An advertisement bidding method based on cross-platform user behavior link

Abstract

This invention discloses an advertising bidding method based on cross-platform user behavior links, specifically relating to the field of internet advertising technology. The method includes the following steps: collecting cross-touchpoint user behavior data and forming a cross-platform link using persistent identifiers; responding to advertising requests and querying the cross-platform profile corresponding to the persistent identifier of the device identifier; receiving the profile, evaluating the advertising value, and making a bid; displaying the advertisement after winning the bid; tracking conversions, calculating cross-device trust entropy values ​​to represent path determinism during attribution, dynamically adjusting touchpoint weights, and allocating conversion value to each touchpoint in the link. This invention constructs a cross-platform user behavior link using persistent user identifiers, overcoming the limitations of single-device identification; dynamically adjusting attribution weights using cross-device trust entropy values, overcoming the static defects of attribution rules; allocating conversion value based on the adjusted weights and providing feedback to optimize the click-through rate prediction model and conversion rate prediction model, forming a bidding attribution closed loop and improving bidding accuracy.

Description

Technical Field

[0001] This invention relates to the field of internet advertising technology, and more specifically, to an advertising bidding method based on cross-platform user behavior links. Background Technology

[0002] With the advent of the mobile internet and the multi-screen era, user behavior is increasingly scattered across multiple terminal devices such as mobile phones, computers, tablets, and smart TVs. Accurately identifying the same user across devices, connecting their behavioral segments on different platforms, and conducting ad bidding and attribution based on the complete user behavior chain has become an important research direction in the field of programmatic advertising.

[0003] In existing technologies, advertising bidding methods typically rely on a single device identifier for user identification and value assessment, such as using a mobile device ID or web browser cookies to determine user identity and calculate ad display value. However, existing methods have the following technical limitations in practical applications: Existing attribution models apply the same attribution logic to all users. For example, for the typical path of "web browsing → mobile click → purchase," the allocation of attribution weights is almost identical for both long-term users with stable behavior patterns and new users with more random behavior. This approach ignores individual differences in user behavior. Some users are accustomed to consistently completing transactions on mobile devices, while others frequently switch between devices randomly. Using a uniform attribution rule makes it difficult to accurately reflect the true impact of different users' behavioral characteristics on their conversion contribution.

[0004] When users exhibit unusual device switching in their cross-platform behavior sequence—for example, a user suddenly clicking an ad on an unused tablet, or the time interval between adjacent touchpoints far exceeding the user's historical habits—existing models still assign credit according to pre-defined rules. In such cases, conversion credit is easily misattributed to accidental channels rather than the key touchpoints that truly drive user decisions, leading to decreased accuracy in attribution results.

[0005] Users' purchase intent dynamically changes across different times and scenarios. The value response to ad reach differs significantly between high-intent moments (like late at night when focused on product research) and low-intent moments (like casual browsing during commutes). Existing attribution methods assign the same value weight to all touchpoints, lacking the ability to perceive fluctuations in user intent and failing to distinguish the actual contribution of ad reach during high-intent and low-intent moments to the final conversion. Therefore, this invention proposes an ad bidding method based on cross-platform user behavior links to address the aforementioned problems. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides the following technical solution: An ad bidding method based on cross-platform user behavior links includes the following steps: By collecting user behavior data at multiple user touchpoints and using persistent user identifiers to associate the behavior data of the same user on different platforms, a cross-platform user behavior chain is formed. In response to an advertising request from the supplier platform, obtain the device identifier corresponding to the current advertising request, and query the data management platform based on the device identifier to obtain a cross-platform unified user profile corresponding to the persistent user identifier associated with the device identifier; The demand-side platform receives a unified user profile from across platforms, evaluates the current ad display value of the user based on the unified user profile from across platforms, and generates a real-time bid based on the current ad display value; The real-time bid is sent to the advertising exchange platform for bidding. If the bid is successful, the corresponding advertising creative will be displayed to the user through the supplier platform. The system tracks subsequent user conversion behaviors, performs cross-platform attribution based on persistent user identifiers, and calculates the cross-device trust entropy value of the user's current conversion path during the attribution process of the preset attribution model. The cross-device trust entropy value is used to characterize the degree of certainty of device switching behavior in the user's current conversion path relative to historical behavior patterns. Then, the weights of each touchpoint in the attribution model are dynamically adjusted according to the cross-device trust entropy value. Finally, the conversion value is allocated to each touchpoint in the user behavior chain according to the adjusted weights.

[0007] In a preferred embodiment, constructing a cross-platform user behavior link further includes: User behavior data collected from multiple user touchpoints is cleaned and standardized to extract the device identifiers corresponding to each touchpoint. The device identifiers corresponding to each touchpoint are clustered and merged using graph algorithms to generate persistent user identifiers. Based on the persistent user identifiers, the behavior data of the same user at multiple user touchpoints are concatenated in chronological order to form a complete cross-platform user behavior chain.

[0008] In a preferred embodiment, the generation of persistent user identifiers by clustering and merging the device identifiers corresponding to each contact point using a graph algorithm specifically includes: The device identifiers corresponding to each contact point are used as points in the graph. The co-occurrence relationship between any two device identifiers within a preset time window is used as the edge connecting the two device identifiers to construct a device relationship graph. The device relationship graph is clustered and merged using a connected graph algorithm, and all device identifiers connected by direct or indirect edges are grouped into the same connected graph. A unique persistent user identifier is assigned to each connected graph, and the root node device identifier in the connected graph is used as the value of the persistent user identifier.

[0009] In a preferred embodiment, in response to an advertising request from a supplier platform, the advertising request includes a device identifier. The device identifier is sent to the data management platform, and a persistent user identifier associated with the device identifier is received from the data management platform. Based on the persistent user identifier, a pre-generated cross-platform unified user profile is obtained from the data management platform. The cross-platform unified user profile includes the user's historical behavior tags and preference information on each platform.

[0010] In a preferred embodiment, the current ad display value of a user is evaluated based on a unified user profile across platforms, and a real-time bid is generated based on the current ad display value, specifically including: The demand-side platform receives and parses the unified user profile across platforms returned by the data management platform; it inputs the unified user profile into the pre-trained click-through rate (CTR) prediction model and conversion rate (CTR) prediction model to obtain the estimated CTR and estimated CTR, respectively; it multiplies the product of the estimated CTR and CTR by the target conversion bid set by the advertiser to calculate the original bid for this ad impression, which is the current ad impression value; and it multiplies the original bid by the expected revenue per thousand impressions conversion factor to generate a real-time bid that meets the bidding requirements of the ad exchange platform.

[0011] In a preferred embodiment, both the click-through rate (CTR) prediction model and the conversion rate (CTR) prediction model are machine learning models based on deep neural networks, sharing the underlying embedding layer parameters. The embedding layer is used to map the features in the cross-platform user unified profile into low-dimensional dense vectors. The CTR prediction model outputs the probability of a user clicking an ad based on the shared embedding layer, while the CTR prediction model outputs the probability of a user converting after clicking an ad based on the shared embedding layer. The CTR prediction model and the CTR prediction model are jointly trained using a multi-task learning framework. The training samples are all exposure data, and the loss function is the weighted sum of the CTR prediction loss and the post-click conversion rate prediction loss. The trained CTR prediction model and CTR prediction model serve as the initial models, which are then used for incremental training and updates based on subsequent attribution results.

[0012] In a preferred embodiment, the calculation process for the expected income conversion factor for thousands of displays is as follows: Obtain historical bidding data from the current ad exchange platform and calculate the average actual revenue per thousand impressions for all winning requests within a preset time window; obtain historical exposure data for the current user on the media corresponding to the current ad request and calculate the user's historical average bid per thousand impressions on that media; multiply the ratio of the average actual revenue per thousand impressions to the historical average bid per thousand impressions by a preset basic conversion factor to calculate the expected revenue per thousand impressions conversion factor.

[0013] In a preferred embodiment, calculating the cross-device trust entropy value of the user's current conversion path specifically includes: Using the device switching stability factor and the time interval anomaly factor as two-dimensional coordinates, the Euclidean distance from the coordinate point to the origin is calculated and used as the cross-device trust entropy value. Among them, the device switching stability factor is based on the device transfer probability matrix constructed from the user's historical behavior data. It calculates the abnormal probability of each device switching in the current path relative to the historical pattern, and then sums them after taking the negative logarithm. The time interval anomaly factor is based on the deviation from the mean in the Gaussian distribution. It calculates the deviation multiple of the time interval between each adjacent touch point in the current path from the user's historical average interval, and then multiplies them after taking the reciprocal. The cross-device trust entropy value is obtained by squaring the device switching stability factor and the time interval anomaly factor respectively, summing them, and then taking the square root.

[0014] In a preferred embodiment, dynamically adjusting the weights of each touchpoint in the attribution model based on the cross-device trust entropy value specifically includes: The initial weights of each touchpoint in the user behavior chain are obtained in the attribution model. The attribution model is a time-decay attribution model, and the initial weights are allocated in an exponential decay manner according to the time interval between each touchpoint and the conversion time. Based on the cross-device trust entropy value, the initial weight of the final touchpoint in the attribution model is reduced according to a preset decay coefficient to obtain the adjusted weight of the final touchpoint. The reduced weight proportion is evenly distributed to each other touchpoint in the user behavior chain except for the final touchpoint to obtain the adjusted weight of each other touchpoint. Based on the adjusted weight of each touchpoint, the monitored conversion value is split and allocated to each touchpoint in the user behavior chain as the attribution contribution value of each touchpoint, which is used to guide the optimization of bidding strategy in subsequent advertising bidding.

[0015] In a preferred embodiment, after obtaining the attribution contribution value of each contact point, the following steps are further included: The attribution contribution values ​​of each touchpoint in the user behavior chain are used as positive sample labels and associated with the cross-platform unified user profiles corresponding to historical ad requests to construct an optimized training dataset. The optimized training dataset is used to incrementally train the click-through rate prediction model and the conversion rate prediction model to update the model parameters. Incremental training is conducted on the basis of a multi-task learning framework. The loss function remains unchanged, but exposure data with attribution contribution values ​​are added to the training samples. Samples with higher attribution contribution values ​​are given higher weights in the loss calculation. In the subsequent advertising bidding process, the click-through rate prediction model and conversion rate prediction model updated by incremental training are used to recalculate the estimated click-through rate and estimated conversion rate, thereby generating new real-time bids and realizing closed-loop feedback optimization of the bidding strategy based on the attribution results.

[0016] The technical effects and advantages of this invention are as follows: This invention constructs a cross-platform user behavior chain, linking the behavioral data of the same user across different platforms. This overcomes the limitations of single-device identification, enabling ad bidding to be evaluated based on the user's complete behavioral trajectory rather than isolated device fragments. This significantly improves the accuracy of user identification and the comprehensiveness of user profiles. Specifically, by collecting behavioral data at multiple user touchpoints and using persistent user identifiers for cross-platform association, this invention can reconstruct the user's entire behavioral trajectory from initial awareness and interest exploration to final decision-making. When an ad request arrives, the unified cross-platform user profile obtained by the demand-side platform is no longer limited to local information from the current device but encompasses the user's historical behavioral tags and preference information across all platforms. This provides a more comprehensive and accurate data foundation for subsequent evaluation of the current ad display value, thereby making real-time bidding more closely aligned with the user's true value.

[0017] This invention quantifies the certainty of device switching behavior relative to historical behavior patterns by calculating the cross-device trust entropy value of the user's current conversion path. Based on this entropy value, it dynamically adjusts the weights of each touchpoint in the attribution model, overcoming the shortcomings of traditional attribution rules that are static, uniform, and lack individual adaptability. This achieves accurate perception of user behavior differences and intent fluctuations. Specifically, the cross-device trust entropy value integrates device switching stability factors and time interval anomaly factors, comprehensively measuring the anomaly degree of the current conversion path from two dimensions: device switching pattern and time interval pattern. When the entropy value is high, it indicates that there are device jumps or time intervals in the current path that deviate from the user's historical habits. In this case, the attribution weight of the final touchpoint is dynamically reduced, and the reduced portion is evenly distributed to other historical touchpoints. This allows the allocation of attribution weights to adaptively adjust according to the certainty of the path, avoiding the erroneous attribution of conversion success to accidental channels, while simultaneously increasing the weight of early touchpoints that truly drive decision-making in the attribution results.

[0018] This invention allocates conversion value to each touchpoint in the user behavior chain based on dynamically adjusted weights. This allows the attribution results to more accurately reflect the actual contribution of each touchpoint to the final conversion, effectively suppressing the interference of noise behaviors such as abnormal device jumps or excessively long intervals on the attribution results. This provides more reliable feedback data for optimizing bidding strategies in subsequent advertising auctions. Specifically, the attribution contribution value obtained from each touchpoint is used as a positive sample label and associated with the corresponding cross-platform user unified profile in historical ad requests. An optimized training dataset is constructed and used for incremental training of the click-through rate prediction model and the conversion rate prediction model. Samples with higher attribution contribution values ​​are given higher weights in the loss calculation, making the model parameter update direction more inclined to optimize the prediction accuracy of high-value users. In the subsequent advertising auction process, the updated model is used to recalculate the estimated click-through rate and estimated conversion rate and generate new real-time bids. This forms a complete closed-loop feedback optimization mechanism from bidding to attribution to model update, driving the advertising strategy to continuously iterate and evolve with each conversion feedback. Attached Figure Description

[0019] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of an advertising bidding method based on cross-platform user behavior links in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 The following examples were obtained: Example 1: By collecting user behavior data at multiple user touchpoints and using persistent user identifiers to associate the behavior data of the same user on different platforms, a cross-platform user behavior chain is formed. This step realizes the unified tracking and recording of user behavior across devices and platforms, splicing together behavior fragments scattered on different platforms into a complete user behavior path, providing a full-chain user behavior data foundation for subsequent advertising bidding.

[0022] In response to an advertising request from the supplier platform, the device identifier corresponding to the current advertising request is obtained, and the data management platform is queried based on the device identifier to obtain a unified cross-platform user profile corresponding to the persistent user identifier associated with the device identifier. This step finds the corresponding persistent user identifier through the device identifier when the advertising request is triggered, and then obtains the unified cross-platform profile of the user, realizing the mapping from a single device request to complete user identification, and providing comprehensive user feature information for advertising value assessment.

[0023] The demand-side platform receives a unified profile of users across platforms, evaluates the current ad display value of the user based on the unified profile, and generates a real-time bid based on the current ad display value. This step inputs the unified profile of users across platforms into the value assessment model, calculates the expected value of the ad display for the advertiser, and converts the value into a specific bid amount, thereby achieving accurate bidding decisions based on the complete user behavior chain.

[0024] The real-time bid is sent to the ad exchange platform for bidding. If the bid is successful, the corresponding ad creative is displayed to the user through the supply-side platform. This step involves submitting the bid generated by the demand-side platform to the ad exchange platform to participate in real-time bidding. After winning the bid, the final display of the ad is completed, realizing the complete execution process of the ad from bidding to placement.

[0025] This process tracks subsequent user conversion behavior, performs cross-platform attribution based on persistent user identifiers, and calculates the cross-device trust entropy value of the user's current conversion path during the attribution process using a pre-defined attribution model. The cross-device trust entropy value characterizes the degree of certainty of the user's device switching behavior in the current conversion path relative to historical behavior patterns. Then, the weights of each touchpoint in the attribution model are dynamically adjusted based on the cross-device trust entropy value. Finally, the conversion value is allocated to each touchpoint in the user behavior chain based on the adjusted weights. This step tracks the complete cross-platform conversion path of the user after conversion, quantifies the degree of certainty of the current path by calculating the cross-device trust entropy value, and dynamically adjusts the weights of each touchpoint in the attribution model accordingly to allocate the conversion value more reasonably to each touchpoint. Finally, the allocated attribution contribution value is used to guide the optimization of bidding strategies in subsequent advertising bidding, forming a complete closed loop from bidding to attribution to bidding optimization.

[0026] The user behavior data collected from multiple user touchpoints is cleaned and standardized to extract the device identifiers corresponding to each touchpoint. This step first unifies the format and filters noise from the original behavior logs scattered across different platforms and terminals to eliminate inconsistencies caused by differences in data sources. For example, the timestamp format of mobile and web terminals is unified to standard time, and invalid records generated by crawlers or abnormal traffic are removed. Then, fields that can uniquely identify the device are extracted from each cleaned behavior record, including the mobile device ID, the web browser cookie identifier, and the smart TV device serial number, thus preparing clean and usable basic data units for subsequent cross-device association.

[0027] By clustering and merging the device identifiers corresponding to each touchpoint using graph algorithms, persistent user identifiers are generated. This step takes the extracted massive number of device identifiers as input and uses graph algorithms to mine the common occurrence relationships between different device identifiers in real user behavior. Multiple device identifiers belonging to the same natural person are merged into the same cluster, and each cluster is assigned a unique persistent user identifier that is independent of any specific device. This identifier is independent of any single device and will not become invalid due to device replacement or data deletion, fundamentally solving the problem of cross-platform user identification and providing a unified user identity anchor for all subsequent steps.

[0028] Based on persistent user identifiers, behavioral data of the same user across multiple user touchpoints is concatenated in chronological order to form a complete cross-platform user behavior chain. This step uses persistent user identifiers as indexes to aggregate behavioral data corresponding to all merged device identifiers under that identifier, and arranges them sequentially from front to back according to the precise timestamp of the behavior. This fully reconstructs the entire process of the user from initial cognition, interest exploration, price comparison consideration to final decision-making. For example, the continuous behaviors of a user watching advertisements on a smart TV, clicking a search on a mobile phone, and completing a purchase on a computer are linked into a traceable chain, providing a global behavioral foundation for subsequent advertising value assessment and attribution analysis.

[0029] The device identifiers corresponding to each touchpoint are used as points in the graph, and the co-occurrence relationship between any two device identifiers within a preset time window is used as the edge connecting the two device identifiers to construct a device relationship graph. This step first abstracts all device identifiers to be processed into nodes in a graph data structure. Each node represents an independent device. For example, user A's mobile device ID, user A's computer device ID, and user A's smart TV device ID correspond to three different nodes. Then, based on the time information recorded in the behavior log, a reasonable time window is set, such as 24 hours. If two different device identifiers appear simultaneously under the same persistent user identifier of the same user within this time window, such as detecting activity records of mobile device ID and computer device ID simultaneously in the same user's login session, it is determined that these two device identifiers have a co-occurrence relationship, and an edge is established between their corresponding nodes. By traversing all behavior data, device identifiers with a co-occurrence relationship are connected pairwise, and finally a device relationship graph with device identifiers as nodes and co-occurrence relationships as edges is constructed. This graph structure completely records the association strength and association path between different devices in real user behavior, providing an intuitive mathematical expression for subsequent clustering and merging.

[0030] A connected graph algorithm is used to cluster and merge device relationship graphs, grouping all device identifiers connected by direct or indirect edges into the same connected graph. This step runs the connected graph algorithm on the constructed device relationship graph, starting from any node and traversing along the edges, marking all nodes that can be reached from each other by one or more edges as the same connected component. For example, in the device relationship graph, the mobile device ID and the computer device ID are directly connected by an edge, and the computer device ID and the smart TV device ID are directly connected by an edge. Although there is no direct edge between the mobile device ID and the smart TV device ID, they are indirectly connected through the computer device ID. Therefore, these three nodes will be grouped into the same connected graph. This process automatically aggregates all device identifiers with direct or indirect behavioral associations, identifying the complete set of devices belonging to the same natural person without manual intervention. Even if the user uses multiple devices and the devices have not appeared simultaneously, as long as there is a connection between the intermediate devices, they can be accurately merged, effectively solving the complexity problem of cross-device identification.

[0031] Each connected graph is assigned a unique persistent user identifier, with the root node device identifier in the graph serving as the value. This step generates a unique identifier for each connected graph after graph partitioning, ensuring that the identifier is independent of any specific device and remains unchanged even if the user changes devices or clears data. This guarantees the long-term stability of user identification. The root node device identifier can be selected using fixed rules, such as choosing the device identifier that first appears in the graph or selecting the smallest device identifier in lexicographical order. This provides a concrete, readable entity representation of the persistent user identifier that is associated with the real device. For example, if user A's mobile device ID in the connected graph is used as the persistent user identifier, all subsequent behavioral data corresponding to other device identifiers associated with that mobile device ID will be uniformly aggregated under this persistent user identifier. This provides a stable and reliable identity anchor for subsequent cross-platform unified user profile construction and cross-platform attribution.

[0032] In response to an ad request from the supply-side platform (SSP), which includes a device identifier, the SSP sends the device identifier to the data management platform and receives a persistent user identifier associated with the device identifier from the data management platform. This step intervenes at the beginning of the real-time ad bidding process. When a user triggers an ad impression opportunity on a media platform, the SSP generates an ad request containing the current device identifier and sends it to the ad exchange platform. The ad exchange platform forwards this request to the demand-side platform (DSP). After extracting the device identifier from the request, the DSP immediately initiates a query to the data management platform. Upon receiving the device identifier, the data management platform stores it in its pre-built mapping database. The search for persistent user identifiers associated with the device identifier is crucial. Device identifiers are temporary or environment-dependent; for example, a mobile device's ID may change when the user resets the advertising identifier, and a web browser's cookie may expire when the user clears it. However, persistent user identifiers are stable identities generated during graph algorithm clustering and merging. They are not dependent on a single device and remain unchanged over a long period. Therefore, through this mapping query, the demand-side platform can associate the current isolated, temporary device request with a stable, cross-platform user identity, thereby breaking through the limitations of a single-device perspective and laying the foundation for obtaining complete user information in the future.

[0033] Based on persistent user identifiers, a pre-generated cross-platform unified user profile is obtained from the data management platform. This profile includes the user's historical behavioral tags and preference information across various platforms. After obtaining the persistent user identifier, the requesting platform sends another request to the data management platform. However, this time, the key value for the query is no longer a temporary device identifier, but a stable and persistent user identity identifier. The data management platform uses this persistent user identifier as an index to retrieve all historical behavioral data associated with that identifier from its stored user profile database. This data is not simply a compilation of behavioral logs from various platforms, but rather structured information that has undergone deep processing and tagging. For example, data based on a user watching car reviews on a smart TV... The video activity generates a "high-intent car buyer" tag, the frequent browsing of luxury goods websites on mobile devices generates a "high-end consumption tendency" tag, and the long hours spent comparing financial product pages on computers late at night generate a "deep financial intention" tag. These tags and data are integrated into a unified profile and returned to the demand-side platform (DSP). This allows the DSP to instantly recognize that the user who initiated the ad request is not a stranger or featureless traffic, but a complete user entity with rich historical behavior records and clear preference characteristics. Even if the current request comes from only a single device, the obtained user profile covers the user's behavior across all other platforms, providing cross-device data support for subsequent ad display value assessment.

[0034] The demand-side platform (DSP) receives and parses the unified user profile across platforms returned by the data management platform. This step is initiated within the real-time bidding window. After the data management platform completes the mapping query of persistent user identifiers and returns the corresponding unified user profile across platforms, the DSP first performs format parsing and field extraction on the profile data. It converts the raw data structure returned by the data management platform into a format that its internal processing engine can recognize, such as parsing JSON format profile data into feature vectors. At the same time, it extracts key fields such as users' historical behavior tags, preference information, and interest intensity on various platforms from the profile, preparing input data for subsequent model prediction. This step ensures that the rich user information from the data management platform can be accurately and efficiently transmitted to the core stage of bidding decision-making.

[0035] The unified cross-platform user profile is input into pre-trained click-through rate (CTR) and conversion rate (CTR) prediction models to obtain estimated CTR and CTR, respectively. This step uses the parsed unified cross-platform user profile as input features and feeds them into two independent but collaborative machine learning models. The CTR prediction model calculates the probability that a user will click on the ad under the current display based on historical click behavior features, user interest tags, and contextual features in the user profile. The CTR prediction model calculates the probability that a user will complete a conversion after clicking the ad based on historical conversion behavior features, purchasing power tags, and decision cycle features in the user profile. The two models output a value between 0 and 1, representing the estimated CTR and CTR, respectively. These two values ​​together constitute the core indicators for evaluating the expected effect of this ad display.

[0036] The original bid for this ad display is calculated by multiplying the product of the estimated click-through rate (CTR) and the estimated conversion rate (CTR) by the advertiser's target conversion bid. The original bid is the current value of the ad display. This step multiplies the estimated CTR and the estimated conversion rate to obtain the estimated click-through rate (CTR). This product represents the overall probability that a user will complete a conversion after seeing the ad. Then, this probability is multiplied by the target conversion bid set by the advertiser before the ad campaign. The target conversion bid represents the maximum cost that the advertiser is willing to pay for a conversion. The product of the two calculates the expected value of this ad display for the advertiser. For example, when the estimated CTR is 1% and the target conversion bid is 100 yuan, the original bid is 1 yuan. This original bid directly reflects the true value of this ad display under the constraints of the current user profile and the advertiser's goals.

[0037] The original bid is multiplied by the expected revenue per thousand impressions (RPI) conversion factor to generate a real-time bid that meets the bidding requirements of the ad exchange. This step converts the calculated original bid into the bidding currency commonly used by the ad exchange. Since ad exchanges typically use RPI as the unified standard for bidding ranking, while the original bid is calculated based on cost per click or cost per conversion, the units of measurement are different. Therefore, it is necessary to convert the units using the RPI conversion factor. This factor converts the original bid by one thousand into a bid based on the RPI dimension. For example, when the original bid is one yuan, multiplying it by the factor of one thousand will generate a real-time bid of one thousand yuan. This real-time bid is packaged in the format of RPI, which can be correctly identified by the ad exchange and participate in bidding ranking with other demand-side platforms, ensuring that the advertising value calculated based on the unified user profile across platforms can be accurately transmitted to the final bidding stage.

[0038] Both the click-through rate (CTR) prediction model and the conversion rate prediction model are machine learning models based on deep neural networks. They share the underlying embedding layer parameters. This setup unifies the two prediction tasks under the same deep learning architecture. The CTR and conversion rate prediction models are no longer trained and deployed as two independent models, but share the same underlying feature representation learning module. During model training, the embedding layer parameters simultaneously receive backpropagation gradient updates from both tasks, allowing the feature representations learned by the embedding layer to simultaneously meet the needs of click prediction and conversion prediction. For example, the behavioral feature of users frequently browsing luxury websites late at night might only be learned as a high click-through rate feature in a standalone CTR model. However, under the shared embedding layer mechanism, this feature is also perceived and learned by the conversion rate model as having a high conversion rate, thus forming a richer feature representation. This sharing mechanism reduces the total number of model parameters, lowers the risk of overfitting, and improves the knowledge transfer effect between the two tasks.

[0039] The embedding layer maps the features in the unified cross-platform user profile to low-dimensional dense vectors. This step transforms the original high-dimensional sparse features into continuous vector representations that the model can efficiently process. The unified cross-platform user profile contains a large number of discrete features. For example, the user's device type includes multiple values ​​such as mobile phone, computer, and smart TV. The user's interest tags may involve hundreds or thousands of categories, and the user's browsing history is high-dimensional sequence data. If these original features are directly input into the neural network, it will lead to parameter explosion and low computational efficiency. The embedding layer uses a trainable embedding matrix to map each discrete feature value to a fixed-length low-dimensional dense vector. For example, the feature "device type is smart TV" is mapped to a 64-dimensional continuous vector, and "interest tag is car" is mapped to another 64-dimensional vector. The relative positions of these dense vectors in the vector space can automatically learn the semantic similarity between features, so that semantically similar features are close to each other in the vector space, providing semantically rich input representations for subsequent deep network layers.

[0040] The click-through rate (CTR) prediction model predicts the probability of a user clicking an ad based on the output of a shared embedding layer. This step involves building a task network specifically for CTR prediction on top of the shared embedding layer. The CTR prediction model receives a low-dimensional dense vector from the embedding layer as input, and performs further feature crossing and abstraction through several fully connected layers and non-linear activation functions. Finally, it outputs a value between 0 and 1 through an output layer with a sigmoid activation function. This value represents the probability that the user will click the ad given the current user profile and ad context. For example, if the input user profile shows that the user has recently frequently searched for digital product information and the current ad displays a new mobile phone, the estimated CTR output by the CTR prediction model may be as high as 8%. However, if the user profile shows that the user has never paid attention to such products, the estimated CTR output may be only 0.1%. This estimated CTR is one of the important inputs for subsequent calculations of ad display value.

[0041] The conversion rate prediction model outputs the probability of a user converting after clicking an ad based on a shared embedding layer. This step is also built on the shared embedding layer but focuses on predicting conversion behavior. The conversion rate prediction model also receives a low-dimensional dense vector output from the embedding layer as input. However, since conversion behavior is sparser than click behavior and usually occurs after a click, the conversion rate prediction model only uses samples with click behavior during training. During prediction, it predicts the probability of a user converting if they click an ad based on the user profile. For example, when the user profile shows that the user has a high purchasing power tag and frequent historical conversion behavior, the estimated conversion rate output by the conversion rate prediction model may reach 5%. However, when the user profile shows that the user is only an information browsing user and has few historical purchase records, the estimated conversion rate may only be 0.5%. This estimated conversion rate and the estimated click-through rate together determine the expected conversion effect of the ad display.

[0042] The click-through rate (CTR) prediction model and the conversion rate (CTR) prediction model are jointly trained using a multi-task learning framework. The training samples are all exposure data, and the loss function is a weighted sum of the CTR prediction loss and the post-click conversion rate prediction loss. This setup optimizes two related tasks simultaneously during the same training process. The core idea of ​​the multi-task learning framework is to leverage the correlation between tasks to improve the learning effect of each task. During training, each exposure sample participates in the loss calculation of both tasks simultaneously. For samples with clicks, both click loss and conversion loss are calculated. For samples without clicks, only click loss is calculated, while the conversion loss is not included in the calculation or is set to zero. The importance of the two tasks is balanced by a weighted sum in the loss function. For example, the weight of click loss is set to 0.3 and the weight of conversion loss is set to 0.7 to emphasize the optimization goal of the conversion task. This joint training method enables the model to utilize the large amount of information contained in the click data to assist in learning the conversion rate, alleviating the overfitting problem caused by the sparsity of conversion data. At the same time, by sharing the underlying representation, the model learns general features that are beneficial to both tasks.

[0043] The trained click-through rate (CTR) and conversion rate (CTR) prediction models serve as initial models, providing an interface for incremental training and updates based on subsequent attribution results. This step reserves an interface for continuous model optimization and closed-loop feedback. While the initial models, trained on historical exposure and conversion data, can reflect general patterns of user behavior to some extent, they cannot fully adapt to the latest market changes due to dynamic shifts in the advertising environment, user preferences, and advertiser objectives. Marking the trained models as initial models signifies that they are not final, static versions, but rather serve as the starting point for subsequent incremental training. Once the attribution process is complete and attribution contribution values ​​for each touchpoint are generated, these values ​​are used as more accurate feedback signals for incremental model training. For example, a user's conversion behavior, after dynamically adjusted attribution weights, results in higher attribution contribution values ​​for advertising channels that reached the user earlier. This information, fed back to the model, adjusts the conversion probability prediction for users originating from that channel, making subsequent bidding more accurate for this type of user. This achieves a complete closed loop from advertising bidding to attribution analysis to model optimization.

[0044] To obtain historical bidding data from the current ad exchange platform, the average actual revenue per thousand impressions (RVR) for all winning bids within a preset time window is calculated. This step first extracts complete bidding records from the ad exchange platform's log system over a past period, such as all bidding data from the past seven days. Then, the winning bids are selected as the analysis objects, as only winning bids truly reflect the transaction price level on the ad exchange platform. For each winning bid, its final transaction price is extracted and converted into a value based on the RVR dimension. For example, if the transaction price of a winning bid is two yuan and corresponds to one single impression, then its RVR for one thousand impressions is two thousand yuan. The RVR for one thousand impressions of all winning bids is summed and divided by the total number of winning bids to obtain the average RVR for one thousand impressions within the preset time window. This average represents the price that advertisers need to pay on average to obtain one thousand ad impressions on the current ad exchange platform, reflecting the overall bidding intensity and market conditions of the platform. For example, if the calculated average RVR for one thousand impressions over the past seven days is fifty yuan, it means that winning one thousand impressions on this platform requires an average cost of fifty yuan.

[0045] The process involves retrieving historical exposure data for the current user on the media corresponding to the current ad request, and calculating the user's historical average bid per thousand impressions on that media. This step shifts the perspective from the platform as a whole to the specific user and media combination. The Demand-Side Platform (DSP) retrieves all historical exposure records for the current user on the current media from its historical database. These records include the bid amount calculated and submitted by the DSP for each ad request triggered by the user on that media over a past period. These historical bid amounts are extracted one by one and converted into values ​​based on the number of impressions. For example, if the DSP bid for a historical exposure was 0.5 yuan and corresponded to one impression, then the bid per thousand impressions would be 500 yuan. The DSP bids for all historical exposure records are summed and divided by the number of historical exposures to obtain the user's historical average bid per thousand impressions on that media. This average reflects the DSP's past assessment of the user's ad display value on that media. For example, if the calculated historical average bid per thousand impressions for the user on that media is 80 yuan, it means that the DSP previously considered each opportunity for the user to display ads on that media to be worth an average of 80 yuan.

[0046] The ratio of actual revenue per thousand impressions to historical average bid per thousand impressions is multiplied by a preset base conversion factor to calculate the expected revenue per thousand impressions conversion factor. This step first calculates the ratio between the two statistical values, using the overall average transaction price of the ad exchange platform as the numerator and the historical average bid for a specific user and media as the denominator, resulting in a dimensionless ratio. If this ratio is greater than one, it indicates that the current user's historical bid level on the current media is lower than the platform's overall average level; if it is less than one, it indicates that the historical bid level is higher than the platform's overall average level. Then, this ratio is multiplied by the preset base conversion factor, which is usually set to one thousand to achieve a benchmark conversion from a single bid to thousands of impressions. For example, the platform average... The actual revenue per thousand impressions is 50 yuan, and the user's historical average bid per thousand impressions is 80 yuan. The ratio between the two is 0.625. Multiplying this by the basic conversion factor of 1000 yields 625, which is the expected revenue conversion factor per thousand impressions. This dynamically calculated conversion factor is used in the real-time bid generation process, ensuring that the final real-time bid submitted to the ad exchange reflects both the overall market conditions of the platform and incorporates the user's historical value assessment on that media. For example, when a user's historical bid is higher than the platform average, the conversion factor is lowered accordingly to avoid wasting budget due to excessively high real-time bids caused by high historical bids. Conversely, when a user's historical bid is lower than the platform average, the conversion factor is increased accordingly to ensure that the real-time bid reaches a reasonable level to win the auction.

[0047] Using the device switching stability factor and time interval anomaly factor as two-dimensional coordinates, the Euclidean distance from this coordinate point to the origin is calculated as the cross-device trust entropy value. This step first constructs a two-dimensional coordinate system, where the horizontal axis represents the device switching stability factor and the vertical axis represents the time interval anomaly factor. Each user's current conversion path is mapped to a point in this coordinate system. The horizontal coordinate of this point is determined by the specific value of the device switching stability factor, and the vertical coordinate by the specific value of the time interval anomaly factor. Then, the straight-line distance from this point to the origin of the coordinate system is calculated, which is done by squaring the values ​​of the two factors separately, summing them, and then taking the root. The distance value obtained by squaring the distance is the cross-device trust entropy value. The larger the distance, the higher the degree of abnormality of the current conversion path in both dimensions. The smaller the distance, the closer the current conversion path is to the historical normal pattern in both dimensions. For example, if a user's current conversion path has a device switching stability factor of three and a time interval abnormality factor of four, then the distance between the two points is five. If another user's current conversion path has a device switching stability factor of one and a time interval abnormality factor of one, then the distance between the two points is 1.414. The cross-device trust entropy value of the former is much higher than that of the latter, indicating that the degree of abnormality of the former's path is significantly higher than that of the latter.

[0048] The device switching stability factor constructs a device transition probability matrix based on user historical behavior data. It calculates the anomaly probability of each device switch in the current path relative to historical patterns, takes the negative logarithm, and sums the results. This step first extracts all device switching events from the user's historical cross-platform behavior chain, counting the frequency of switching from one device type to another. For example, it counts the number of times the user switched from a mobile phone to a computer, from a computer to a mobile phone, and from a mobile phone to a tablet. These frequencies are normalized to form a device transition probability matrix. Each element in the matrix represents the conditional probability of switching to another device type given the current device type. Then, for the current conversion path... For each device switch in the path, the corresponding historical transition probability is looked up in the transition probability matrix. If the current switch rarely occurs in history, the corresponding probability value is very small. For example, if a switch from a smart TV to a mobile phone occurs in the current path, but this switching pattern has never occurred in the user's historical data, the corresponding historical transition probability is close to zero. The negative logarithm of this probability value is taken. Taking the negative logarithm of the low probability yields a large positive value, and taking the negative logarithm of the high probability yields a small positive value. The negative logarithm values ​​corresponding to all device switches in the current path are summed up to obtain the device switch stability factor. The larger the factor, the further the device switch sequence in the current path deviates from the user's historical habits, and the more unstable the device switch behavior is.

[0049] The time interval anomaly factor is based on the deviation from the mean in a Gaussian distribution. It calculates the deviation factor of the time interval between each adjacent touchpoint in the current path from the user's historical average interval, takes the reciprocal, and multiplies them together. This step first extracts the time interval data between all adjacent touchpoints from the user's historical cross-platform behavior chain, such as the time interval from the first click to the second click, the time interval from the second click to the third click, etc., and calculates the mean and standard deviation of these historical time intervals. Assuming that the user's historical behavior is regular, the time interval distribution approximately follows a Gaussian distribution. Then, for each pair of adjacent touchpoints in the current conversion path, the deviation factor of its time interval from the historical average interval is calculated, i.e., the current time interval divided by the historical average interval. A larger deviation factor indicates a longer delay in the current behavior compared to historical habits, while a smaller deviation factor indicates a longer delay in the current behavior compared to historical habits. The more habit acceleration occurs, for example, if the user's historical average interval is two hours, and a certain interval in the current path is six hours, then the deviation multiple is three. Taking the reciprocal of this deviation multiple, three is one-third, which is 0.33. The larger the deviation multiple, the smaller its reciprocal. Multiplying the reciprocals of the deviation multiples of all adjacent touchpoint pairs in the current path yields the time interval anomaly factor. The smaller this factor, the more significant the deviation from historical habits in the current path, indicating excessively long or short intervals, and the greater the fluctuation in user behavior intent. For example, if a path has three adjacent touchpoint pairs with reciprocals of deviation multiples of 0.5, 0.2, and 0.8, the product is 0.08. Another path has three adjacent touchpoint pairs with reciprocals of deviation multiples of 0.9, 0.95, and 0.92, the product is 0.78. The former's time interval anomaly factor is much smaller than the latter, indicating that the former path has time intervals that seriously deviate from historical habits.

[0050] The cross-device trust entropy value is obtained by squaring the device switching stability factor and the time interval anomaly factor, summing them, and then taking the square root. This step uses the two factors calculated above as two coordinate values ​​of a two-dimensional coordinate point. Each coordinate value is squared to eliminate the influence of positive and negative signs. The two squared values ​​are added together to obtain a comprehensive sum of squares. This sum of squares is then square-rooted to restore the distance value to the same dimension as the original coordinates. This distance value is the final cross-device trust entropy value. This entropy value integrates anomaly information from two dimensions, merging device switching mode anomalies and time interval mode anomalies into a unified quantitative indicator. For example, if a user's current conversion path has a device switching stability factor of 2.5 and a time interval anomaly factor of 1... If the cross-device trust entropy is 0.8, then the cross-device trust entropy is 2.5 squared (6.25) plus 1.8 squared (3.24), which equals 9.49. The square root of this value is approximately 3.08. For another user, the device switching stability factor of the current conversion path is 0.8, and the time interval anomaly factor is 0.6. Therefore, the cross-device trust entropy is 0.8 squared (0.64) plus 0.6 squared (0.36), which equals 1.00. The square root of this value is 1.00. The former's cross-device trust entropy of 3.08 is significantly higher than the latter's 1.00, indicating that the overall certainty of the former's current conversion path is much lower than the latter's. This entropy value will serve as the core basis for dynamically adjusting the attribution weight. The higher the entropy value, the less trustworthy the current path is, and the greater the reduction in the attribution weight of the final touchpoint.

[0051] The initial weights of each touchpoint in the user behavior chain are obtained in the attribution model. The attribution model is a time-decay attribution model, and the initial weights are assigned in an exponential decay manner according to the time interval between each touchpoint and the conversion time. This step first determines that the basic model on which the attribution calculation is based is a time-decay attribution model. The core assumption of this model is that the closer the user's interaction is to the conversion time, the greater its contribution to the final conversion, and the further away the interaction is from the conversion time, the smaller its contribution, and the contribution decays exponentially with the extension of the time interval. In specific implementation, for each touchpoint in the user behavior chain, the time interval between the occurrence time of that touchpoint and the final conversion time is calculated. For example... If a user saw an ad on a smart TV three days ago, clicked a search on a computer two days ago, clicked again on a mobile phone one day ago, and completed the purchase today, then the time intervals between the three touchpoints and the conversion time are three days, two days, and one day, respectively. Then, the initial weight of each touchpoint is calculated using an exponential decay function with the time interval as the independent variable. The shorter the time interval, the higher the weight, and the longer the time interval, the lower the weight. The sum of the weights is one. For example, the initial weight of a touchpoint one day ago may be 60%, the touchpoint two days ago may be 30%, and the touchpoint three days ago may be 10%. This initial weight allocation scheme reflects the biased perception of recent interaction behavior in traditional attribution logic.

[0052] Based on the cross-device trust entropy value, the initial weight of the final touchpoint in the attribution model is adjusted downwards according to a preset decay coefficient, resulting in the adjusted weight of the final touchpoint. This step introduces the calculated cross-device trust entropy value as the basis for dynamic adjustment. The cross-device trust entropy value quantifies the degree of certainty of device switching behavior in the current conversion path relative to historical behavior patterns. The higher the entropy value, the more the device switching sequence and time interval in the current path deviate from the user's historical habits, the higher the degree of abnormality of the path, and the less trustworthy the corresponding user behavior. At this time, the preset decay coefficient is activated. The decay coefficient is a value between zero and one, used to control the impact of the entropy value on the user's behavior. The reduction in the final touchpoint weight, for example, with a preset attenuation coefficient of 0.1, means that each unit of cross-device trust entropy value corresponds to a 10% reduction in weight. When the cross-device trust entropy value is calculated to be 3.08, the initial weight reduction of the final touchpoint is 3.08 multiplied by 0.1, which equals 0.308, or a reduction of 30.8%. Assuming that the initial weight of the final touchpoint is 60%, the adjusted weight is 60% multiplied by 1 minus 0.308, which equals 60% multiplied by 0.692, which equals 41.52%. This adjusted weight is significantly lower than the initial weight, reflecting a cautious attitude towards the contribution of the final touchpoint in abnormal paths.

[0053] The reduced weight percentage is evenly distributed to all touchpoints in the user behavior chain except the final touchpoint, resulting in the adjusted weights for each touchpoint. This step then redistributes the total weight reduction of the final touchpoint to other historical touchpoints in the chain. The reduced weight percentage is the difference between the initial weight of the final touchpoint and the adjusted weight. For example, if the initial weight of the final touchpoint is 60%, and the adjusted weight is 41.52%, then the reduced weight percentage is 18.48 percentage points. These reduced weights are then evenly distributed to all other touchpoints except the final touchpoint. Assuming there are two historical touchpoints in the user behavior chain besides the final touchpoint, each historical touchpoint receives a weight gain of 9.24 percentage points. The historical touchpoints with initial weights of 30% and 10% will have adjusted weights of 39.24% and 19.24%, respectively. This even distribution method reflects compensatory trust in historical touchpoints when uncertainty increases. That is, when the credibility of the final touchpoint is questioned, the credit is evenly traced back to the earlier touchpoints in the path, avoiding the negation of the groundwork laid by the channel due to a single abnormal behavior.

[0054] Based on the adjusted weights of each touchpoint, the monitored conversion value is broken down and allocated to each touchpoint in the user behavior chain, serving as the attribution contribution value for each touchpoint. This value guides the optimization of bidding strategies in subsequent ad bidding. This step transforms the theoretical weight allocation into actual numerical contributions. When the system detects a user completing a conversion and records the corresponding conversion value—for example, detecting a user completing a purchase with an order amount of 100 yuan—this 100 yuan is allocated to each touchpoint in the user behavior chain according to the adjusted weights. Ultimately, the first historical touchpoint receives 41.52 yuan, the first historical touchpoint receives 39.24 yuan, and the second historical touchpoint receives 19.2 yuan. The four elements—attribution contribution value, ad request value, and touchpoint value—are written into a data warehouse and associated with the corresponding user identifier, ad request, and touchpoint information. This data becomes an important data asset for subsequent optimization. In subsequent ad bidding processes, when similar users or similar link characteristics are encountered again, the system can adjust the bidding strategy based on historical attribution contribution values. For example, it can appropriately increase the bid for channel source users with high historical attribution contribution values ​​and be cautious in bidding for channel source users with low historical attribution contribution values. This achieves a complete data closed loop from attribution analysis to bidding decision-making, enabling the ad placement strategy to be continuously iterated and optimized with each conversion feedback.

[0055] The attribution contribution value of each touchpoint in the user behavior chain is used as a positive sample label and associated with the corresponding cross-platform unified user profile in historical ad requests to construct an optimized training dataset. This step is initiated after the conversion value allocation is completed. The system extracts the attribution contribution value obtained by each touchpoint in the user behavior chain. For example, in a certain conversion path, the smart TV touchpoint earns 40 yuan, the computer touchpoint earns 30 yuan, and the mobile phone touchpoint earns 30 yuan. These three values ​​are used as positive sample labels for the corresponding touchpoints. Then, the system traces back the historical ad request records corresponding to these touchpoints, finds the cross-platform unified user profile obtained by the system at the time of each ad request, and associates the attribution contribution value label with the profile one-to-one to form a structured training sample. The profile contains the user's interest tags and historical behavior at that time. Information such as features and device environment, along with labels, provides optimization target values ​​for subsequent model learning. By traversing historical attribution data, a large number of such samples are aggregated to form an optimized training dataset. This dataset differs from the exposure dataset used in the initial training in that the initial training only uses binary labels such as whether it was clicked or converted, while the optimized training dataset introduces continuous numerical labels such as attribution contribution values, which more precisely reflect the actual contribution of each touchpoint to the final conversion value. For example, exposures that were originally marked as not converting in the initial training were found through attribution analysis to have played an early role in introducing users to subsequent conversions and obtained a certain attribution contribution value. These samples are given non-zero positive labels in the optimized training dataset, allowing valuable exposures that were originally ignored to participate in model training.

[0056] Incremental training of the click-through rate (CTR) and conversion rate (CTR) prediction models is performed using an optimized training dataset to update model parameters. This step involves inputting the completed optimized training dataset into the model training process. Incremental training refers to continuing training based on the existing initial model, rather than retraining from scratch. The initial model has already learned the general patterns of user behavior based on massive exposure data. Incremental training uses attribution contribution value labels for fine-tuning based on this. During training, the model parameters are gradually updated to fit the new label distribution. For example, the model may have initially considered a certain user group to have low conversion value, but under the guidance of attribution contribution value labels, it is found that this type of user made a significant contribution in the early introduction stage. The model will adjust the weight allocation of the features of this type of user accordingly. Incremental training is carried out iteratively using mini-batch gradient descent. Each time, a batch of samples is drawn from the optimized training dataset to calculate the loss and backpropagate to update the parameters. After multiple iterations, the model parameters gradually converge to the new optimal value. This training method avoids the waste of computational resources caused by retraining, while maintaining the model's memory of historical data, allowing the model to absorb new knowledge from attribution feedback while retaining its original generalization ability.

[0057] Incremental training is conducted within a multi-task learning framework. The loss function remains unchanged, but exposure data with attribution contribution values ​​is added to the training samples. Samples with higher attribution contribution values ​​are given higher weights in the loss calculation. This step clarifies the specific technical implementation details of incremental training. Maintaining the multi-task learning framework means that the click-through rate (CTR) prediction and conversion rate (CTR) prediction tasks still share the underlying embedding layer parameters. The loss function remains a weighted sum of the CTR prediction loss and the post-click conversion rate prediction loss, maintaining the stability of the training objective. The composition of the training samples has changed; in addition to the original exposure data, samples with attribution contribution values ​​are added. These samples are specially processed during training. The prediction task still uses a binary label for whether a click occurred, but the label for the conversion rate prediction task has been replaced with a continuous numerical label for attribution contribution value. For example, a sample that originally had a conversion label of 1 may now have an attribution contribution value of 50 yuan, 30 yuan, or other different values. This allows the model to learn the differences between different conversion values. Samples with higher attribution contribution values ​​are given higher weights in the loss calculation. For example, a sample with an attribution contribution value of 50 yuan has five times the weight of a sample with an attribution contribution value of 10 yuan when calculating the loss. This weighting mechanism forces the model to pay more attention to the feature patterns of high-value conversion paths, making the model parameter update direction more inclined to optimize the prediction accuracy of high-value users.

[0058] In subsequent ad bidding, the updated click-through rate (CTR) and conversion rate (CTR) prediction models are incrementally trained to recalculate the estimated CTR and CTR, thereby generating new real-time bids. This achieves closed-loop feedback optimization of the bidding strategy based on attribution results. This step deploys the incrementally trained models back to the online bidding environment, completing the full closed loop from bidding to attribution to model updates. When new ad requests arrive, the demand-side platform obtains a unified user profile across platforms and inputs it into the updated CTR and CTR prediction models. Since the models have absorbed feedback information from historical attribution contributions, the output estimated CTR and CTR are significantly higher than before the update. More precise, for example, if a certain user characteristic is found to have a high contribution value in historical attribution, the updated model will increase the predicted conversion rate of this type of user, which will lead to a higher real-time bid, making these high-value users have a higher probability of winning in subsequent bids. Conversely, user characteristics with low historical attribution contribution values ​​will have their predictions lowered by the model to avoid wasting budget. This cycle repeats itself, and the attribution contribution value brought by each conversion continuously optimizes the model. The model, in turn, guides more accurate bidding decisions, forming a self-iteratory and continuously evolving intelligent advertising system, ensuring that the advertising budget always flows to the user group most likely to generate high-value conversions.

[0059] The above algorithms or formulas are all dimensionless and numerical calculations, and the results are obtained by software simulation based on a large amount of collected data to obtain the most recent real-world results. The preset parameters are set by those skilled in the art according to the actual situation.

[0060] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0061] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0062] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0063] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An advertising bidding method based on cross-platform user behavior links, characterized in that, Includes the following steps: By collecting user behavior data at multiple user touchpoints and using persistent user identifiers to associate the behavior data of the same user on different platforms, a cross-platform user behavior chain is formed. In response to an advertising request from the supplier platform, obtain the device identifier corresponding to the current advertising request, and query the data management platform based on the device identifier to obtain a cross-platform unified user profile corresponding to the persistent user identifier associated with the device identifier; The demand-side platform receives a unified user profile from across platforms, evaluates the current ad display value of the user based on the unified user profile from across platforms, and generates a real-time bid based on the current ad display value; The real-time bid is sent to the advertising exchange platform for bidding. If the bid is successful, the corresponding advertising creative will be displayed to the user through the supplier platform. The system tracks subsequent user conversion behaviors, performs cross-platform attribution based on persistent user identifiers, and calculates the cross-device trust entropy value of the user's current conversion path during the attribution process of the preset attribution model. The cross-device trust entropy value is used to characterize the degree of certainty of device switching behavior in the user's current conversion path relative to historical behavior patterns. Then, the weights of each touchpoint in the attribution model are dynamically adjusted according to the cross-device trust entropy value. Finally, the conversion value is allocated to each touchpoint in the user behavior chain according to the adjusted weights.

2. The advertising bidding method based on cross-platform user behavior links according to claim 1, characterized in that, Building a cross-platform user behavior chain further includes: User behavior data collected from multiple user touchpoints is cleaned and standardized to extract the device identifiers corresponding to each touchpoint. The device identifiers corresponding to each touchpoint are clustered and merged using graph algorithms to generate persistent user identifiers. Based on the persistent user identifiers, the behavior data of the same user at multiple user touchpoints are concatenated in chronological order to form a complete cross-platform user behavior chain.

3. The advertising bidding method based on cross-platform user behavior links according to claim 2, characterized in that, The persistent user identifier is generated by clustering and merging the device identifiers corresponding to each contact point using a graph algorithm. Specifically, this includes: The device identifiers corresponding to each contact point are used as points in the graph. The co-occurrence relationship between any two device identifiers within a preset time window is used as the edge connecting the two device identifiers to construct a device relationship graph. The device relationship graph is clustered and merged using a connected graph algorithm, and all device identifiers connected by direct or indirect edges are grouped into the same connected graph. A unique persistent user identifier is assigned to each connected graph, and the root node device identifier in the connected graph is used as the value of the persistent user identifier.

4. The advertising bidding method based on cross-platform user behavior links according to claim 3, characterized in that, When responding to an advertising request from a supplier platform, the device identifier is sent to the data management platform, and a persistent user identifier associated with the device identifier is received from the data management platform. Based on the persistent user identifier, a pre-generated cross-platform unified user profile is obtained from the data management platform. The cross-platform unified user profile contains the user's historical behavior tags and preference information on each platform.

5. The advertising bidding method based on cross-platform user behavior links according to claim 4, characterized in that, The system assesses the current ad display value of a user based on a unified user profile across platforms, and generates a real-time bid based on that value. Specifically, this includes: The demand-side platform receives and parses the unified user profile across platforms returned by the data management platform; it inputs the unified user profile into the pre-trained click-through rate (CTR) prediction model and conversion rate (CTR) prediction model to obtain the estimated CTR and estimated CTR, respectively; it multiplies the product of the estimated CTR and CTR by the target conversion bid set by the advertiser to calculate the original bid for this ad impression, which is the current ad impression value; and it multiplies the original bid by the expected revenue per thousand impressions conversion factor to generate a real-time bid that meets the bidding requirements of the ad exchange platform.

6. The advertising bidding method based on cross-platform user behavior links according to claim 5, characterized in that, Both the click-through rate (CTR) prediction model and the conversion rate (CTR) prediction model are machine learning models based on deep neural networks, sharing the underlying embedding layer parameters. The embedding layer is used to map the features in the unified user profile across platforms into low-dimensional dense vectors. The CTR prediction model outputs the probability of a user clicking an ad based on the shared embedding layer, while the CTR prediction model outputs the probability of a user converting after clicking an ad based on the shared embedding layer. The CTR and CTR prediction models are jointly trained using a multi-task learning framework. The training samples are all exposure data, and the loss function is the weighted sum of the CTR prediction loss and the post-click conversion rate prediction loss. The trained CTR and CTR prediction models serve as the initial models.

7. The advertising bidding method based on cross-platform user behavior links according to claim 5, characterized in that, The calculation process for the expected income conversion factor per thousand impressions is as follows: Get the historical bidding data of the current ad exchange platform and calculate the average actual revenue per thousand impressions for all winning requests within the preset time window; get the historical exposure data of the current user on the media corresponding to the current ad request and calculate the user's historical average bid per thousand impressions on that media. The ratio of the average actual revenue per thousand impressions to the historical average bid per thousand impressions is multiplied by a preset basic conversion factor to calculate the expected revenue per thousand impressions conversion factor.

8. The advertising bidding method based on cross-platform user behavior links according to claim 1, characterized in that, The calculation of the cross-device trust entropy value of the user's current conversion path specifically includes: Using the device switching stability factor and the time interval anomaly factor as two-dimensional coordinates, the Euclidean distance from the coordinate point to the origin is calculated and used as the cross-device trust entropy value. Among them, the device switching stability factor is based on the device transfer probability matrix constructed from the user's historical behavior data. It calculates the abnormal probability of each device switching in the current path relative to the historical pattern, and then sums them after taking the negative logarithm. The time interval anomaly factor is based on the deviation from the mean in the Gaussian distribution. It calculates the deviation multiple of the time interval between each adjacent touch point in the current path from the user's historical average interval, and then multiplies them after taking the reciprocal. The cross-device trust entropy value is obtained by squaring the device switching stability factor and the time interval anomaly factor respectively, summing them, and then taking the square root.

9. The advertising bidding method based on cross-platform user behavior links according to claim 6, characterized in that, The dynamic adjustment of the weights of each touchpoint in the attribution model based on cross-device trust entropy specifically includes: The initial weights of each touchpoint in the user behavior chain in the attribution model are obtained. The attribution model is a time-decay attribution model, and the initial weights are allocated in an exponential decay manner according to the time interval between each touchpoint and the conversion time. Based on the cross-device trust entropy value, the initial weight of the final touchpoint in the attribution model is reduced according to a preset decay coefficient to obtain the adjusted weight of the final touchpoint. The reduced weight ratio is evenly distributed to each other touchpoint in the user behavior chain except for the final touchpoint to obtain the adjusted weight of each other touchpoint. Based on the adjusted weight of each touchpoint, the detected conversion value is split and allocated to each touchpoint in the user behavior chain as the attribution contribution value of each touchpoint.

10. The advertising bidding method based on cross-platform user behavior links according to claim 9, characterized in that, After obtaining the attribution contribution value of each contact point, the following steps are also included: The attribution contribution values ​​of each touchpoint in the user behavior chain are used as positive sample labels and associated with the cross-platform unified user profiles corresponding to historical ad requests to construct an optimized training dataset. The optimized training dataset is used to incrementally train the click-through rate prediction model and the conversion rate prediction model and update the model parameters. Incremental training is conducted on the basis of a multi-task learning framework, with the loss function remaining unchanged. In subsequent advertising bidding processes, the click-through rate prediction model and conversion rate prediction model updated by incremental training are used to recalculate the estimated click-through rate and estimated conversion rate, and generate new real-time bids.

Images