Multi-factor Marketing Data Correlation Analysis Methods and Systems

Abstract

This invention provides a method and system for multi-factor marketing data correlation analysis. The method includes: collecting and integrating multi-dimensional marketing data from advertising platforms, user behavior tracking systems, and sales transaction systems to generate a raw dataset containing timestamps, channel tags, user profiles, and numerical indicators; calculating time fluctuation factors, normalizing market size, and allocating channel weights on the raw dataset to generate a normalized and weighted dataset; constructing a correlation matrix, calculating correlation coefficients, and screening core factors on the normalized dataset to generate a set of core factors; and using the ranking results of the core factors and the normalized and weighted dataset for model training, prediction, strategy simulation, and suggestion generation to generate marketing decision optimization suggestions. This application integrates multi-source data to solve compatibility issues, dynamically processes and accurately analyzes data, and generates optimization suggestions in a closed loop to improve the efficiency of marketing decision-making.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of data processing, and in particular to a method and system for multi-factor marketing data correlation analysis. Background Technology

[0002] With the rapid development of the digital economy, marketing data generated in e-commerce platforms, social media, and brand marketing scenarios is experiencing explosive growth. Efficiently analyzing multi-source, heterogeneous marketing data and uncovering its inherent correlations has become a key issue in improving marketing decision-making efficiency. Applications such as video montage, targeted advertising, and user behavior analysis place higher demands on the real-time and intelligent integration, dynamic modeling, and strategy generation of marketing data. Traditional data processing solutions are no longer sufficient to meet the needs of large-scale personalized marketing.

[0003] Existing marketing data analysis technologies typically employ a "fixed template + manual parameter tuning" approach: marketing data is initially processed using a pre-set data analysis template, and parameters are then adjusted based on human experience to complete correlation analysis. This approach often relies on a single-dimensional data processing logic, lacking the ability to semantically understand and dynamically integrate multi-source data. For example, when faced with multi-source data such as advertising placement, user behavior, and sales transactions, current technologies struggle to automatically identify key factors and potential correlations within the data, resulting in analysis results that fail to accurately reflect market dynamics and user needs.

[0004] Further analysis reveals the following technical shortcomings in existing solutions: First, the multi-source data integration process lacks a standardized mechanism, resulting in poor compatibility with heterogeneous data from advertising platforms, user behavior systems, and sales databases, easily leading to data processing errors. Second, the data analysis process relies on human experience or fixed rules, failing to dynamically generate analytical logic based on the semantic features of the data (such as scenario type and user profile), making it difficult to adapt to complex and ever-changing marketing scenarios. Third, the correlation analysis stage lacks an intelligent optimization mechanism, with coarse selection and weight allocation of core factors leading to low analysis efficiency and insufficient accuracy. Fourth, the strategy generation stage lacks multi-dimensional simulation and dynamic optimization capabilities, failing to provide accurate decision-making suggestions based on marketing objectives and scenario changes. These technical bottlenecks make it difficult for existing solutions to achieve efficient data-driven decision-making when faced with massive amounts of marketing data, necessitating more intelligent and flexible technical solutions to overcome these limitations. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and system for multi-factor marketing data correlation analysis to at least partially solve the above problems.

[0006] A multi-factor marketing data correlation analysis method, comprising:

[0007] The system collects and integrates multi-dimensional marketing data from advertising platforms, user behavior tracking systems, and sales transaction systems to generate a raw dataset containing timestamps, channel tags, user profiles, and numerical metrics.

[0008] The original dataset is processed by calculating time fluctuation factors, normalizing market size, and allocating channel weights to generate a normalized and weighted dataset.

[0009] The normalized dataset is subjected to correlation matrix construction, correlation coefficient calculation and core factor screening to generate a core factor set;

[0010] This tool is used to train, predict, simulate strategies, and generate suggestions based on the ranking results of core factors and the normalized and weighted dataset, in order to generate marketing decision optimization recommendations.

[0011] Optionally, the method further includes: performing outlier filtering and missing value completion on the original dataset to generate a cleaned dataset.

[0012] Optionally, the step of performing outlier filtering and missing value completion on the original dataset to generate a cleaned dataset includes:

[0013] Perform statistical analysis on the original dataset to generate a set of outlier labels.

[0014] The original dataset with marked outliers is subjected to data removal to generate a preliminary cleaned dataset.

[0015] The initial cleaned dataset is scanned for null and incomplete data to generate a list of missing value locations.

[0016] The initial cleaned dataset containing missing values ​​is interpolated and filled to generate the cleaned dataset.

[0017] Optionally, the method further includes: performing exponential smoothing on historical daily / monthly data in the original dataset to generate a seasonal fluctuation coefficient.

[0018] Optionally, exponential smoothing is applied to the historical daily / monthly data in the original dataset to generate seasonal fluctuation coefficients, including:

[0019] The historical daily / monthly data in the original dataset is segmented into time series and decomposed into trends to generate trend component and periodic component datasets.

[0020] The periodic component dataset is subjected to exponential weighting to generate weighted periodic data;

[0021] The weighted periodic data is normalized and coefficients are extracted to generate seasonal fluctuation coefficients.

[0022] Optionally, the method further includes: visualizing the correlation matrix and supporting manual selection of factor combinations to generate a factor importance chart.

[0023] Optionally, the correlation matrix can be visualized, and manual selection of factor combinations can be supported to generate factor importance charts, including:

[0024] Semantic label embedding and weight pre-calculation are performed on the relevance matrix to generate enhanced matrix data;

[0025] Enhanced matrix data is processed by integrating interactive logic weaving and dynamic filtering engines to generate an intelligent interactive visualization framework;

[0026] Based on an intelligent interactive visualization framework, topological mapping and importance visualization rendering are performed on user-selected factor combinations and enhanced matrix data to generate factor importance charts.

[0027] Optionally, the method further includes: performing an error assessment on the prediction results; when the error rate exceeds a threshold, readjusting the weights of the core factors; and performing a strategy simulation to generate a strategy adjustment suggestion.

[0028] Optionally, the prediction results are evaluated for error. When the error rate exceeds a threshold, the weights of the core factors are readjusted, and a strategy simulation is performed to generate strategy adjustment suggestions, including:

[0029] The prediction results are subjected to multi-index cross-validation and error distribution analysis to generate an error assessment report;

[0030] When the error assessment report shows that the error rate exceeds the preset threshold, the weights of the core factors are optimized and adaptively adjusted using game theory to generate a corrected weight set.

[0031] Monte Carlo simulation and strategy extrapolation are performed on the revised weight set and historical marketing data to generate strategy adjustment suggestions.

[0032] A multi-factor marketing data correlation analysis system, comprising:

[0033] The data aggregation unit is used to collect and integrate raw data from advertising platform APIs, user behavior tracking systems, and sales databases to generate raw datasets containing timestamps, channel tags, user profiles, and numerical metrics.

[0034] The factor calculation unit is used to calculate time fluctuation factors, normalize market size, and allocate channel weights to the original dataset to generate a normalized and weighted dataset.

[0035] The correlation analysis unit is used to construct a correlation matrix, calculate correlation coefficients, and screen core factors from the normalized dataset to generate a set of core factors.

[0036] The prediction and optimization unit is used to perform model training, prediction, strategy simulation, and suggestion generation on the core factor ranking results and the normalized and weighted dataset to generate marketing decision optimization suggestions.

[0037] The multi-factor marketing data correlation analysis method proposed in this application provides a systematic solution for the entire data processing process, and has the following technical advantages:

[0038] 1. By collecting and integrating multi-dimensional data from advertising platforms, user behavior tracking systems, and sales transaction systems, a unified raw dataset containing information such as timestamps and channel tags is generated. This effectively solves the problem of heterogeneous data compatibility and avoids data processing deviations caused by inconsistent formats and standards, providing an accurate and standardized data foundation for subsequent analysis.

[0039] 2. The original dataset is processed by calculating time fluctuation factors, normalizing market size, and allocating channel weights to generate a normalized and weighted dataset. This process automatically adapts to the time characteristics of the data, changes in the market environment, and channel differences, breaking through the limitations of traditional reliance on fixed templates and human experience. It dynamically adjusts the analysis logic based on the semantic features of the data, improving the flexibility and adaptability of data processing in complex marketing scenarios.

[0040] 3. By constructing a correlation matrix, calculating correlation coefficients, and screening core factors, a core factor set is formed. Compared to the crude factor screening in existing technologies that lack intelligent optimization, this solution can automatically identify key influencing factors and their potential correlations in multi-source data, accurately allocate weights, significantly improve the efficiency and accuracy of correlation analysis results, and accurately reflect market dynamics and user needs.

[0041] 4. Based on the core factor set and normalized dataset, the system performs model training, prediction, strategy simulation, and suggestion generation, outputting optimized marketing decision-making recommendations. This process generates precise strategies through multi-dimensional simulation and dynamic optimization, combined with changes in marketing objectives and scenarios. It compensates for the shortcomings of existing solutions in the strategy generation stage, achieving a complete closed loop from data collection and analysis to decision-making recommendations, effectively improving the efficiency of data-driven marketing decision-making. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0043] Figure 1 This is a flowchart illustrating a multi-factor marketing data correlation analysis method according to an embodiment of this application.

[0044] Figure 2 This invention provides a schematic diagram of the structure of a multi-factor marketing data correlation analysis system. Detailed Implementation

[0045] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0046] It should be understood that the terms "first," "second," and "third," etc., in the claims, specification, and drawings of this disclosure are used to distinguish different objects, not to describe a specific order. The terms "comprising" and "including" as used in the specification and claims of this disclosure indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or sets thereof.

[0047] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any and all combinations of one or more of the associated listed items, and includes such combinations.

[0048] Figure 1 This is a flowchart illustrating a multi-factor marketing data correlation analysis method according to an embodiment of this application. Figure 1 As shown, a multi-factor marketing data correlation analysis method includes:

[0049] The system collects and integrates multi-dimensional marketing data from advertising platforms, user behavior tracking systems, and sales transaction systems to generate a raw dataset containing timestamps, channel tags, user profiles, and numerical metrics.

[0050] The original dataset is processed by calculating time fluctuation factors, normalizing market size, and allocating channel weights to generate a normalized and weighted dataset.

[0051] The normalized dataset is subjected to correlation matrix construction, correlation coefficient calculation and core factor screening to generate a core factor set;

[0052] This tool is used to train, predict, simulate strategies, and generate suggestions based on the ranking results of core factors and the normalized and weighted dataset, in order to generate marketing decision optimization recommendations.

[0053] Preferably, during the collection and structured integration of multi-source heterogeneous data, the raw marketing data from advertising platforms (such as Google Ads, Douyin's massive search engine), user behavior tracking systems (such as event tracking logs and heatmap data), and sales transaction systems (such as POS terminals and e-commerce order data) are collected in real time through API interfaces and ETL tools, supporting formats such as CSV, JSON, and XML. The raw data is then preprocessed using data cleaning algorithms (such as outlier filtering and missing value interpolation). The data is then structured and integrated according to a unified data model (such as a star schema), adding timestamps, channel tags (such as "information flow ads" and "organic traffic"), user profiles (such as age, region, and consumption level), and numerical indicators (such as exposure, click-through rate, and GMV) to each data point, thereby obtaining the original dataset (i.e., a standardized, structured, multi-dimensional marketing data set).

[0054] Preferably, during dynamic factor weighted normalization, the Holt-Winters seasonality index smoothing algorithm is used on the original dataset to decompose historical time series data (such as sales revenue counted daily / monthly), extract seasonal fluctuation coefficients, and quantify the periodic impact of holidays, promotional activities, etc.; further, industry total market size data is obtained through third-party data platforms (such as iResearch Consulting, Statista), and the original numerical indicators are converted into relative values ​​(such as the proportion of brand sales revenue in the industry); then, based on the reinforcement learning model, weights are dynamically allocated according to the historical ROI performance of each marketing channel and current business objectives (such as 0.6 weight for information flow advertising and 0.4 weight for SEO), thereby obtaining a normalized and weighted dataset (standardized data that eliminates differences in data dimensions and highlights the contribution of factors).

[0055] Preferably, during intelligent factor selection and correlation modeling, the Pearson correlation coefficients (such as advertising investment and conversion rate) between factors are calculated for the normalized and weighted dataset to generate an N×N-dimensional correlation matrix. Then, by combining Lasso regression and genetic algorithms and optimizing through cross-validation, the core factors that contribute the top 30% to the target indicator (such as ROI) are selected. Finally, natural language processing (NLP) technology is used to analyze marketing text data (such as advertising copy and user reviews) to mine implicit correlation factors (such as the relationship between public opinion heat and sales), thereby obtaining a core factor set (containing key influencing factors and their weights, and a subset of correlation indicators).

[0056] When performing predictive optimization and intelligent decision generation, a gradient boosting tree (GBDT) or Transformer architecture is used for the core factor set and the normalized and weighted dataset. The predictive model is trained based on historical data, inputting the core factors and normalized features, and outputting the predicted values ​​of marketing indicators for future periods. Then, Monte Carlo simulation is used to generate more than 1,000 marketing strategy combinations (such as budget reallocation and channel priority adjustment), and the expected effect of each strategy is evaluated in combination with the predictive model. Finally, a multi-objective optimization algorithm (such as Pareto optimality) is used to select 3-5 high-cost-performance strategies and generate marketing decision optimization suggestions that include adjustment direction, expected returns, and risk warnings (which can directly guide action plans for resource allocation and strategy adjustment).

[0057] Optionally, the method further includes: performing outlier filtering and missing value completion on the original dataset to generate a cleaned dataset.

[0058] Optionally, the step of performing outlier filtering and missing value completion on the original dataset to generate a cleaned dataset includes:

[0059] Perform statistical analysis on the original dataset to generate a set of outlier labels.

[0060] The original dataset with marked outliers is subjected to data removal to generate a preliminary cleaned dataset.

[0061] The initial cleaned dataset is scanned for null and incomplete data to generate a list of missing value locations.

[0062] The initial cleaned dataset containing missing values ​​is interpolated and filled to generate the cleaned dataset.

[0063] Preferably, in multi-dimensional outlier detection and labeling, for the original dataset (containing multi-dimensional marketing data, such as ad impressions, click-through rate, conversion rate, etc.), the interquartile range (IQR), Z-score, and DBSCAN density clustering algorithms are combined to perform anomaly detection on numerical fields. For example, for the ad impressions field, both IQR (outside Q3+1.5IQR is considered anomaly) and Z-score (|Z|>3 is considered anomaly) are calculated, and the intersection is taken as the initial anomaly set. A business rule engine is designed based on the characteristics of the marketing data. For example, if the click-through rate of a certain channel exceeds 3 times the industry average and the conversion rate is less than 50% of the industry average, it is marked as a suspected fraudulent activity. For time-series data (such as daily active users), the Prophet model is used to predict the normal value range, and points where the deviation between the actual value and the predicted value exceeds a threshold (such as 2σ) are marked as anomalies, thus obtaining an outlier label set (containing the location index and anomaly type label of the outlier data in each dimension).

[0064] Preferably, during intelligent data removal and repair, different processing is performed on the original dataset marked with outliers, based on the severity of the anomalies. Severe anomalies (such as inflated traffic data) are directly deleted; minor anomalies (such as exceeding the IQR but not triggering business rules) are corrected using a local regression model (e.g., replacing outliers with curves fitted based on data from adjacent time points). When removing outlier rows, data integrity is protected through correlation analysis. For example, if a user's click behavior is marked as abnormal, but the user subsequently has genuine purchase records, the purchase data is retained while the click data is corrected. For small amounts of data from newly launched channels or products, industry benchmark values ​​are used to replace outliers to avoid misjudgments caused by data sparsity, thus obtaining a preliminary cleaned dataset (a dataset with severely abnormal data removed and minor anomalies repaired).

[0065] When locating multimodal missing values, a comprehensive scan is performed on the initially cleaned dataset, considering the field dimension (which fields are missing), the record dimension (which records are missing), and the time dimension (the temporal pattern of missing values). For example, it was found that the missing rate of the user age field increased significantly during a certain promotional activity. Potential patterns of missing values ​​are then mined using association rules. For instance, if the "region" field is missing, the "average order value" field is also likely to be missing, marking it as a strongly correlated missing pattern. Based on the missing mechanism, values ​​are categorized into MCAR (completely random missing), MAR (random missing), and MNAR (non-random missing), providing a basis for subsequent imputation strategies and resulting in a list of missing value locations (including the coordinates, type, and association pattern information of the missing values).

[0066] Preferably, during adaptive missing value imputation, the optimal imputation algorithm is automatically selected for the initially cleaned dataset containing missing values ​​based on the missing value type (numerical / categorical), missing mechanism (MCAR / MAR / MNAR), and field importance (e.g., core indicators are given priority for more complex methods).

[0067] Numerical fields: KNN nearest neighbor interpolation is used for MCAR; random forest regression is used for MAR; and variational autoencoder (VAE) is used to generate MNAR.

[0068] Categorical fields: BERT model imputation based on contextual semantics (e.g., predicting missing interest tags based on user history behavior).

[0069] For the same missing value, the results of multiple imputation methods are combined and assigned different weights. For example, for the user age field, the KNN imputation value (weight 0.4), the regression prediction value (weight 0.4), and the business average value (weight 0.2) are combined. A posterior simulation method is used to perform self-consistency checks on the imputation results. For example, after imputing the "purchase amount" field, the logical relationship between it and "purchase quantity" and "average order price" is verified to be reasonable, thus obtaining a cleaned dataset (a complete dataset without outliers and missing values, and maintaining the inherent logical consistency of the data).

[0070] Optionally, the method further includes: performing exponential smoothing on historical daily / monthly data in the original dataset to generate a seasonal fluctuation coefficient.

[0071] Optionally, exponential smoothing is applied to the historical daily / monthly data in the original dataset to generate seasonal fluctuation coefficients, including:

[0072] The historical daily / monthly data in the original dataset is segmented into time series and decomposed into trends to generate trend component and periodic component datasets.

[0073] The periodic component dataset is subjected to exponential weighting to generate weighted periodic data;

[0074] The weighted periodic data is normalized and coefficients are extracted to generate seasonal fluctuation coefficients.

[0075] Preferably, in the adaptive time series segmentation and trend decomposition, for the historical daily / monthly data in the original dataset (such as sales revenue and advertising exposure over the past 36 months), a change point detection algorithm (such as the Pelt algorithm) is used to automatically identify structural change points (such as business strategy adjustments and sudden market events) in the data, dividing the time series into multiple homogeneous intervals; then, wavelet transform is used to perform multi-resolution analysis on the data in each interval, separating the long-term trend (low-frequency component), seasonal cycle (mid-frequency component), and random noise (high-frequency component); finally, for the trend component of each interval, Kalman filter is used to perform dynamic parameter estimation to adapt to the changing characteristics of non-stationary time series, thereby obtaining the trend component and cycle component dataset (separated seasonal fluctuation characteristic data, eliminating long-term trend and random noise interference).

[0076] Preferably, when using time-varying parameters for exponential weighting, the α, β, and γ parameters of the Holt-Winters exponential smoothing model are dynamically adjusted using reinforcement learning (Q-learning) for the periodic component dataset, enabling the model to automatically select the optimal smoothing coefficient under different data patterns. Furthermore, a time decay factor is introduced to assign higher weights to recent data (e.g., 0.6 for data from the last month and 0.1 for data from three months ago), while preserving historical periodic characteristics. During the weighting process, outliers are robustly estimated using the Huber loss function to avoid abnormal fluctuations distorting the seasonal pattern, thus obtaining weighted periodic data (a periodic data sequence that reinforces recent seasonal patterns and suppresses noise interference).

[0077] Preferably, during dynamic normalization and coefficient generation, for weighted periodic data, for each period segment (e.g., monthly, quarterly), its Z-score standardized value is calculated, mapping the data to the [-1,1] interval, preserving the relative fluctuation range; the power spectral density of the periodic components is calculated through spectral analysis to quantify the energy proportion of each frequency component and extract the dominant periodic frequency; the standardized periodic data is converted into seasonal fluctuation coefficients, and by comparing the deviation between historical actual values ​​and predicted values, a calibration factor (e.g., industry average fluctuation level) is introduced for fine-tuning, thereby obtaining the seasonal fluctuation coefficient (a set of numerical values ​​that quantifies the seasonal characteristics of each period, such as a coefficient of 1.8 during the "618" period, indicating that sales are usually 1.8 times the average).

[0078] In summary, the Pelt algorithm, by automatically identifying structural changes (such as new product launches or policy adjustments), can more accurately segment homogeneous intervals compared to traditional fixed-window decomposition. For example, if an e-commerce platform launches a new live-streaming sales strategy in Q2 2024, the algorithm can automatically identify this time point, avoiding the impact of data overlap before and after the strategy change on trend analysis. This makes seasonal analysis more aligned with actual business cycles and improves the responsiveness of subsequent predictive models to strategy adjustments.

[0079] In this application, the multi-resolution analysis capability of wavelet transform can simultaneously separate long-term trends (such as industry growth curves), seasonal cycles (such as the Spring Festival shopping peak), and random noise (such as sudden public opinion events). For example, after decomposition, monthly sales data of a fast-moving consumer goods product can clearly separate quarterly promotional cycles and annual consumption upgrade trends. Traditional STL decomposition can only handle addition or multiplication models, while wavelet transform can adapt to nonlinear and non-stationary data, improving the decomposition accuracy of complex marketing data by more than 30% (based on industry test data).

[0080] In this application, for non-stationary trend components, the Kalman filter dynamically updates parameters through a state-space model to adapt to changes in the trend slope. For example, when a new energy vehicle company's sales data changes due to policy subsidies, the Kalman filter can quickly adjust the trend prediction parameters, reducing lag error. Compared to the traditional moving average method, the Kalman filter reduces prediction error by 50% in scenarios with abrupt trend changes (refer to related research in IEEE Signal Processing Letters).

[0081] In this application, the Q-learning algorithm dynamically adjusts the Holt-Winters model parameters based on historical prediction errors to achieve "intelligent smoothing." For example, during peak seasons in the beauty industry (such as Singles' Day), the algorithm automatically increases the α value (recent data weight), enabling the model to capture the surge in sales trends more quickly. Traditional exponential smoothing methods require manual parameter tuning, while this solution achieves parameter self-optimization, improving model training efficiency by 40% and prediction accuracy by 15%-20% (based on A / B testing results from a leading e-commerce platform).

[0082] In this application, data from the most recent month is assigned a weight of 0.6, while data from three months ago is weighted at 0.1, effectively balancing the pursuit of newer trends with the preservation of older ones. For example, the sales cycle of a certain 3C product is short (average 2 months), and the decay mechanism can quickly weaken the impact of outdated data, making the seasonality coefficient closer to the current market rhythm. Compared with equal-weighted or fixed-decay models, this solution improves the prediction accuracy in the fast-moving consumer goods sector by 25%.

[0083] In this application, the Huber loss function provides robust estimation of outliers, preventing occasional outliers (such as a surge in daily sales due to celebrity endorsements) from distorting seasonal patterns. For example, if a sports brand experiences a tenfold increase in daily sales due to an Olympic champion's endorsement, the seasonality coefficient, after Huber processing, still reflects the true monthly fluctuation pattern. In scenarios involving frequent influencer marketing and unexpected events, the model's stability is improved by 60%, reducing resource waste caused by misjudgments.

[0084] In this application, Z-score standardization maps data from each period to the [-1,1] interval, preserving relative fluctuations. For example, the sales fluctuation of a luxury brand is much greater than that of a mass-market brand. Standardization allows for comparison of seasonal characteristics on a unified scale, preventing large numbers from masking small fluctuations. This makes marketing metrics of different magnitudes (such as exposure and conversion rate) comparable, improving the accuracy of cross-dimensional analysis.

[0085] In this application, power spectral density analysis is used to quantify the energy proportion of each frequency component and automatically identify the dominant cycle. For example, after spectral analysis, traffic data from a travel platform can clearly identify multiple cycles such as weekly (weekend travel), monthly (holiday effect), and quarterly (winter and summer vacations), extracting the most significant cycle frequency as the dominant factor. Compared with traditional Fourier transform, power spectral density analysis improves the accuracy of cycle identification for non-stationary and nonlinear data by 40% (refer to ACM Transactions on Intelligent Systems and Technology).

[0086] This application introduces the industry average volatility level as a calibration factor to correct for company-specific biases. For example, after calibration with the industry average, the seasonal volatility coefficient of an emerging brand can more objectively reflect the overall market trend and avoid coefficient deviations caused by its own immature business. This makes the seasonality coefficient more universal and supports cross-company and cross-industry marketing effectiveness benchmarking analysis.

[0087] Optionally, the method further includes: visualizing the correlation matrix and supporting manual selection of factor combinations to generate a factor importance chart.

[0088] Optionally, the correlation matrix can be visualized, and manual selection of factor combinations can be supported to generate factor importance charts, including:

[0089] Semantic label embedding and weight pre-calculation are performed on the relevance matrix to generate enhanced matrix data;

[0090] Enhanced matrix data is processed by integrating interactive logic weaving and dynamic filtering engines to generate an intelligent interactive visualization framework;

[0091] Based on an intelligent interactive visualization framework, topological mapping and importance visualization rendering are performed on user-selected factor combinations and enhanced matrix data to generate factor importance charts.

[0092] Preferably, during the semantic processing of the knowledge graph-enhanced matrix, for the original relevance matrix (containing the correlation coefficients between various marketing factors), the factor names in the matrix (such as "ad placement" and "user retention rate") are aligned with the industry knowledge graph (such as a marketing domain ontology), and semantic descriptions (such as category and business definition) are added to each factor. Combining the Shapley value algorithm and business objectives (such as maximizing ROI), importance weights are calculated for each factor. For example, during promotional activities, the weight of the "discount strength" factor is dynamically increased. Causal inference (such as Do-calculus) is used to identify the causal direction between factors, upgrading the correlation coefficient matrix to a causal strength matrix. For example, it is discovered that "ad impressions" are a cause of "click-through rate," rather than a simple correlation, thus obtaining enhanced matrix data (a multi-dimensional matrix containing semantic descriptions, dynamic weights, and causal relationships).

[0093] Preferably, when building the 3D interactive visualization engine, for enhanced matrix data, a 3D interactive scene is constructed based on Three.js, mapping factors to nodes in space, and correlation coefficients to line thickness and color (e.g., red indicates positive correlation, blue indicates negative correlation). Based on a multi-dimensional filter, real-time filtering is performed by factor category (e.g., channel category, user behavior category), importance threshold (e.g., only displaying factors with a weight > 0.5), and time range (e.g., data from the last 30 days). Based on user historical operation records, high-frequency factor combinations are recommended through a collaborative filtering algorithm. For example, when a new user opens the interface, the classic combination of "advertising placement - conversion rate - average order value" is automatically recommended, thereby generating an intelligent interactive visualization framework (an interactive interface supporting 3D browsing, dynamic filtering, and intelligent recommendations).

[0094] Preferably, during adaptive topology mapping and visualization rendering, for the user-selected factor combinations and enhanced matrix data, a force-directed graph algorithm (such as ForceAtlas2) is used to automatically lay out nodes based on the causal relationships and correlation strength between factors, bringing strongly correlated factors closer together and making the causal chain clearer. Simultaneously, heatmaps (matrix form), network diagrams (topology form), and sunburst diagrams (hierarchical importance) are displayed, allowing users to freely switch between views. For example, network diagrams are suitable for observing complex relationships between factors, while sunburst diagrams are suitable for quickly grasping hierarchical importance. When a user clicks on a factor node, the causal effect value of that factor (such as its direct impact on ROI) is calculated in real time, and its upstream and downstream related factors are highlighted. When a node is dragged, the topology structure is automatically updated and the layout is recalculated, thus obtaining a factor importance chart (including multi-view display, dynamic interaction, and visualization results of causal effect analysis).

[0095] In summary, traditional visualizations only display numerical relationships, while this solution injects business semantics through knowledge graphs, enabling marketers to quickly understand the business implications behind factors. For example, when viewing the "CTR" factor, the system automatically associates it with the industry standard definition and optimization suggestions for "Click-Through Rate." Furthermore, based on Shapley value-based weight calculations, compared to traditional fixed-weight methods, factor importance can be dynamically adjusted according to business objectives. An e-commerce platform applied this technology during "Double 11," increasing the weight of the "Promotion Sensitivity" factor by 30%, providing more precise guidance for budget allocation. Moreover, through causal inference technology, traditional correlation analysis is upgraded to causal analysis, helping companies identify the true driving factors. After applying this technology, a financial institution found that the causal effect of "Customer Complaint Handling Speed" on "Customer Loyalty" was stronger than the correlation coefficient indicated, thereby optimizing customer service processes.

[0096] In this application, WebGL-driven 3D visualization, compared to traditional 2D heatmaps, can more clearly display high-dimensional data relationships. When an automotive brand used this technology to analyze 20+ marketing factors, it discovered a hidden correlation path between "social media topic popularity" and "offline test drive appointments" in 3D space. Factor combination recommendations based on collaborative filtering improved marketers' efficiency in discovering high-value combinations by 60%. A FMCG brand optimized its influencer collaboration strategy by using the system-recommended combination of "KOL fan activity - product repurchase rate," resulting in a 25% increase in ROI. The multi-dimensional filtering engine supports quick switching of analytical perspectives. A cross-border e-commerce company, allowing free filtering across three dimensions ("region-channel-product"), discovered that the "influencer livestreaming" factor had a significantly greater impact on sales in the Southeast Asian market than in the European and American markets.

[0097] In this application, the force-directed graph algorithm automatically optimizes node layout to ensure that strongly correlated factors remain close together. In the analysis of an online education platform, this technology keeps the spatial distance between "course completion rate" and "renewal rate" within the visual focal point, facilitating an intuitive understanding of the relationship. It also provides heatmaps, network diagrams, and sunburst diagrams to meet different analytical needs. A beauty brand's marketing department used heatmaps to identify the strength of factor correlations, sunburst diagrams to determine budget allocation priorities, and network diagrams to identify potential cooperation opportunities. Clicking on nodes calculates causal effect values ​​in real time, providing quantitative evidence for decision-making. When clicking on the "ad placement" node on a travel platform, the system immediately displays the expected increase in bookings (e.g., 0.87%) from a 1% increase in placement, along with a confidence interval.

[0098] Optionally, the method further includes: performing an error assessment on the prediction results; when the error rate exceeds a threshold, readjusting the weights of the core factors; and performing a strategy simulation to generate a strategy adjustment suggestion.

[0099] Optionally, the prediction results are evaluated for error. When the error rate exceeds a threshold, the weights of the core factors are readjusted, and a strategy simulation is performed to generate strategy adjustment suggestions, including:

[0100] The prediction results are subjected to multi-index cross-validation and error distribution analysis to generate an error assessment report;

[0101] When the error assessment report shows that the error rate exceeds the preset threshold, the weights of the core factors are optimized and adaptively adjusted using game theory to generate a corrected weight set.

[0102] Monte Carlo simulation and strategy extrapolation are performed on the revised weight set and historical marketing data to generate strategy adjustment suggestions.

[0103] Preferably, during multi-dimensional error diagnosis and attribution analysis, for the predicted results (such as predicted values ​​of marketing indicators like future sales and conversion rates), MAE (Mean Absolute Error), RMSE (Root Mean Square Error), MAPE (Mean Absolute Percentage Error), and R... 2 12+ evaluation indicators, including the coefficient of determination, quantify prediction bias from different dimensions. For example, in predicting sales for a promotional activity, MAE reflects absolute error, MAPE reflects relative error, and R... 2 The model's goodness of fit is evaluated. Kernel density estimation (KDE) is used to analyze the probability distribution of errors and identify anomalous error clusters. For example, if the prediction error for a certain channel shows a peak distribution over a specific time period, it indicates a systematic bias. The contribution of each input factor to the prediction error is quantified by combining SHAP (SHapley Additive exPlanations) values ​​and LIME (Local Interpretable Model-agnostic Explanations) techniques. For example, it is determined that the "advertising volume" factor accounts for 35% of the prediction error, thus generating an error assessment report (a diagnostic document including multi-indicator assessment results, error distribution maps, and factor attribution analysis).

[0104] Preferably, when performing game theory-driven dynamic weight adjustment, each core factor is treated as a game participant, and its marginal contribution to reducing error is calculated using the Shapley value, taking into account the error assessment report and the core factor weights (the influence weights of each marketing factor in the original model). For example, the impact of the "user retention rate" factor on the overall error under different weight configurations is calculated. A weight optimization agent is constructed based on the Q-learning algorithm to dynamically adjust factor weights with the goal of minimizing error. For example, when the error contribution of the "promotional discount" factor is found to be high, the agent automatically reduces its initial weight. Weight adjustment upper limits (e.g., a single adjustment cannot exceed 20% of the original weight) and business logic constraints (e.g., the "advertising investment" weight must be positive) are set to avoid overfitting, thereby obtaining a corrected weight set (the optimal factor weight configuration that balances prediction accuracy and business logic).

[0105] Preferably, during Monte Carlo strategy simulation and elasticity analysis, based on the historical data distribution and the adjusted weight set, over 10,000 Monte Carlo samples are taken from key marketing parameters (such as budget allocation and channel combination). For example, 25 combinations of "increasing the budget for feed ads by 10%-30%" and "reducing the budget for search ads by 5%-15%" are simulated. A causal graph model is constructed using a Do-calculus framework to calculate the counterfactual results under different strategies. For example, "how would sales change if live-streaming e-commerce were not conducted in Q1 2024?" The elasticity coefficients of each factor (such as the impact of a 1% increase in advertising investment on sales) are calculated, and their effective ranges are determined. For example, it is found that "discount strength" has the most significant effect on conversion rate improvement within the 5%-15% range, thus obtaining strategy adjustment recommendations (including a decision support document containing high-potential strategy combinations, expected effect ranges, and risk warnings).

[0106] In summary, traditional methods rely solely on RMSE or MAE, while this solution employs cross-validation with 12+ indicators. In sales forecasting on an e-commerce platform, it was discovered that a single RMSE indicator masked systematic prediction biases in high-value orders, and multi-indicator analysis improved error localization accuracy by 40%. In this application, kernel density estimation technology reveals the probability distribution characteristics of errors, helping to identify abnormal patterns. A FMCG brand, through error distribution analysis, found that prediction errors during the new product launch period were significantly higher than those during the maturity period, and optimized the new product prediction model accordingly. Furthermore, the attribution technique combining SHAP and LIME can more accurately quantify factor contributions compared to traditional feature importance analysis. After application by a financial institution, it was found that the error contribution of the "customer age" factor in high-net-worth customer groups was severely underestimated; after adjustment, the model accuracy improved by 15%.

[0107] In this application, the Shapley value algorithm ensures that the contribution of each factor to error reduction is fairly evaluated, avoiding the problem of strong factors squeezing out weak factors in traditional methods. After application on a travel platform, the weight of the "user reviews" factor increased from being ignored to the third most important, optimizing the word-of-mouth marketing strategy. Furthermore, the Q-learning agent can respond to market changes in real time. In a 3C product promotion campaign, when a surge in prediction error was detected in the "social media popularity" factor, the agent adjusted the weights within 2 hours, improving subsequent prediction accuracy by 22%. Finally, the weight adjustment constraint mechanism prevents model overfitting. In the long-term prediction of a cosmetics brand, compared to unconstrained optimization, robust constraints reduced the model's prediction error by 28% after 6 months.

[0108] This application, with over 10,000 parameter samples covering more potential scenarios, demonstrated that a supermarket chain, through simulating promotional combinations in different regions, found that "discounts + gifts" were 40% more effective in third-tier cities than in first-tier cities, leading to targeted adjustments to its regional strategies. Furthermore, the Do-calculus framework supports precise calculations of causal effects; an online education platform, through counterfactual analysis, discovered that the true impact of "free trials" on long-term paid conversion rates was overestimated by 35%, resulting in optimized customer acquisition strategies.

[0109] The elasticity coefficient analysis in this application provides a quantitative boundary for decision-making. A certain car brand determined that increasing the "advertising budget" by 15% on the existing basis is the optimal solution. Exceeding this value will result in a sharp decrease in marginal benefits, thus avoiding waste of resources.

[0110] Figure 2 This is a schematic diagram of the structure of a multi-factor marketing data correlation analysis system according to an embodiment of this application. Figure 2 The above includes:

[0111] The data aggregation unit is used to collect and integrate raw data from advertising platform APIs, user behavior tracking systems, and sales databases to generate raw datasets containing timestamps, channel tags, user profiles, and numerical metrics.

[0112] The factor calculation unit is used to calculate time fluctuation factors, normalize market size, and allocate channel weights to the original dataset to generate a normalized and weighted dataset.

[0113] The correlation analysis unit is used to construct a correlation matrix, calculate correlation coefficients, and screen core factors from the normalized dataset to generate a set of core factors.

[0114] The prediction and optimization unit is used to perform model training, prediction, strategy simulation, and suggestion generation on the core factor ranking results and the normalized and weighted dataset to generate marketing decision optimization suggestions.

[0115] The above Figure 2 In the embodiments, an exemplary explanation of the technical processing procedures performed by each unit can be found above. Figure 1 The records.

[0116] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims. The application, device, module, or unit described in the above embodiments can be specifically implemented by a computer chip or entity, or by a product with a certain function.

[0117] For ease of description, the above apparatus is described by dividing it into various functional units. Of course, in implementing this invention, the functions of each unit can be implemented in one or more software and / or hardware components.

[0118] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, this application, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0119] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (this application), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0120] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0121] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0122] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, a network interface, and memory. Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0123] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0124] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0125] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, this application, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific transactions or implement specific abstract data types. This invention can also be practiced in distributed computing environments where transactions are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0127] The various embodiments in this specification are described in a progressive manner, with identical or similar parts between the embodiments referred to or substituted for each other. For the embodiments of this application, since they are basically similar to the method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0128] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims. The present application, device, module, or unit described in the above embodiments is specifically implemented by a computer chip or entity, or by a product having a certain function.

[0129] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, this application, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A multi-factor marketing data correlation analysis method, characterized in that, include: Multi-dimensional marketing data from advertising platforms, user behavior tracking systems, and sales transaction systems are collected and integrated to generate a raw dataset containing timestamps, channel tags, user profiles, and numerical metrics. The original dataset is processed by calculating time fluctuation factors, normalizing market size, and allocating channel weights to generate a normalized and weighted dataset; The normalized dataset is subjected to correlation matrix construction, correlation coefficient calculation, and core factor screening to generate a core factor set; The model is trained, predicted, simulated, and suggestions are generated based on the ranking results of core factors and the normalized and weighted dataset to generate marketing decision optimization suggestions. The method further includes: Semantic label embedding and weight pre-calculation are performed on the relevance matrix to generate enhanced matrix data; Enhanced matrix data is processed by integrating interactive logic weaving and dynamic filtering engines to generate an intelligent interactive visualization framework; Based on an intelligent interactive visualization framework, topological mapping and importance visualization rendering are performed on user-selected factor combinations and enhanced matrix data to generate factor importance charts.

2. The method according to claim 1, characterized in that, The method further includes: performing outlier filtering and missing value completion on the original dataset to generate a cleaned dataset.

3. The method according to claim 2, characterized in that, The process of filtering outliers and imputing missing values ​​in the original dataset to generate a cleaned dataset includes: Perform statistical analysis on the original dataset to generate a set of outlier labels; Data removal is performed on the original dataset with marked outliers to generate a preliminary cleaned dataset; The initially cleaned dataset is scanned for null and incomplete data to generate a list of missing value locations; The initial cleaned dataset containing missing values ​​is interpolated and filled to generate the cleaned dataset.

4. The method according to claim 1, characterized in that, The method further includes: performing exponential smoothing on historical daily / monthly data in the original dataset to generate a seasonal fluctuation coefficient.

5. The method according to claim 4, characterized in that, Exponential smoothing is applied to historical daily / monthly data in the original dataset to generate seasonal fluctuation coefficients, including: The historical daily / monthly data in the original dataset is segmented into time series and decomposed into trends to generate trend component and periodic component datasets. The periodic component dataset is subjected to exponential weighting to generate weighted periodic data; The weighted periodic data is normalized and coefficients are extracted to generate seasonal fluctuation coefficients.

6. The method according to claim 1, characterized in that, The method further includes: evaluating the prediction results for error; when the error rate exceeds a threshold, readjusting the weights of the core factors and performing strategy simulation to generate strategy adjustment suggestions.

7. The method according to claim 6, characterized in that, The prediction results are evaluated for error. When the error rate exceeds a threshold, the weights of the core factors are readjusted, and a strategy simulation is performed to generate strategy adjustment suggestions, including: The prediction results are subjected to multi-index cross-validation and error distribution analysis to generate an error assessment report; When the error assessment report shows that the error rate exceeds the preset threshold, the weights of the core factors are optimized and adaptively adjusted using game theory to generate a corrected weight set. Monte Carlo simulation and strategy extrapolation are performed on the revised weight set and historical marketing data to generate strategy adjustment suggestions.

Images