User journey attribution and budget self-optimization method for e-commerce global placement

Abstract

The present application relates to the technical field of electronic commerce, and more particularly to a user journey attribution and budget self-optimization method for e-commerce global investment, comprising the following steps: step 1: collecting multi-channel user behavior event streams, taking the difference between the cumulative amount of conversion event condition strength values of the original sequence and the counterfactual residual sequence in the unified prediction interval as the unified window counterfactual conversion risk drop; step 2: for each target channel, obtaining the net causal attribution coefficient through multi-fold cross fitting double machine learning; step 3: constructing a cumulative attribution revenue concave utility function for each channel, forming an iterative closed loop of attribution modeling, causal debiasing and budget allocation. The present application can finely capture the timing incentive effect and decay characteristics of cross-channel touchpoints, effectively eliminate the confounding interference of user self-selection bias on attribution estimation, and ensure that the long-term cumulative revenue of the budget allocation strategy gradually approaches the theoretical optimal level in a non-stationary investment environment.

Description

Technical Field

[0001] This invention belongs to the field of e-commerce technology, specifically relating to a user journey attribution and budget self-optimization method for e-commerce full-domain advertising. Background Technology

[0002] Early attribution methods were primarily rule-driven, typically including last-click attribution, first-click attribution, and linear equal-distribution attribution. Last-click attribution attributes all conversion credit to the channel that made the last click before the user placed the order; first-click attribution attributes credit to the channel that first attracted the user's attention; and linear equal-distribution attribution distributes credit evenly across all touchpoints in the user journey. These methods are simple to implement and have very low computational overhead, and were widely used in the early stages of e-commerce advertising. However, the fundamental limitation of rule-driven attribution is that its allocation logic is pre-defined and does not rely on the statistical relationship between touchpoints and conversions in actual data. In today's advertising environment, where user journeys are increasingly complex and inter-channel synergies are significantly enhanced, rule-driven attribution is no longer able to provide accurate metrics of channel contribution.

[0003] To overcome the limitations of rule-driven attribution, academia and industry have proposed data-driven multi-touchpoint attribution methods. A representative approach is Markov chain-based attribution models. This method models the user journey as a first- or higher-order Markov chain, treating each channel touchpoint as a state in the chain. It describes the journey's evolution by statistically analyzing the transition probabilities between states and then calculates the marginal contribution of each channel to the conversion probability through removal effects. The advancement of Markov chain attribution over rule-driven attribution lies in its contribution estimation, which is learned from actual journey data. However, its implicit Markov property assumption requires that the transition probability from the current state to the next state depends only on the current state itself and is independent of earlier historical paths. In real-world e-commerce advertising scenarios, whether a user clicks on an ad is often influenced by the cumulative effects of earlier touchpoints. For example, the probability of a user clicking an ad in a search channel may be closely related to a brand exposure ad seen on social media several days prior. The Markov property assumption is insufficient for modeling such long-term temporal dependencies, easily leading to biased attribution results. Furthermore, the classic Markov chain attribution model only focuses on discrete transitions between states and does not explicitly model the impact of time intervals between touch events. However, time intervals often carry important information in e-commerce scenarios because the incentive effect produced by users continuously touching multiple touch points in a short period of time is significantly different from that produced by sporadic touch points over a longer period of time. Summary of the Invention

[0004] The main objective of this invention is to provide a user journey attribution and budget self-optimization method for e-commerce full-domain advertising. It captures the temporal incentive effect of cross-channel touchpoints through structural decomposition of basic reach intensity and touchpoint-induced incremental intensity. It effectively eliminates the attribution overestimation problem caused by user self-selection bias through dual machine learning residual processing. It ensures that the long-term cumulative benefit of the budget allocation strategy in non-stationary advertising environment approaches the theoretical optimal level with a sublinear regret rate through online concave utility maximization and negative entropy regularization.

[0005] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0006] A user journey attribution and budget self-optimization method for e-commerce full-domain advertising includes the following steps:

[0007] Step 1: Collect user behavior event streams from multiple channels and map them into a journey event sequence using user identifiers; construct an event embedding vector for each event in the journey event sequence and input it into a recurrent neural network; calculate the base arrival strength and touchpoint activation increment strength based on the hidden state vectors, and add the two to obtain the conversion event conditional strength value; define a unified prediction time domain endpoint, remove touchpoint events from each channel, and re-input the remaining sequence into the recurrent neural network; use the difference between the cumulative conversion event conditional strength values ​​of the original sequence and the counterfactual remaining sequence within the unified prediction interval as the unified window counterfactual conversion risk reduction amount;

[0008] Step 2: For each target channel, the unified window counterfactual conversion risk reduction amount is used as the processing variable, whether the journey event sequence contains conversion events before the unified prediction time domain end is used as the outcome variable, and user characteristics, environmental characteristics and the unified window counterfactual conversion risk reduction amount of other channels are used as confounding covariates. The net causal attribution coefficient is obtained through multi-fold cross-fit dual machine learning.

[0009] Step 3: Construct a concave utility function for cumulative attribution revenue for each channel. With the daily total budget, the upper limit of each channel's investment, and the minimum exploration investment ratio as constraints, the follow-regularized leader algorithm is used to update the concave utility function for cumulative attribution revenue in each decision cycle through steps 1 and 2. The budget allocation ratio vector is then adjusted by gradient and projected into the convex feasible region, forming an iterative closed loop of attribution modeling, causal bias removal, and budget allocation.

[0010] Furthermore, each event record in the user behavior event stream includes a user identifier, channel identifier, ad group identifier, event type, and event timestamp. Event types cover impressions, clicks, landing page visits, add-to-cart, order placement, and repeat purchases.

[0011] Furthermore, the event embedding vector is constructed as follows: the channel feature vector is obtained by retrieving the channel embedding lookup table with the channel identifier of the event, the type feature vector is obtained by retrieving the event type embedding lookup table with the event type of the event, the time interval between the current event and the preceding adjacent event is input into a single-layer linear projection layer and mapped into a time encoding vector, and the channel feature vector, type feature vector and time encoding vector are concatenated to form the event embedding vector.

[0012] Furthermore, the recurrent neural network is a gated recurrent unit network; the hidden state vector is input into the first single-layer feedforward network and passed through the softplus activation function to obtain the basic arrival strength; the hidden state vector is input into the second single-layer feedforward network and passed through the softplus activation function to obtain the touch point excitation increment strength; the basic arrival strength represents the spontaneous arrival tendency of the conversion event without considering the previous touch point excitation, and the touch point excitation increment strength represents the additional increase in the arrival probability of the current touch point event on the subsequent conversion event.

[0013] Furthermore, the joint training objective of the gated recurrent unit network, the first single-layer feedforward network, and the second single-layer feedforward network is to maximize the transformation event conditional strength value of all observed events in the journey event sequence at the time of their occurrence, while minimizing the piecewise cumulative summation of the transformation event conditional strength values ​​along the time axis within the time interval between adjacent events.

[0014] Furthermore, the unified prediction time domain endpoint is the time corresponding to the last event timestamp in the journey event sequence plus a fixed-length observation extension window; the unified prediction interval is the time interval from the last event timestamp to the unified prediction time domain endpoint.

[0015] Furthermore, the user characteristics in the confounding covariates include user demographics, device type, access source, duration of stay per session, and new or returning customer markers; the environmental characteristics in the confounding covariates include promotion cycle markers, weekday and time period markers, product price range markers, and discount level markers.

[0016] Furthermore, the specific process of multi-fold cross-fit dual machine learning is as follows: All journey event sequence samples are randomly divided into 5 equal and independent subsets; for the k-th subset, k takes values ​​from 1 to 5, and the remaining 4 subsets are used as training data to train a first auxiliary gradient boosting tree model to predict the treatment variable using confounding covariates, and a second auxiliary gradient boosting tree model to predict the outcome variable using confounding covariates; for each journey event sequence sample in the k-th subset, the deviation between the actual value of the treatment variable and the predicted output of the first auxiliary gradient boosting tree model is recorded as the treatment deviation, and the deviation between the actual value of the outcome variable and the predicted output of the second auxiliary gradient boosting tree model is recorded as the outcome deviation; after traversing all 5 subsets, the treatment deviation and outcome deviation of all samples are combined, and linear regression fitting is performed with the treatment deviation as the explanatory variable and the outcome deviation as the explained variable. The regression coefficient is the net causal attribution coefficient of the target channel.

[0017] Furthermore, for journey event sequences that have not yet included transformation events before the end of the unified prediction time domain, the Cox proportional hazards model is used with confounding covariates as input to estimate their cumulative transformation probability at the end of the unified prediction time domain. The cumulative transformation probability is used to replace the binary transformation indicator as the outcome variable to participate in the multi-fold cross-fit dual machine learning process to obtain the net causal attribution coefficient corrected for delayed transformation.

[0018] Furthermore, the Follow Regularized Leader Algorithm performs the following operations in each decision cycle: reads the actual conversion feedback data returned by each channel in the previous decision cycle, processes the actual conversion feedback data through steps 1 and 2, and updates the net causal attribution coefficient and cumulative attribution benefit concave utility function of each channel; uses the negative entropy function as a regularization term to adjust the budget allocation ratio vector along the gradient direction of the cumulative attribution benefit concave utility function, and projects the adjusted budget allocation ratio vector into the convex feasible region composed of global constraints, upper bound constraints, and lower bound constraints to obtain the optimal budget allocation ratio vector for the current decision cycle and issues budget instructions to each channel for execution; constrains the change amplitude of the budget allocation ratio vector between adjacent decision cycles through the negative entropy function, so that the regret value between the sum of the cumulative attribution benefit concave utility function values ​​of all channels and the ex-post optimal fixed allocation strategy in a continuous decision cycle sequence of arbitrary length approaches zero at a sublinear rate as the total number of decision cycles increases.

[0019] The user journey attribution and budget self-optimization method for e-commerce full-domain advertising of this invention has the following beneficial effects: This invention uses a neural Hawkes process to perform time-series incentive modeling on cross-channel touchpoint event sequences, explicitly decomposing the conditional intensity value of conversion events into two independent components: basic arrival intensity and touchpoint excitation increment intensity. The former characterizes the spontaneous changing trend of users' intrinsic purchase tendency, while the latter characterizes the external incentive effect exerted by advertising touchpoints on conversion probability and its decay process over time. This structural decomposition allows the model to accurately capture the differentiated incentive intensity of different channel touchpoints at different time positions on conversion events, overcoming the limitations of traditional Markov chain attribution that ignores time interval information and long-range dependencies. Simultaneously, this invention calculates the counterfactual conversion risk drop of each channel by defining a unified prediction time domain endpoint and a unified prediction interval, ensuring that the construction method of processing variables is completely consistent for both converted and non-converted journeys, avoiding the contamination of processing variable definitions by result information, and eliminating the risk of tag leakage from the source.

[0020] This invention employs multi-fold cross-fit dual machine learning to causally correct the counterfactual conversion risk reduction. Through two-stage residual processing, the combined effects of confounding covariates on the processing and outcome variables are separated, ensuring that the final estimated net causal attribution coefficient reflects only the pure causal contribution of channel touchpoints to the conversion event. This effectively eliminates the problem of systematically overestimating the attribution contribution of channels capturing high-intent users due to user self-selection bias. Furthermore, this invention incorporates the counterfactual conversion risk reduction of other channels into the confounding covariates, controlling for the interference of inter-channel synergy and collinearity on the single-channel attribution estimation, making the net causal attribution coefficients of each channel closer to the true independent causal contribution.

[0021] This invention constructs a concave utility function for the cumulative attribution revenue of each channel based on the net causal attribution coefficient, and employs a follow-regularized leader algorithm to solve for online concave utility maximization in each decision cycle. This tightly couples attribution modeling, causal bias correction, and budget allocation into a continuously iterative closed-loop system. The conversion feedback data for each decision cycle is updated in real-time after undergoing neural Hawkes process modeling and dual machine learning bias correction, enabling the budget allocation strategy to respond promptly to dynamic changes in the deployment environment. The follow-regularized leader algorithm constrains the change in budget allocation ratios between adjacent decision cycles through a negative entropy regularization term, theoretically ensuring that the cumulative regret value grows at a sublinear rate, meaning that the average regret per decision cycle approaches zero with iteration. This ensures that the long-term cumulative revenue of the budget allocation strategy in a non-stationary deployment environment gradually approaches the theoretical optimum. Attached Figure Description

[0022] Figure 1 A schematic diagram of the decomposition of the conditional intensity value of the conversion event along the time axis in cross-channel contact excitation modeling based on neural Hawkes processes provided in an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram illustrating the calculation principle of the risk reduction amount of counterfactual sequence perturbation and unified window counterfactual transformation provided in an embodiment of the present invention;

[0024] Figure 3 The scatter plot of residual-to-residual linear regression based on dual machine learning cross-fit is provided for an embodiment of the present invention. Detailed Implementation

[0025] A user journey attribution and budget self-optimization method for e-commerce full-domain advertising includes the following steps:

[0026] Step 1: Collect user behavior event streams from multiple channels and map them into a journey event sequence using user identifiers; construct an event embedding vector for each event in the journey event sequence and input it into a recurrent neural network; calculate the base arrival strength and touchpoint activation increment strength based on the hidden state vectors, and add the two to obtain the conversion event conditional strength value; define a unified prediction time domain endpoint, remove touchpoint events from each channel, and re-input the remaining sequence into the recurrent neural network; use the difference between the cumulative conversion event conditional strength values ​​of the original sequence and the counterfactual remaining sequence within the unified prediction interval as the unified window counterfactual conversion risk reduction amount;

[0027] Step 2: For each target channel, the unified window counterfactual conversion risk reduction amount is used as the processing variable, whether the journey event sequence contains conversion events before the unified prediction time domain end is used as the outcome variable, and user characteristics, environmental characteristics and the unified window counterfactual conversion risk reduction amount of other channels are used as confounding covariates. The net causal attribution coefficient is obtained through multi-fold cross-fit dual machine learning.

[0028] Step 3: Construct a concave utility function for cumulative attribution revenue for each channel. With the daily total budget, the upper limit of each channel's investment, and the minimum exploration investment ratio as constraints, the follow-regularized leader algorithm is used to update the concave utility function for cumulative attribution revenue in each decision cycle through steps 1 and 2. The budget allocation ratio vector is then adjusted by gradient and projected into the convex feasible region, forming an iterative closed loop of attribution modeling, causal bias removal, and budget allocation.

[0029] The following section provides a detailed explanation of the complete implementation process of cross-channel touchpoint incentive modeling and unified window counterfactual conversion risk calculation based on a specific e-commerce omni-channel advertising scenario.

[0030] In the actual operation of e-commerce omnichannel advertising, users often experience a complex behavioral path spanning multiple channels from their first encounter with an ad to their final purchase. For example, a user might first see a brand exposure ad in a social media feed, then actively search for and click on a paid ad on a search engine the next day to browse products on a landing page, and several hours later see the same product again through a recommended ad on a short video platform and add it to their shopping cart, finally completing the purchase the following day through an in-app push notification from the e-commerce platform. This complete behavioral path spans four channels and includes five different types of events: exposure, click, landing page visit, adding to cart, and placing an order, with time intervals between each event ranging from minutes to tens of hours. Traditional last-click attribution or linear average attribution cannot distinguish the differentiated incentive effects of each channel touchpoint on the conversion probability over time, and this method is designed to solve this problem.

[0031] First, the system collects and processes user behavior event streams from multiple channels. By connecting to data interfaces with various advertising channels, the system acquires user behavior event records in real-time or near real-time. Each event record contains at least five fields: user identifier, channel identifier, ad group identifier, event type, and event timestamp. The user identifier uniquely identifies the user and may exist in different forms across different channels; for example, one channel might use a device fingerprint identifier while another uses a phone number hash. Therefore, the system maintains a cross-channel user identifier mapping table, associating the identifiers of the same individual across different channels with a unified global user ID. The event type field covers six typical behaviors in e-commerce advertising scenarios: impressions, clicks, landing page visits, adding to cart, placing orders, and repeat purchases. Event timestamps are uniformly calibrated to Coordinated Universal Time (UTC) to eliminate time offsets caused by channel servers being deployed in different time zones. After completing the identifier mapping and time calibration, all event records under the same global user ID are sorted in ascending order of event timestamps, forming a complete journey event sequence. In a typical e-commerce platform with 5 million daily active users, approximately 30 million to 80 million event records can be generated daily. After identification and mapping, approximately 5 million journey event sequences can be formed, with each sequence having an average length of approximately 6 to 16 events.

[0032] Next, an event embedding vector is constructed for each event in the journey event sequence. The purpose of the event embedding vector is to encode discrete channel categories, event types, and continuous time interval information into the same dense real-number vector space, enabling the subsequent recurrent neural network to simultaneously perceive the source attribute and temporal structure of the event. Specifically, the system pre-initializes two embedding lookup tables: a channel embedding lookup table and an event type embedding lookup table. The number of rows in the channel embedding lookup table equals the total number of delivery channels. In a scenario with 8 delivery channels, the channel embedding lookup table contains 8 rows, each row being a real-number vector with a dimension of 32. The event type embedding lookup table has 6 rows, corresponding to the aforementioned 6 event types, and each row is also a real-number vector with a dimension of 32. For a given event in the journey event sequence, the corresponding row vector is retrieved from the channel embedding lookup table using its channel identifier as the index, thus obtaining the channel feature vector for that event; similarly, the corresponding row vector is retrieved from the event type embedding lookup table using its event type as the index, thus obtaining the type feature vector for that event.

[0033] The time-encoded vector is generated as follows: The time interval between the current event and the immediately preceding event in the journey event sequence is calculated and expressed as a scalar value in seconds. For the first event in the sequence, since there is no preceding event, its time interval is set to 0. This scalar time interval is then input into a single-layer linear projection layer. The linear projection layer has an input dimension of 1 and an output dimension of 32, thus mapping the one-dimensional time interval scalar to a time-encoded vector with a dimension of 32. In an optional implementation, the output dimension of the linear projection layer can be adjusted to 16 or 64, and the embedding dimensions of the channel embedding lookup table and the event type embedding lookup table are adjusted accordingly to maintain consistency among the three dimensions. Finally, the channel feature vector, type feature vector, and time-encoded vector are concatenated sequentially to form an event embedding vector with a dimension of 96. In the aforementioned optional implementation, if the dimension of each sub-vector is 16, the event embedding vector dimension is 48; if the dimension of each sub-vector is 64, the event embedding vector dimension is 192.

[0034] The reason for choosing concatenation instead of summation or element-wise multiplication to fuse the three sub-vectors is that the channel feature, event type feature, and time interval feature respectively characterize three orthogonal dimensions of the event: the source of the touchpoint, the nature of the behavior, and the temporal position. The concatenation operation preserves the complete information of each of the three dimensions, avoiding information confusion caused by the superposition of vector spaces, and enabling the subsequent recurrent neural network to autonomously learn the interaction patterns between the three dimensions during training.

[0035] After the event embedding vectors are constructed, the event embedding vectors of all events in the same journey event sequence are arranged strictly in ascending order according to the event timestamps and sequentially input into a gated recurrent unit (GRU) network. At each event position, the GRU network receives the event embedding vector of the current event as input and combines it with the hidden state vector passed from the previous event position for update calculation, outputting a new hidden state vector for the current event position. The hidden state dimension of the GRU network is set to 128, meaning that the hidden state vector output at each event position is a 128-dimensional real-valued vector. In optional implementations, the hidden state dimension can be set to 64 or 256. A larger hidden state dimension results in a stronger ability to express complex sequence patterns, but also increases the computational overhead of training and inference. Through its internal update and reset gate mechanisms, the GRU network can selectively remember or forget touchpoint information at earlier positions in the journey event sequence, thus possessing the ability to model both short-range and long-range dependencies. In optional implementations, the recurrent neural network can also be replaced with a long short-term memory (LSTM) network, which can also achieve the modeling of sequence dependencies.

[0036] Starting from the latent state vector at each event location, two independent scalar outputs need to be calculated: the base arrival strength and the touchpoint evoked incremental strength. The base arrival strength characterizes the probability that a conversion event might occur spontaneously due to the user's own purchasing inclination even without any advertising touchpoint incentives. The touchpoint evoked incremental strength measures the additional boost effect exerted by the presence of the current touchpoint event on the probability of subsequent conversion events. The fundamental purpose of separating these two, rather than directly outputting a single strength value, is to explicitly decouple the user's intrinsic conversion trend from the external incentive effect of advertising touchpoints at the model structure level. This decoupling is crucial for causal attribution and bias removal in subsequent steps. Without decoupling, the model's output strength value will be a mixture of components including the user's own purchasing inclination and the true contribution of advertising touchpoints, naturally contaminating the attribution results with confounding factors.

[0037] The calculation process for the base arrival strength is as follows: The hidden state vector is input into the first single-layer feedforward network. The input dimension of the first single-layer feedforward network is equal to the hidden state dimension, which is 128, and the output dimension is 1, generating a real scalar. This real scalar is then input into the softplus activation function. The softplus activation function is calculated as follows: ,in For the real scalar output of the first single-layer feedforward network, It is a natural constant. As a natural logarithm operation, the softplus activation function always outputs a positive and smooth value, ensuring that the base arrival strength is always numerically positive. The calculation process for the contact excitation increment strength is completely symmetrical: the same hidden state vector is input into the second single-layer feedforward network. The structure of the second single-layer feedforward network is the same as that of the first single-layer feedforward network, but the parameters are independent. After outputting a real scalar, it is also processed by the softplus activation function to obtain the contact excitation increment strength. The base arrival strength and the contact excitation increment strength are directly added together to obtain the conditional strength value of the transition event at that event location. The reason for using softplus instead of exponential or modified linear unit functions is that exponential functions are prone to numerical overflow when the input value is large, and modified linear unit functions always output zero and are not differentiable at zero when the input value is negative. The softplus function, on the other hand, ensures that the output is always positive and is smooth and differentiable over the entire domain, taking into account both numerical stability and optimization friendliness.

[0038] All trainable parameters, including the channel embedding lookup table, event type embedding lookup table, single-layer linear projection layer, gated recurrent unit network, first single-layer feedforward network, and second single-layer feedforward network, are jointly optimized end-to-end. The training objective consists of two parts: Part 1 aims to maximize the conditional intensity values ​​of the transition events at the actual occurrence times of all observed events in the journey event sequence. Intuitively, this objective prompts the model to give the highest possible intensity prediction at the actual occurrence times, meaning the model considers it reasonable for events to occur at these times. Part 2 aims to minimize the piecewise cumulative summation of the conditional intensity values ​​of the transition events along the time axis between adjacent events in the journey event sequence. This objective constrains the model to avoid giving excessively high intensity predictions during periods without events, preventing the model from inflating the intensity across the entire time axis to appease the degenerate behavior of Part 1. These two objectives together constitute the standard log-likelihood training criterion for neural Hawkes processes. The training process uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 256 journey event sequences. Early stopping is determined by whether the log-likelihood value on the validation set fails to improve for three consecutive epochs. On the aforementioned dataset of 5 million journey event sequences, convergence typically occurs after 15 to 30 training epochs. In an optional implementation, the learning rate can be set to 0.0005 or 0.002, the batch size can be adjusted to 128 or 512, and the patience epoch for early stopping can be set to 5.

[0039] Reference 1, Figure 1The upper part of the chart displays a complete timeline of user journey events, with the horizontal axis representing time in hours. Eight events are arranged chronologically on the timeline: exposure event on channel A, click event on channel A, exposure event on channel B, click event on channel B, landing page visit event on channel C, add-to-cart event on channel C, exposure event on channel A, and the final order conversion event. Time intervals are marked between adjacent events; for example, the interval between the first exposure event and the second click event is 4 hours, reflecting the user's cross-time behavior across different channels. Figure 1 The lower half of the graph shows the curves of the conversion event conditional strength value over time, corresponding to the event sequence in the upper half. The vertical axis represents the conditional strength value, and the horizontal axis is aligned with the time axis of the upper half. The graph decomposes the conditional strength value into two superimposed components: the lower region represents the base arrival strength, characterizing the tendency for conversion events to arrive spontaneously without considering any advertising touchpoint incentives. This component exhibits a slow and steady change along the time axis, reflecting the slow drift of users' intrinsic purchase intentions; the upper superimposed region represents the touchpoint activation increment strength, characterizing the additional boost effect exerted on the probability of conversion event arrival after each touchpoint event occurs. It can be observed that whenever a touchpoint event occurs in the journey event sequence, the touchpoint activation increment strength experiences a steep jump at that moment, followed by a gradual decline in an exponential manner. This "jump-decay" pattern is precisely the temporal incentive characteristic of touchpoints on subsequent events in the Hawkes process. The upper contour obtained by adding the base arrival strength and the touchpoint activation increment strength is the conversion event conditional strength value, which is marked with a solid dot at each event occurrence. Figure 1 As can be clearly seen, as the user journey progresses to the add-to-cart stage, the incentive effects of multiple channel touchpoints are superimposed layer by layer, the intensity value of the conversion event condition rises to a high level, and finally reaches the peak area within the sequence at the moment of the order placement event, which reflects the gradual cumulative driving effect of cross-channel touchpoint collaborative incentives on conversion probability.

[0040] After model training, the process moves to the counterfactual sequence perturbation and unified window transformation risk calculation stage. First, a unified prediction time domain endpoint is defined for each journey event sequence. The unified prediction time domain endpoint is set as follows: take the timestamp of the last event in the journey event sequence, and extend it by a fixed-length observation extension window. The end time of this extension is the unified prediction time domain endpoint. The length of the observation extension window is determined based on the typical delay time from the user's last touchpoint to completing the order in the business scenario. In fast-moving consumer goods e-commerce scenarios, it can be set to 72 hours; in large appliances or high-priced goods scenarios, it can be set to 168 hours or 336 hours. The time span from the last event timestamp to the unified prediction time domain endpoint constitutes the unified prediction interval. The reason for choosing a unified prediction interval instead of using the timestamps of actual transformation events in each journey as the calculation benchmark is that the latter approach would cause the definition of processing variables to depend on whether the transformation event has occurred, resulting in outcome leakage. For transformed journeys, there are explicit transformation timestamps available, but for untransformed journeys, there are no such timestamps. The construction mechanism of processing variables for the two types of journeys is unequal, leading to systematic bias in subsequent causal estimations. With a unified prediction interval, regardless of whether the journey ultimately transforms, the processing variables are derived using the same time window and the same calculation method, eliminating the contamination of processing variable definitions by outcome information.

[0041] The counterfactual sequence perturbation process is as follows: Assume a journey event sequence involves three delivery channels: Channel A, Channel B, and Channel C. When performing counterfactual perturbation on Channel A, all touchpoint events in the journey event sequence identified as Channel A are removed, retaining only touchpoint events from Channels B and C, as well as all non-touchpoint events, while keeping the original timestamps of the retained events unchanged. The remaining event sequence after removal is then re-input into the same pre-trained gated recurrent unit network and intensity output layer in chronological order, and the conditional intensity values ​​of the conversion events under counterfactual conditions are calculated step-by-step within a unified prediction interval.

[0042] The cumulative conversion risk values ​​for the original sequence and the counterfactual residual sequence are calculated separately within the unified prediction interval. Specifically, the unified prediction interval is divided into several discrete time steps. In practice, the step size can be set to 1 hour, resulting in 72 time steps for a 72-hour observation window. At each time step, the conditional strength value of the conversion event is read, and the conditional strength values ​​of the conversion event for all time steps within the unified prediction interval are accumulated and summed segment by segment at intervals equal to the step size. This accumulation and summation operation is essentially a discrete approximation of the definite integral of the continuous time intensity function over the unified prediction interval. This operation is performed on the original sequence to obtain the unified window cumulative conversion risk value for the original sequence, and on the counterfactual residual sequence to obtain the unified window cumulative conversion risk value for the counterfactual residual sequence. The difference between the unified window cumulative conversion risk value of the original sequence and the unified window cumulative conversion risk value of the counterfactual residual sequence is the unified window counterfactual conversion risk reduction amount for channel A on this user journey. The physical meaning of this dropout amount is: how much the risk of a conversion event for the user would decrease within the uniform prediction interval if all touchpoints of channel A were completely removed from the user's actual behavior path. A larger dropout amount indicates a stronger incentive effect of channel A on conversion events during the user's journey.

[0043] The counterfactual sequence perturbation and unified window conversion risk drop calculation process is repeated for each channel involved in the same journey event sequence. In the example of the three channels mentioned above, one counterfactual perturbation is required for each of channels A, B, and C, resulting in three unified window counterfactual conversion risk drop amounts. In actual deployment, if the total number of channels is eight, then eight counterfactual perturbations are required for each journey event sequence. To reduce computational overhead, optional implementation methods include: performing counterfactual perturbations only on channels that have actually had contact points in the journey event sequence, skipping channels not involved in the journey and directly setting their unified window counterfactual conversion risk drop amounts to 0; or using a batch parallel computing method, grouping multiple counterfactual residual sequences of the same journey event sequence into a batch and simultaneously inputting them into a gated recurrent unit network for forward inference, utilizing the parallel computing capabilities of the graphics processor to accelerate processing.

[0044] Following the above process, a uniform window counterfactual conversion risk scalar value is obtained for each channel involved in each journey event sequence. This scalar value will be used as a processing variable input in the subsequent channel causal attribution debiasing estimation step to further isolate the influence of confounding factors and estimate the net causal contribution of each channel to the conversion event.

[0045] refer to Figure 2 The horizontal axis represents time in hours, and the vertical axis represents the intensity value of the transformation event condition. Figure 2The left side of the graph displays the conditional intensity changes of the user journey event sequence during the actual observation period. Multiple event markers are distributed along the curves, distinguishing touchpoint events for channels A, B, and C using circles, squares, and triangles, respectively, with each marker indicating its channel affiliation. At the last event timestamp of the journey event sequence, a vertical dividing line is drawn. The interval between this line and another vertical dividing line at the end of the unified prediction time domain is the unified prediction interval, marked with a double-headed arrow and text labeling. The end of the unified prediction time domain is determined by a fixed-length observation extension window extending backward from the last event timestamp. Within the unified prediction interval, two intensity curves are plotted: the solid line represents the decay trend of the conversion event conditional intensity value of the original journey event sequence within the unified prediction interval over time; the dashed line represents the trend of the conversion event conditional intensity value of the remaining event sequence within the same unified prediction interval after removing all touchpoint events from channel A. Due to the removal of the incentive contribution from channel A, the dashed line lies entirely below the solid line, forming a sandwich region that expands along the time direction between the two curves. The area of ​​this mezzanine region represents the unified window counterfactual conversion risk reduction for Channel A on this user journey. Its physical meaning is: if all touchpoints of Channel A were completely removed from the user's actual behavioral path, the cumulative risk of conversion events occurring within the unified prediction interval would decrease. In the diagram, an arrow pointing from the mezzanine region to the text label clearly indicates the unified window counterfactual conversion risk reduction represented by this area. Figure 2 It can be seen that the strength difference between the original sequence and the counterfactual residual sequence is largest at the beginning of the unified prediction interval. As time progresses, the two curves gradually converge, indicating that the incentive effect of channel A touchpoint decays over time. Using a unified prediction interval instead of the actual conversion event timestamp as the calculation benchmark ensures that regardless of whether the journey ultimately results in a conversion, the unified window counterfactual conversion risk drop is obtained using the same time window and the same calculation method, avoiding contamination of the processing variable definition by the result information.

[0046] The following section, using the unified window counterfactual conversion risk reduction figure generated in step 1, provides a detailed explanation of the complete implementation process of channel causal attribution debiasing estimation based on dual machine learning cross-fitting.

[0047] In e-commerce omnichannel advertising, the user groups reached by different channels inherently differ systematically. For example, search engine bidding ads naturally capture users with a clear purchase intention; these users have a high probability of spontaneous conversion even without clicking the ads. Social media feed ads, on the other hand, often reach users still exploring their interests, whose spontaneous conversion probability is relatively low. If the uniform window counterfactual conversion risk reduction output in step 1 is directly considered as the channel's attribution contribution, the contribution of the search engine channel will be systematically overestimated because its reduction is mixed with the influence of users' own high purchase intention. The core value of dual machine learning lies in using two-stage residual processing to separate the combined effects of confounding factors on the processing and outcome variables, ensuring that the final estimated attribution coefficient only reflects the net causal effect of the channel touchpoint on the conversion event.

[0048] First, let's define the three types of variables. The processing variable is the uniform window counterfactual conversion risk drop for the target channel on each user journey, obtained in step 1, which is a continuous non-negative real number. The outcome variable is whether the journey event sequence contains a conversion event before the end of the uniform prediction time domain. A value of 1 indicates that an order placement event occurred in the journey event sequence before the end of the uniform prediction time domain, and a value of 0 indicates that it did not occur. The confounding covariates consist of three sets of features: The first set comprises user characteristics, including user demographic features such as gender, age group, and city level; device type to distinguish between mobile and desktop devices; access source to distinguish between organic and paid traffic; continuous values ​​for single session dwell time in seconds; and a new / returning customer flag (0 or 1). The second set comprises environmental characteristics, including a promotional period flag (0 or 1) to indicate whether the current journey is within a major promotional period; weekday codes (1 to 7); time period codes (0 to 23 corresponding to the 24 hourly intervals of the day); product price range flags categorizing products into low, mid, and high price ranges based on average order value; and discount percentage flags representing continuous values. The third set represents the counterfactual conversion risk reduction for all channels except the target channel. In a scenario with eight advertising channels, when the target channel is channel A, the third set of confounding covariates includes seven continuous values ​​from channel B to channel H. The purpose of including the counterfactual conversion risk reduction of other channels in the confounding covariates is to control for the linkage effect and collinearity interference between channels. In e-commerce advertising, there are often synergistic or substitutive relationships between channels. For example, exposure on a product seeding channel may increase the click probability on a search channel. If the incentive effect of other channels is not controlled, the attribution coefficient of the target channel will include indirect spillover effects from other channels, causing the estimated results to deviate from the true net causal contribution of a single channel.

[0049] After concatenating the three sets of features, the total dimensions of the confounding covariate vector in the above eight channel scenarios are: five dimensions of user features, five dimensions of environmental features, and seven dimensions of the counterfactual conversion risk drop from the unified window of the remaining seven channels, totaling 17 dimensions. In an optional implementation, if the business needs to introduce additional confounding control variables, such as the cumulative purchase frequency of users in the past 30 days or the distribution vector of browsing categories in the past 7 days, these can be added to the confounding covariates, and the total dimensions will increase accordingly.

[0050] The specific execution process of multi-fold cross-fit dual machine learning is as follows. The entire set of user journey samples is randomly shuffled and uniformly divided into five equal-sized, non-overlapping subsets, denoted as subset 1 to subset 5. The fundamental reason for using a 5-fold partition, rather than simply using all samples simultaneously for both auxiliary model training and deviation calculation, is that if the auxiliary model is trained and deviation is calculated on the same batch of samples, the auxiliary model will tend to overfit the training samples, causing the deviation to be artificially compressed, resulting in regularization bias in the final net causal attribution coefficient estimate. Cross-fit eliminates this overfitting propagation path by strictly ensuring that the training data of the auxiliary model and the data for deviation calculation do not overlap. In optional implementations, the number of folds can be set to 3 or 10 folds. More folds result in a larger proportion of the training set per fold and higher fitting accuracy of the auxiliary model, but the computational cost also increases exponentially.

[0051] Taking the first subset as the currently retained fold as an example, the single-fold operation process is illustrated. Subsets 2, 3, 4, and 5 are merged into a training set, on which two independent auxiliary prediction models are trained. The first auxiliary gradient boosting tree model uses a 17-dimensional vector of confounding covariates as input features and is trained with the processing variable, i.e., the uniform window counterfactual conversion risk drop of the target channel, as the prediction label. The second auxiliary gradient boosting tree model also uses a 17-dimensional vector of confounding covariates as input features, but is trained with the outcome variable, i.e., whether the journey event sequence contains a conversion event before the unified prediction time domain endpoint, as the prediction label. The hyperparameters of the two auxiliary gradient boosting tree models are set independently: the maximum tree depth is set to 6, the minimum number of leaf nodes is set to 20, the learning rate is set to 0.05, the number of iterations is set to 500, and the early stopping condition is that the loss value on the validation set does not decrease for 50 consecutive iterations. In alternative implementations, the auxiliary prediction model can also be replaced by a random forest model or a multilayer perceptron model. Gradient boosting tree models typically have both good fitting ability and robustness to missing values ​​when dealing with mixed covariates of medium dimensionality, and are therefore preferred in this method.

[0052] After the two auxiliary gradient boosting tree models are trained, predictions are made for each journey sample in the first subset. The difference between the actual value of the processing variable for each journey sample and the prediction output of the first auxiliary gradient boosting tree model for that sample is the processing deviation for that sample, denoted as . ,in Assign the index number of the journey sample in the first subset. This refers to the remaining unpredictable variance in the processed variables after removing the explainable portion of confounding covariates. Similarly, the difference between the actual value of the outcome variable for each journey sample and the predicted output of the second auxiliary gradient boosting tree model for that sample is the outcome deviation for that sample, denoted as . ,in This indicates the remaining unpredictable variance in the outcome variable after removing the explanatory portion of the confounding covariates.

[0053] The above operation is repeated for subsets 2 through 5. Each time, the current subset is used as the retained fold, and the remaining 4 subsets are combined into the training set to retrain two auxiliary gradient boosting tree models. The processing deviation and the result deviation are calculated for the retained fold samples. After traversing all 5 subsets, each journey sample has exactly one processing deviation and one result deviation calculated, and no sample participates in both auxiliary model training and deviation calculation simultaneously.

[0054] Processing deviation of all journey samples Deviation from the result The model is then fitted using ordinary least squares linear regression with the deviation from the outcome as the dependent variable and the deviation from the treatment variable as the explanatory variable. This linear regression model does not include an intercept term and only fits a slope coefficient. The fitting objective is to make all samples fit. The sum of the squares reaches its minimum value. The solution obtained... This is the net causal attribution coefficient of the target channel. This means that, after controlling for the effects of all confounding covariates, for every unit increase in the uniform window counterfactual conversion risk reduction for the target channel, the average change in the probability that the user journey includes a conversion event before the end of the uniform prediction time domain is: Units. A positive value and a larger value indicate a stronger net causal promotion effect of the target channel on the conversion event.

[0055] The reason for using linear regression without an intercept term in the final stage instead of a more complex nonlinear model stems from the core design of the dual machine learning theory: two-stage residualization has fully absorbed the nonlinear effects of confounding factors into the auxiliary gradient boosting tree model. The residual relationship between the deviation and the result deviation is theoretically close to linear. At this point, a simple linear regression can be used to obtain a model with... A consistent estimator of the convergence rate, where This represents the total number of user journey samples. Forcing a non-linear model for the final fit may introduce additional estimation variance without improving bias.

[0056] With a data scale of 5 million journey event sequences, each fold in the above 5-fold cross-fit process contains approximately 4 million samples in the training set, and the retained fold contains approximately 1 million samples. The training time for a single auxiliary gradient boosting tree model on 4 million samples and 17-dimensional features is approximately 3 to 8 minutes. The complete 5-fold cross-fit process requires training 10 auxiliary gradient boosting tree models. Including deviation calculation and final linear regression fitting, the total processing time for a single target channel is approximately 30 to 80 minutes. If there are a total of 8 delivery channels, the above process needs to be performed on each of the 8 channels, with a total time of approximately 4 to 11 hours. In an optional implementation, the cross-fit process for the 8 channels can be executed in parallel to reduce the total time.

[0057] refer to Figure 3 The horizontal axis represents the treatment deviation, and the vertical axis represents the outcome deviation. The treatment deviation is the difference between the actual value of the unified window counterfactual conversion risk drop of the target channel and the predicted output of the first auxiliary gradient boosting tree model, representing the residual variance in the treatment variable after removing the explainable components of confounding covariates. The outcome deviation is the difference between the actual value of whether the journey event sequence includes a conversion event before the unified prediction time domain endpoint and the predicted output of the second auxiliary gradient boosting tree model, representing the residual variance in the outcome variable after removing the explainable components of confounding covariates. The scatter points in the figure are divided into 5 groups, corresponding to the 5 subsets in the 5-fold crossfit, distinguished by different shaped symbols: circle, square, triangle, rhombus, and inverted triangle, respectively. The 5 groups of scatter points are approximately symmetrically distributed around the origin on both the horizontal and vertical axes, indicating that the two auxiliary gradient boosting tree models have effectively absorbed the predictive power of confounding covariates on the treatment and outcome variables, and the residual deviation no longer carries the systematic bias of confounding factors. The figure shows a straight line passing through the origin, representing the fitted line obtained by performing ordinary least squares linear regression without an intercept term, using the treatment deviation of all samples as the explanatory variable and the outcome deviation as the explained variable. The slope of this fitted line is the net causal attribution coefficient of the target channel. , The numerical meaning is: after controlling for the effects of all confounding covariates, for every unit increase in the uniform window counterfactual conversion risk reduction of the target channel, the average change in the probability that the journey event sequence contains a conversion event before the end of the uniform prediction time domain changes. The dispersion of the scatter points around the fitted line reflects the noise level of the estimate; the smaller the dispersion, the higher the accuracy of the net causal attribution coefficient estimate.

[0058] For journey event sequences within the delayed conversion observation period that have not yet included a conversion event before the unified prediction time end, the observed value of the outcome variable is 0. However, this observation value does not mean that the user will absolutely not convert, but only that no conversion event has been observed at the current observation deadline. If the outcome variable is directly set to 0 and used in the dual machine learning process, the net causal attribution coefficient of each channel will be systematically underestimated because some journeys that should have converted in the future are incorrectly included in the unconverted samples. To correct this delayed conversion bias, a Cox proportional hazards model is used to compensate for the conversion probability estimation for such journeys. Specifically, a 17-dimensional vector of confounding covariates is used as the covariate input, the remaining observation time between the last event timestamp in the journey event sequence and the unified prediction time end is used as the time variable, and whether a conversion event has been observed is used as the state indicator variable. A Cox proportional hazards model is fitted on all journey samples. After fitting, for each journey event sequence that has not yet included a conversion event before the unified prediction time end, the cumulative conversion probability at the unified prediction time end is estimated using the Cox proportional hazards model. The cumulative conversion probability is a continuous value between 0 and 1, which replaces the original binary conversion indicator of 0 as the outcome variable in the multi-fold cross-fit dual machine learning process. The regression coefficient obtained after this replacement is the net causal attribution coefficient corrected for delayed conversion, which more accurately reflects the true causal contribution of channel touchpoints throughout the complete conversion cycle compared to the uncorrected version. In an optional implementation, the Cox proportional hazards model can also be replaced by a parameterized Weibull accelerated failure time model or a discrete-time hazard rate model. The Cox proportional hazards model is preferred in this method because it does not require prior assumptions about the specific form of the baseline hazard function and has strong versatility.

[0059] After completing the above process for all eight advertising channels, each channel obtains one net causal attribution coefficient corrected for delayed conversion. These eight net causal attribution coefficients together constitute a channel-level causal attribution vector, which will serve as the core input for the subsequent online concave utility maximization budget dynamic allocation step, guiding the dynamic reallocation of the budget among the channels.

[0060] The following section, using the net causal attribution coefficients of each channel after delay conversion correction generated in step 2, provides a detailed explanation of the complete implementation process of online concave utility maximization budget dynamic allocation based on regret boundary convergence guarantee.

[0061] Budget allocation for e-commerce advertising across all channels is essentially a dynamic decision-making problem: the advertising environment constantly changes with promotional pace, competitive landscape, and user behavior patterns. Any fixed allocation strategy based on a single offline calculation will gradually become ineffective as the environment drifts. At the same time, advertisers cannot afford large-scale budget waste during the process of exploring the optimal allocation solution. This contradiction requires budget allocation algorithms to adapt to environmental changes online while ensuring that long-term cumulative returns are not significantly lower than the optimal strategy in retrospect. The Follow Regularized Leader algorithm is designed to address these needs. It constrains the magnitude of strategy changes by introducing a regularization term into the optimization objective of each decision cycle, achieving a balance between exploration and utilization. Theoretically, it guarantees that the cumulative regret value grows at a sublinear rate, meaning that the average regret per decision cycle approaches zero over time.

[0062] First, construct the concave utility function for the cumulative attribution revenue of each channel. (Based on channel...) For example, Values ​​range from 1 to , In this embodiment, the total number of distribution channels is [number]. Equals 8. Channel The budget amount obtained within a certain decision-making cycle is denoted as The unit is yuan. Channel The cumulative attribution benefit concave utility function is denoted as Defined as ,in The channel obtained in step 2 Net causal attribution coefficient after delayed transformation correction, For channels The average number of touchpoints generated per unit budget, i.e., the average number of impressions or clicks that can be generated per dollar of budget. For channels The saturation scaling parameter is used to control the rate of diminishing returns. It is the natural logarithm function. The intrinsic properties of the natural logarithm function determine that... about growth rate with The marginal returns decrease as the budget increases, meaning that the incremental revenue from each additional dollar of budget gradually shrinks. This is consistent with the diminishing marginal returns phenomenon that is prevalent in e-commerce advertising: when the advertising volume on a certain channel is already large, the quality of incremental users reached by continuing to increase the budget decreases, the pressure to control the frequency increases, and the bidding costs rise, resulting in diminishing marginal returns. It is about The strictly concave function is a mathematical premise that subsequent online optimization problems are solvable and have regret bound guarantees.

[0063] Saturation scaling parameters The value of this parameter is related to the channel's traffic capacity and the bidding environment. In actual deployment, This can be achieved by analyzing channels over multiple past decision-making cycles. The actual touchpoint output data under different budget levels was obtained through nonlinear least squares fitting. In a typical scenario, the search engine bidding channel... It could be 5000 yuan, from social media information flow channels. It might cost 20,000 yuan, through short video platform channels. It could be 15,000 yuan. In an optional implementation, the cumulative attribution benefit concave utility function can also be in the form of a power function. ,in The concavity index is between 0 and 1. The smaller the value, the more drastic the diminishing returns. Both logarithmic and power functions satisfy the strict concavity requirement. The logarithmic form shows a more gradual decrease when the budget is large, while the power function form grows faster when the budget is close to zero. The choice can be made based on the characteristics of the business scenario.

[0064] The marginal revenue estimation function for each channel is an incremental approximation of the cumulative attribution revenue concave utility function with respect to the budget allocation. Taking the logarithmic form as an example, the channel... The marginal revenue estimation function is right The derivative, i.e. This indicates the current budget amount. The incremental causal attribution benefit obtained by adding 1 unit of budget at the current level. This derivative value follows... It increases but decreases strictly, consistent with the properties of concave utility functions.

[0065] Next, we define the constraints for the online concave utility maximization problem. Let the decision cycle number be denoted as... , Starting from 1 and increasing incrementally, the time span of each decision cycle is set to 1 day in this embodiment. In each decision cycle... In this process, the system needs to determine a budget allocation ratio vector. ,in To be allocated to channels The budget allocation ratio is the proportion of the total daily budget. The budget allocation ratio vector needs to simultaneously satisfy three types of constraints: the global constraint requires that the sum of the allocation ratios for all channels equals 1, i.e. This ensures that the total daily budget is fully allocated without leaving any unused balance; the upper bound constraint requires that the allocation ratio for each channel does not exceed the ratio corresponding to the daily spending limit for that channel, denoted as... ,in For channels The daily spending cap is divided by the total daily budget. In a scenario with a total daily budget of 100,000 yuan, if the daily spending cap for channel A is 30,000 yuan, then... Equals 0.3; the lower bound constraint requires that the allocation ratio for each channel is not less than the minimum exploration deployment ratio for that channel, denoted as . ,in For channels The minimum exploration allocation ratio is set to ensure that each channel receives a certain amount of budget in each decision-making cycle to continuously generate conversion feedback data. This prevents a channel from losing its data update capability due to extremely low allocation ratios for several consecutive cycles, which could lead to a long-term inability to correct its net causal attribution coefficient and cumulative attribution benefit concave utility function, resulting in an "information starvation" state. In actual deployment, the minimum exploration allocation ratio is usually set between 0.02 and 0.05, meaning each channel receives at least 2% to 5% of the total daily budget. The above three types of constraints together define a convex polyhedron as the convex feasible region of the budget allocation ratio vector.

[0066] In each decision cycle At the start time, follow the regularized leader algorithm to perform the following operations. First, read the previous decision cycle. Actual conversion feedback data returned from various channels, including data from each channel during the decision-making cycle. The actual budget allocated to ad placement, the number of impressions and clicks generated, and the number of order conversions observed to date are used as inputs into the neural Hawkes process modeling workflow in step 1 for new journey event sequences. This calculates the updated counterfactual conversion risk drop for each channel within a unified window, and then, through a multi-fold cross-fit dual machine learning process in step 2, obtains the updated net causal attribution coefficients for each channel after delayed conversion correction. Then with the updated Recalculate the concave utility function of cumulative attribution benefits for each channel. And the corresponding marginal revenue estimation function. In actual deployment, since the complete retraining of steps 1 and 2 is time-consuming, an incremental update strategy can be adopted: the gated recurrent unit network is only fine-tuned a few rounds on the newly added journey event sequences instead of being completely retrained, and the auxiliary gradient boosting tree model is incrementally trained on the newly added samples, thereby controlling the update time of each decision cycle within an acceptable range. In an optional implementation, the decision cycle can also be set to half a day or 1 hour. The shorter the cycle, the more sensitive the budget allocation is to environmental changes, but the higher the requirements for the update speed of steps 1 and 2.

[0067] After the utility function is updated, the optimal budget allocation ratio vector for the current decision period is calculated. The core idea of ​​the Follow Regularized Leader algorithm is: in each decision period, select the allocation ratio vector that maximizes the sum of all observed cumulative utilities up to the current time plus the sum of the regularization term. The regularization term uses a negative entropy function, defined as follows: ,in The regularization intensity coefficient is . For channels Budget allocation ratio, for The natural logarithm of . Because The value is between 0 and 1. The negative entropy term is always non-positive; the entire term is consistently non-positive. Adding a negative entropy term to the maximization objective penalizes overly concentrated allocation schemes—when the allocation ratio vector is highly concentrated in a few channels, the absolute value of the negative entropy term is larger, resulting in a greater drag on the objective function; conversely, when the allocation ratio vector is more evenly distributed, the absolute value of the negative entropy term is smaller. This penalty mechanism naturally encourages a moderately dispersed allocation scheme across channels, consistent with the business requirement to maintain exploration diversity in budget allocation. Regularization strength coefficient. The balance between exploration and exploitation was controlled: The larger the value, the stronger the negative entropy penalty, the more evenly the distribution ratio tends to be distributed, and the greater the exploration effort. The smaller the value, the more the algorithm tends to concentrate the budget on the most efficient channels, resulting in greater utilization. In actual deployments... The initial value is usually set between 0.01 and 0.1, and can be adjusted according to the number of decision cycles. Decreasing in a manner that, among which The initial regularization strength coefficient is , Number the current decision-making cycle. for The square root of the value. This decreasing strategy involves a greater initial exploration effort to fully collect feedback from various channels, followed by a gradual increase in utilization to improve returns as data accumulates and attribution estimates stabilize.

[0068] The specific gradient adjustment and projection operations are as follows. Based on the decision period up to... Given all the cumulative utility gradient information up to this point and the currently updated cumulative attribution benefit concave utility function, calculate the gradient of the objective function with respect to the budget allocation ratio vector. Under the logarithmic form of the utility function, the objective function with respect to... The gradient components are ,in This represents the total daily budget. The budget allocation ratio vector from the previous period is adjusted along the gradient direction to obtain a preliminary updated ratio vector. This preliminary updated ratio vector may not satisfy one or more of the global constraints, upper bound constraints, or lower bound constraints, therefore it needs to be projected into the convex feasible region. The projection operation finds the point within the convex feasible region closest to the preliminary updated ratio vector, using the Bregman divergence (Kullback-Leibler divergence) corresponding to the negative entropy regularization term as the distance metric. The projection is specifically implemented by iteratively adjusting the Lagrange multipliers to simultaneously satisfy the global constraints and upper and lower bound constraints. In a scenario with 8 channels, it typically converges to a feasible solution satisfying all constraints after 10 to 20 iterations. The ratio vector obtained after projection is the current decision-making cycle. Optimal budget allocation ratio vector .

[0069] Allocation ratio vector according to the optimal budget Budget directives were issued to various channels for execution. (Channels) In the decision-making cycle The budget amount obtained within is In a scenario with a daily total budget of 100,000 yuan and 8 channels, assuming the optimal budget allocation ratio vector for the current decision-making cycle is calculated as follows: The eight channels will receive budgets of 18,000 yuan, 25,000 yuan, 12,000 yuan, 15,000 yuan, 8,000 yuan, 7,000 yuan, 5,000 yuan, and 10,000 yuan respectively. After the campaign is completed, each channel will send back actual conversion feedback data, and the system will enter the next decision-making cycle. The update process forms a continuous iterative closed loop between attribution modeling, causal bias correction, and budget allocation.

[0070] The follower-regularized leader algorithm, under negative entropy regularization, has the following convergence guarantee: For any length of... In a continuous decision-making cycle sequence, the sum of the concave utility function values ​​of the cumulative attribution benefits across all channels and the ex-post optimal fixed allocation strategy, i.e., reviewing all... The difference between the sum of cumulative utility obtained from the single optimal fixed-proportion vector selected after a certain number of cycles is the regret value. growth The rate of growth, of which The upper bound notation indicates an asymptotic upper bound. The total number of decision-making cycles. Total number of channels. Divide the regret value by... That is, the average regret per cycle is This value follows The increase approaches zero, indicating that the long-term average performance of the budget allocation strategy gradually approaches the ex-post optimal level. In an 8-channel scenario with a daily decision cycle, after approximately 90 days (90 decision cycles) of iteration, the average regret per cycle is about one-tenth of the initial regret value. The budget allocation strategy is basically stable within a near-optimal range, while still retaining its adaptability to environmental changes.

[0071] In an optional implementation, the regularization term can also be replaced with the square of the Euclidean norm, i.e. In this case, the projection operation simplifies to standard Euclidean projection. However, negative entropy regularization is more natural in scenarios where the budget allocation ratio vector lies on the probability simplex, because the negative entropy function naturally adapts to the geometry of the probability distribution space. The resulting projection update has a multiplicative adjustment characteristic, meaning that channels with higher ratios receive larger absolute adjustments while channels with lower ratios receive smaller absolute adjustments, avoiding the problem of small-ratio channels being easily compressed to the lower bound under Euclidean projection. Furthermore, in optional implementations, the follow-regularized leader algorithm can be replaced with an online mirror descent algorithm. Both have the same update rule form under negative entropy regularization, differing only slightly in the way gradient information is used. In the application scenario of this method, their actual performance is basically the same.

[0072] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A user journey attribution and budget self-optimization method for e-commerce full-domain advertising, characterized by: Includes the following steps: Step 1: Collect user behavior event streams from multiple channels and form a journey event sequence by mapping user identifiers; construct an event embedding vector for each event in the journey event sequence and input it into a recurrent neural network; calculate the basic arrival intensity and touchpoint activation increment intensity based on the hidden state vectors, and add the two to obtain the conversion event conditional intensity value. Define a unified prediction time domain endpoint, remove touchpoint events through each channel, and re-input the remaining sequence into a recurrent neural network. Use the difference between the cumulative value of the transformation event condition intensity of the original sequence and the counterfactual remaining sequence within the unified prediction interval as the unified window counterfactual transformation risk reduction amount. Step 2: For each target channel, the unified window counterfactual conversion risk reduction amount is used as the processing variable, whether the journey event sequence contains conversion events before the unified prediction time domain end is used as the outcome variable, and user characteristics, environmental characteristics and the unified window counterfactual conversion risk reduction amount of other channels are used as confounding covariates. The net causal attribution coefficient is obtained through multi-fold cross-fit dual machine learning. Step 3: Construct a concave utility function for cumulative attribution revenue for each channel. With the daily total budget, the upper limit of each channel's investment, and the minimum exploration investment ratio as constraints, the follow-regularized leader algorithm is used to update the concave utility function for cumulative attribution revenue in each decision cycle through steps 1 and 2. The budget allocation ratio vector is then adjusted by gradient and projected into the convex feasible region, forming an iterative closed loop of attribution modeling, causal bias removal, and budget allocation.

2. The method according to claim 1, characterized in that, Each event record in the user behavior event stream includes a user identifier, channel identifier, ad group identifier, event type, and event timestamp. Event types cover impressions, clicks, landing page visits, add-to-cart, order placement, and repeat purchases.

3. The method according to claim 1, characterized in that, The event embedding vector is constructed as follows: the channel feature vector is obtained by retrieving the channel embedding lookup table with the channel identifier of the event; the type feature vector is obtained by retrieving the event type embedding lookup table with the event type; the time interval between the current event and the preceding adjacent event is input into a single-layer linear projection layer and mapped into a time encoding vector; the channel feature vector, type feature vector and time encoding vector are concatenated to form the event embedding vector.

4. The method according to claim 1, characterized in that, The recurrent neural network is a gated recurrent unit network; the hidden state vector is input into the first single-layer feedforward network and then activated by the softplus activation function to obtain the basic arrival strength; the hidden state vector is input into the second single-layer feedforward network and then activated by the softplus activation function to obtain the touch point excitation increment strength; the basic arrival strength represents the spontaneous arrival tendency of the conversion event without considering the previous touch point excitation, and the touch point excitation increment strength represents the additional increase in the arrival probability of the current touch point event on the subsequent conversion event.

5. The method according to claim 4, characterized in that, The joint training objective of the gated recurrent unit network, the first single-layer feedforward network, and the second single-layer feedforward network is to maximize the transformation event conditional strength value of all observed events in the journey event sequence at the time of their occurrence, while minimizing the piecewise cumulative summation of the transformation event conditional strength values ​​along the time axis within the time interval between adjacent events.

6. The method according to claim 1, characterized in that, The unified prediction time domain endpoint is the time corresponding to the last event timestamp in the journey event sequence plus a fixed-length observation extension window; the unified prediction interval is the time interval from the last event timestamp to the unified prediction time domain endpoint.

7. The method according to claim 1, characterized in that, User characteristics in the confounding covariates include user demographics, device type, access source, duration of stay per session, and new or returning customer markers; environmental characteristics in the confounding covariates include promotion cycle markers, weekday and time period markers, product price range markers, and discount level markers.

8. The method according to claim 1, characterized in that, The specific process of multi-fold cross-fit dual machine learning is as follows: all journey event sequence samples are randomly divided into 5 subsets of equal size and independent of each other; for the kth subset, k takes values ​​from 1 to 5 in turn, and the other 4 subsets are used as training data to train the first auxiliary gradient boosting tree model to predict the processing variable with confounding covariates, and to train the second auxiliary gradient boosting tree model to predict the outcome variable with confounding covariates. For each journey event sequence sample in the k-th subset, the deviation between the actual value of the processing variable and the predicted output of the first auxiliary gradient boosting tree model is recorded as the processing deviation, and the deviation between the actual value of the result variable and the predicted output of the second auxiliary gradient boosting tree model is recorded as the result deviation. After traversing all 5 subsets, the treatment deviation and outcome deviation of all samples are combined. A linear regression is performed with the treatment deviation as the explanatory variable and the outcome deviation as the explained variable. The regression coefficient is the net causal attribution coefficient of the target channel.

9. The method according to claim 1, characterized in that, For journey event sequences that do not yet include transformation events before the end of the unified prediction time domain, the Cox proportional hazards model is used with confounding covariates as input to estimate the cumulative transformation probability at the end of the unified prediction time domain. The cumulative transformation probability is used instead of the binary transformation indicator as the outcome variable to participate in the multi-fold cross-fit dual machine learning process to obtain the net causal attribution coefficient corrected for delayed transformation.

10. The method according to claim 1, characterized in that, The regularized leader algorithm performs the following operations in each decision cycle: It reads the actual conversion feedback data from each channel in the previous decision cycle, processes the actual conversion feedback data through steps 1 and 2, and updates the net causal attribution coefficient and cumulative attribution benefit concave utility function of each channel; it uses the negative entropy function as a regularization term to adjust the budget allocation ratio vector along the gradient direction of the cumulative attribution benefit concave utility function, and projects the adjusted budget allocation ratio vector onto the convex feasible region composed of global constraints, upper bound constraints, and lower bound constraints to obtain the optimal budget allocation ratio vector for the current decision cycle, and issues budget instructions to each channel for execution; it constrains the change amplitude of the budget allocation ratio vector between adjacent decision cycles through the negative entropy function, so that the regret value between the sum of the cumulative attribution benefit concave utility function values ​​of all channels and the ex-post optimal fixed allocation strategy in a continuous decision cycle sequence of arbitrary length approaches zero at a sublinear rate as the total number of decision cycles increases.

Images