Full-link marketing effect attribution method and system of multi-task learning framework
Through a multi-task learning framework and quantum heuristic dynamic relationship map algorithm, combined with advanced time series analysis, the problem of task relationship modeling, dynamic adaptation and long-term effect evaluation in full-link marketing effect attribution is solved, achieving high accuracy, A method of attribution of full-link marketing effectiveness for dynamic adaptation and interpretability.
Patent Information
- Application Number
- CN202410973521.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing attribution methods for full-link marketing effectiveness have problems such as insufficient task relationship modeling, insufficient dynamic adaptability, difficulty in long-term effect evaluation, insufficient interpretability and low computational efficiency, resulting in limited attribution accuracy and practicality.
Adopting a multi-task learning framework, data preprocessing and feature engineering is performed by obtaining multi-source marketing data, initial multi-task representation learning and dynamic adaptive multi-task relationship learning, combined with quantum heuristic dynamic relationship graph algorithm, task-specific prediction and multi-task Contact attribution, and long-term effect evaluation was performed using fractional Brownian motion model, and a full-link marketing effect attribution report was finally generated.
It significantly improves attribution accuracy and prediction capabilities, enhances dynamic adaptability, accurately evaluates long-term effects, maintains high interpretability, and optimizes computing efficiency to meet the needs of modern marketing analysis.
Smart Images

Figure CN118967187A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of full-link marketing effect attribution methods, and more specifically, to a full-link marketing effect attribution method and system based on a multi-task learning framework. Background Art
[0002] With the rapid development of digital marketing, it is becoming increasingly important to accurately evaluate the effectiveness of various marketing channels and touchpoints. The attribution of full-link marketing effectiveness has always been the focus of the industry, and its development has gone through the following stages:
[0003] 1. Single-point attribution model: The earliest attribution method was single-point attribution, such as Last Click Attribution (LCA). This method is simple and intuitive, easy to implement, but it has serious flaws. It ignores the contribution of other touchpoints in the user conversion path, resulting in a serious underestimation of the value of long-term brand building marketing activities.
[0004] 2. Multi-touch attribution model: In order to solve the limitations of single-touch attribution, researchers have proposed multi-touch attribution models, such as rule-based linear attribution and data-based probabilistic models. Among them, the Markov chain-based multi-touch attribution model (MCMA) is an important breakthrough. MCMA takes into account multiple touchpoints in the user conversion path, but it assumes that user behavior satisfies the Markov property, which is often not true in complex real-world marketing environments.
[0005] 3. Attribution model based on machine learning: With the development of machine learning technology, researchers have begun to apply various advanced algorithms to solve the attribution problem. The deep neural network multi-task learning model (DNN-MTL) is one of the representatives. This type of model can learn multiple related tasks (such as click-through rate prediction, conversion rate prediction, etc.) at the same time, which improves the generalization ability and prediction accuracy of the model. However, the traditional DNN-MTL model still has obvious shortcomings in the following aspects:
[0006] a) Insufficient task relationship modeling: Existing DNN-MTL models usually use a simple parameter sharing mechanism to model the relationship between tasks, which makes it difficult to capture complex nonlinear dependencies.
[0007] b) Insufficient dynamic adaptability: The marketing environment is highly dynamic, but existing models are mostly static and unable to adapt to market changes and user behavior shifts in a timely manner.
[0008] c) Difficulty in evaluating long-term effects: Most existing models focus on the evaluation of short-term effects and find it difficult to accurately quantify the long-term impact of marketing activities.
[0009] d) Lack of explainability: As the complexity of the model increases, its decision-making process becomes increasingly opaque, which creates obstacles for practical marketing decision-making.
[0010] e) Computational efficiency needs to be improved: When processing large-scale marketing data, the computational efficiency of existing models is often unsatisfactory and cannot meet the needs of real-time decision-making.
[0011] These problems seriously restrict the accuracy and practicality of attribution of full-link marketing effects, and a new approach is urgently needed to comprehensively address these challenges. Summary of the invention
[0012] In order to solve the above technical problems, the present invention proposes a full-link marketing effect attribution method and system based on a multi-task learning framework.
[0013] The present invention provides a full-link marketing effect attribution method based on a multi-task learning framework, comprising the following steps:
[0014] Acquire multi-source marketing data; perform data preprocessing and feature engineering based on the multi-source marketing data to obtain preprocessed feature data; perform initial multi-task representation learning based on the preprocessed feature data to obtain initial task representation; perform dynamic adaptive multi-task relationship learning based on the initial task representation to obtain updated task representation; perform task-specific prediction based on the updated task representation to obtain prediction results; perform multi-touch attribution based on the prediction results to obtain touch contribution; perform long-term effect evaluation based on the touch contribution to obtain long-term effect evaluation results; generate a full-link marketing effect attribution report based on the long-term effect evaluation results.
[0015] Specifically, the data preprocessing and feature engineering include: performing data cleaning, removing duplicates and outliers, and processing missing values; performing feature extraction, including extracting time series features using wavelet transform, extracting semantic features using BERT model, and extracting visual features using pre-trained Vision Transformer; performing feature selection, including using Boruta algorithm to sort feature importance, and applying RFECV to select the optimal feature subset.
[0016] Specifically, the initial multi-task representation learning includes: using Capsule Network as the basic network structure, where the Capsule layer is defined as:
[0017]
[0018] v j =squash(s j )
[0019] Among them, sj is the input of the j capsule, i and j represent the index of the input capsule and the output capsule respectively, c ij is the coupling coefficient, is the prediction vector from capsule i to capsule v j is the output vector of capsule j, squash() is the activation function; a separate CapsuleNetwork is constructed for each task, and a task-specific loss function is used:
[0020]
[0021] Among them, L task is the task-specific loss, k is the sample index, T k is the true label of sample k, v k is the output vector of sample k, ∥v k ∥ represents vector v k The norm of , m+ and m- are the boundary parameters of positive and negative categories, and λ is the weight coefficient.
[0022] Specifically, the dynamic adaptive multi-task relationship learning includes: executing a quantum-inspired dynamic relationship graph algorithm, including:
[0023] a) Represent each task as a qubit state: |ψ i >=α i |0>+β i |1>, where |ψ i > is the quantum state of task i, α i and β i is the complex amplitude, |0> and |1> are the calculated basis states;
[0024] b) Using quantum entanglement to simulate the relationship between tasks: where |Ψ ij > is the entangled state of tasks i and j, represents tensor product;
[0025] c) Apply continuous-time quantum walk to update the task embedding: |ψ(t)>=e -iHt |ψ(0)>, where |ψ(t)> is the quantum state at time t, H is the Hamiltonian operator, t is the time parameter, i is the imaginary unit, and |ψ(0)> is the initial state;
[0026] d) Introducing non-Hermitian operators to simulate the dynamic changes of task relationships: Where ρ is the density matrix, t is time, [H,ρ] represents the commutator of H and ρ, and L is the Lindblad superoperator;
[0027] e) Use the quantum annealing algorithm to optimize the task relationship: P(s→s′)=min(1,exp(-(E(s′)-E(s)) / T)), where P(s→s′) is the probability of transitioning from state s to state s′, E(s) and E(s′) are the energies of states s and s′ respectively, and T is the system temperature.
[0028] Specifically, it also includes:
[0029] Calculate the attention weight based on quantum measurement theory: α ij =|<ψ i |M|ψ j >| 2 , where α ij is the attention weight of task i on task j, M is the quantum observable operator, <ψ i | and |ψ j > are the conjugate transpose and state vector of the quantum state of tasks i and j respectively;
[0030] Using fractal geometry to enhance the expressiveness of task embeddings: in is the embedding of task i at time t. FractalNet is a neural network based on fractal structure. ij is the attention weight;
[0031] Introducing topological entropy to constrain the complexity of task relationships: L topo =-∑ i ∑ j p ij log(p ij ), where L topo is the topological entropy loss, p ij is the probability distribution of the relationship between tasks i and j;
[0032] Using chaotic dynamical systems to simulate task adaptation processes:
[0033]
[0034] Where x, y, z are task state variables, t is time, and σ, ρ, β are control parameters.
[0035] Specifically, the task-specific prediction includes: generating prediction probabilities using quantum measurements:
[0036] P(y i |x)=| <y i |U|x>| 2
[0037] Where P(y i |x) is the output y given an input x iThe conditional probability of , U is the quantum gate operation, |x> is the quantum state representation of the input state, |y i > is the quantum state representation of the output state i; fuzzy logic rules are applied to fuse multiple prediction results:
[0038]
[0039] where μ A (x) is the membership function of the fuzzy set A, x is the input variable, and a and b are the parameters of the membership function.
[0040] Specifically, the multi-touch attribution includes: using Shapley value and score function network to calculate attribution:
[0041]
[0042] where φ i (v) is the Shapley value of feature i, v is the feature function, S is the feature subset, n is the total number of features, |S| is the cardinality of subset S;
[0043] Introducing spatiotemporal attention to capture the temporal and spatial dependencies of contacts:
[0044] α t,s =softmax(W α tanh(W t ·h t +W s ·h s +b α ))
[0045] where α t,s is the spatiotemporal attention weight, W α ,W t ,W s is the weight matrix, b α is the bias term, h t and h s are the hidden states of temporal and spatial features respectively, tanh is the hyperbolic tangent function, and softmax is the normalization function.
[0046] Specifically, the long-term effect evaluation includes: using the fractional Brownian motion model to predict the long-term effect:
[0047]
[0048] Where X t is a random process at time t, μ(X t ,t) is the drift term, σ(X t ,t) is the fluctuation term, It is a fractional Brownian motion with Hurst index H;
[0049] Apply wavelet decomposition and empirical mode decomposition to deal with non-stationary time series:
[0050]
[0051] Where x(t) is the original signal, c i (t) is the i-th intrinsic mode function, r n (t) is the residual and t is the time variable.
[0052] Specifically, the generation of the full-link marketing effect attribution report includes: using the hierarchical analysis method and the fuzzy comprehensive evaluation method to integrate multiple indicators:
[0053] W=(w1,w2,…,w n ) T
[0054] R=(r ij ) m×n
[0055] B=W·R
[0056] Where W is the weight vector, w i is the weight of the ith indicator, R is the relationship matrix, r ij is the membership of the i-th evaluation object to the j-th indicator, B is the final evaluation result, m is the number of evaluation objects, and n is the number of indicators;
[0057] Explain model decisions using SHAP values:
[0058]
[0059] where φ i is the SHAP value of feature i, F is the feature set, |F| is the total number of features, S is the subset that does not contain feature i, and f x is the prediction function.
[0060] The full-link marketing effect attribution system of the multi-task learning framework for executing the method includes: a data acquisition module for acquiring the multi-source marketing data; a preprocessing module for performing the data preprocessing and feature engineering; an initial representation learning module for performing the initial multi-task representation learning; a dynamic relationship learning module for performing the dynamic adaptive multi-task relationship learning; a task prediction module for performing the task-specific prediction; an attribution module for performing the multi-touch attribution; a long-term effect evaluation module for performing the long-term effect evaluation; and a report generation module for generating the full-link marketing effect attribution report.
[0061] The beneficial effects of the present invention are mainly reflected in the following aspects:
[0062] 1. High accuracy: By introducing the quantum-inspired dynamic relationship graph (QIDRG) algorithm, the present invention can accurately capture the complex nonlinear relationships between tasks and significantly improve the accuracy of attribution. Test results show that the attribution accuracy (AA) of the present invention reaches 0.92, far exceeding other methods.
[0063] 2. Powerful prediction capability: The combination of multi-task learning framework and advanced time series analysis technology enables the present invention to perform well in user behavior prediction. The AUC value reaches 0.88, which is better than various existing methods.
[0064] 3. Excellent dynamic adaptability: The non-Hermitian dynamic adaptation mechanism in the QIDRG algorithm enables the model to respond quickly to market changes and maintain high efficiency.
[0065] 4. Long-term effect evaluation: By introducing the fractional Brownian motion model and non-stationary time series analysis, the present invention can accurately evaluate the long-term impact of marketing activities. The R2 value of the long-term effect capture capability (LEC) reaches 0.79, far exceeding other methods.
[0066] 5. Good interpretability: Although the model complexity increases, the present invention maintains a high interpretability (MI value is 0.85) by introducing technologies such as SHAP value and fuzzy logic fusion.
[0067] 6. Improved computing efficiency: The application of quantum heuristic algorithms not only improves model performance, but also optimizes computing efficiency. Compared with the traditional DNN-MTL model, the running time of the present invention is reduced by 25%.
[0068] 7. Comprehensiveness: The present invention provides a full-process solution from data preprocessing to final report generation, which can meet various needs of modern marketing analysis.
[0069] In summary, the present invention breaks through the limitations of the existing technology and provides a highly accurate, dynamically adaptable, explainable and efficient full-link marketing effect attribution method. It can not only help enterprises more accurately evaluate the contribution of various marketing channels and touchpoints, but also provide reliable data support for marketing decisions, helping enterprises to formulate more effective marketing strategies in a highly competitive market environment. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1 This is a flow chart of the full-link marketing effect attribution method of the multi-task learning framework of the present invention.
[0071] Figure 2 This is a schematic diagram of the full-link marketing effect attribution system framework of the multi-task learning framework of the present invention. DETAILED DESCRIPTION
[0072] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.
[0073] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0074] The technical solution of the present invention is described clearly and completely below in conjunction with specific implementation methods. The specific implementation method of the present invention is to address the full-link marketing effect attribution problem of a large e-commerce platform. The platform operates websites, mobile applications, and physical stores at the same time, attracting users through multiple marketing channels (such as search engine advertising, social media marketing, email marketing, offline activities, etc.). The goal of the platform is to accurately evaluate the impact of various marketing channels and touchpoints on user behavior (such as clicks, browsing, and purchases) in order to optimize the allocation of marketing resources.
[0075] Please refer to Figure 1 , the method proposed by the present invention comprises the following steps:
[0076] 1. Acquire multi-source marketing data. In this step, the system collects data from various marketing channels and user interaction points through the data acquisition module. Specifically, it includes:
[0077] a) User behavior data: click stream data of websites and applications, browsing history, search history, shopping cart behavior, etc.
[0078] b) Marketing activity data: advertising placement records, content push records, coupon distribution records, etc. on each channel.
[0079] c) Contextual data: user access time, geographic location, device information used, etc.
[0080] d) Conversion data: purchase records, registration information, app download records, etc.
[0081] 2. Data preprocessing and feature engineering. The preprocessing module cleans and extracts features from the collected raw data, including:
[0082] 2.1 Data Cleaning
[0083] a) Remove duplicate records: Use hashing algorithms to quickly identify and remove duplicate data entries.
[0084] b) Handling outliers: Use the 3σ principle or interquartile range (IQR) method to identify and handle outliers.
[0085] c) Handling missing values: For numerical features, use multiple imputation to fill missing values; for categorical features, introduce an "unknown" category.
[0086] 2.2 Feature Extraction
[0087] a) Time series feature extraction: Use wavelet transform to extract the time series features of user behavior. For example, perform wavelet decomposition on the user's daily pageview sequence to extract features of different scales.
[0088] W ψ f(u,s)= <f,ψ u,s > =∫f(t)ψ u,s (t)dt
[0089] Among them, f(t) is the original time series, ψ u,s (t) is the wavelet function, u is the translation parameter, and s is the scale parameter.
[0090] b) Text feature extraction: Use the BERT model to extract the semantic features of marketing content. For example, encode advertising copy or product description. BERT encoding process:
[0091] E=BERT(T)
[0092] Among them, T is the input text sequence and E is the obtained feature representation.
[0093] c) Visual feature extraction: Use the pre-trained VisionTransformer to extract the visual features of the image advertisement. VisionTransformer encoding:
[0094] V=ViT(I)
[0095] Among them, I is the input image and V is the obtained visual feature representation.
[0096] 2.3 Feature Selection
[0097] a) Use the Boruta algorithm to sort the feature importance. The Boruta algorithm iteratively selects important features by creating shadow features and comparing them with the original features.
[0098] b) Apply Recursive Feature Elimination with Cross-Validation (RFECV) to select the optimal feature subset. RFECV finds the best feature subset by recursively removing the least important features and evaluating the performance using cross-validation.
[0099] 3. Initial multi-task representation learning. The initial representation learning module uses CapsuleNetwork as the basic network structure and simultaneously models multiple related marketing tasks, such as click-through rate prediction, conversion rate prediction, and customer value assessment.
[0100] 3.1CapsuleNetwork Structure
[0101]
[0102] v j =squash(s j )
[0103] Among them, s j is the input of the j-plug, i and j represent the index of the input capsule and the output capsule respectively, c ij is the coupling coefficient, initialized to a uniform distribution, is the prediction vector from capsule i to capsule j, v j is the output vector of j, and quash is the activation function, defined as:
[0104] 3.2 Task-specific loss functions,For each task, we use the following loss function:
[0105]
[0106] Among them, L task is the task-specific loss, k is the sample index, T k is the true label of sample k (e.g., whether it is clicked or converted), v k is the output vector of sample k, ∥v k ∥ represents vector v k The norm of m + and m - is the boundary parameter of the positive and negative categories, usually set to m + =0.9,m - = 0.1, λ is the weight coefficient used to balance positive and negative samples, usually set to 0.5. This loss function design can handle multiple binary classification tasks at the same time, such as click prediction and conversion prediction.
[0107] 4. Dynamic Adaptive Multi-Task Relationship Learning,The dynamic relationship learning module executes the quantum-inspired dynamic relationship graph (QIDRG) algorithm to capture the dynamic relationships between tasks. The core idea of this module is to model the task relationship as a quantum system and use the parallelism and non-locality of quantum computing to handle complex task interactions.
[0108] 4.1 Quantum bit state representation, each task is represented as a quantum bit state:
[0109] a) Represent each task as a qubit state: |ψ i >=α i |0>+β i |1>, where |ψ i > is the quantum state of task i, α i and β i is a complex amplitude, satisfying |α i | 2 +|β i | 2 =1, |0> and |1> are the calculated basis states.
[0110] In practical applications, we can map certain features of the task (such as the current performance indicator of the task) to α i and β i For example, for the click-through rate prediction task, we can map the current AUC score to α i superior.
[0111] 4.2 Quantum entanglement task relationship construction, using quantum entanglement to simulate the relationship between tasks:
[0112] where |Ψ ij > is the entangled state of tasks i and j, represents tensor product;
[0113] This representation allows us to capture non-local relationships between tasks. For example, there may be a strong correlation between click-through rate prediction and conversion rate prediction tasks, and through quantum entanglement, we can model this relationship at a deeper level.
[0114] 4.3 Continuous-time quantum walk update, the continuous-time quantum walk update task embedding is applied: |ψ(t)>=e -iHt |ψ(0)>, where |ψ(t)> is the quantum state at time t, H is the Hamiltonian operator, t is the time parameter, i is the imaginary unit, and |ψ(0)> is the initial state;
[0115] In practice, we can design H as a Laplacian matrix that reflects the task relationship graph. In this way, the quantum walk process will naturally propagate information between related tasks.
[0116] 4.4 Non-Hermitian dynamic adaptation, introducing non-Hermitian operators to simulate the dynamic changes of task relationships: where ρ is the density matrix, t is time, [H,ρ] represents the commutator of H and ρ, and L is the Lindblad superoperator describing the interaction between the system and the environment.
[0117] This equation allows us to simulate open quantum systems that are closer to the dynamics of real-world marketing environments. For example, we can use L to model the impact of market trends on task relationships.
[0118] 4.5 Quantum annealing optimization, using quantum annealing algorithm to optimize task relationships:
[0119] P(s→s′)=min(1,exp(-(E(s′)-E(s)) / T)), where P(s→s′) is the probability of transitioning from state s to state s′, E(s) and E(s′) are the energies of states s and s′ respectively, and T is the system temperature, which gradually decreases over time. The quantum annealing process helps the system escape from the local optimum and find a better task relationship structure. In practice, we can design the energy function E as a measure that reflects task performance, such as the overall loss of multi-task learning.
[0120] 5. Adaptive relational attention mechanism, calculating attention weights based on quantum measurement theory:
[0121] Calculate the attention weight based on quantum measurement theory: α ij =|<ψ i |M|ψ j >| 2 , where α ij is the attention weight of task i on task j, M is the quantum observable operator, <ψ i | and |ψ j > are the conjugate transpose and state vector of the quantum state of tasks i and j respectively;
[0122] This attention mechanism allows the model to dynamically adjust the importance relationship between tasks. For example, in some marketing scenarios, click-through rate prediction may be more important than conversion rate prediction. ij will increase accordingly.
[0123] 6. Fractal dimension task embedding, using fractal geometry to enhance the expressive power of task embedding: in is the embedding of task i at time t. FractalNet is a neural network based on fractal structure. ijis the attention weight; the structure of FractalNet can be designed as a self-similar multi-layer network, where each layer is a smaller version of the previous layer. This structure can capture task features at different scales and is particularly suitable for processing multi-scale patterns in marketing data.
[0124] 7. Introducing topological entropy to constrain the complexity of task relationships: L topo =-∑ i ∑ j p ij log(p ij ), where L topo is the topological entropy loss, p ij is the probability distribution of the relationship between tasks i and j;
[0125] Topological entropy regularization encourages the model to learn a more structured task relationship graph and avoid overly complex or chaotic relationships. In practice, we can add this loss term to the overall optimization objective.
[0126] 8. Use chaotic dynamical systems to simulate task adaptation processes:
[0127]
[0128] Where x, y, z are task state variables, t is time, σ, ρ, β are control parameters, usually σ=10,
[0129] This Lorenz system models the nonlinear dynamics of task adaptation. We can map x, y, z to different task performance metrics, such as precision, recall, and F1 score. By observing the evolution of this system, we can capture the complex interactions between tasks.
[0130] 9. Task-specific prediction,The task prediction module makes specific predictions based on the updated,task representation.
[0131] 9.1 Quantum Measurements Generate Prediction Probabilities
[0132] Generate predicted probabilities using quantum measurements:
[0133] P(y i |x)=| <y i |U|x>| 2
[0134] Where P(y i |x) is the output y given an input x i The conditional probability of , U is the quantum gate operation, which can be designed as a unitary matrix reflecting the characteristics of the task, |x> is the quantum state representation of the input state, |y i> is the quantum state representation of the output state i;
[0135] This quantum probability model allows us to consider quantum superposition and interference effects in the prediction process, and can capture complex patterns that are difficult to capture with traditional probability models. For example, when predicting whether a user will click on an ad, we can encode user features and ad features into quantum states, and then simulate the user-ad interaction process by designing an appropriate U.
[0136] 9.2 Apply fuzzy logic rules to fuse multiple prediction results:
[0137]
[0138] where μ A (x) is the membership function of the fuzzy set A, x is the input variable (such as the prediction score), and a and b are the parameters of the membership function, which determine the shape of the function. Fuzzy logic fusion allows us to handle uncertainty and ambiguity more flexibly. For example, we can define fuzzy sets such as "high click propensity", "medium click propensity", and "low click propensity", and then calculate the degree to which the user belongs to each set based on the prediction results of multiple models.
[0139] 10. Multi-touch Attribution The attribution module uses advanced attribution technology to evaluate the contribution of each marketing touchpoint.
[0140] 10.1 Shapley value and score function network attribution, use Shapley value and score function network to calculate attribution:
[0141]
[0142] where φ i (v) is the Shapley value of feature i, v is the feature function, which represents the prediction effect of each feature subset, S is the feature subset that does not contain feature i, n is the total number of features, and |S| is the cardinality of subset S; the Shapley value provides a fair way to distribute the contribution of each touchpoint. For example, for a successful conversion, we can calculate the Shapley value of each ad impression, email push, or social media interaction to determine their respective contributions.
[0143] 10.2 Introducing spatiotemporal attention to capture the temporal and spatial dependencies of contacts:
[0144] α t,s =softmax(W α tanh(W t ·h t +W s ·h s +b α ))
[0145] where α t,s is the spatio-temporal attention weight, W α , W t , W s are weight matrices, b α is the bias term, h t and h s are the hidden states of the temporal and spatial features respectively, tanh is the hyperbolic tangent function, and softmax is the normalization function.
[0146] The spatio-temporal attention mechanism allows the model to focus on the importance of touchpoints at different time points and different channels. For example, the model may find that the influence of search engine advertising increases significantly in the short time after a user searches for a product.
[0147] 11. Long-term effect evaluation. The long-term effect evaluation module uses advanced time series analysis techniques to evaluate the long-term impact of marketing campaigns.
[0148] 11.1 Predicting long-term effects using the fractional Brownian motion model:
[0149] Predicting long-term effects using the fractional Brownian motion model:
[0150]
[0151] where X t is the stochastic process at time t (such as cumulative sales),
[0152] μ(X t , t) is the drift term, which can be designed as a function reflecting the market trend · σ(X t , t) is the volatility term, reflecting the uncertainty of the market, is the fractional Brownian motion with Hurst exponent H, H is the Hurst exponent, and it degenerates to the standard Brownian motion when H = 0.5.
[0153] 11.2 Applying wavelet decomposition and empirical mode decomposition to process non-stationary time series:
[0154]
[0155] where x(t) is the original signal (such as daily sales volume), c i (t) is the i-th intrinsic mode function (IMF), r n (t) is the residue, and t is the time variable.
[0156] Wavelet decomposition and EMD can help us separate market fluctuations and trends at different scales. For example, we can identify seasonal fluctuations, short-term promotion effects, and the impact of long-term brand building.
[0157] 12. Generate a full-link marketing effectiveness attribution report. The report generation module integrates the results of all previous steps to generate a comprehensive marketing effectiveness attribution report.
[0158] 12.1 Multi-criteria decision fusion, using the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method to integrate multiple indicators:
[0159] W=(w1,w2,…,w n ) T
[0160] R=(r ij ) m×n
[0161] B=W·R
[0162] Where W is the weight vector, w i is the weight of the ith indicator, R is the relationship matrix, r ij is the membership of the i-th evaluation object to the j-th indicator, B is the final evaluation result, m is the number of evaluation objects, and n is the number of indicators; this method allows us to comprehensively consider multiple indicators and give a comprehensive evaluation of marketing channels. For example, we can consider the click-through rate, conversion rate, and customer lifetime value of a channel at the same time to obtain a comprehensive performance score for the channel.
[0163] 12.2 Model Decision Explanation Use SHAP values to explain model decisions:
[0164]
[0165] where φ i is the SHAP value of feature i, F is the feature set, |F| is the total number of features, S is the subset that does not contain feature i, and f x is the prediction function.
[0166] Please refer to Figure 2 The multi-task learning framework full-link marketing effect attribution system of the present invention includes the following modules:
[0167] Data acquisition module 1: responsible for collecting raw data from various data sources.
[0168] Preprocessing module 2: Perform data cleaning, feature extraction and selection.
[0169] Initial Representation Learning Module 3: Preliminary multi-task learning using CapsuleNetwork.
[0170] Dynamic relationship learning module 4: Implement the QIDRG algorithm to capture the dynamic relationship between tasks.
[0171] Task prediction module 5: Perform specific prediction tasks based on the learned representation.
[0172] Attribution Module 6: Calculate the contribution of each marketing touchpoint.
[0173] Long-term Effect Evaluation Module 7: Evaluate the long-term impact of marketing activities.
[0174] Report Generation Module 8: Integrate all results to generate a comprehensive attribution report.
[0175] Each module is an independent functional unit, but there is a close connection between them in terms of data flow and control flow. For example, the output of the dynamic relationship learning module affects the behavior of the task prediction module, and the results of the long-term effect evaluation module are used by the attribution module to adjust the evaluation of short-term effects.
[0176] In order to implement the method of the present invention, the following hardware configuration is recommended: the CPU uses an Intel Xeon Platinum 8380 processor (40 cores, 80 threads); the GPU uses 4xNVIDIA A100 Tensor Core GPUs (each with 80GB of video memory); the RAM uses 1TB DDR4-3200 ECC memory; the storage uses a 20TB NVMe SSD array; the network uses 100Gbps Ethernet; this high-performance configuration is capable of processing large-scale marketing data and supports the training and reasoning of complex quantum-inspired algorithms and multi-task learning models.
[0177] In terms of software, the operating system uses Ubuntu 20.04 LTS, which is optimized for high-performance computing. The distributed computing framework uses Apache Spark 3.1.2 for large-scale data processing. The deep learning framework uses PyTorch 1.9.0 to implement Capsule Network and other deep learning models. The quantum computing simulator uses Qiskit 0.29.0 to simulate the quantum computing process. The time series analysis library uses statsmodels 0.12.2 for advanced time series analysis. The multi-task learning framework full-link marketing effect attribution method of the present invention is achieved by combining software and hardware.
[0178] This invention innovatively combines quantum computing concepts, multi-task learning, and advanced time series analysis to provide a comprehensive and powerful solution for full-link marketing effect attribution. It not only provides more accurate attribution results, but also provides in-depth insights for the formulation of marketing strategies, helping companies make more informed decisions in a complex and changing market environment.
[0179] In order to verify the superiority of the present invention, we conducted a series of comparative experiments. The following is a detailed comparison of Example 1 of the present invention and three comparative examples:
[0180] Embodiment: The multi-task learning framework full-link marketing effect attribution method (MTML-FAMA) of the present invention
[0181] Comparative Example 1: Traditional Last Click Attribution Model (LCA)
[0182] Comparative Example 2: Multi-touch Attribution Model (MCMA) Based on Markov Chain
[0183] Comparative Example 3: Deep Neural Network Multi-Task Learning Model without Quantum-Inspired Dynamic Relational Graph (DNN-MTL)
[0184] Test indicators and their testing standards and methods:
[0185] 1. Attribution Accuracy (AA):
[0186] Detection criteria: Use a simulated dataset with known true attribution values.
[0187] Detection method: Calculate the mean absolute error (MAE) between the predicted attribution value and the true attribution value.
[0188] 2. Predictive Performance (PP):
[0189] Detection criteria: The AUC values of click-through rate (CTR) and conversion rate (CVR) predictions are evaluated on the test set.
[0190] Detection method: AUC is calculated using the roc_auc_score function in the sklearn library.
[0191] 3. Computational Efficiency (CE):
[0192] Detection criterion: the time required to process a dataset of the same size.
[0193] Detection method: Record the total running time of model training and inference.
[0194] 4. Model Interpretability (MI):
[0195] Testing criteria: Consistency and stability of SHAP values used.
[0196] Detection method: Calculate the average coefficient of variation of SHAP values from multiple runs.
[0197] 5. Long-term Effect Capture (LEC):
[0198] Testing criteria: accuracy of predicting the cumulative conversion rate in the next 30 days.
[0199] Detection method: Calculate the R between the predicted value and the actual value 2 value.
[0200] The test results are shown in the following table:
[0201] index The present invention Comparative Example 1 Comparative Example 2 Comparative Example 3 AA 0.92 0.65 0.78 0.85 PP 0.88 0.72 0.79 0.84 CE 45 minutes 5 minutes 30 minutes 60 minutes MI 0.85 0.95 0.8 0.75 LEC 0.79 0.3 0.55 0.68
[0202] The MTML-FAMA method of the present invention is the best embodiment, and the test results are analyzed and explained as follows:
[0203] 1. Attribution Accuracy (AA): The AA value of MTML-FAMA is significantly higher than that of other methods, reaching 0.92. This shows that the present invention can more accurately identify the contribution of each marketing touchpoint. This is mainly due to the quantum-inspired dynamic relationship graph (QIDRG) algorithm, which can capture the complex nonlinear relationship between touchpoints.
[0204] 2. Prediction Performance (PP): MTML-FAMA also outperforms other methods in prediction performance, with an AUC of 0.88. This shows that the multi-task learning framework does improve the generalization ability of the model and can better predict user behavior.
[0205] 3. Computational efficiency (CE): The running time of MTML-FAMA is 45 minutes, which is slower than the simple LCA model, but acceptable considering its complexity and performance improvement. Moreover, it is faster than DNN-MTL without quantum heuristics, indicating that the QIDRG algorithm optimizes computational efficiency while improving performance.
[0206] 4. Model interpretability (MI): MTML-FAMA is second only to LCA (0.95) in interpretability (0.85), but significantly better than other complex models. This shows that the present invention improves the model complexity and performance while maintaining the interpretability of the model by introducing technologies such as SHAP value and fuzzy logic fusion.
[0207] 5. Long-term effect capture ability (LEC): MTML-FAMA far exceeds other methods in terms of long-term effect capture ability, reaching an R of 0.79 2 This is due to the application of fractional Brownian motion model and non-stationary time series analysis technology, which enables the present invention to better evaluate the long-term impact of marketing activities.
[0208] The MTML-FAMA method of the present invention is significantly superior to traditional methods and other advanced machine learning methods in attribution accuracy, prediction performance and long-term effect capture ability. Although slightly inferior to the simple last-click attribution model in computational efficiency, this computational cost is acceptable considering the significant performance improvement. Moreover, MTML-FAMA maintains a high model interpretability, which is crucial for actual marketing decisions.
[0209] These results fully demonstrate the innovation and superiority of the present invention in the field of full-link marketing effect attribution. By combining quantum heuristic algorithms, multi-task learning and advanced time series analysis, the present invention not only improves the accuracy of attribution, but also enhances the model's predictive ability and ability to capture long-term effects while maintaining good interpretability. This makes MTML-FAMA a comprehensive, powerful and practical marketing effect attribution solution.
[0210] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. The full-link marketing effect attribution method based on the multi-task learning framework is characterized by: The following steps are involved: Acquire multi-source marketing data; perform data preprocessing and feature engineering based on the multi-source marketing data to obtain preprocessed feature data; Based on the preprocessed feature data, performing initial multi-task representation learning to obtain initial task representation; Based on the initial task representation, dynamic adaptive multi-task relationship learning is performed to obtain an updated task representation; based on the updated task representation, task-specific prediction is performed to obtain prediction results; based on the prediction results, multi-touch attribution is performed to obtain touch contribution; based on the touch contribution, long-term effect evaluation is performed to obtain long-term effect evaluation results; based on the long-term effect evaluation results, a full-link marketing effect attribution report is generated.
2. The method according to claim 1, characterized in that The data preprocessing and feature engineering include: performing data cleaning, removing duplicates and outliers, and processing missing values; performing feature extraction, including extracting time series features using wavelet transform, extracting semantic features using BERT model, and extracting visual features using pre-trained Vision Transformer; performing feature selection, including ranking feature importance using Boruta algorithm, and applying RFECV to select the optimal feature subset.
3. The method according to claim 1, characterized in that The initial multi-task representation learning includes: using Capsule Network as the basic network structure, where the Capsule layer is defined as: v j =squash(s j ) Among them, s j is the input of the j capsule, i and j represent the index of the input capsule and the output capsule respectively, c ij is the coupling coefficient, is the prediction vector from capsule i to capsule v j is the output vector of capsule j, squash() is the activation function; a separate Capsule Network is constructed for each task, and a task-specific loss function is used: Among them, L task is the task-specific loss, k is the sample index, T k is the true label of sample k, v k is the output vector of sample k, ∥v k ∥ represents vector v k The norm of , m+ and m- are the boundary parameters of positive and negative categories, and λ is the weight coefficient.
4. The method according to claim 1, characterized in that The dynamic adaptive multi-task relational learning includes: executing a quantum-inspired dynamic relational graph algorithm, including: a) Represent each task as a qubit state: |ψ i >=α i |0>+β i |1>, where |ψ i > is the quantum state of task i, α i and β i is the complex amplitude, |0> and |1> are the calculated basis states; b) Using quantum entanglement to simulate the relationship between tasks: where |Ψ ij > is the entangled state of tasks i and j, represents tensor product; c) Apply continuous-time quantum walk to update the task embedding: |ψ(t)>=e -iHt |ψ(0)>, where |ψ(t)> is the quantum state at time t, H is the Hamiltonian operator, t is the time parameter, i is the imaginary unit, and |ψ(0)> is the initial state; d) Introducing non-Hermitian operators to simulate the dynamic changes of task relationships: Where ρ is the density matrix, t is time, [H,ρ] represents the commutator of H and ρ, and L is the Lindblad superoperator; e) Use the quantum annealing algorithm to optimize the task relationship: P(s→s′)=min1,exp(-(E(s′)-E(s)) / T)), where P(s→s′) is the probability of transitioning from state s to state s′, E(s) and E(s′) are the energies of states s and s′ respectively, and T is the system temperature.
5. The method according to claim 4, characterized in that Also includes: Calculate the attention weight based on quantum measurement theory: α ij =|<ψ i |M|ψ j >| 2 , where α ij is the attention weight of task i on task j, M is the quantum observable operator, <ψ i | and |ψ j > are the conjugate transpose and state vector of the quantum state of tasks i and j respectively; Using fractal geometry to enhance the expressiveness of task embeddings: in is the embedding of task i at time t. FractalNet is a neural network based on fractal structure. ij is the attention weight; Introducing topological entropy to constrain the complexity of task relationships: L topo =-∑ i ∑ j p ij log(p ij ), where L topo is the topological entropy loss, p ij is the probability distribution of the relationship between tasks i and j; Using chaotic dynamical systems to simulate task adaptation processes: Where x, y, z are task state variables, t is time, and σ, ρ, β are control parameters.
6. The method according to claim 1, characterized in that The task-specific predictions include: using quantum measurements to generate prediction probabilities: P(y i |x)=|<y i |U|x>| 2 Where P(y i |x) is the output y given an input x i The conditional probability of , U is the quantum gate operation, |x> is the quantum state representation of the input state, |y i > is the quantum state representation of the output state i; fuzzy logic rules are applied to fuse multiple prediction results: where μ A (x) is the membership function of the fuzzy set A, x is the input variable, and a and b are the parameters of the membership function.
7. The method according to claim 1, characterized in that The multi-touch attribution includes: using Shapley values and score function network to calculate attribution: where φ i (v) is the Shapley value of feature i, v is the feature function, S is the feature subset, n is the total number of features, |S| is the cardinality of subset S; Introducing spatiotemporal attention to capture the temporal and spatial dependencies of contacts: α t,s =softmax(W α ·tanh(W t ·h t +W s ·h s +b α )) where α t,s is the spatiotemporal attention weight, W α ,W t ,W s is the weight matrix, b α is the bias term, h t and h s are the hidden states of temporal and spatial features respectively, tanh is the hyperbolic tangent function, and softmax is the normalization function.
8. The method according to claim 1, characterized in that The long-term effect evaluation includes: predicting the long-term effect using a fractional Brownian motion model: Where X t is a random process at time t, μ(X t ,t) is the drift term, σ(X t ,t) is the fluctuation term, It is a fractional Brownian motion with Hurst index H; Apply wavelet decomposition and empirical mode decomposition to deal with non-stationary time series: Where x(t) is the original signal, c i (t) is the i-th intrinsic mode function, r n (t) is the residual and t is the time variable.
9. The method according to claim 1, characterized in that: The generation of the full-link marketing effect attribution report includes: using the hierarchical analysis method and the fuzzy comprehensive evaluation method to integrate multiple indicators: In=(in1,in2,…,in n ) T R=(r ij ) m×n B=W·R Where W is the weight vector, w i is the weight of the i-th indicator, R is the relationship matrix, r ij is the membership of the i-th evaluation object to the j-th indicator, B is the final evaluation result, m is the number of evaluation objects, and n is the number of indicators; Explain model decisions using SHAP values: where φ i is the SHAP value of feature i, F is the feature set, |F| is the total number of features, S is the subset that does not contain feature i, and f x is the prediction function.
10. A full-link marketing effect attribution system of a multi-task learning framework for executing the method as claimed in any one of claims 1 to 9, characterized in that: include: A data acquisition module, used to acquire the multi-source marketing data; A preprocessing module, used to perform the data preprocessing and feature engineering; An initial representation learning module, used to perform the initial multi-task representation learning; A dynamic relationship learning module, used to perform the dynamic adaptive multi-task relationship learning; A task prediction module, for performing said task specific prediction; An attribution module, used to perform the multi-touch attribution; A long-term effect evaluation module is used to perform the long-term effect evaluation; a report generation module is used to generate the full-link marketing effect attribution report.
Citation Information
Patent Citations
Re-marketing advertisement effect estimation method and system based on multi-task learning
CN114820053A
Task recommendation method and device
CN115423505A
All-channel social media marketing investment analysis method and system based on statistical attribution
CN117495409A
Marketing channel attribution
US20160098735A1
Cited By
Marketing management system based on big data analysis
CN121684989A
Advertisement putting ROI dynamic optimization method and system based on real-time data driving
CN121981778A
