance evaluation system. The set of candidate indicators includes at least two of the following: exposure completion rate, click-through rate, form submission rate, or transaction ROI.

[0040] In this embodiment, each candidate metric includes a calculation formula, attribution window, data source table, and minimum effective sample size requirement. Based on the "transaction conversion category" determined in S21, the corresponding key "transaction_conversion" is retrieved from the knowledge base to obtain a set of candidate metrics: ["click_conversion_rate", "roi"] (i.e., "click conversion rate" and "transaction ROI"). The system simultaneously loads the metadata for these two metrics: Click conversion rate = number of successfully paid orders / number of ad clicks, with an attribution window of 7 days; Transaction ROI = total GMV / ad spending, only counting non-returned orders.

[0041] S23: Obtain key performance indicator (KPI) achievement data for the advertising campaign to be evaluated over the historical campaign period.

[0042] In this embodiment, Key Performance Indicator (KPI) achievement data refers to core business outcome data used to measure the overall effectiveness of the advertising campaign. The selection of core business outcomes must align with the marketing objectives. For transaction conversion campaigns, KPIs are typically Total Merchandise Volume (GMV) or the number of valid orders, as these directly reflect commercial value.

[0043] S24: Calculate the Pearson correlation coefficient between each candidate indicator in the candidate indicator set and the KPI achievement data, and use it as the correlation score of each candidate indicator.

[0044] In this embodiment, the Pearson correlation coefficient is used to quantify the linear correlation strength between the daily series of candidate indicators and the daily series of KPIs. Its value ranges from -1 to 1, with a value closer to 1 indicating a stronger correlation. The calculation formula is as follows:

[0045] in Let i be the candidate index value for day i. Let n be the KPI value (GMV) for day i, where n=30.

[0046] For example, the system first aggregates candidate metrics by day: Click-through rate (CVR) daily series: [2.1%, 2.3%, ..., 2.4%]; Transaction ROI daily series: [5.8, 6.1, ..., 6.3]; GMV daily series: [6.2M, 7.1M, ..., 8.5M].

[0047] Using the pearsonr function from the SciPy library, we calculated the correlation coefficients between CVR and GMV: r1 = 0.872; and r2 = 0.935.

[0048] The system records r1 and r2 as relevance scores for "click conversion rate" and "transaction ROI" respectively, which will be used for the next step of screening.

[0049] S25: Compare the relevance score with the preset relevance threshold, and select candidate indicators with a relevance score higher than the preset relevance threshold as the performance evaluation indicators for the advertising campaign to be evaluated.

[0050] In this embodiment, the preset correlation threshold is an empirical threshold, set to 0.80. When |r|≥0.8, the indicator change can explain more than 64% of the KPI change (because...). =0.64), which has strong predictive power; below this value, it is easily affected by noise and is not suitable as an attribution target.

[0051] Specifically, the relevance score obtained in S24 is compared with the preset relevance threshold of 0.80: the click conversion rate score of 0.872 > 0.80, so it is retained; the transaction ROI score of 0.935 > 0.80, so it is retained.

[0052] Ultimately, the system outputs the following performance metrics for this advertising campaign: {Click-through rate, ROI}. This result is written into the task context and used as the target variable for factor ranking in S3. If all candidate metric scores are below a preset relevance threshold (e.g., r_max = 0.75 for a certain brand campaign), the system triggers an alert and recommends manual review.

[0053] Furthermore, in step S2, extracting the performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-constructed marketing performance evaluation system, and locating the correlation factors of the indicators, also includes: S201: Based on the standardized attribution dataset, construct a user-level conversion path sequence. The user-level conversion path sequence records multiple ad touchpoints that the user encounters before conversion in chronological order.

[0054] In this embodiment, the user-level conversion path sequence refers to an ordered list formed by arranging all advertising touchpoints encountered by a single user in ascending order of timestamps before a conversion event occurs (e.g., setting the attribution window to 7 days). Each touchpoint is identified by a unique exposure ID or click ID and carries multiple metadata such as time, channel, and creative.

[0055] For example, all records with `conversion_flag=1` are filtered from the standardized attribution dataset generated in Example 1, totaling 421,867 conversion samples. For each user ID (e.g., user_8a3f9c), its most recent conversion time (e.g., 2025-06-15 14:23:18) is extracted, and all ad touchpoints within the 7 days prior to this time are traced back. For example, the path sequence for user_8a3f9c is: [(At 09:12 on June 10, 2025, exposure ID = e101, channel = Platform A), (At 20:05 on June 12, 2025, click ID = c205, channel = Platform 2), (At 18:33 on June 14, 2025, click ID = c312, channel = Platform 3)]. The system stores each path as a JSON object.

[0056] S202: For each advertising touchpoint, extract the corresponding delivery element feature vector. The delivery element feature vector includes channel identifier, material type code, delivery time segment label, and bid level.

[0057] In this embodiment, the delivery element feature vector is a numerical or categorical representation after structuring and encoding the multi-dimensional attributes of the advertising touchpoint, used for input into the machine learning model. One-Hot encoding is adopted. For example, Platform A is encoded as [1, 0, 0], Platform B as [0, 1, 0], and Platform C as [0, 0, 1]; the material type code adopts a predefined mapping table method. For example, short video is mapped to 1, graphic text to 2, and live broadcast slice to 3; the delivery time segment label means dividing 24 hours into multiple segments. In this embodiment, it is divided into 6 segments: 00–04, 04–08, 08–12, 12–16, 16–20, 20–24. Take the index of the segment when the touchpoint occurs. For example, 20:05 is mapped to 5; the bid level is divided into three levels according to the daily industry average: "low" (<P30), "medium" (P30–P70), "high" (>P70), and are encoded as 1, 2, 3 respectively.

[0058] Exemplarily, traverse each touchpoint in each user-level conversion path sequence in S201, and call the feature extraction service to generate the feature vector. Take the touchpoint c312 (Baidu SEM, material = short video, time = 18:33, bid = 2.8 yuan) as an example: Channel identifier: One-Hot [0, 0, 1] of Platform C; Material type: short video, encoded 1; Time segment label: 18:33 belongs to the 16–20 segment, label 4; Bid level: The P70 of Platform C on that day = 2.5 yuan, 2.8 > 2.5, level "high", encoded 3.

[0059] Finally, the feature vector of this touchpoint is [0, 0, 1, 1, 4, 3] (a total of 6 dimensions). For the entire user path, the system splices the feature vectors of all touchpoints in chronological order and fills them to a fixed length (such as at most 10 touchpoints, padded with 0 if insufficient) to form a single-sample input vector (dimension = 60).

[0060] S203: Use the observed values ​​of the effect evaluation indicators as the target variable and the concatenation result of the feature vectors of all deployed elements as the input feature matrix to train the gradient boosting tree model.

[0061] In this embodiment, the observed values ​​of the performance evaluation indicators refer to the actual values ​​of the indicators selected in the above steps. In this example, these are the transaction ROI (continuous value) and click-through rate (binary classification: whether or not a conversion occurs). Since ROI is more business-sensitive, this embodiment uses ROI as the primary target variable.

[0062] Specifically, the system constructs a training dataset: the input is a 60-dimensional user path feature vector generated by S202, and the output is the actual ROI value for each user (e.g., ROI for user_8a3f9c = GMV 1200 yuan / advertising expenditure 200 yuan = 6.0). XGBoostRegressor is used for training. The XGBoost model training dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Mean squared error (MSE) and mean absolute error (MAE) are used as evaluation metrics for the regression model. Key hyperparameter settings include: max_depth=8 (controlling tree depth to prevent overfitting); learning_rate=0.1; n_estimators=500; subsample=0.8; objective='reg:squarederror'. Training is performed on a Spark on YARN cluster. Missing value handling methods include: filling numerical missing values ​​with the median and filling categorical missing values ​​with the mode; dividing frequency features into 5 intervals (0 times, 1-2 times, 3-5 times, 6-10 times, >10 times) using equal frequency binning; and representing the first / last occurrence position using relative position.

[0063] S204: Calculate the average information gain value of each dimension feature in the prediction performance evaluation index by using the built-in feature importance evaluation rules of the gradient boosting tree model.

[0064] In this embodiment, the built-in feature importance evaluation rule of the gradient boosting tree model refers to the quantification mechanism within the XGBoost model after training to assess the role of each input feature in predicting the target variable. This mechanism is based on the statistical summary of the reduction in the objective function (i.e., the loss function) resulting from selecting features for node splitting when constructing each decision tree. The average information gain value refers to the weighted average information gain (gain) brought by a certain feature when used for node splitting across all decision trees in the gradient boosting tree (XGBoost) model, reflecting the contribution of that feature to the prediction of the target variable. XGBoost provides the get_score(importance_type='gain') interface to directly output the average information gain value, expressed in terms of the reduction in information entropy.

[0065] For example, after training, the system calls the model API to obtain importance scores for 60 input dimensions. Because features are repeated based on touchpoint location (e.g., channels for touchpoint 1, channels for touchpoint 2, etc.), the system aggregates them by feature type: The average gain across all "channel" dimensions (10 in total, corresponding to 10 touchpoint locations) is 0.32. The average gain for all "material types" dimensions is 0.21; The average gain for all "time period labels" dimensions is 0.09; The average gain across all “bid tiers” dimensions is 0.18.

[0066] Furthermore, the system identified the first-touch channel (i.e., the channel of touchpoint 1) as having a gain of 0.41, which is significantly higher than other locations, indicating that the first-touch channel has the greatest impact on ROI.

[0067] S205: The advertising touchpoints or delivery elements corresponding to feature dimensions whose average information gain value is greater than the preset importance threshold are identified as correlation factors related to the performance evaluation indicators.

[0068] In this embodiment, the preset importance threshold is set to 0.15. The preset importance value is determined based on historical model analysis: when the feature gain ≥ 0.15, its average absolute contribution to the prediction result in SHAP value analysis exceeds 5%.

[0069] For example, iterate through the average gain of all feature dimensions in S204: Channel (overall): 0.32 > 0.15, retain; Material type: 0.21 > 0.15, keep; Bid tier: 0.18 > 0.15, hold; Time period label: 0.09 < 0.15, needs to be removed.

[0070] Furthermore, the system refines the values ​​to specific values: in the "Channel" category, the average SHAP value for "Platform A" is +1.2 (positive), for "Platform C" it is +0.8, and for "Platform B" it is -0.3 (negative). Therefore, the final determined correlation factors are: {Platform A, Platform C, Short Video Material, High Bid Tier}. These factors will serve as inputs to the "Initial Correlation Factors" in S4.

[0071] S3: Obtain historical interaction data of the advertising campaign to be evaluated. Based on the historical interaction data and standardized attribution dataset, identify the statistical characteristics of the factors associated with the indicators. Based on the statistical characteristics, construct the factor levels of the factors associated with the indicators with respect to the performance evaluation indicators.

[0072] In this embodiment, statistical features include the mean, variance, quantiles, and conditional expectation of each factor dimension relative to the target indicator. Factor levels are discrete levels, such as L1–L5, assigned to each factor dimension based on the intensity of its influence on the performance indicator.

[0073] The specific quantitative standards for the classification are as follows: The criteria for a significantly positive L1 level are: "The factor accounts for >60% of the high ROI group and the average CVR is >150% of the industry average." The criteria for L2 (moderately positive) are: "The factor accounts for 40%-60% of the high ROI group and the average CVR / ROI is >120% of the industry average." The criteria for L3 (neutral) are: "The difference in the proportion of the factor between the high and low ROI groups is <10% and the average CVR / ROI is close to the industry average." The criteria for L4 (negative) are: "The factor accounts for >40% of the low ROI group and the average CVR / ROI is <80% of the industry average." The criteria for L5 (significantly negative) are: "The factor accounts for >60% of the low ROI group and the average CVR / ROI is <50% of the industry average." Specific quantitative standards can be customized as needed.

[0074] For example, the system extracts all "transaction conversion" samples (i.e., converted user paths) from the standardized attribution dataset, totaling 420,000. For each campaign element dimension (e.g., "channel"), it calculates the distribution difference between high ROI (ROI>3.0) and low ROI (ROI<1.5) groups. For instance, "Platform A" accounts for 68% in the high ROI group, but only 22% in the low ROI group; "Platform D" is the opposite. Furthermore, it calculates the average CVR for each factor value: Platform A's short video content has a CVR of 4.2%, while text / image content has a CVR of only 1.8%. Based on this, it assigns levels to the "channel" dimension: Platform A=L1, Platform D=L3, others=L4; and to "content type": short video=L1, live stream clips=L2, text / image=L4. The level classification rules are: L1 (significantly positive), L2 (moderately positive), L3 (neutral), L4 (negative), L5 (significantly negative).

[0075] S4: Based on the factor level, screen the index-related factors to obtain the initial related factors. After filtering the initial related factors, perform dimensionality reduction processing to obtain the dimensionality-reduced related factors.

[0076] In this embodiment, the initial association factor refers to the element dimension with a factor level of L1 or L2. Filtering refers to weight evaluation based on the similarity of historical samples, and dimensionality reduction refers to removing collinearity and retaining path-significant factors.

[0077] Specifically, such as Figure 2 As shown, in step S4, the initial correlation factors are filtered, including: S401: Obtain the historical advertising campaign sample set corresponding to the initial association factor from the preset advertising database. The historical advertising campaign sample set includes the path coverage and average interaction depth of each initial association factor in the historical conversion path.

[0078] In this embodiment, the preset advertising database refers to an internally built historical warehouse of advertising performance (such as an attribution analysis data lake based on HBase or ClickHouse), where each record represents a statistical summary of a historical advertising campaign. The historical advertising campaign sample set is a collection of similar campaigns selected based on the initial correlation factors involved in the current campaign to be evaluated. Path coverage is defined as the proportion of conversion paths containing initial correlation factors to the total number of conversion paths, reflecting the reach of the initial correlation factors; average interaction depth is defined as the average position number of the initial correlation factor in the user conversion path (first touchpoint is 1, second touchpoint is 2, and so on), reflecting the timing of its intervention in the user journey, with smaller values ​​indicating closer proximity to the conversion starting point.

[0079] For example, starting from the initial association factor list {Platform A, Platform C, Short Video Material, High Bid Tier} output in Example 3, the system retrieves all e-commerce advertising campaigns from the past 12 months that simultaneously contain at least two of the aforementioned initial association factors from the advertising database, resulting in 87 matched historical campaigns. For each campaign, the system calculates the path coverage and average interaction depth of the four initial association factors. For example, in the campaign "A_202503_smartphone": Platform A had a path coverage rate of 72.3% (meaning 72.3% of the converted user paths included touchpoints on Platform A), with an average interaction depth of 1.8. The path coverage rate of "short video materials" was 68.5%, and the average interaction depth was 2.1. The "higher bid tier" had a path coverage rate of 54.2% and an average interaction depth of 3.4.

[0080] These values ​​constitute the feature vector of each historical activity [coverage_douyin, depth_douyin, coverage_video, depth_video, ...], for a total of 87 samples.

[0081] S402: Randomly select a target sample from the historical advertising campaign sample set, and based on path coverage and average interaction depth, identify samples of the same class with a similarity higher than a preset high threshold and samples of different classes with a similarity lower than a preset low threshold.

[0082] In this embodiment, the target sample refers to a reference activity randomly selected from the historical sample set, used to construct a local discriminative context. Similarity is calculated using cosine similarity, based on a feature vector composed of path coverage and average interaction depth. A preset high threshold is set to 0.85, and a preset low threshold is set to 0.30, both determined through historical A / B testing: when similarity ≥ 0.85, the standard deviation of KPI fluctuation between activities is < 10%; when similarity ≤ 0.30, there is no significant overlap in KPI distributions.

[0083] For example, the system randomly selects the activity "A_202505_home_appliance" from 87 samples as the target sample, with a feature vector of [0.76, 1.7, 0.71, 2.0, 0.58, 3.2, 0.49, 4.1] (representing the coverage and depth of platform A / short video / platform C / high bid, respectively). Then, the system calculates the cosine similarity between this sample and the remaining 86 samples. For instance, the activity "B_202504_kitchen" has a similarity of 0.89, classifying it as a similar sample; the activity "C_202502_auto" has a similarity of 0.28, classifying it as a different sample. Ultimately, the system identifies 12 similar samples (sim ≥ 0.85) and 23 different samples (sim ≤ 0.30).

[0084] S403: Configure the initial factor weights of the initial correlation factors.

[0085] In this embodiment, the initial factor weights are the prior confidence levels assigned to each initial correlation factor, reflecting its importance in a general scenario. The initial factor weights can be derived from historical global feature importance (such as the XGBoost gain value in Embodiment 3), expert experience, or uniform initialization. This embodiment uses the latter to avoid prior bias, setting all initial factor weights to 1.0.

[0086] S404: Extract a preset number of nearest neighbor samples of the same class and the nearest neighbor samples of different classes from samples of the same class and samples of different classes, respectively.

[0087] In this embodiment, the nearest neighbor samples of the same class and the nearest neighbor samples of different classes are representative subsets further filtered from the set of samples of the same class / different classes identified in S402. The preset number is set to k=5, that is, 5 nearest neighbor samples of each class are selected, and the selection is based on Euclidean distance in the feature space.

[0088] For example, the system sorts 12 samples of the same class in ascending order of their Euclidean distance from the target sample and selects the top 5 as the nearest neighbors of the same class; similarly, it selects the 5 samples with the smallest distance from 23 samples of different classes as the nearest neighbors of different classes. For example, nearest neighbors of the same class include "B_202504_kitchen" (dist=0.12), "D_202505_beauty" (dist=0.15), etc.; nearest neighbors of different classes include "C_202502_auto" (dist=0.31), "E_202412_travel" (dist=0.33), etc. Each nearest neighbor sample carries its complete factor coverage and depth vector.

[0089] S405: Calculate the association factor weights of the initial association factor based on the distance relationships in the feature space between the initial factor weights, the target sample, the nearest neighbor samples of the same class, and the nearest neighbor samples of different classes.

[0090] In this embodiment, the distance relationship in the feature space refers to the difference between the target sample and its nearest neighbor samples across various factor dimensions. All feature dimensions in the feature space are subjected to Min-Max normalization, uniformly mapping path coverage (0-1) and average interaction depth (1-N) to the 0-1 interval. The calculation logic is based on the discrimination principle of "same category should be close, different category should be far".

[0091] S406: When the weight of the correlation factor is not less than the preset weight threshold, the corresponding initial correlation factor is determined as the filtered correlation factor.

[0092] In this embodiment, the preset weight threshold is set to 1.5. This value is determined through cross-validation: when the weight is ≥1.5, the SHAP value of the association factor in the subsequent attribution model is significantly non-zero (p<0.01).

[0093] For example, the system compares the correlation factor weights of each factor: Platform A: 2.68 ≥ 1.5, retained; Short video footage: 2.31 ≥ 1.5, retain; Platform C: 1.71 > 1.5, retain; High bid: 1.05 < 1.5, needs to be eliminated.

[0094] Ultimately, the filtered association factors are {Platform A, short video material, Platform C}.

[0095] In another embodiment, specifically, in step S4, the association factor weights of the initial association factor are calculated based on the distance relationships in the feature space between the initial factor weights, the target sample, similar neighbor samples, and different class neighbor samples, including: S4001: Calculate the first distance difference between the target sample and each of its nearest neighbors of the same class in the feature space, and calculate the second distance difference between the target sample and each of its nearest neighbors of different classes in the feature space.

[0096] In this embodiment, the feature space refers to a multi-dimensional vector space composed of the path coverage and average interaction depth of each initial correlation factor. For a single factor dimension (such as "Platform A"), its feature space is one-dimensional (only coverage or depth), which facilitates local discriminant analysis. The first distance difference reflects the consistency of the target sample's behavior with efficient activities (of the same type) on this factor, and the second distance difference reflects its difference from inefficient activities (of different types). Together, they constitute the basis of the discriminant signal.

[0097] Specifically, let's take the path coverage dimension of the correlation factor "Platform A" as an example. Let the coverage of the target samples be... =0.76, and the coverage rates of the 5 nearest neighbor samples of the same type are [0.72, 0.78, 0.75, 0.79, 0.74], then the first distance difference is calculated using the mean absolute error (MAE):

[0098] Similarly, the coverage rates of the 5 different class nearest neighbor samples are [0.45, 0.38, 0.51, 0.42, 0.47], and the second distance difference is: The system performs this calculation for each initial correlation factor (such as short video, Baidu SEM, etc.) to obtain its respective first distance difference and second distance difference.

[0099] S4002: Obtain the distribution ratio of samples of the same class and samples of different classes in the historical advertising campaign sample set, as the prior probability of the class.

[0100] In this embodiment, the category prior probability is used to weight the distance difference, reflecting the natural distribution ratio of efficient and inefficient activities in historical data, thus avoiding bias caused by sample imbalance. The category prior probability is calculated based on the total number of all similar / different class samples identified in S402.

[0101] For example, in S402, the system identified 12 samples of the same class and 23 samples of different classes, for a total of 35 valid samples. Therefore, the prior probability of being of the same class is:

[0102] The prior probabilities of different classes are: .

[0103] S4003: Based on the first distance difference and the prior probability of the same category, a positive update term is obtained; based on the second distance difference and the prior probability of different categories, a negative update term is obtained.

[0104] In this embodiment, the stronger the consistency of the positive update term in similar samples, the better. The smaller the value, the greater the positive incentive should be; the more significant the difference in the factor among different classes of samples, the greater the positive incentive should be. The larger the value, the stronger the negative suppression that should be obtained. The two are weighted by prior probabilities to form a biased estimate.

[0105] For example, for the correlation factor "Platform A", the positive update term is calculated as follows:

[0106] The negative update term is calculated as follows:

[0107] Note: Although this is called "positive update item", since it is proportional to the distance, the smaller the actual value, the better.

[0108] S4004: Perform a difference operation on the positive and negative update terms to obtain the weight update amount for a single round.

[0109] In this embodiment, the single-round weight update amount is defined as the negative update item minus the positive update item.

[0110] S4005: Accumulate the single-round weight update amounts obtained from multiple rounds of random sampling, and calculate the correlation factor weights of the initial correlation factors based on the accumulation results and the initial factor weights.

[0111] In this embodiment, multi-round random sampling refers to repeatedly executing S402–S4054 for a total of T=10 times, randomly selecting different target samples each time to cover more historical scenarios and improve the robustness of weight estimation. The initial factor weight is set to 1.0.

[0112] For example, the system performs 10 rounds of sampling to obtain the single-round update sequence of factor "Platform A":

[0113] Total number of updates accumulated:

[0114] The final correlation factor weights are calculated as follows: The system rounds the value to two decimal places, recording it as 2.94.

[0115] S5: Calculate the factor state data of the dimensionality-reduced correlation factors corresponding to the historical interaction data, and determine the attribution analysis model of the advertising campaign to be evaluated based on the factor state data.

[0116] In this embodiment, the attribution analysis model is an adaptive attribution model built based on the fusion architecture of Markov chain and gradient boosting tree XGBoost; the factor state data refers to the value sequence of each user path on the dimension-reduced correlation factor, which is used to construct the Markov chain.

[0117] For example, the system encodes each user path as a sequence of factor states, such as: [Platform A, Short Video, 12:00-14:00, High] encoded as [A, B, C, D]. A state transition graph is constructed based on 12.4 million paths, and the steady-state probability is P(Platform) = 0.62. Combined with the actual conversion rate of 5.1%, counterfactual inference suggests that if the "Platform" touchpoint is removed, the expected conversion rate drops to 3.4%, thus its path contribution baseline is +1.7%. The pre-trained XGBoost model (already trained on the entire platform's advertising data) is called, and the factor features of the current activity are input; the original predicted contribution is +1.2%. After bias correction (+0.5%), the calibrated attribution prediction function is obtained.

[0118] S6: Obtain real-time conversion data of the advertising campaign to be evaluated, and determine the real-time attribution results of the advertising campaign to be evaluated based on the real-time conversion data and attribution analysis model.

[0119] Specifically, step S6 includes: S61: Construct a state transition graph of the user conversion path based on factor state data. The nodes of the state transition graph are advertising touchpoint types, and the edge weights are the transition frequencies of adjacent touchpoints in the historical path.

[0120] In this embodiment, factor state data refers to the user path sequence encoded by dimensionality-reduced correlation factors, with each path represented as a time series of touchpoint types. The state transition graph is a directed weighted graph used to model user switching behavior between different advertising touchpoints, where nodes represent touchpoint types and edge weights represent the historical frequency of transitions from one touchpoint to another, reflecting path dependency patterns.

[0121] Specifically, all user conversion paths are extracted from the standardized attribution dataset, and touchpoints in each path are mapped to touchpoint types. The frequency of occurrence of all adjacent touchpoint pairs is then counted. Simultaneously, a virtual conversion absorbing state, denoted as "C," is introduced, with all final touchpoints pointing to "C." Unconverted paths (those not converted within the attribution window) are truncated, do not point to the absorbing state, and are not included in the steady-state probability calculation of the Markov chain. The system constructs an adjacency matrix from the above statistical results. Where n is the number of contact types (Platform A, Platform B, Platform C), for example, if the value is 3, the 4th row / column is "C". The adjacency matrix T is the basis for the transition probabilities of the Markov chain.

[0122] S62: Using the state transition diagram and the preset transformation absorption state, solve the steady-state probability distribution of the Markov chain to obtain the long-term access probability of each advertising touchpoint.

[0123] In this embodiment, "steady-state probability distribution" refers to the probability vector of the user's state at each touch point after an infinite number of transitions. ,satisfy and Since "C" is in an absorption state (once it enters, it does not leave), in steady state... However, the access frequency of each non-absorption state can be determined by solving the fundamental matrix. We obtain, where Q is the non-absorbing state submatrix and I is the identity matrix.

[0124] Specifically, the system extracts the non-absorbing state transition submatrix Q (3×3):

[0125] (After normalization, for example, there is a 65% probability of switching from platform A to WeChat, a 25% probability to platform B, and a 10% probability of direct conversion.) Calculate the fundamental matrix The expected number of visits vector is t = N⋅1: Platform A: Platform B: Platform C: Normalize it to the probability of long-term access: .

[0126] Similarly, we get platform B = 0.325 and platform C = 0.291.

[0127] S63: Based on the deviation between long-term visit probability and actual conversion rate, use counterfactual inference to verify: estimate the expected change in overall conversion rate after removing any ad touchpoint, and determine the expected change as the path-level contribution benchmark value of the ad touchpoint.

[0128] In this embodiment, counterfactual inference refers to simulating a virtual scenario of "if a certain touchpoint had never been deployed" and estimating the change in conversion rate by modifying the transition matrix.

[0129] Specifically, the current overall conversion rate is Taking platform A as an example, construct the counterfactual transfer matrix T′: set all incoming edges pointing to "platform A" to 0, and proportionally redistribute the traffic originally from other touchpoints to platform A to the remaining touchpoints (keeping the row sum to 1). Recalculate the basic matrix N′ and the new conversion rate. Therefore, the contribution benchmark value of platform A is: Similarly, the path-level contribution baseline value for platform B is 0.32%, and the path-level contribution baseline value for platform C is 0.26%.

[0130] S64: Call the pre-built general XGBoost attribution model, which takes the features of the delivery elements as input and the attribution contribution value as output.

[0131] In this embodiment, the general XGBoost attribution model refers to a gradient boosting tree regression model specifically designed for advertising performance attribution tasks. Its core function is to predict the marginal contribution of each element to the conversion target based on the combination of delivery elements in the user path.

[0132] Specifically, the construction and training process of the general XGBoost attribution model is as follows: We extracted 480 million user conversion path samples from 120,000 advertising campaigns over the past 24 months from the enterprise data lake. Each sample includes the user path, the delivery elements of each touchpoint (channel, creative type, time period, bid tier), and whether a conversion occurred. For each path, we extracted the binary existence (e.g., "whether platform A is included"), frequency, first occurrence position, and last occurrence position of each dimensionality-reduced correlation factor, and performed One-Hot encoding or binning to form a 50-dimensional input feature vector. The target variable uses attribution labels based on Shapley values ​​(generated through historical Markov counterfactual simulation) to ensure path fairness. We can use Monte Carlo sampling to approximate the Shapley value, randomly sampling 200 subsets of elements for each user path and calculating the mean marginal contribution of each element as the attribution label.

[0133] The model structure uses XGBoost Regressor, with key hyperparameters being max_depth=7 (to control the complexity of a single tree and avoid overfitting sparse element combinations), learning_rate=0.05, n_estimators=800, subsample=0.8, and colsample_bytree=0.7. Training was performed in a distributed manner on Spark MLlib using an early stopping strategy (patience=50 rounds), with a final validation set MAE of 0.082%. The model was exported in ONNX format and loaded into the TensorRT inference engine, with a single prediction latency of <5ms.

[0134] S65: Using the path-level contribution benchmark as a supervision signal, the prediction output of the general XGBoost attribution model in the reduced-dimensional correlation factor feature space of the current advertising campaign is biased to obtain the calibrated real-time attribution prediction function.

[0135] In this embodiment, the deviation correction adopts the linear offset method: the deviation between the predicted mean of the general model in the current activity and the Markov baseline value is calculated, and subsequent predictions are uniformly corrected.

[0136] For example, the system runs general XGBoost on 1,000 validation paths of the current activity and obtains the average prediction contribution. The Markov benchmark value ,deviation .

[0137] The calibrated prediction function is defined as:

[0138] Therefore, for any new input, the output automatically increases by 0.17%.

[0139] S66: When real-time conversion data of the advertising campaign to be evaluated is received, extract the real-time feature vector of the dimension-reduced correlation factor in the corresponding user conversion path and input it into the calibrated real-time attribution prediction function.

[0140] In this embodiment, real-time conversion data refers to payment success events captured through a message queue (such as Kafka), carrying the user ID and timestamp. The real-time feature vector is dynamically generated by tracing back the user's touchpoint logs over the past 7 days.

[0141] For example, when the system receives a payment event from user U987654 (2025-06-18 15:22:10), it immediately queries the user's behavior log and reconstructs the path: [Platform A (2025-06-16 20:05), Baidu SEM (2025-06-17 14:30)]. It then extracts whether the dimensionality reduction correlation factor "Platform A" appears → feature vector x = [1] (if only Platform A is the dimensionality reduction factor). Finally, it inputs x into the calibrated function. The predicted contribution score is 0.52% + 0.17% = 0.69%.

[0142] S67: Receive the predicted contribution scores of each dimension-reduced correlation factor output by the calibrated real-time attribution prediction function, and aggregate them according to the dimensions of advertising touchpoints and delivery elements to generate real-time attribution results that include the contribution ratio of channel, creative, and time period dimensions.

[0143] Specifically, aggregation refers to summing up the contributions of a single user by dimension and normalizing them into a percentage form.

[0144] In one embodiment, after filtering and dimensionality reduction of the initial correlation factors to obtain dimensionality-reduced correlation factors, the method further includes: S41: Calculate the variance inflation factor (VIF) for the feature columns corresponding to the initial correlation factors, and remove the top K feature columns with larger VIF values ​​to obtain the set of decollinearized correlation factors, where K is an integer.

[0145] In this embodiment, K is set to 10, and the dimensionality reduction process targets the feature dimensions rather than individual values. The Variance Inflation Factor (VIF) is a statistical indicator that measures multicollinearity and is used to detect whether there is a high linear correlation between feature columns. VIF is defined as:

[0146] in VIF is the coefficient of determination obtained by performing linear regression with the j-th feature as the dependent variable and the remaining features as independent variables. VIF > 10 indicates that the feature has severe collinearity with other features, which may lead to unstable model parameters or distorted attribution results.

[0147] Specifically, for ease of explanation, assume that the filtered set of associated factors contains 5 associated factors, corresponding to 5 feature columns (such as coverage or depth values). The system uses the Python statsmodels library to calculate the VIF for each factor sequentially: Platform A: VIF=8.2; "Short video material": VIF=9.1; "20:00-22:00 time period": VIF=12.7; "Higher bid tier": VIF=6.5; "Live Stream Slice": VIF=10.3; Based on a preset threshold of 10, the "20:00-22:00 time slot" (VIF=12.7) and "live stream slices" (VIF=10.3) are removed, leaving the remaining three factors to form a decollinearized factor set {Platform A, short video materials, high bid tier}. This step effectively eliminates attribution bias caused by strong coupling between channels and time slots (e.g., Platform A mostly advertises in the evening).

[0148] S42: Statistically measure the path coverage and average interaction depth of decollinearized factors in historical conversion paths. The path coverage is the proportion of conversion paths containing decollinearized factors to the total number of conversion paths, and the average interaction depth is the average position number of the factor in the conversion path.

[0149] In this embodiment, path coverage reflects the reach of the factor; a higher value indicates greater prevalence. Average interaction depth reflects the timing of the factor's intervention in the user journey; a smaller value (e.g., close to 1) indicates that it frequently appears at the beginning of the path (e.g., the first touchpoint), and has a stronger effect on conversion guidance. Both together characterize the salience of the factor's path behavior.

[0150] For example, the system performs statistical analysis on each factor in the decollinearity set based on the standardized attribution dataset (containing 421,867 transformation paths) generated in Implementation Example 1: Platform A: Appears in 304,215 paths, path coverage. The sum of its position indices in each path is 548,320, therefore the average interaction depth is... The coverage of short video materials is approximately 0.686 (289,450 / 421,867), with an average depth of 2.15. The coverage of higher bid tiers is approximately 0.541 (228,100 / 421,867), with an average depth of 3.42.

[0151] S43: Based on path coverage and average interaction depth, the path significance score S of each decollinearity correlation factor is calculated using a weighted scoring formula. The weighted scoring formula is as follows: S=α×Coverage+β×(1 / Depth), Where Coverage is the path coverage rate, Depth is the average interaction depth, and α and β are preset weight coefficients with α+β=1.

[0152] Specifically, the weighted scoring formula balances "coverage" and "depth influence" (1 / Depth) through a linear combination. Since shallower depth is more important, its inverse form is used to make it positively correlated with the score. The weighting coefficients α and β reflect business preferences: if the focus is on reach scale, α is larger; if the focus is on path guidance (such as first-touch value), β is larger. In this embodiment, α = 0.6 and β = 0.4.

[0153] S44: Sort the path significance scores in descending order and select the top M decollinearity correlation factors as the dimension reduction correlation factors, where M is a positive integer.

[0154] In this embodiment, M is greater than or equal to 3, or M ≤ 70% of the total number of collinearity-free factors. This was verified by A / B testing: in e-commerce advertising scenarios, the Top-50% factors can explain more than 90% of the path contribution variance.

[0155] Specifically, the system sorts the path significance scores of the three factors in descending order: Platform A: 0.6548 Short video footage: 0.5976 High bid tier: 0.4416 The total number of collinearity-reducing factors is 3, and 70% of them is 2.1. Rounding down, we get M=2 (or rounding up to 3 according to business needs; this embodiment adopts a conservative strategy, taking M=⌊3×0.7⌋=2). Therefore, only the top 2 factors are selected as the final dimensionality-reduced correlation factors.

[0156] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0157] In one embodiment, an advertising performance attribution analysis system based on multi-source data fusion is provided, which corresponds to an advertising performance attribution analysis method based on multi-source data fusion in the above embodiment.

[0158] An advertising performance attribution analysis system based on multi-source data fusion includes a data acquisition and graph construction module, a feature recognition module, a filtering and dimensionality reduction module, a model optimization module, and an attribution result output module. Detailed descriptions of each functional module are as follows: The data acquisition and graph construction module is used to collect multi-source advertising data of the advertising campaign to be evaluated and preprocess it, and build a deep correlation graph based on the relationship between touchpoints, elements and conversions to generate a standardized attribution dataset. The feature recognition module is used to extract the performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-built marketing performance evaluation system, and to locate the factors related to the indicators. The filtering and dimensionality reduction module is used to acquire historical interaction data of the advertising campaign to be evaluated. Based on the historical interaction data and the standardized attribution dataset, it identifies the statistical characteristics of the index-related factors. According to the statistical characteristics, it constructs the factor levels of the index-related factors with respect to the performance evaluation index. Based on the factor levels, it filters the index-related factors to obtain initial related factors. After filtering the initial related factors, it performs dimensionality reduction processing to obtain dimensionality-reduced related factors. The model optimization module is used to calculate the factor state data of the dimensionality-reduced correlation factors corresponding to the historical interaction data, and to determine the attribution analysis model of the advertising campaign to be evaluated based on the factor state data. The attribution results output module is used to obtain real-time conversion data of the advertising campaign to be evaluated, and to determine the real-time attribution results of the advertising campaign to be evaluated based on the real-time conversion data and attribution analysis model.

[0159] For specific limitations regarding the advertising performance attribution analysis system based on multi-source data fusion, please refer to the limitations of an advertising performance attribution analysis method based on multi-source data fusion mentioned above, which will not be repeated here. Each module in the aforementioned advertising performance attribution analysis system based on multi-source data fusion can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or it can be stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0160] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores standardized attribution datasets, historical interaction data, etc. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements an advertising effectiveness attribution analysis method based on multi-source data fusion.

[0161] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the advertising effectiveness attribution analysis method based on multi-source data fusion as described above.

[0162] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the advertising effectiveness attribution analysis method based on multi-source data fusion as described above.

[0163] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0164] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0165] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for attribution analysis of advertising effectiveness based on multi-source data fusion, characterized in that, include: Collect and preprocess multi-source advertising data from advertising campaigns to be evaluated, and build a deep correlation graph based on the relationship between touchpoints, elements, and conversions to generate a standardized attribution dataset; Extract the performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-constructed marketing performance evaluation system, and locate the correlation factors of the indicators; Obtain historical interaction data of the advertising campaign to be evaluated; based on the historical interaction data and the standardized attribution dataset, identify the statistical characteristics of the index-related factors; and construct the factor levels of the index-related factors with respect to the performance evaluation index based on the statistical characteristics. Based on the factor level, the index correlation factors are screened to obtain initial correlation factors. After filtering the initial correlation factors, dimensionality reduction processing is performed to obtain dimensionality-reduced correlation factors. Calculate the factor state data corresponding to the dimensionality-reduced correlation factor of the historical interaction data, and determine the attribution analysis model of the advertising campaign to be evaluated based on the factor state data; Obtain real-time conversion data of the advertising campaign to be evaluated, and determine the real-time attribution result of the advertising campaign to be evaluated based on the real-time conversion data and the attribution analysis model.

2. The advertising effectiveness attribution analysis method based on multi-source data fusion according to claim 1, characterized in that, The extracted performance evaluation indicators for the advertising campaign to be evaluated within the pre-constructed marketing performance evaluation system specifically include: Identify the marketing objective category to which the advertising campaign to be evaluated belongs, including brand awareness, lead generation, or transaction conversion; Based on the marketing objective category, retrieve a set of candidate indicators associated with the category from a pre-built marketing performance evaluation system. The set of candidate indicators includes at least two of the following: exposure completion rate, click conversion rate, form submission rate, or transaction ROI. Obtain the key performance indicator (KPI) achievement data of the advertising campaign to be evaluated within the historical campaign period; Calculate the Pearson correlation coefficient between each candidate indicator in the candidate indicator set and the KPI achievement data, and use it as the correlation score of each candidate indicator; The relevance score is compared with a preset relevance threshold, and candidate indicators with relevance scores higher than the preset relevance threshold are selected as the performance evaluation indicators for the advertising campaign to be evaluated.

3. The advertising effectiveness attribution analysis method based on multi-source data fusion according to claim 1, characterized in that, The step of extracting the performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-constructed marketing performance evaluation system, and locating the correlation factors of the indicators, also includes: Based on the standardized attribution dataset, a user-level conversion path sequence is constructed, which records multiple ad touchpoints that the user encounters before conversion in chronological order. For each ad touchpoint, the corresponding delivery element feature vector is extracted. The delivery element feature vector includes channel identifier, material type code, delivery time segment tag and bid tier. Using the observed values ​​of the aforementioned performance evaluation indicators as the target variable, and the concatenated result of the feature vectors of all deployed elements as the input feature matrix, a gradient boosting tree model is trained. The average information gain value of each dimension feature in predicting the performance evaluation index is calculated using the built-in feature importance evaluation rule of the gradient boosting tree model. The advertising touchpoints or delivery elements corresponding to feature dimensions whose average information gain value is greater than a preset importance threshold are identified as correlation factors related to the performance evaluation index.

4. The advertising effectiveness attribution analysis method based on multi-source data fusion according to claim 1, characterized in that, The filtering of the initial correlation factors includes: Obtain a historical advertising campaign sample set corresponding to the initial association factor from a preset advertising database. The historical advertising campaign sample set includes the path coverage and average interaction depth of each initial association factor in the historical conversion path. A target sample is randomly selected from the historical advertising campaign sample set, and based on the path coverage and average interaction depth, samples of the same class with a similarity higher than a preset high threshold and samples of different classes with a similarity lower than a preset low threshold are identified. Configure the initial factor weights of the initial correlation factors; Extract a predetermined number of nearest neighbor samples of the same type and a predetermined number of nearest neighbor samples of different types from the samples of the same type and the samples of different types, respectively. The association factor weights of the initial association factor are calculated based on the distance relationships in the feature space between the initial factor weights, the target sample, the similar nearest neighbor samples, and the different near neighbor samples; When the weight of the correlation factor is not less than a preset weight threshold, the corresponding initial correlation factor is determined as the filtered correlation factor.

5. The advertising effectiveness attribution analysis method based on multi-source data fusion according to claim 4, characterized in that, The step of calculating the association factor weight of the initial association factor based on the distance relationship in the feature space of the initial factor weight, the target sample, the nearest neighbor samples of the same class, and the nearest neighbor samples of different classes includes: Calculate the first distance difference between the target sample and each of its nearest neighbors of the same class in the feature space, and calculate the second distance difference between the target sample and each of its nearest neighbors of different classes in the feature space; Obtain the distribution ratio of samples of the same type and samples of different types in the historical advertising campaign sample set, as the category prior probability; Based on the first distance difference and the prior probability of the same category, a positive update term is obtained; based on the second distance difference and the prior probability of different categories, a negative update term is obtained. The difference between the positive update term and the negative update term is calculated to obtain the single-round weight update amount; The single-round weight update amounts obtained from multiple rounds of random sampling are accumulated, and the correlation factor weight of the initial correlation factor is calculated based on the accumulation result and the initial factor weight.

6. The advertising effectiveness attribution analysis method based on multi-source data fusion according to claim 4, characterized in that, The step of filtering and dimensionality reduction of the initial correlation factors to obtain dimensionality-reduced correlation factors further includes: Calculate the variance inflation factor (VIF) for the feature columns corresponding to the initial correlation factors, and remove the top K feature columns with larger VIF values ​​to obtain the set of decollinearity correlation factors, where K is an integer; The path coverage and average interaction depth of the decollinearized factors in historical conversion paths are statistically analyzed. The path coverage is the proportion of the number of conversion paths containing the decollinearized factors to the total number of conversion paths, and the average interaction depth is the average position index of the factor in the conversion path. Based on the path coverage and average interaction depth, the path significance score S of each decollinearity correlation factor is calculated using a weighted scoring formula, which is: S=α×Coverage+β×(1 / Depth), Where Coverage is the path coverage rate, Depth is the average interaction depth, and α and β are preset weight coefficients with α+β=1; The path significance scores are sorted in descending order, and the top M decollinearity correlation factors are selected as the dimension reduction correlation factors, where M is a positive integer.

7. The advertising effectiveness attribution analysis method based on multi-source data fusion according to claim 1, characterized in that, The step of determining the attribution analysis model for the advertising campaign to be evaluated based on the factor state data includes: Based on the factor state data, a state transition graph of the user conversion path is constructed. The nodes of the state transition graph are advertising touchpoint types, and the edge weights are the transition frequencies of adjacent touchpoints in the historical path. Using the state transition diagram and the preset transition absorption state, the steady-state probability distribution of the Markov chain is solved to obtain the long-term access probability of each advertising touchpoint. Based on the deviation between the long-term access probability and the actual conversion rate, counterfactual inference is used to verify: estimate the expected change in the overall conversion rate after removing any ad touchpoint, and determine the expected change as the path-level contribution benchmark value of the ad touchpoint; Invoke a pre-built general XGBoost attribution model, which takes the characteristics of the delivery elements as input and the attribution contribution value as output; Using the path-level contribution benchmark value as a supervision signal, the prediction output of the general XGBoost attribution model in the dimensionality-reduced correlation factor feature space of the current advertising campaign is biased to obtain the calibrated real-time attribution prediction function. Determining the real-time attribution results of the advertising campaign to be evaluated includes: When the real-time conversion data of the advertising campaign to be evaluated is received, the real-time feature vector of the dimension-reduced correlation factor in the corresponding user conversion path is extracted and input into the calibrated real-time attribution prediction function. The system receives the predicted contribution scores of each dimension-reduced correlation factor output by the calibrated real-time attribution prediction function, and aggregates them according to the dimensions of advertising touchpoints and delivery elements to generate real-time attribution results that include the contribution ratio of channel, creative, and time period dimensions.

8. An advertising effectiveness attribution analysis system based on multi-source data fusion, characterized in that, The system includes: The data acquisition and graph construction module is used to collect multi-source advertising data of the advertising campaign to be evaluated and preprocess it, and build a deep correlation graph based on the relationship between touchpoints, elements and conversions to generate a standardized attribution dataset. The feature recognition module is used to extract the performance evaluation indicators corresponding to the advertising campaign to be evaluated within the pre-built marketing performance evaluation system, and to locate the factors related to the indicators. The filtering and dimensionality reduction module is used to acquire historical interaction data of the advertising campaign to be evaluated, identify the statistical characteristics of the indicator correlation factors based on the historical interaction data and the standardized attribution dataset, construct the factor levels of the indicator correlation factors with respect to the effect evaluation indicator according to the statistical characteristics, filter the indicator correlation factors according to the factor levels to obtain initial correlation factors, and perform dimensionality reduction processing on the initial correlation factors to obtain dimensionality-reduced correlation factors. The model optimization module is used to calculate the factor state data of the dimensionality-reduced correlation factors corresponding to the historical interaction data, and determine the attribution analysis model of the advertising campaign to be evaluated based on the factor state data. The attribution result output module is used to obtain the real-time conversion data of the advertising campaign to be evaluated, and to determine the real-time attribution result of the advertising campaign to be evaluated based on the real-time conversion data and the attribution analysis model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the advertising effectiveness attribution analysis method based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the advertising effect attribution analysis method based on multi-source data fusion as described in any one of claims 1 to 7.

Images