Data attribution processing method and system based on multi-model fusion and electronic device

Abstract

The application provides a data attribution processing method and system based on multi-model fusion and an electronic device, and relates to the technical field of data processing. The method acquires multi-source touch data and constructs an interaction sequence; a causal inference model is used to strip off confusing factors to obtain a causal benchmark parameter, and based on the parameter, the marginal contribution weight of each data source is calculated by sampling approximation; the marginal contribution weight is taken as a global weight of a batch processing layer, a sliding window of a flow processing layer is initialized, and when a conversion event is detected, the interaction sequence is updated incrementally for attribution, to obtain updated marginal contribution weight; in combination with the updated weight, a multi-objective evolutionary algorithm accelerated by a proxy model is used to solve resource allocation parameters. The method solves the technical problems of low attribution accuracy and high calculation delay existing in the prior art data attribution processing method, realizes the effect of giving consideration to attribution accuracy and calculation real-time performance, and dynamically and efficiently solves resource allocation parameters.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a data attribution processing method, system, and electronic device based on multi-model fusion. Background Technology

[0002] In the field of multi-source data attribution processing, such as in digital marketing attribution scenarios, traditional methods generally employ attribution methods based on preset rules, such as last-click attribution and linear attribution. These methods ignore the interactive influence between different data sources and the confounding factors in the user's decision-making process, resulting in low attribution accuracy and an inability to reflect the true contribution of each data source. Some improved solutions introduce game theory models for contribution allocation, but their computational complexity increases exponentially with the number of data sources, becoming slow and impractical for large-scale data. Furthermore, existing systems often employ offline batch processing architectures, resulting in long attribution weight update cycles and an inability to respond in real-time to dynamic changes in the data stream. When solving resource allocation optimization problems, traditional algorithms suffer from high latency in solving nonlinear, multi-constraint objective functions, making it difficult to meet the timeliness requirements of dynamic adjustments. These problems mean that existing technologies cannot simultaneously achieve attribution accuracy and computational real-time performance under high-dimensional nonlinear objectives, hindering the dynamic and efficient solution of resource allocation parameters. In summary, existing data attribution processing methods suffer from low attribution accuracy and high computational latency. Summary of the Invention

[0003] The purpose of this invention is to provide a data attribution processing method, system, and electronic device based on multi-model fusion, so as to solve the technical problems of low attribution accuracy and high computational latency in existing data attribution processing methods.

[0004] In a first aspect, embodiments of the present invention provide a data attribution processing method based on multi-model fusion. The method includes: acquiring multi-source touchpoint data of a target object under multiple data sources, and constructing an interaction sequence of the target object; processing the interaction sequence based on a causal inference-based confounding factor stripping model to obtain causal baseline parameters, and approximately calculating the marginal contribution weight of each of the aforementioned data sources by sampling based on the aforementioned causal baseline parameters; using the marginal contribution weight of each of the aforementioned data sources as the global weight of the batch processing layer, initializing a sliding window of the stream processing layer, and in response to the detection of a conversion event, incrementally updating the interaction sequence within the aforementioned sliding window to obtain the updated marginal contribution weight; and solving resource allocation parameters based on a multi-objective evolutionary algorithm accelerated by a surrogate model, combined with the updated marginal contribution weight.

[0005] In some optional implementations, the aforementioned confounding factor stripping model based on causal inference processes the aforementioned interaction sequence to obtain causal baseline parameters, including: processing the aforementioned interaction sequence through a first prediction model to obtain the exposure tendency residual of the aforementioned data source; processing the aforementioned interaction sequence through a second prediction model to obtain the baseline conversion result residual; and performing regression analysis based on the aforementioned exposure tendency residual and the aforementioned baseline conversion result residual to obtain the aforementioned causal baseline parameters.

[0006] In some optional implementations, the marginal contribution weight of each of the aforementioned data sources is approximately calculated by sampling based on the aforementioned causal benchmark parameters, including: performing hierarchical sampling according to the historical contribution of each of the aforementioned data sources to generate multiple data source arrangement orders; for each of the aforementioned data source arrangement orders, sequentially traversing each of the aforementioned data sources, calculating the first predicted transformation value after the current data source is added to the previous data source set, and the second predicted transformation value before the current data source is added, and taking the difference between the first predicted transformation value and the second predicted transformation value as the marginal contribution of the current data source under the aforementioned data source arrangement order; repeating the steps of generating multiple data source arrangement orders and calculating marginal contributions until the change in the aforementioned marginal contribution is less than a preset convergence threshold in a consecutive preset number of iterations; averaging the aforementioned marginal contributions under all the aforementioned data source arrangement orders to obtain the aforementioned marginal contribution weight of each of the aforementioned data sources; wherein the aforementioned first predicted transformation value and the aforementioned second predicted transformation value are calculated based on the aforementioned causal benchmark parameters and a preset transformation prediction function.

[0007] In some optional implementations, the marginal contribution weights of each of the aforementioned data sources are used as global weights for the batch processing layer to initialize a sliding window for the stream processing layer. In response to the detection of a conversion event, incremental attribution updates are performed on the interaction sequence within the sliding window to obtain updated marginal contribution weights. This includes: calculating and storing global precise attribution weights by the batch processing layer in a first cycle; receiving a real-time event stream by the stream processing layer and initializing the sliding window based on the global precise attribution weights; and in response to the detection of a conversion event in the real-time event stream, performing incremental attribution updates on the interaction sequence within the sliding window based on a preset approximate weighting algorithm to obtain the updated marginal contribution weights.

[0008] In some optional implementations, the multi-objective evolutionary algorithm accelerated by the surrogate model, combined with the updated marginal contribution weights to solve for resource allocation parameters, includes: initializing a population containing multiple individuals, encoding each individual as a candidate resource allocation vector for each of the aforementioned data sources; during the iteration of the multi-objective evolutionary algorithm, using a Gaussian process surrogate model to evaluate the fitness of each individual in the population, the fitness being calculated based on the updated marginal contribution weights and preset resource constraints; calibrating the Gaussian process surrogate model using a real objective function model at a preset calibration period, and updating the hyperparameters of the Gaussian process surrogate model; generating a progeny population based on crossover and mutation operations, and using an elite retention strategy to preserve the non-dominated solutions of the Pareto front; when a preset convergence condition is met, selecting a target individual from the Pareto front, and using the candidate resource allocation vector corresponding to the target individual as the resource allocation parameter; wherein, the real objective function model is used to calculate the expected total conversion and resource utilization efficiency under a given candidate resource allocation vector.

[0009] In some optional implementations, the above-mentioned acquisition of multi-source touchpoint data of the target object under multiple data sources and construction of the interaction sequence of the target object includes: acquiring multiple identifier type data in the multi-source touchpoint data; constructing a probability factor graph containing identifier nodes and observation association edges, wherein the identifier nodes correspond to the identifier type data and the observation association edges correspond to the co-occurrence relationship of identifiers in the same event; performing iterative calculations in the probability factor graph through a message propagation algorithm to obtain the edge probability of each identifier node; fusing cross-source identifiers based on the edge probabilities to determine the unified identifier of the target object; and sorting the multi-source touchpoint data by time based on the unified identifier to construct the interaction sequence.

[0010] In some optional implementations, the above method further includes: real-time monitoring of the operational status indicators of each of the aforementioned data sources, including the number of exposure events, the number of interaction events, and the number of conversion events; calculating the real-time contribution change of each of the aforementioned data sources based on the operational status indicators and the updated marginal contribution weights; triggering the multi-objective evolutionary algorithm to re-solve the resource allocation parameters when the real-time contribution change of any of the aforementioned data sources exceeds a preset fluctuation threshold; generating task scheduling control instructions for each of the aforementioned data sources based on the re-solved resource allocation parameters; and issuing the task scheduling control instructions to the corresponding data source nodes to adjust the resource output strategy of the aforementioned data source nodes; wherein the task scheduling control instructions include a target resource allocation quota, resource allocation time window parameters, and target object feature tags, and the aforementioned data source nodes dynamically adjust the call frequency and concurrency of the resource allocation interface according to the task scheduling control instructions.

[0011] Secondly, embodiments of the present invention provide a data attribution processing system based on multi-model fusion, comprising: a sequence construction module, used to acquire multi-source touchpoint data of a target object under multiple data sources, and construct an interaction sequence of the target object; an attribution calculation module, used to process the interaction sequence based on a causal inference confounding factor stripping model, acquire causal baseline parameters, and approximate the marginal contribution weight of each of the aforementioned data sources by sampling based on the aforementioned causal baseline parameters; an incremental update module, used to use the marginal contribution weight of each of the aforementioned data sources as the global weight of the batch processing layer, initialize the sliding window of the stream processing layer, and in response to the detection of a conversion event, perform incremental attribution update on the interaction sequence within the aforementioned sliding window to obtain the updated marginal contribution weight; and a resource solving module, used to solve resource allocation parameters based on a multi-objective evolutionary algorithm accelerated by a surrogate model, combined with the updated marginal contribution weight.

[0012] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0013] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to perform the method described in any of the first aspects above.

[0014] This invention provides a data attribution processing method, system, and electronic device based on multi-model fusion. The method acquires multi-source touchpoint data of a target object across multiple data sources to construct an interaction sequence. It then processes the interaction sequence using a confounding factor stripping model based on causal inference to obtain causal baseline parameters. Based on these parameters, it approximates the marginal contribution weights of each data source through sampling. These marginal contribution weights are used as global weights in the batch processing layer to initialize a sliding window in the stream processing layer. In response to the detection of a conversion event, the interaction sequence is incrementally updated within the sliding window to obtain the updated marginal contribution weights. Finally, a multi-objective evolutionary algorithm accelerated by a surrogate model is used to solve resource allocation parameters in conjunction with the updated marginal contribution weights. This solves the technical problems of low attribution accuracy and high computational latency in existing data attribution processing methods, achieving a balance between attribution accuracy and computational real-time performance, and dynamically and efficiently solving resource allocation parameters. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a data attribution processing method based on multi-model fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a data attribution processing system based on multi-model fusion provided in an embodiment of the present invention; Figure 3 A flowchart illustrating another data attribution processing method based on multi-model fusion provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention provides a data attribution processing method based on multi-model fusion. This method can be applied to scenarios that require quantifying the contribution of multiple data sources to a target event, such as digital marketing attribution, web traffic source analysis, and multi-channel communication effect evaluation. For ease of understanding, the following embodiment uses the quantification of the contribution of each marketing touchpoint to the conversion event in a digital marketing scenario as an example, but the method of this invention is also applicable to other similar multi-source data attribution scenarios.

[0019] Example 1: Figure 1 This diagram illustrates a data attribution processing method based on multi-model fusion provided by an embodiment of the present invention. The method includes the following steps: Step S102: Obtain multi-source touchpoint data of the target object under multiple data sources, and construct the interaction sequence of the target object.

[0020] The target object refers to the entity unit to be analyzed, such as a user, device, or account. The data source refers to the channel or platform that can interact with the target object, such as search advertising platforms, feed advertising platforms, social media platforms, official websites, applications, email systems, and SMS platforms. Multi-source touchpoint data refers to the behavioral records of the target object across various data sources, including but not limited to exposure records, click records, access records, dwell time records, interaction records, and conversion records. Each touchpoint data entry may include information such as data source identifier, behavior type, timestamp, and target object identifier.

[0021] After acquiring multi-source touchpoint data scattered across multiple data sources, it is necessary to correlate and integrate touchpoint data belonging to the same target object, and arrange them in chronological order to form the interaction sequence (i.e., user journey) for that target object. For example, the interaction sequence of a target object can be represented as: [Time T1: Search ad click] → [Time T2: Official website visit] → [Time T3: Feed ad exposure] → [Time T4: Application download] → [Time T5: Push notification click] → [Time T6: Conversion event]. This interaction sequence fully records the touchpoints experienced by the target object from the first touchpoint to the final conversion and their temporal relationships, providing basic data for subsequent attribution calculations.

[0022] In one embodiment, the steps described above for acquiring multi-source touchpoint data of the target object from multiple data sources and constructing the interaction sequence of the target object can be specifically implemented through cross-data source identifier fusion using a probabilistic graphical model to ensure the integrity and accuracy of the interaction sequence. For a detailed implementation, please refer to Embodiment Six below.

[0023] Step S104: The confounding factor stripping model based on causal inference processes the interaction sequence, obtains the causal baseline parameters, and calculates the marginal contribution weight of each data source by sampling based on the causal baseline parameters.

[0024] Confounding factors refer to those factors that influence both the target audience's access to a specific data source and their eventual conversion behavior. These include the target audience's activity level, inherent preferences for specific products, and time-cycle effects. Failure to eliminate these confounding factors will lead to a false correlation between data source exposure and conversion results, resulting in an inaccurate assessment of the data source's true contribution.

[0025] A causal baseline parameter can be a parameter obtained through causal inference methods, after removing the influence of confounding factors, that characterizes the true causal effect between each data source and the conversion result. For example, a causal baseline parameter can be expressed as the net increment in the conversion probability of a target object after exposure to a specific data source.

[0026] Marginal contribution weights refer to the weight values ​​that fairly allocate the final transformation result of the target object to each data source in its interaction sequence, reflecting the relative contribution of each data source to the transformation result. In this step, sampling approximation calculation refers to the method of estimating marginal contribution weights through random sampling when the complexity of precise calculation is too high, such as approximating the Shapley Value based on the Monte Carlo sampling method.

[0027] This embodiment first utilizes causal inference to remove confounding factors and obtain an accurate causal baseline. Then, using this causal baseline as input, it combines a sampling approximation method to efficiently calculate the marginal contribution weights of each data source, thereby significantly reducing computational complexity while ensuring attribution accuracy. Causal inference and game theory attribution form a close data synergy here—the causal baseline parameters output by causal inference provide biased baseline predictions for the calculation of game theory marginal contributions, ensuring that the calculated marginal contribution weights reflect causal effects rather than simple correlations.

[0028] In one embodiment, the confounding factor removal model for causal inference described above can be implemented using a dual machine learning framework, as detailed in Embodiment 2 below. In another embodiment, the marginal contribution weights calculated by sampling based on causal baseline parameters can be implemented using a hierarchical sampling Monte Carlo method, as detailed in Embodiment 3 below.

[0029] Step S106: Use the marginal contribution weights of each data source as the global weights of the batch processing layer, initialize the sliding window of the stream processing layer, and in response to the detection of a transformation event, perform incremental attribution updates on the interaction sequence within the sliding window to obtain the updated marginal contribution weights.

[0030] The batch processing layer refers to the processing layer that performs periodic calculations on the full dataset offline. For example, it uses distributed computing frameworks such as Spark to perform attribution calculations on the historical full dataset on a daily (T+1) basis. The global weight refers to the marginal contribution weight of each data source calculated by the batch processing layer on the historical full dataset, reflecting the average contribution level of each data source over a longer period.

[0031] The stream processing layer refers to the processing layer that performs incremental computation on real-time data streams online. For example, stream processing frameworks such as Flink are used to process real-time event streams one by one or in micro-batch. A sliding window refers to a time window with a preset time span (e.g., 24 hours) maintained in the stream processing layer. The window slides forward continuously over time, and the window retains the sequence of touch points of the target object within that time span.

[0032] Initializing the sliding window of the stream processing layer means using the global weights calculated by the batch processing layer as the initial weight state of the streaming processing layer's sliding window. This allows the streaming processing layer to obtain an accurate initial state based on precise offline calculation results upon startup or restart, rather than starting from scratch. Incremental attribution update refers to the process where, when a new conversion event is detected, instead of recalculating all historical data, the marginal contribution weights are incrementally adjusted and updated based on the current sliding window state, combined with the newly triggered conversion event and its associated touchpoint sequence.

[0033] This embodiment uses a batch-stream hybrid architecture, which utilizes the accurate offline calculation results of the batch processing layer to provide a high-precision initial state for the stream processing layer. The stream processing layer then performs incremental updates on the real-time event stream, thereby balancing attribution accuracy and real-time performance, and achieving near real-time attribution weight updates.

[0034] In one embodiment, the specific implementation of the above batch-stream mixed incremental calculation can be found in Embodiment 4 below.

[0035] Step S108: Solve the resource allocation parameters using a multi-objective evolutionary algorithm accelerated by the surrogate model, combined with the updated marginal contribution weights.

[0036] Resource allocation parameters refer to control parameters used to guide resource allocation among various data sources, such as the target resource allocation amount, resource allocation ratio, and resource allocation priority for each data source. Multi-objective evolutionary algorithms are evolutionary computation methods capable of simultaneously optimizing multiple conflicting objective functions, such as the NSGA-III algorithm. Surrogate models are simplified models used to approximate the evaluation of the true objective function, such as Gaussian process models, whose evaluation cost is much lower than that of the true model, significantly accelerating the optimization process.

[0037] In this step, the updated marginal contribution weights serve as a key input for fitness evaluation in the multi-objective evolutionary algorithm. They are used to measure the relative importance of each data source in the current state, guiding the solution of resource allocation parameters. The fitness function can be designed to comprehensively consider multiple objectives, such as maximizing the expected total conversion and maximizing resource utilization efficiency, while being constrained by preset resource constraints (such as the total resource ceiling, the upper and lower limits of resources for each data source, and the diminishing marginal returns effect).

[0038] This embodiment uses a surrogate model to replace the real objective function for evaluation, which significantly reduces the computational cost of each iteration, thereby accelerating the convergence of the multi-objective evolutionary algorithm. At the same time, the periodic calibration of the surrogate model ensures the accuracy of the optimization results, achieving a balance between solution efficiency and accuracy.

[0039] In one embodiment, the specific implementation of the multi-objective evolutionary algorithm accelerated by the surrogate model can be found in Embodiment 5 below.

[0040] The steps S102 to S108 described above constitute a complete technical closed loop: step S102 constructs the complete interaction sequence of the target object; step S104 efficiently calculates accurate marginal contribution weights based on causal inference and sampling approximation; step S106 achieves near real-time incremental weight updates through a batch-stream hybrid architecture; and step S108 solves for the optimal resource allocation parameters based on the updated weights. There is a close data flow and collaborative relationship between each step—causal inference provides a bias-free benchmark for game theory attribution, the batch processing layer provides an accurate initial state for the stream processing layer, and the attribution weights provide a fitness evaluation basis for the evolutionary algorithm, thus forming a complete processing chain from data acquisition, attribution calculation, real-time updates to resource optimization.

[0041] Example 2: In one embodiment, step S104 above, which involves processing the interaction sequence using a confounding factor stripping model based on causal inference to obtain causal baseline parameters, includes: processing the interaction sequence using a first prediction model to obtain the exposure tendency residual of the data source; processing the interaction sequence using a second prediction model to obtain the baseline conversion result residual; and performing regression analysis based on the exposure tendency residual and the baseline conversion result residual to obtain the causal baseline parameters.

[0042] The exposure propensity residual refers to the difference between whether the target audience actually comes into contact with a specific data source and the predicted probability of contact. The baseline conversion result residual refers to the difference between whether the target audience actually converts and the predicted probability of conversion. By extracting the residuals, the influence of confounding factors on the relationship between data source exposure and conversion results can be effectively eliminated.

[0043] Specifically, this embodiment can be implemented using a dual machine learning framework: In the first stage, a first prediction model (e.g., a random forest model) is used to predict the probability of the target object's tendency to contact various data sources based on its characteristic data (such as historical behavioral characteristics, demographic characteristics, etc.). The difference between the predicted value and the actual exposure is calculated to obtain the exposure tendency residual. Simultaneously, a second prediction model (e.g., a random forest model) is used to predict the baseline conversion probability based on the target object's characteristic data. The difference between the predicted value and the actual conversion result is calculated to obtain the baseline conversion result residual.

[0044] In the second stage, the exposure bias residual is used as the independent variable, and the baseline conversion result residual is used as the dependent variable for regression analysis. The regression coefficient is the causal baseline parameter. This parameter characterizes the net causal effect of a specific data source on the conversion result after removing confounding factors. Preferably, a cross-fitting method can be used, dividing the sample data into multiple parts, using one part of the samples to train the prediction model in turn, and using another part of the samples to calculate the residuals, thus avoiding overfitting bias.

[0045] For example, the causal baseline parameter can be represented as θ. If θ is positive and statistically significant, it indicates that the data source has a positive causal effect on the conversion; if θ approaches zero, it indicates that the relationship between the data source and the conversion is only a correlation rather than a causal relationship.

[0046] Example 3: In one embodiment, step S104 above, which approximates the marginal contribution weight of each data source based on causal benchmark parameters through sampling, includes: performing hierarchical sampling based on the historical contribution of each data source to generate multiple data source arrangement orders; for each data source arrangement order, sequentially traversing each data source, calculating the first predicted transformation value after the current data source is added to the previous data source set, and the second predicted transformation value before the current data source is added, and taking the difference between the first predicted transformation value and the second predicted transformation value as the marginal contribution of the current data source under the data source arrangement order; repeatedly executing the steps of generating multiple data source arrangement orders and calculating marginal contributions until the change in marginal contribution is less than a preset convergence threshold in a preset number of iterations; averaging the marginal contributions under all data source arrangement orders to obtain the marginal contribution weight of each data source; wherein, the first predicted transformation value and the second predicted transformation value are calculated based on the causal benchmark parameters and a preset transformation prediction function.

[0047] Historical contribution refers to the cumulative contribution weight assigned to each data source in historical attribution calculations, reflecting its relative importance. Stratified sampling involves dividing data sources into multiple tiers based on historical contribution, and then randomly sampling within each tier to generate the data source ranking. Compared to completely random sampling, stratified sampling ensures that both important and unimportant data sources are adequately represented in the sampling, reducing estimation variance.

[0048] Specifically, this embodiment uses the Monte Carlo sampling method to approximate the Shapley Value as the marginal contribution weight. Accurate calculation of the Shapley Value requires traversing all possible data source permutations and combinations, resulting in a computational complexity of O(2^3). n ), where n is the number of data sources. When n is large, the computational complexity increases exponentially. This embodiment approximates the estimation by randomly sampling a finite number of permutations, thus reducing the computational complexity to a controllable range while ensuring accuracy.

[0049] The algorithm flow is as follows: First, the data sources are divided into high contribution layer, medium contribution layer and low contribution layer according to the historical contribution of each data source. Random sampling is performed in each layer according to the proportion to generate M data source order (M is the preset number of sampling, such as 10,000 times).

[0050] The second step is to iterate through each data source in order for each permutation order π. Let the current data source be i, and its preceding data source set be S (i.e., all data sources that are ranked before i in the permutation order π). Then: the first predicted transformation value is the joint predicted transformation value V(S∪{i}) after adding data source i to S; the second predicted transformation value is the predicted transformation value V(S) based solely on S without adding data source i; the marginal contribution of data source i under this permutation order is V(S∪{i})-V(S).

[0051] The conversion prediction value V(S) is calculated using a pre-defined conversion prediction function. This function takes the data source set S and the causal baseline parameter θ as input and outputs the expected conversion probability. The causal baseline parameter θ, as the core parameter of the prediction function, ensures that the predicted value reflects the causal effect after removing confounding factors, rather than spurious correlations.

[0052] The third step involves repeating the permutation generation and marginal contribution calculation steps until the convergence condition is met—the change in the marginal contribution estimate of each data source is less than a preset convergence threshold (e.g., 1%) in a series of preset iterations (e.g., 100). By setting the convergence condition, the number of sampling times can be dynamically controlled while ensuring accuracy, avoiding unnecessary computational overhead.

[0053] The fourth step is to average the marginal contributions of each data source i across all M permutations to obtain the marginal contribution weight of that data source, which is an approximate estimate of the Shapley Value.

[0054] Preferably, the above calculation process can be parallelized using a distributed computing framework (such as Spark), which distributes the M permutations to multiple computing nodes to process the marginal contribution calculation of each permutation in parallel, and finally sums and averages the results to further shorten the calculation time.

[0055] Example 4: In one embodiment, step S106 above uses the marginal contribution weights of each data source as the global weights of the batch processing layer, initializes the sliding window of the stream processing layer, and, in response to the detection of a conversion event, performs incremental attribution updates on the interaction sequence within the sliding window to obtain the updated marginal contribution weights. This includes: calculating and storing the global precise attribution weights by the batch processing layer according to a first cycle; receiving the real-time event stream by the stream processing layer and initializing the sliding window based on the global precise attribution weights; and, in response to the detection of a conversion event in the real-time event stream, performing incremental attribution updates on the interaction sequence within the sliding window based on a preset approximate weighting algorithm to obtain the updated marginal contribution weights.

[0056] The first cycle refers to the calculation and update cycle of the batch processing layer, which can be set to once a day (T+1) or other cycles depending on business needs and data volume. The global precise attribution weight refers to the marginal contribution weight of each data source calculated by the batch processing layer based on the full historical data using the causal inference and sampling approximation methods described in step S104. This weight reflects the average contribution level of each data source over a longer time period and has high accuracy.

[0057] Specifically, this embodiment uses the Lambda architecture to implement incremental attribution calculation for batch and stream processing. The batch processing layer is implemented using a distributed batch processing framework (such as Spark): daily, it performs complete steps S102 to S104 on all historical multi-source touchpoint data from day T-1 and earlier, calculates the global accurate attribution weight of each data source, and stores the calculation results in a distributed storage system (such as HBase) for the stream processing layer to read and use.

[0058] The stream processing layer is implemented using a distributed stream processing framework (such as Flink): it continuously receives real-time event streams from various data sources (e.g., via Kafka message queues), including exposure events, click events, conversion events, etc. When the stream processing layer starts, it reads the latest globally accurate attribution weights calculated by the batch processing layer from the storage system, using them as the initial weight state of the sliding window to complete the initialization.

[0059] The sliding window maintains the interaction sequence of each target object and its corresponding data source weight within a preset time window (e.g., 24 hours). When the stream processing layer detects a conversion event from the real-time event stream, it triggers an incremental attribution update process: it obtains the interaction sequence of the target object corresponding to the conversion event within the sliding window, incrementally adjusts the global precise attribution weights based on a preset approximate weighting algorithm (e.g., the Last-N-Touch weighting algorithm), and outputs the updated marginal contribution weights.

[0060] Preferably, the approximate weighted algorithm can adopt a weighting method based on the most recent N touchpoints, which reflects the trend of recent touchpoint contribution changes while maintaining computational efficiency. The stream processing layer uses Exactly-Once semantics to ensure that data is not lost or processed repeatedly; it uses RocksDB state backend storage to store sliding window states, supporting efficient management of large states; and it uses asynchronous I / O and operator chain optimization to reduce processing latency, achieving near real-time attribution with a P99 latency of less than 500 milliseconds.

[0061] Through the above batch-stream hybrid architecture, the batch processing layer provides a high-precision attribution benchmark, and the stream processing layer implements incremental updates at the minute or second level on this basis, which takes into account both attribution accuracy and real-time performance. This enables the system to capture fluctuations in the effect of the data source within a day and adjust the attribution weights in a timely manner.

[0062] Example 5: In one embodiment, the multi-objective evolutionary algorithm based on the surrogate model in step S108 above, combined with the updated marginal contribution weights to solve for resource allocation parameters, includes: initializing a population containing multiple individuals, encoding each individual as a candidate resource allocation vector for each data source; during the iteration of the multi-objective evolutionary algorithm, using a Gaussian process surrogate model to evaluate the fitness of each individual in the population, the fitness being calculated based on the updated marginal contribution weights and preset resource constraints; calibrating the Gaussian process surrogate model using a real objective function model according to a preset calibration period, and updating the hyperparameters of the Gaussian process surrogate model; generating a progeny population based on crossover and mutation operations, and using an elite retention strategy to retain the non-dominated solutions of the Pareto front; when a preset convergence condition is met, selecting a target individual from the Pareto front, and using the candidate resource allocation vector corresponding to the target individual as the resource allocation parameter; wherein, the real objective function model is used to calculate the expected total conversion and resource utilization efficiency under a given candidate resource allocation vector.

[0063] Here, the candidate resource allocation vector refers to the combination scheme of resource allocation quotas for each data source, encoded as a chromosome in the evolutionary algorithm. For example, assuming there are n data sources, the chromosome can be represented as [x1, x2, ..., x...]. n ], where x i This represents the resource quota allocated to the i-th data source.

[0064] Specifically, this embodiment uses the NSGA-III algorithm as the framework for a multi-objective evolutionary algorithm, combined with a Gaussian process surrogate model to accelerate fitness evaluation. The specific implementation steps are as follows: (1) Initializing the population. The initial population is generated using the Latin hypercube sampling method, and the population size can be set to 200 individuals. Latin hypercube sampling can generate diverse initial solutions that are uniformly distributed in the solution space, avoiding the initial population from being concentrated in a local area, which is beneficial for global search.

[0065] (2) Fitness Assessment. In each iteration of the evolutionary algorithm, a Gaussian process surrogate model is used to predict the fitness of each individual in the population. Fitness is calculated based on two objective functions: the first objective is to maximize the expected total conversion, and the second objective is to maximize resource utilization efficiency. The calculation of both objective functions uses the updated marginal contribution weights as input parameters, so that the resource allocation scheme reflects the latest contribution level of each data source. At the same time, fitness assessment must meet the preset resource constraints, including the total resource upper limit constraint (∑x i Constraints include total budget, upper and lower limits for resources from each data source, and data source saturation effect constraints (diminishing marginal returns). For individuals that do not meet the constraints, a penalty function method can be used to reduce their fitness, or a repair operator can be used to project them into the feasible region.

[0066] Among them, the Gaussian process surrogate model is a probability-based nonparametric regression model that can predict the uncertainty of the prediction while predicting the input. Compared with directly calling the true objective function (which requires complex transformation prediction calculations), the evaluation cost of the surrogate model is extremely low, which can be completed in milliseconds, significantly accelerating the iterative process of evolutionary algorithms.

[0067] (3) Proxy model calibration. Since the proxy model is an approximation of the real objective function, the prediction error of the proxy model may accumulate as the evolutionary iteration proceeds. Therefore, it is necessary to select representative individuals in the population according to a preset calibration period (e.g., every 10 generations), calculate their true fitness using the real objective function model, add the results to the training set of the proxy model, and update the hyperparameters of the Gaussian process proxy model to ensure the prediction accuracy of the proxy model.

[0068] Among them, the true objective function model refers to the model that directly calculates the actual expected total conversion amount and resource utilization efficiency under a given candidate resource allocation vector based on the attribution model and the conversion prediction model. It has high calculation accuracy but large calculation cost.

[0069] (4) Genetic operations. A simulated binary crossover operator (SBX, distribution index η_c=20) is used to crossover the parent individuals to generate offspring individuals; a multinomial mutation operator (distribution index η_m=20) is used to mutate the offspring individuals to increase population diversity. After merging the parent and offspring populations, an elite retention strategy is adopted to select the non-dominated solution of the next generation's Pareto front based on non-dominated sorting and crowding distance, ensuring that excellent solutions are not lost during the evolutionary process.

[0070] (5) Convergence Judgment. Preset convergence conditions may include: the change in hypervolume index is less than a preset threshold (e.g., 0.5%), or the preset maximum number of iterations is reached (e.g., 500 generations). When the convergence conditions are met, the evolutionary process terminates, and the target individual is selected from the final Pareto front. The selection strategy can be determined according to decision preferences, such as selecting the individual that maximizes the expected total conversion, or selecting the individual that maximizes resource utilization efficiency, or making a trade-off between two objectives. The candidate resource allocation vector corresponding to the target individual is used as the resource allocation parameter for the final solution.

[0071] Compared to traditional gradient-based optimization algorithms, NSGA-III, combined with a Gaussian process surrogate model, can effectively handle the non-convexity of the objective function and the multi-objective trade-off problem. The optimal resource allocation scheme found significantly improves the expected total conversion, while providing a variety of Pareto front solutions for decision-makers to choose flexibly according to actual needs.

[0072] Example 6: In one embodiment, step S102 above, which involves acquiring multi-source touchpoint data of the target object from multiple data sources and constructing an interaction sequence for the target object, includes: acquiring various identifier type data from the multi-source touchpoint data; constructing a probability factor graph containing identifier nodes and observation association edges, where identifier nodes correspond to identifier type data and observation association edges correspond to identifier co-occurrence relationships in the same event; performing iterative calculations on the probability factor graph using a message propagation algorithm to obtain the edge probability of each identifier node; fusing cross-source identifiers based on edge probabilities to determine a unified identifier for the target object; and sorting the multi-source touchpoint data by time based on the unified identifier to construct an interaction sequence.

[0073] Multiple identifier types refer to different types of identifiers used by the target object in different data sources, such as, but not limited to: browser cookie identifiers, mobile device identifiers, login account identifiers, IP addresses, behavioral characteristic identifiers, etc. Since the same target object may use different identifiers at different times, on different devices, and on different platforms, its touchpoint data is scattered under multiple identifiers, requiring identifier fusion to construct a complete interaction sequence.

[0074] Specifically, this embodiment achieves cross-data source identifier fusion by constructing a probabilistic graphical model. A probabilistic factor graph is a probabilistic graphical model composed of nodes and edges: identifier nodes correspond to specific identifier values ​​for each identifier type (such as Cookie_A, DeviceID_B, UserID_C, etc.), and observation association edges connect multiple identifier nodes that appear together in the same event, indicating that there is an association between them (for example, if Cookie_A, DeviceID_B, and IP_D are recorded simultaneously in a certain access event, then an observation association edge is established between them).

[0075] After the factor graph is constructed, a factor function (i.e., a potential function) is defined to describe the association strength between adjacent identifier nodes. For example, the higher the frequency of co-occurrence of two identifier nodes, the stronger their association. The factor function can take the form of a log-linear model: ψ(i, j) = exp(w·f(i, j)), where f(i, j) is the feature vector between identifier nodes i and j (including features such as co-occurrence frequency, time interval, and IP overlap), and w is the weight parameter.

[0076] A message propagation algorithm (such as Loopy Belief Propagation) is used for iterative calculation in the factor graph: each node updates its own confidence based on the messages passed by its neighboring nodes, and the message update rule is based on the sum-product algorithm. Through multiple rounds of iteration until convergence or the preset maximum number of iterations is reached, the marginal probability of each labeled node is obtained. The marginal probability represents the confidence that the labeled node and its neighboring labeled nodes belong to the same target object.

[0077] Identifier fusion based on edge probability: Identifier nodes with edge probabilities higher than a preset threshold (e.g., 0.8) are clustered into a set of identifiers for the same target object, a unified identifier for the target object is determined, and a mapping relationship between the unified identifier and each original identifier is established and stored in a graph database for real-time query and incremental update.

[0078] After obtaining the unified identifier, the data of each touch point is arranged in chronological order according to the timestamp information in the multi-source touch point data associated with the unified identifier, forming a complete interaction sequence of the target object.

[0079] Preferably, to reduce the computational complexity of identifier fusion, the following acceleration techniques can be adopted: By using blocking techniques, pre-blocking is performed based on the hash values ​​of strong identifier information such as name, mobile phone number, and email address, limiting matching candidate pairs to the same block and reducing unnecessary matching calculations; MinHash and Locality Sensitive Hashing are used to accelerate similarity calculations and quickly filter candidate identifier pairs with high similarity; The EM algorithm is used to iteratively optimize the weight parameter w in the factor function to improve the accuracy of edge probability estimation.

[0080] Example 7: In one embodiment, the method further includes: real-time monitoring of the operational status indicators of each data source, including the number of exposure events, the number of interaction events, and the number of conversion events; calculating the real-time contribution change of each data source based on the operational status indicators and the updated marginal contribution weight; triggering a multi-objective evolutionary algorithm to re-solve the resource allocation parameters when the real-time contribution change of any data source exceeds a preset fluctuation threshold; generating task scheduling control instructions for each data source based on the re-solved resource allocation parameters; and issuing the task scheduling control instructions to the corresponding data source nodes to adjust the resource output strategy of the data source nodes; wherein the task scheduling control instructions include the target resource allocation quota, resource allocation time window parameters, and target object feature tags, and the data source nodes dynamically adjust the call frequency and concurrency of the resource allocation interface according to the task scheduling control instructions.

[0081] This embodiment describes a closed-loop dynamic optimization mechanism based on real-time monitoring feedback.

[0082] Among them, operational status metrics refer to quantifiable indicators generated in real time by each data source during operation, reflecting its current performance status. Exposure events refer to the number of times a target object sees resource content on that data source; interaction events refer to the number of times a target object clicks or engages in deep interaction; and conversion events refer to the number of times a target object completes the target action. These metrics can be continuously collected and aggregated from the real-time event stream through the aforementioned stream processing layer.

[0083] The real-time contribution change is used to measure the deviation between the current actual contribution level and the expected contribution level of each data source. The specific calculation method is as follows: calculate the actual conversion contribution of each data source based on the operational status indicators (e.g., the weighted result of the number of actual conversion events and the attribution weight), compare it with the expected conversion contribution calculated based on the updated marginal contribution weight, and take the difference or ratio of the two as the real-time contribution change.

[0084] When the real-time contribution change of any data source exceeds a preset fluctuation threshold, it indicates a significant change in the performance of that data source. Reasons for this change may include: a decline in the effectiveness of the data source itself (e.g., audience fatigue), changes in the external environment (e.g., increased costs due to intensified competition), or the failure of attribution weights. At this point, a multi-objective evolutionary algorithm is triggered, using the latest marginal contribution weights as input to recalculate the resource allocation parameters and obtain the optimal resource allocation scheme adapted to the current market conditions.

[0085] Based on the re-solved resource allocation parameters, the system generates task scheduling control instructions for each data source. These instructions are machine-executable parameterized instructions and include the following control information: target resource allocation quota (used to specify the upper limit of resource usage for the data source in the next cycle), resource allocation time window parameters (used to specify the effective time period for resource delivery), and target object feature tags (used to specify the target audience attributes for resource delivery, such as region, age, interest category, etc.).

[0086] The system sends task scheduling control commands to each data source node via the network interface. Upon receiving the commands, the data source nodes parse the control parameters and dynamically adjust the calling behavior of their resource allocation interfaces, including but not limited to: adjusting the interface call frequency (e.g., requests per second), adjusting the number of concurrent connections, and enabling or pausing resource delivery for specific target object groups. Through this closed-loop control mechanism, the system can automatically adjust its resource allocation strategy in a timely manner when abnormal fluctuations in data source performance are detected, achieving agile resource optimization and loss mitigation.

[0087] In summary, the embodiments of the present invention solve the problem of the separation of these three elements in the prior art through the deep synergy of causal inference, game theory approximation calculation and batch-stream hybrid architecture. This achieves the technical effect of balancing attribution accuracy and computational real-time performance under high-dimensional nonlinear objectives, thereby dynamically and efficiently solving resource allocation parameters.

[0088] Based on the same inventive concept, this invention also provides a data attribution processing system based on multi-model fusion, see [link to relevant documentation]. Figure 2 As shown, the system mainly includes the following parts: Sequence construction module 210 is used to acquire multi-source touchpoint data of the target object under multiple data sources and construct the interaction sequence of the target object; The attribution calculation module 220 is used to process the interaction sequence of the confounding factor stripping model based on causal inference, obtain causal baseline parameters, and approximate the marginal contribution weight of each data source by sampling based on the causal baseline parameters. The incremental update module 230 is used to use the marginal contribution weights of each data source as the global weights of the batch processing layer, initialize the sliding window of the stream processing layer, and in response to the detection of a transformation event, perform incremental attribution updates on the interaction sequence within the sliding window to obtain the updated marginal contribution weights. Resource solving module 240 is used for a multi-objective evolutionary algorithm based on a proxy model to solve resource allocation parameters by combining updated marginal contribution weights.

[0089] The data attribution processing system based on multi-model fusion provided in this invention embodiment can be specific hardware on a device or software or firmware installed on the device. The system provided in this invention embodiment has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment section can be referred to the corresponding content in the aforementioned method embodiment. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0090] The above embodiments will now be described in detail with a specific application example. Preferably, this embodiment provides a data attribution processing system based on multi-model fusion, which can execute the steps in any of the above method embodiments. In an exemplary application scenario, the system can be deployed in a multi-source data attribution processing environment to quantify the contributions of multiple data sources and optimize resource allocation. Its overall processing flow includes... Figure 3 The following eight steps are shown: S301: Multi-source contact data acquisition and integration; The system acquires multi-source touchpoint data generated by the target object across various data sources through multiple data acquisition interfaces. These data sources include, but are not limited to, search and display data sources, news feed data sources, social media platform data sources, proprietary platform data sources (such as official websites, applications, WeChat official accounts, and email systems), and third-party platform data sources. The collected touchpoint data includes various behavior types such as exposure, interaction, and target behavior (conversion events), as well as information such as data source identifiers, timestamps, and target object identifiers. Through technologies such as unified resource location identifier parameters, data source encoding, and cross-domain tracking, the system achieves unified identification and aggregation of multi-source data, providing a complete data foundation for subsequent processing.

[0091] S302: Target object transformation journey reconstruction and interaction serialization; Based on the collected multi-source touchpoint data, the system reconstructs a complete conversion journey for each target object where a conversion event occurs. Specifically, this stage can employ the probabilistic graphical model identifier fusion method described in step S102 and Example 6. Through cross-source identifier integration technology (such as probability factor graphs and message propagation algorithms), all touchpoints of the same target object at different data sources and times are associated and arranged chronologically to form an interaction sequence. For example, a typical interaction sequence can be represented as: search data source click → browsing on proprietary platform → leaving → information flow data source exposure → application download → push notification click → conversion event. This stage can handle complex journeys across data sources and long cycles, and identify key conversion paths and high-frequency touchpoint combinations.

[0092] S303: Multi-model attribution analysis and contribution calculation; The system can deploy multiple attribution models in parallel to calculate the contribution of each data source. As a benchmark, the system can include attribution models based on preset rules (such as last-touch attribution, first-touch attribution, linear attribution, time-decay attribution, etc.). As the core attribution engine, the system preferentially adopts the attribution method based on causal inference and sampling approximation as described in the above embodiments of this application (corresponding to step S104). First, it obtains causal benchmark parameters by removing confounding factors through dual machine learning, and then uses a hierarchical sampling Monte Carlo method to approximate the Shapley Value marginal contribution weight of each data source. Furthermore, the system can also integrate the output of data-driven attribution models (such as logistic regression, XGBoost, and deep learning models) as an aid, and output the final attribution weight based on the fusion of multiple model results to balance the accuracy and robustness of the attribution.

[0093] S304: Incremental Value Assessment and Counterfactual Analysis; Attribution calculation not only identifies which data sources participated in the conversion journey but also quantifies the incremental value of each data source, i.e., answering the question, "Would the conversion event have occurred without the touchpoint of this data source?" The system employs counterfactual inference techniques to construct a control group of similar target objects who did not come into contact with a specific data source and an experimental group that did come into contact with that data source. By comparing the conversion rate differences between the two groups, the net incremental contribution of that data source is quantified. In this process, the system uses causal inference methods (such as propensity score matching, difference-in-differences, instrumental variables, etc.) to handle confounding factors such as selection bias and time effects, ensuring that the attribution results reflect true causal effects rather than spurious correlations. The processing logic at this stage is integrated into the causal baseline parameter acquisition process in step S104 mentioned above.

[0094] S305: Solving for resource allocation parameters; Based on the contribution of each data source obtained from attribution calculations, the system constructs a resource allocation optimization model. The optimization objective is to maximize the expected total conversion volume or resource utilization efficiency, among other target conversion indicators. The decision variables are the resource allocation quotas for each data source, and the constraints include the total resource ceiling, the upper and lower limits of resource allocation for each data source, and the data source saturation effect (diminishing marginal returns). The system uses a multi-objective evolutionary algorithm accelerated by a surrogate model (corresponding to step S108, Example 5) to solve the problem, while considering the synergistic and substitution effects among data sources to achieve a globally optimal resource allocation scheme.

[0095] S306: Near real-time attribution and dynamic resource adjustment; The system achieves near real-time attribution weight updates through a batch-stream hybrid architecture (corresponding to step S106, Example 4), rather than simply performing post-hoc analysis. The stream processing layer monitors the operational status indicators (exposure, interaction, conversion, etc.) of each data source in real time. When the contribution fluctuation of any data source exceeds a preset threshold, dynamic resource adjustment is triggered: resource allocation is increased for data sources with increased contribution, and resource allocation is reduced for data sources with decreased contribution, achieving agile resource optimization and timely mitigation of inefficient resources. For the specific closed-loop control process, please refer to the embodiment described in Example 7.

[0096] S307: Experimental verification and model calibration; The accuracy of the attribution model can be verified through the system's embedded experimental module. The system supports randomized experiments at the data source level: randomly selecting a subset of target objects and pausing access to a specific data source (experimental group), comparing the conversion differences with those of continuously accessed target objects (control group), thus verifying the true incremental value of the data source. Multiple experiments are used to calibrate the parameters of the attribution model, improving prediction accuracy. Simultaneously, the system can test the actual effects of different resource allocation strategies, iteratively optimizing the resource allocation model.

[0097] S308: Attribution Insights and Continuous Strategy Optimization; The system generates multi-dimensional attribution analysis reports, including conversion contribution rankings of various data sources, resource utilization efficiency analysis, synergistic effects between data sources, high-frequency conversion paths, and data source preferences for different target audiences. Based on these analysis results, efficient data sources are identified for increased resource investment, while inefficient and replaceable data sources are eliminated to reduce resource waste. Simultaneously, common characteristics of successful conversion paths are extracted to guide adjustments to touchpoint content and delivery strategies. The system continuously inputs feedback from actual delivery performance and environmental changes into the attribution model, forming a spiral improvement loop of "attribution-optimization-validation-iteration" to continuously optimize algorithms and parameters.

[0098] This embodiment organically integrates data acquisition, journey reconstruction, attribution modeling, resource optimization, and effect verification into a closed-loop system, achieving intelligent processing across the entire data-to-decision chain. Deep collaboration exists between each stage: the interaction sequence output from stage two provides high-quality input for stage three; the attribution weights of stage three directly drive resource allocation solutions in stage five; and real-time monitoring in stage six and experimental verification in stage seven provide the model with continuous calibration and dynamic adjustment capabilities. This comprehensively solves the technical problems of low attribution accuracy, high computational latency, and the inability to dynamically and efficiently solve resource allocation parameters in existing technologies.

[0099] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, specifically, the electronic device includes a processor and a storage device; the storage device stores a computer program, and the computer program, when run by the processor, executes the method described in any of the above embodiments.

[0100] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 400 includes: a processor 410, a memory 420, a communication interface 430, and a bus 440. The memory 420 stores machine-readable instructions that can be executed by the processor 410. When the electronic device is running, the processor 410 communicates with the memory 420 through the bus 440. The processor 410 executes the machine-readable instructions to perform the steps of the method described above.

[0101] Specifically, the memory 420 and processor 410 can be general-purpose memory and processor, without any specific limitations. When the processor 410 runs the computer program stored in the memory 420, it can execute the above method.

[0102] Processor 410 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 410 or by instructions in software form. The processor 410 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 420, and processor 410 reads the information from memory 420 and, in conjunction with its hardware, completes the steps of the above method.

[0103] Corresponding to the above method, this embodiment of the invention also provides a computer-readable storage medium storing machine-executable instructions. When the computer-executable instructions are called and run by a processor, the computer-executable instructions cause the processor to perform the steps of the above method.

[0104] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0105] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0106] Furthermore, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0107] It should be noted that if the functionality is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0109] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data attribution processing method based on multi-model fusion, characterized in that, include: Acquire multi-source touchpoint data of the target object from multiple data sources, and construct the interaction sequence of the target object; The confounding factor stripping model based on causal inference processes the interaction sequence, obtains causal baseline parameters, and calculates the marginal contribution weight of each data source by sampling based on the causal baseline parameters. The marginal contribution weights of each data source are used as the global weights of the batch processing layer. The sliding window of the stream processing layer is initialized, and in response to the detection of a transformation event, the interaction sequence is incrementally attributed and updated within the sliding window to obtain the updated marginal contribution weights. A multi-objective evolutionary algorithm based on a proxy model is used to solve for resource allocation parameters by combining the updated marginal contribution weights.

2. The method according to claim 1, characterized in that, The confounding factor stripping model based on causal inference processes the interaction sequence to obtain causal baseline parameters, including: The interaction sequence is processed by a first prediction model to obtain the exposure tendency residual of the data source; The interaction sequence is processed by a second prediction model to obtain the baseline transformation result residual; Regression analysis was performed based on the exposure tendency residual and the baseline transformation result residual to obtain the causal baseline parameters.

3. The method according to claim 2, characterized in that, The step of approximating the marginal contribution weights of each data source based on the causal baseline parameters through sampling includes: Based on the historical contribution of each data source, hierarchical sampling is performed to generate a ranking order of multiple data sources; For each of the data source arrangements, each data source is traversed in sequence, and the first predicted conversion value after the current data source is added to the previous data source set and the second predicted conversion value before the current data source is added are calculated. The difference between the first predicted conversion value and the second predicted conversion value is taken as the marginal contribution of the current data source under the data source arrangement. Repeat the steps of generating the order of multiple data sources and calculating the marginal contribution until the change in the marginal contribution is less than a preset convergence threshold in a preset number of consecutive iterations. The marginal contribution weight of each data source is obtained by averaging the marginal contributions of all data sources in the order of arrangement. The first predicted conversion value and the second predicted conversion value are calculated based on the causal benchmark parameters and the preset conversion prediction function.

4. The method according to claim 1, characterized in that, The step of using the marginal contribution weights of each of the data sources as global weights for the batch processing layer, initializing the sliding window of the stream processing layer, and incrementally updating the attribution of the interaction sequence within the sliding window in response to the detection of a transformation event to obtain the updated marginal contribution weights includes: The batch processing layer calculates and stores the global accurate attribution weights in the first cycle. The streaming layer receives real-time event streams and initializes the sliding window based on the global accurate attribution weights. In response to the detection of the transformation event in the real-time event stream, the interaction sequence is incrementally attributed and updated within the sliding window based on a preset approximate weighting algorithm to obtain the updated marginal contribution weight.

5. The method according to claim 4, characterized in that, The multi-objective evolutionary algorithm based on the surrogate model, which combines the updated marginal contribution weights to solve for resource allocation parameters, includes: Initialize a population containing multiple individuals, and encode each individual as a candidate resource allocation vector for each of the data sources; During the iterative process of the multi-objective evolutionary algorithm, the fitness of each individual in the population is evaluated using a Gaussian process surrogate model, and the fitness is calculated based on the updated marginal contribution weight and preset resource constraints. According to the preset calibration cycle, the Gaussian process proxy model is calibrated using the real objective function model, and the hyperparameters of the Gaussian process proxy model are updated. The offspring population is generated based on crossover and mutation operations, and an elite preservation strategy is used to preserve the non-dominated solutions of the Pareto front. When the preset convergence condition is met, a target individual is selected from the Pareto front, and the candidate resource allocation vector corresponding to the target individual is used as the resource allocation parameter. The true objective function model is used to calculate the expected total conversion and resource utilization efficiency under a given candidate resource allocation vector.

6. The method according to claim 1, characterized in that, The step of acquiring multi-source touchpoint data of the target object from multiple data sources and constructing the interaction sequence of the target object includes: Obtain various identifier types from the multi-source contact data, and construct a probability factor graph containing identifier nodes and observation association edges. The identifier nodes correspond to the identifier type data, and the observation association edges correspond to the co-occurrence relationship of identifiers in the same event. The edge probability of each of the identified nodes is obtained by iteratively calculating the probability factor graph using a message propagation algorithm. Based on the edge probability fusion cross-source identifier, a unified identifier for the target object is determined; The interaction sequence is constructed by sorting the multi-source touch data by time based on the unified identifier.

7. The method according to claim 6, characterized in that, The method further includes: Real-time monitoring of the operational status metrics of each of the aforementioned data sources, including the number of exposure events, the number of interaction events, and the number of conversion events; Based on the operational status indicators and the updated marginal contribution weights, the real-time contribution change of each of the data sources is calculated; When the change in the real-time contribution of any of the data sources exceeds a preset fluctuation threshold, the multi-objective evolutionary algorithm is triggered to re-solve the resource allocation parameters. Based on the resource allocation parameters obtained by re-solution, task scheduling control instructions are generated for each of the data sources. The task scheduling control command is sent to the corresponding data source node to adjust the resource output strategy of the data source node; The task scheduling control instruction includes a target resource allocation quota, a resource allocation time window parameter, and a target object feature tag. The data source node dynamically adjusts the call frequency and concurrency of the resource allocation interface according to the task scheduling control instruction.

8. A data attribution processing system based on multi-model fusion, characterized in that, include: The sequence construction module is used to acquire multi-source touchpoint data of the target object from multiple data sources and construct the interaction sequence of the target object. The attribution calculation module is used to process the interaction sequence based on the confounding factor stripping model of causal inference, obtain causal baseline parameters, and approximate the marginal contribution weight of each of the data sources by sampling based on the causal baseline parameters. The incremental update module is used to take the marginal contribution weight of each of the data sources as the global weight of the batch processing layer, initialize the sliding window of the stream processing layer, and in response to the detection of a transformation event, perform incremental attribution update on the interaction sequence within the sliding window to obtain the updated marginal contribution weight. The resource solution module is used to solve resource allocation parameters by combining the updated marginal contribution weights with a multi-objective evolutionary algorithm based on the surrogate model.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 7.

Images