Attention-weighted advertising amplification attribution method

Abstract

The present application relates to the technical field of advertisement efficiency, in particular to an advertisement efficiency attribution method based on attention weighting, comprising: first, aggregating cross-screen user logs based on household identification, collecting multi-dimensional interaction features such as visual duration, sound and picture state in the advertisement exposure process, and calculating the attention score of single exposure. Secondly, the attention score is introduced into the multi-touch attribution model as a weight to calculate the preliminary contribution value of each touch point. At the same time, the real incremental conversion value obtained by small flow training is used to construct a calibration coefficient, and the preliminary attribution result is modified by a multi-dimensional bucketing strategy. The present application solves the problem of underestimating the value of non-explicit interaction in the traditional model.

Description

Technical Field

[0001] This invention relates to the field of advertising effectiveness technology, specifically to an attention-weighted attribution method for advertising effectiveness. Background Technology

[0002] With the evolution of terminal technology, user behavior logs are typically scattered across multiple heterogeneous devices such as smart TVs, mobile terminals, and PCs, forming a complex cross-screen data transmission and processing chain. In the field of big data-based attribution computing technology, existing data processing methods suffer from the following technical shortcomings:

[0003] First, the completeness and granularity of interactive data collection are insufficient. Existing computer systems typically use triggering mechanisms based on explicit, discrete events such as "clicks" to record user behavior logs. However, in scenarios such as OTT large-screen power-on and video streaming playback, user interactions with the terminal often manifest as continuous, unstructured states such as viewing duration and audio-visual status (e.g., mute, picture-in-picture). The lack of effective means to capture and vectorize these unobvious state characteristics leads to a missing data dimension at the system input layer, resulting in a broken information chain in the cross-screen user behavior path, making it impossible for the computer to reconstruct the true human-computer interaction process.

[0004] Secondly, large-scale data processing architectures face technical bottlenecks in terms of computational accuracy and system resource consumption. To improve the fitting accuracy of multi-touch attribution (MTA) models, causal inference or complex machine learning algorithms are typically introduced. However, with massive amounts of ad impression logs (often reaching hundreds of millions of data points), performing highly complex calibration operations on the entire dataset generates enormous computational load and storage I / O overhead, leading to excessively high system processing latency. Conversely, using simple statistical models results in low logical confidence of the outputs due to the inability to remove data noise. Currently, the industry lacks a data processing architecture that can achieve efficient throughput of large-scale data with extremely low system overhead, while also enabling dynamic correction of model parameters through sparse data feedback.

[0005] To address this, an attribution method for enhanced advertising effectiveness based on attention weighting is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an attention-weighted attribution method for enhanced advertising effectiveness. First, it aggregates cross-screen user logs based on family identifiers, collecting multi-dimensional interaction features such as viewing duration and audio-visual status during ad exposure to calculate the attention score for a single exposure. Second, the attention score is used as a weight to introduce a multi-touchpoint attribution model to calculate the initial contribution value of each touchpoint. Simultaneously, a calibration coefficient is constructed using real incremental conversion values ​​obtained from small-volume training, and a multi-dimensional bucketing strategy is used to correct the initial attribution results. This invention solves the problem of traditional models underestimating the value of indirect interactions.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The attention-weighted attribution method for enhanced advertising effectiveness includes: aggregating user behavior logs from different terminals based on the household identifier dimension; collecting multi-dimensional interaction features associated with user exposure behavior during the ad exposure process based on the user behavior logs; and vectorizing the multi-dimensional interaction features to generate attention feature vectors.

[0009] Based on the logistic regression model, the attention feature vector is analyzed to calculate the attention score for a single ad exposure; the attention score is then used as a weight and input into the multi-touchpoint attribution (MTA) model to calculate the preliminary attribution contribution value of each ad touchpoint.

[0010] Triggering the incremental parameter correction training process: Sparsely sample a specified advertising campaign to construct a calibration sample set, and obtain the actual incremental conversion value of the calibration sample set within a specific bucket; use the actual incremental conversion value as an anchor point; calculate the predicted incremental conversion value based on the multi-touchpoint attribution MTA model for sample users in the calibration sample set; calculate the calibration coefficient based on the actual incremental conversion value and the predicted incremental conversion value; and associate the calibration coefficient with the multi-dimensional bucketing strategy and store it as a correction parameter in the calibration coefficient knowledge base;

[0011] The calibration coefficients are applied to correct the initial attribution contribution value, and incremental advertising returns and incremental reach are output with evidence level and timeliness markers.

[0012] Preferably, the multi-dimensional interactive features include at least one of the following: visible duration feature, audio-visual status feature, screen ratio feature, and active interaction feature; the visible duration feature is the ratio of the actual playback time of the advertisement on the screen to the total duration of the advertisement; the audio-visual status feature indicates whether the device is in a silent state and whether the device is in a picture-in-picture playback state; the screen ratio feature is the proportion of the area occupied by the advertisement content on the display screen; the active interaction feature indicates the active operation behavior generated by the user during the exposure period and within a preset time window after the exposure, and the active operation behavior includes at least one of the following: voice search, cross-screen interaction, fast forward operation, rewind operation, and pause operation.

[0013] Preferably, the construction process of the logistic regression model includes: defining positive samples as events in which a preset conversion behavior occurs on a mobile device associated with a family identifier within a preset time window after an advertisement is exposed; defining negative samples as exposure events in which the preset conversion behavior does not occur; training the model using a supervised learning classification algorithm to learn the functional relationship between the attention feature vector and the probability of the preset conversion behavior; and using the probability of occurrence output by the model as the attention score for a single advertisement exposure.

[0014] Preferably, the step of using attention scores as weights and inputting them into the multi-touchpoint attribution (MTA) model to calculate the preliminary attribution contribution value of each ad touchpoint includes: constructing the MTA model using the Shapley value algorithm; defining the conversion path value, which is the aggregate value of the attention scores of all touchpoints on the path; and reconstructing the feature function in the Shapley value algorithm. ,in For channel alliance, Defined as all containing only The sum of path values ​​of the conversion paths in the middle channel; based on the reconstructed feature function The marginal contribution of each touchpoint is calculated, and the marginal contribution is positively correlated with the attention score.

[0015] Preferably, the process of triggering incremental parameter correction training and constructing a calibration sample set includes: dividing the target traffic into training set traffic and baseline set traffic based on a preset sampling strategy; executing ad delivery logic for the training set traffic and airdrop delivery logic for the baseline set traffic to form a homogeneous calibration sample set; calculating the difference between the conversion rate in the training set traffic and the natural conversion rate in the baseline set traffic to obtain the true incremental conversion value; the dimensions based on the multi-dimensional bucketing strategy include at least one of ad placement type, marketing objective, audience tag, and industry category.

[0016] Preferably, the incremental reach number is calculated as follows: obtain the total number of converted users in the training set traffic of the calibration sample set; calculate the natural conversion rate based on the baseline set traffic; calculate the number of natural converted users in the training set traffic, wherein the number of natural converted users is equal to the total number of users in the training set traffic multiplied by the natural conversion rate; subtract the number of natural converted users from the total number of converted users to obtain the incremental reach number.

[0017] Preferably, the generation logic for the evidence level and timeliness marker includes: obtaining the generation time of the calibration coefficient matched by the current attribution request, and calculating the interval between the generation time and the current time; when the attribution request triggers the incremental parameter correction training process for the advertising campaign, it is marked as a Level 1 evidence level; when the attribution request does not trigger the training process, and simultaneously satisfies the three conditions of the interval being less than a first preset time threshold, the calibration coefficient being within its effective lifespan, and the binning dimension of the calibration coefficient being completely matched with the advertising campaign attributes, it is marked as a Level 2 evidence level; when any of the following conditions are met, it is marked as a Level 3 evidence level: First condition: the binning dimension of the calibration coefficient... The calibration coefficient's binning dimension matches the advertising campaign attributes only partially; the second case is that the binning dimension of the calibration coefficient matches the advertising campaign attributes completely, but the interval duration is greater than the first preset time threshold, and it is determined to be usable after downgrading; the third case is that the binning dimension of the calibration coefficient matches the advertising campaign attributes completely, but the calibration coefficient exceeds its valid lifespan, and it is determined to be usable after downgrading; when any of the following cases are true, it is marked as a level four evidence level: the first case is that there is no matching binning key in the calibration coefficient knowledge base; the second case is that the matched calibration coefficient is determined to be severely expired after downgrading; when outputting the results, the generation timestamp of the calibration coefficient is also output.

[0018] Preferably, the method further includes a lifecycle maintenance step for the calibration coefficient knowledge base: for each calibration coefficient stored in the calibration coefficient knowledge base Set an effective lifespan; periodically scan the knowledge base, and when a certain calibration coefficient is detected... When the storage time exceeds the effective lifetime, a downgrade process is performed, and the corresponding evidence level is lowered; the evidence level status of a specific bucket is monitored, and when the evidence level is detected to be lower than the preset level threshold, a new incremental parameter correction training request is automatically generated.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1. This invention, through the collection of multi-dimensional interactive features and the linkage with an attention scoring model, transforms non-explicit data (such as viewing duration and audio-visual status) that were previously unanalyzable by computers during video streaming into high-dimensional quantified weighting factors. This technique effectively fills the data gaps caused by traditional models relying solely on click events. By increasing the dimensionality and granularity of the input data, it repairs the information gaps in the user conversion path at the underlying data level, thereby significantly improving the completeness and computational accuracy of the Multi-Touch Attribution (MTA) model in reconstructing complex interactive logic.

[0021] 2. This invention triggers an incremental parameter correction training process, consuming computational resources only on a local sparse sample set to obtain high-precision calibration parameters. Then, a multi-dimensional bucketing strategy is used to generalize and map these parameters to the global data model. This "local high-precision computation + global parameter correction" processing architecture effectively avoids the massive data processing load caused by comparative calculations on the entire dataset. It achieves accurate calibration and systematic bias correction of large-scale attribution results with extremely low system overhead, thus providing a continuous, reliable, and interpretable attribution solution for enhanced advertising effectiveness.

[0022] 3. This invention improves the Shapley value algorithm at the algorithmic level, upgrading its feature function from a simple statistical logic based on "touchpoint frequency" to a weighted aggregation logic based on "total path value." This algorithmic optimization enables the computing system to automatically identify and filter noise interference from low-quality interactive data (such as muting or background playback), ensuring that attribution weights dynamically tilt towards nodes with high feature scores. This guarantees a positive correlation between weight allocation and actual traffic quality from an algorithmic perspective, significantly improving the model's logical robustness in the face of complex noisy data, thereby providing a continuous, reliable, and interpretable attribution solution for enhanced advertising effectiveness.

[0023] 4. This invention utilizes a calibration coefficient knowledge base and TTL lifecycle management technology to construct an automatic aging and degradation maintenance mechanism for model parameters. The system can monitor the timestamps of calibration parameters in real time, automatically adjust the confidence level markers of output results based on data freshness, and automatically trigger new parameter correction training requests when parameters become invalid. This closed-loop feedback control mechanism effectively prevents model prediction failures caused by changes in data distribution over time, ensuring the stability of the attribution system and the reliability of output data during long-term operation, thereby providing a continuous, reliable, and interpretable attribution solution for enhanced advertising effectiveness.

[0024] 5. This invention enhances the system's perception dimension and sensitivity to unstructured user interaction data by introducing attention feature vectors at the data input end; it reconstructs the Shapley value feature function at the algorithm inference end, ensuring the logical correctness of multi-touchpoint weight allocation; and it introduces sparse ground truth feedback through incremental parameter correction training at the output end, achieving dynamic correction of model prediction results. This end-to-end closed-loop architecture enables the system to maintain high-confidence attribution calculation capabilities without needing to collect all ground truth data, achieving a globally optimal balance between computational resource efficiency and data restoration accuracy, thereby providing a continuous, reliable, and interpretable advertising enhancement attribution solution. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating an attention-weighted attribution method for enhancing advertising effectiveness, as provided in an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram illustrating a multidimensional bucket matching and evidence level generation method provided in an embodiment of the present invention. Detailed Implementation

[0027] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1:

[0029] Figure 1 The flowchart of an attention-weighted attribution method for enhancing advertising effectiveness is shown in an embodiment of the present invention, including: aggregating user behavior logs from different terminals based on the family identification dimension;

[0030] Specifically, to address the challenge of cross-screen device identification, this embodiment uses the household as the smallest attribution unit to construct a unique household identifier (HomeID). To ensure the accuracy of HomeID generation, the system employs the following hierarchical matching logic: First, it acquires the network fingerprint information of smart TVs (OTT) and mobile devices (phones, tablets) within a preset period (e.g., 7 days), including a list of IP addresses and Wi-Fi SSIDs; Second, it performs strong association matching, calculating the co-occurrence frequency between devices. When the number of days two devices co-occur under the same Wi-Fi SSID exceeds a preset threshold (e.g., 3 days), they are determined to belong to the same household device cluster; Finally, it performs weak association correction, combining Geo-Hash geographic location features to handle non-Wi-Fi connection scenarios (e.g., 4G / 5G status). Considering the physical accuracy limitations of mobile network base station positioning, this step adopts a spatiotemporal co-occurrence probability algorithm: when the Geo-Hash accuracy of the mobile device is above 8 bits, and its overlap with the TV device's active time period within a preset period is higher than 80% (indicating that the two have highly consistent family routines), it is determined that the two have a high probability of spatiotemporal co-occurrence relationship, and they are classified as weakly associated devices into the same HomeID probability cluster to supplement the attribution path.

[0031] Based on the unique HomeID generated above as the cross-screen identity index (Key), the system will associate and aggregate the advertising exposure logs, interaction logs and conversion logs scattered across different terminals (smart TVs, mobile phones, PCs, etc.), and sort them by timestamp to form complete end-to-end user conversion path data.

[0032] Based on user behavior logs, multi-dimensional interaction features associated with user exposure behavior are collected during the ad exposure process; these multi-dimensional interaction features are vectorized to generate attention feature vectors.

[0033] Furthermore, the multi-dimensional interactive features include at least one of the following: visible duration feature, audio-visual status feature, screen ratio feature, and active interaction feature; the visible duration feature is the ratio of the actual playback time of the advertisement on the screen to the total duration of the advertisement; the audio-visual status feature indicates whether the device is in a silent state and whether the device is in a picture-in-picture playback state; the screen ratio feature is the proportion of the area occupied by the advertisement content on the display screen; the active interaction feature indicates the active operation behavior generated by the user during the exposure period and within a preset time window after the exposure, and the active operation behavior includes at least one of the following: voice search, cross-screen interaction, fast forward operation, rewind operation, and pause operation.

[0034] The above features are normalized and then concatenated in a fixed order to form an attention feature vector (hereinafter referred to as X_att), which specifically includes, but is not limited to:

[0035] visible_ratio (visible time ratio) = actual ad viewing time / total ad duration;

[0036] isMuted (whether to mute, 0 = sound, 1 = mute);

[0037] isPiP (whether to use picture-in-picture, 0=no, 1=ye);

[0038] screen_ratio (screen ratio = ad area / total screen area);

[0039] voice_search, cross_screen, fast_forward, rewind, pause (whether the above active interactions occur within 30 minutes after exposure, 0=no, 1=yes).

[0040] The final result is a high-dimensional vector X_att with at least 10 fields.

[0041] In addition, to capture users' dynamic behavior patterns, the system also extracts temporal statistical features from the original operation sequence, including: 1. High-intent operation frequency: the total number of pause and replay operations performed per unit time (e.g., 1 minute), used to measure the intensity of deep interaction; 2. Noise operation frequency: the frequency of frequent channel switching and rapid mute / unmute operations per unit time, used to identify non-focused "noise" viewing behavior; 3. Focus interval duration: the time interval between the last remote control operation and the end of the advertisement. The larger this value, the more likely the user is to be in a focused viewing state after putting down the remote control. After normalizing the above continuous numerical features, they are concatenated with the following binary features;

[0042] Specifically, this includes: In each ad exposure event, the system records in real time whether the user has engaged in any of the following five distinct proactive behaviors within a time window from the start of the exposure to 30 minutes after the exposure. Each behavior corresponds to a binary feature (0 = not occurred, 1 = occurred), and this feature is used as an important basis for judging the user's true attention level:

[0043] voice_search

[0044] Within 30 minutes of exposure, users can trigger a search using voice commands on any home-related device (TV, mobile phone, tablet), and the search keywords must be highly relevant to the brand, category, or creative content of this advertisement (e.g., if the advertisement is for a milk powder brand, the user can say "×× milk powder" or "baby milk powder recommendation").

[0045] cross_screen (cross-screen interaction / cross-screen continuation)

[0046] Within 30 minutes of seeing an advertisement on an OTT television, users may continue to browse brand / product content that is the same as or highly related to the advertisement on other devices such as mobile phones, tablets, and PCs under the same HomeID. This could include opening the same advertiser's app, entering the details page of the same product, or watching a longer version of the same video.

[0047] fast_forward (fast forward operation)

[0048] During the ad playback, the user actively clicks or uses the remote control to perform the "fast forward" operation, and the fast forward range is less than 80% of the remaining ad duration (that is, the user does not skip the entire ad directly, but chooses to speed up the viewing, indicating that they still maintain a certain level of attention).

[0049] rewind (back / rewind operation)

[0050] During or after an ad is played, users may actively drag the progress bar back to rewatch part or all of the ad content (a typical scenario is when a user becomes interested in the ad content and actively chooses to "watch it again").

[0051] pause (pause operation)

[0052] Users actively click the pause button during the ad playback and then resume playback within 30 minutes (indicating that users may have temporarily interrupted the playback due to reasons such as answering a phone call or discussing the ad content, but then actively resumed watching, showing a high level of attention engagement).

[0053] All five types of proactive interactive behaviors are collected in real time via the terminal SDK or log reporting, with the judgment time window uniformly set to the exposure start time + 30 minutes. If any of the above behaviors occurs within this time window, the corresponding feature value is set to 1; otherwise, it is set to 0.

[0054] The five binary features mentioned above are normalized together with other interaction features (visible_ratio, isMuted, isPiP, screen_ratio, etc.) and then concatenated in a fixed order to form a complete attention feature vector X_att.

[0055] Attention feature vectors are analyzed based on logistic regression models to calculate the attention score for a single ad exposure.

[0056] Furthermore, the construction process of the logistic regression model includes: defining positive samples as events in which a preset conversion behavior occurs on a mobile device associated with a family identifier within a preset time window after an ad exposure; defining negative samples as exposure events in which the preset conversion behavior does not occur; training the model using a supervised learning classification algorithm to learn the functional relationship between the attention feature vector and the probability of the preset conversion behavior; and using the probability of occurrence output by the model as the attention score for a single ad exposure.

[0057] The preset time window is dynamically adjusted based on the ad type. For performance-based ads, the time window is set to 30 minutes; for brand-related ads (such as OTT startup ads), the time window is set to 24 hours to 7 days. When constructing the logistic regression model, positive samples are defined as exposure events that result in conversion within the above time window. In this embodiment, to quickly capture the impact of users' immediate attention on conversion, 30 minutes is selected as an example as a short-term validation window for attention features, but in actual applications, it should cover the complete attribution period. Negative samples are exposure events that do not result in conversion. The model output probability is the attention score (hereinafter referred to as AttentionScore), with a value range of (0,1). A higher value indicates higher actual user attention.

[0058] This embodiment transforms non-explicit data (such as viewing duration and audio-visual status) that was previously unanalyzable by computers during video streaming (such as viewing time and audio-visual status) into high-dimensional quantified weighting factors by collecting multi-dimensional interactive features and linking them with an attention scoring model. This technique effectively fills the data gaps caused by traditional models relying solely on click events. By increasing the dimensionality and granularity of the input data, it repairs the information gaps in the user conversion path at the underlying data level, thereby significantly improving the completeness and computational accuracy of the Multi-Touch Attribution (MTA) model in reconstructing complex interactive logic.

[0059] The attention score is used as a weight and input into the multi-touchpoint attribution MTA model to calculate the preliminary attribution contribution value of each ad touchpoint.

[0060] Furthermore, the step of using attention scores as weights and inputting them into the multi-touchpoint attribution (MTA) model to calculate the preliminary attribution contribution value of each ad touchpoint includes: constructing the MTA model using the Shapley value algorithm; defining the conversion path value, which is the aggregate value of the attention scores of all touchpoints on the path; and reconstructing the feature function in the Shapley value algorithm. ,in For channel alliance, Defined as all containing only The sum of path values ​​of the conversion paths in the middle channel; based on the reconstructed feature function The marginal contribution of each touchpoint is calculated, and the marginal contribution is positively correlated with the attention score.

[0061] Specifically, characteristic function The calculation follows the above logic: for any channel alliance (A channel alliance can be understood as a "combination" selected from all possible advertising touchpoints.) Traverse all conversion paths; if a certain path... All contact points involved in satisfy Then the path value of the path. Accumulate to In this process, based on this characteristic function, the Shapley value formula is used to calculate the value of each contact point. marginal contribution ;

[0062] ;

[0063] in For set The total number of contacts in the system; For the Alliance The number of contacts in the system; The total number of possible arrangements of all contacts; Indicates advertising touchpoints In the league In this context, it brings additional (new) contribution value;

[0064] The marginal contributions (Shapley values) of all contacts are normalized to obtain the preliminary attribution contribution values ​​of each contact.

[0065] To more intuitively illustrate how the reconstructed feature function distinguishes traffic quality, a calculation example is given here: Assume there are two conversion paths: Path A (high-quality path): The user first watches an OTT startup ad (attention score 0.9), then clicks on a search (attention score 1.0) and converts. Its path value W(A) = (0.9 + 1.0) / 2 = 0.95. Path B (low-quality path): The user plays an ad muted on an OTT device (attention score 0.2), then clicks on a search (attention score 1.0) and converts. Its path value W(B) = (0.2 + 1.0) / 2 = 0.60. When calculating the Shapley value, the algorithm considers the "1 conversion" contributed by path A as 0.95 value units, while considering path B as 0.60 value units. Therefore, the attribution weight obtained by the OTT touchpoint in path A will be significantly higher than that in path B, thus achieving the correct reward for high-quality exposure.

[0066] Triggering the incremental parameter correction training process: Sparsely sample a specified advertising campaign to construct a calibration sample set, and obtain the actual incremental conversion value of the calibration sample set within a specific bucket; use the actual incremental conversion value as an anchor point; calculate the predicted incremental conversion value based on the multi-touchpoint attribution MTA model for sample users in the calibration sample set; calculate the calibration coefficient based on the actual incremental conversion value and the predicted incremental conversion value; and associate the calibration coefficient with the multi-dimensional bucketing strategy and store it as a correction parameter in the calibration coefficient knowledge base;

[0067] Furthermore, the process of triggering incremental parameter correction training and constructing a calibration sample set includes: dividing the target traffic into training set traffic and baseline set traffic based on a preset sampling strategy; executing ad delivery logic for the training set traffic and airdrop delivery logic for the baseline set traffic to form a homogeneous calibration sample set; calculating the difference between the conversion rate in the training set traffic and the natural conversion rate in the baseline set traffic to obtain the true incremental conversion value; the dimensions based on the multi-dimensional bucketing strategy include at least one of ad placement type, marketing objective, audience tag, and industry category.

[0068] When a new advertising campaign is launched, has been running for more than 7 days without calibration, or the highest evidence level of the corresponding bucket in the calibration coefficient knowledge base drops to level three or below, the system will automatically or manually trigger the incremental parameter correction training process.

[0069] Upon triggering, (using a pre-defined sampling strategy) a sparse sampling ratio of 1%–5% is employed, with hierarchical random segmentation based on HomeID or Geo-Hash. This divides the target audience traffic into two completely homogeneous groups: a training set (Treatment group) and a baseline set (Control group). The training set traffic is served with real ads, while the baseline set traffic is served with public service announcements (PSA) or airdrops (without displaying any commercial ads), ensuring that the two groups are identical in all aspects except for whether or not ads are displayed.

[0070] The training cycle typically lasts 7–14 days. After the cycle ends:

[0071] 1. Calculate the true incremental conversion value

[0072] ;

[0073] in: Indicates the conversion rate of the training set; Represents the natural conversion rate of the baseline set; The calculation method is: total number of converted users in the training set ÷ number of users in the training set; The calculation method is: Total number of converted users in the baseline set ÷ Number of users in the baseline set (i.e., natural conversion rate).

[0074] On the exact same calibration sample set, the predicted incremental conversion rate was recalculated using the constructed weighted Shapley value MTA model;

[0075] The calculation method is: attribution conversion rate of MTA model on the training set − attribution conversion rate of MTA model on the benchmark set; Indicates the predicted incremental conversion value;

[0076] Calculating calibration coefficients under multi-dimensional binning ;

[0077] Typical bucket example: Any combination of the following four dimensions: "Ad placement type = OTT boot-up ad", "Marketing objective = Brand exposure", "Audience tag = Women aged 25-45", and "Industry = FMCG".

[0078] The revised initial attribution contribution value is calculated as the product of the initial attribution contribution value and the calibration coefficient.

[0079] This embodiment triggers an incremental parameter correction training process, consuming computational resources only on a local sparse sample set to obtain high-precision calibration parameters. Then, a multi-dimensional bucketing strategy is used to generalize and map these parameters to the global data model. This "local high-precision computation + global parameter correction" processing architecture effectively avoids the massive data processing load caused by performing comparative calculations on the entire dataset. It achieves accurate calibration and systematic bias correction of large-scale attribution results with extremely low system overhead, thereby providing a continuous, reliable, and interpretable attribution solution for enhanced advertising effectiveness.

[0080] The calibration coefficients are applied to correct the initial attribution contribution value, and incremental advertising returns and incremental reach are output with evidence level and timeliness markers.

[0081] Furthermore, the incremental reach number is calculated as follows: obtain the total number of converted users in the training set traffic of the calibration sample set; calculate the natural conversion rate based on the benchmark set traffic; calculate the number of natural conversions in the training set traffic, wherein the number of natural conversions is equal to the total number of users in the training set traffic multiplied by the natural conversion rate; subtract the number of natural conversions from the total number of converted users to obtain the incremental reach number.

[0082] The incremental reach (iReach) is calculated as follows: Total conversions in the training set − Number of users in the training set × Natural conversion rate in the baseline set;

[0083] For example: if the training set has 1 million users and 1200 conversions, and the baseline set has 1 million users and 800 conversions naturally occur, then...

[0084] iReach = 1200 − 1,000,000 × (800 ÷ 1,000,000) = 400 people;

[0085] The incremental advertising return is characterized using the internationally recognized incremental return on investment (iROAS) metric, and its calculation formula is as follows:

[0086] iROAS = iReach × Average Customer Lifetime Value (LTV) ÷ Actual Ad Spending (Cost) of the Campaign

[0087] Where: iReach is the calculated incremental reach number (i.e., the number of new unique conversion users actually driven by the ad).

[0088] Average lifetime value (LTV) is the statistical average of historical data for the industry / advertiser (e.g., 90–180 days LTV is commonly used in the FMCG industry, and 180–365 days LTV is commonly used in the gaming industry); Cost is the total cost of the advertising campaign within the statistical period.

[0089] Further, the generation logic for the evidence level and timeliness label includes: obtaining the generation time of the calibration coefficient matched by the current attribution request, and calculating the interval between the generation time and the current time; when the attribution request triggers the incremental parameter correction training process for the advertising campaign, it is labeled as a Level 1 evidence level; when the attribution request does not trigger the training process, and simultaneously satisfies the three conditions of the interval being less than a first preset time threshold, the calibration coefficient being within its effective lifespan, and the binning dimension of the calibration coefficient being completely matched with the advertising campaign attributes, it is labeled as a Level 2 evidence level; when any of the following conditions are met, it is labeled as a Level 3 evidence level: the first condition: the binning dimension of the calibration coefficient... The dimensions and advertising campaign attributes only partially match; the second case: the binning dimension of the calibration coefficient completely matches the advertising campaign attributes, but the interval duration is greater than the first preset time threshold, and it is determined to be usable after downgrading; the third case: the binning dimension of the calibration coefficient completely matches the advertising campaign attributes, but the calibration coefficient exceeds its valid lifespan, and it is determined to be usable after downgrading; when any of the following cases are true, it is marked as a level four evidence level: the first case: there is no matching binning key in the calibration coefficient knowledge base; the second case: the matched calibration coefficient is determined to be severely expired after downgrading; when outputting the results, the generation timestamp of the calibration coefficient is also output.

[0090] This embodiment improves the Shapley value algorithm at the algorithmic level, upgrading its feature function from a simple statistical logic based on "touchpoint frequency" to a weighted aggregation logic based on "total path value". This algorithmic optimization enables the computing system to automatically identify and filter noise interference from low-quality interactive data (such as mute or background playback), ensuring that attribution weights dynamically tilt towards nodes with high feature scores. This guarantees a positive correlation between weight allocation and actual traffic quality from the algorithmic principle, significantly improving the model's logical robustness when facing complex noisy data, thereby providing a continuous, reliable, and interpretable attribution solution for enhanced advertising effectiveness.

[0091] Furthermore, the method also includes a lifecycle maintenance step for the calibration coefficient knowledge base: for each calibration coefficient stored in the calibration coefficient knowledge base Set an effective lifespan; periodically scan the knowledge base, and when a certain calibration coefficient is detected... When the storage time exceeds the effective lifetime, a downgrade process is performed, and the corresponding evidence level is lowered; the evidence level status of a specific bucket is monitored, and when the evidence level is detected to be lower than the preset level threshold, a new incremental parameter correction training request is automatically generated.

[0092] The calibration coefficients are stored in a structured format in a calibration coefficient knowledge base. The record structure includes, but is not limited to:

[0093] {

[0094] Bucket Dimensions: {Ad Placement Type: "OTT Startup Ad", Marketing Objective: "Brand Exposure", ...}

[0095] Value: 1.35

[0096] Generation timestamp: 2025-09-27 10:30:00

[0097] Evidence levels: Level 1 (training triggered in this instance) / Level 2 / Level 3

[0098] Effective lifespan (TTL): 90 days

[0099] };

[0100] The rules for classifying the level of evidence are as follows:

[0101] Level 1 (First-level evidence): This attribution request directly triggered the incremental parameter correction training process for this advertising campaign;

[0102] Level 2 (Secondary Evidence): The training process has not been triggered, and three conditions are met simultaneously: the interval between the generation time and the current time is less than the first preset time threshold, the calibration coefficient is within the effective time-to-live (TTL), and the bucket dimension is completely matched with the advertising campaign attributes.

[0103] Level 3 (Third-level evidence): When any one of the following conditions is met:

[0104] The bucketing dimension only partially matches;

[0105] If the bucket dimensions are completely matched, but the interval duration is greater than or equal to the first preset time threshold, it is determined to be usable after downgrading.

[0106] The bucket dimensions are perfectly matched, but the calibration coefficients are marked as TTL invalid (soft failure). After downgrading, they are determined to still be usable as historical references.

[0107] Level 4 (Fourth Level Evidence): When any one of the following conditions is met:

[0108] There is no corresponding bucket key in the knowledge base;

[0109] The matched calibration coefficients are downgraded and determined to be severely expired (i.e., completely invalid). At this point, only the raw prediction result of the MTA model (k=1.0) is output and marked as "pure model prediction". Further, the effective survival period is preferably 90 days, the periodic scanning cycle is once daily, and the preset level threshold is level two (i.e., triggered when the highest usable evidence level drops to level three or below).

[0110] Furthermore: The first preset time threshold is preferably 90 days (used to determine whether the data is fresh), and the severely expired threshold is preferably 180 days (i.e., the hard upper limit of TTL). The periodic scanning cycle is once a day.

[0111] A separate lifecycle management service is deployed (performing a full database scan once daily from 00:00 to 01:00), and the specific maintenance process is as follows:

[0112] The calibration coefficient knowledge base uses a distributed key-value store. Each record is timestamped upon creation. The system is deployed with a separate lifecycle management service that performs the following maintenance process:

[0113] Daily automatic scan: Traverse all calibration coefficient records in the knowledge base and calculate the interval between the current time and the generated timestamp.

[0114] Perform downgrade processing (corresponding) Figure 2 (Degradation nodes in the middle)

[0115] Path A (Fresh to Stale): If the interval is greater than or equal to the first preset time threshold (90 days) but less than the severe expiration threshold (180 days), it is judged as a normal downgrade and the evidence level is marked as Level 3;

[0116] Path B (Outdated to Expired): If the interval is greater than or equal to the severe expiration threshold (180 days), it is judged as severely expired, the evidence level is marked as Level 4, and it is no longer used in attribution;

[0117] Path C (Manual / Abnormal Intervention): If the interval is less than the threshold, but the TTL field is manually marked as invalid, it will also enter the downgrade process. The default mark is Level 3 (for reference). If it is verified to be unusable, it will be marked as Level 4.

[0118] Triggering closed loop: When the highest evidence level of a certain bucket drops to Level 3 (i.e. there are no more Level 1 / 2 records below the first preset time threshold), a new incremental parameter correction training request is immediately and automatically triggered.

[0119] Example: Suppose a bucket generates a Level 1 entry with k=1.37 on August 1, 2025, with the first preset time threshold set to 90 days and the severe expiration threshold set to 180 days:

[0120] From August 1, 2025 to October 29, 2025 (interval < 90 days): If the time frame meets the requirement of being "less than the first preset time threshold" and the TTL is valid, it can be used as Level 2.

[0121] 2025-10-30 (Day 90): The scan found that the interval duration was no less than the first preset time threshold, and entered the "perform downgrade processing" stage. It was determined that it was not seriously expired and was downgraded to Level 3; at this time, a new training request was triggered.

[0122] 2026-01-27 (Day 180): If no new coefficients are generated during this period, a second scan will find that the interval has reached the severe expiration threshold, and the system will be judged as "severely expired" and downgraded to Level 4.

[0123] During this process, when it is detected that there are no records of ≥Level 2 at the highest evidence level in the bucket, a new training request is automatically generated, and a new round of sampling training begins.

[0124] Figure 2 This is a schematic diagram illustrating a multi-dimensional bucket matching and evidence level generation method provided in an embodiment of the present invention. For example... Figure 2 As shown, the multi-dimensional bucket matching and evidence level generation process for processing a single attribution request in this invention is as follows: Step 1: Feature extraction and key generation. The system first parses the attributes of the advertising campaign to be attributed, extracts four-dimensional features: "ad placement type, marketing objective, audience tag, and industry category," and generates a unique bucket key. Step 2: Knowledge base query and matching. The system uses the bucket key to query the calibration coefficient knowledge base. If the corresponding calibration coefficient record is found directly, it is judged as a "dimensional perfect match"; if not found, the partial matching logic described above is executed. Step 3: Evidence level determination. The system makes real-time judgments based on the search results: Judgment 1 (Whether training is triggered): If the current request directly corresponds to the ongoing incremental parameter correction training process, then Level 1 evidence is output; Judgment 2 (Accuracy and timeliness): If training is not triggered, but a perfect dimensional match is achieved, and the interval between the generation time and the current time is less than a first preset time threshold (e.g., 90 days) and the TTL is valid, then Level 2 evidence is output; Judgment 3 (Catch-all judgment): If a partial dimensional match occurs, or although a perfect match occurs, the interval is too old, then Level 3 evidence is output; Judgment 4 (No calibration): If all the above matches fail or the TTL has severely expired, then Level 4 evidence is output. Finally, the system outputs calibration coefficients with corresponding evidence level labels.

[0125] This embodiment utilizes a calibration coefficient knowledge base and TTL lifecycle management technology to construct an automatic aging and degradation maintenance mechanism for model parameters. The system can monitor the timestamps of calibration parameters in real time, automatically adjust the confidence level of output results based on data freshness, and automatically trigger new parameter correction training requests when parameters become invalid. This closed-loop feedback control mechanism effectively prevents model prediction failures caused by changes in data distribution over time, ensuring the stability of the attribution system and the reliability of output data during long-term operation, thereby providing a continuous, reliable, and interpretable attribution solution for enhanced advertising effectiveness.

[0126] This embodiment first aggregates cross-screen user logs based on family identifiers to collect multi-dimensional interaction features such as viewing duration and audio-visual status during ad exposure, calculating the attention score for a single exposure. Second, the attention score is used as a weight to introduce a multi-touchpoint attribution model to calculate the initial contribution value of each touchpoint. Simultaneously, calibration coefficients are constructed using real incremental conversion values ​​obtained from small-volume training, and a multi-dimensional bucketing strategy is used to correct the initial attribution results. This invention solves the problem of traditional models underestimating the value of indirect interactions.

[0127] Example 2:

[0128] The attention-weighted attribution method for enhanced advertising effectiveness includes: aggregating user behavior logs from different terminals based on the household identifier dimension; collecting multi-dimensional interaction features associated with user exposure behavior during the ad exposure process based on the user behavior logs; and vectorizing the multi-dimensional interaction features to generate attention feature vectors.

[0129] Based on the logistic regression model, the attention feature vector is analyzed to calculate the attention score for a single ad exposure; the attention score is then used as a weight and input into the multi-touchpoint attribution (MTA) model to calculate the preliminary attribution contribution value of each ad touchpoint.

[0130] Triggering the incremental parameter correction training process: Sparsely sample a specified advertising campaign to construct a calibration sample set, and obtain the actual incremental conversion value of the calibration sample set within a specific bucket; use the actual incremental conversion value as an anchor point; calculate the predicted incremental conversion value based on the multi-touchpoint attribution MTA model for sample users in the calibration sample set; calculate the calibration coefficient based on the actual incremental conversion value and the predicted incremental conversion value; and associate the calibration coefficient with the multi-dimensional bucketing strategy and store it as a correction parameter in the calibration coefficient knowledge base;

[0131] The calibration coefficients are applied to correct the initial attribution contribution value, and incremental advertising returns and incremental reach are output with evidence level and timeliness markers.

[0132] Furthermore, the multi-dimensional interactive features include at least one of the following: visible duration feature, audio-visual status feature, screen ratio feature, and active interaction feature; the visible duration feature is the ratio of the actual playback time of the advertisement on the screen to the total duration of the advertisement; the audio-visual status feature indicates whether the device is in a silent state and whether the device is in a picture-in-picture playback state; the screen ratio feature is the proportion of the area occupied by the advertisement content on the display screen; the active interaction feature indicates the active operation behavior generated by the user during the exposure period and within a preset time window after the exposure, and the active operation behavior includes at least one of the following: voice search, cross-screen interaction, fast forward operation, rewind operation, and pause operation.

[0133] Furthermore, the construction process of the logistic regression model includes: defining positive samples as events in which a preset conversion behavior occurs on a mobile device associated with a family identifier within a preset time window after an ad exposure; defining negative samples as exposure events in which the preset conversion behavior does not occur; training the model using a supervised learning classification algorithm to learn the functional relationship between the attention feature vector and the probability of the preset conversion behavior; and using the probability of occurrence output by the model as the attention score for a single ad exposure.

[0134] Furthermore, the step of using attention scores as weights and inputting them into the multi-touchpoint attribution (MTA) model to calculate the preliminary attribution contribution value of each ad touchpoint includes: constructing the MTA model using the Shapley value algorithm; defining the conversion path value, which is the aggregate value of the attention scores of all touchpoints on the path; and reconstructing the feature function in the Shapley value algorithm. ,in For channel alliance, Defined as all containing only The sum of path values ​​of the conversion paths in the middle channel; based on the reconstructed feature function The marginal contribution of each touchpoint is calculated, and the marginal contribution is positively correlated with the attention score.

[0135] Furthermore, the process of triggering incremental parameter correction training and constructing a calibration sample set includes: dividing the target traffic into training set traffic and baseline set traffic based on a preset sampling strategy; executing ad delivery logic for the training set traffic and airdrop delivery logic for the baseline set traffic to form a homogeneous calibration sample set; calculating the difference between the conversion rate in the training set traffic and the natural conversion rate in the baseline set traffic to obtain the true incremental conversion value; the dimensions based on the multi-dimensional bucketing strategy include at least one of ad placement type, marketing objective, audience tag, and industry category.

[0136] Furthermore, the incremental reach number is calculated as follows: obtain the total number of converted users in the training set traffic of the calibration sample set; calculate the natural conversion rate based on the benchmark set traffic; calculate the number of natural conversions in the training set traffic, wherein the number of natural conversions is equal to the total number of users in the training set traffic multiplied by the natural conversion rate; subtract the number of natural conversions from the total number of converted users to obtain the incremental reach number.

[0137] Further, the generation logic for the evidence level and timeliness label includes: obtaining the generation time of the calibration coefficient matched by the current attribution request, and calculating the interval between the generation time and the current time; when the attribution request triggers the incremental parameter correction training process for the advertising campaign, it is labeled as a Level 1 evidence level; when the attribution request does not trigger the training process, and simultaneously satisfies the three conditions of the interval being less than a first preset time threshold, the calibration coefficient being within its effective lifespan, and the binning dimension of the calibration coefficient being completely matched with the advertising campaign attributes, it is labeled as a Level 2 evidence level; when any of the following conditions are met, it is labeled as a Level 3 evidence level: the first condition: the binning dimension of the calibration coefficient... The dimensions and advertising campaign attributes only partially match; the second case: the binning dimension of the calibration coefficient completely matches the advertising campaign attributes, but the interval duration is greater than the first preset time threshold, and it is determined to be usable after downgrading; the third case: the binning dimension of the calibration coefficient completely matches the advertising campaign attributes, but the calibration coefficient exceeds its valid lifespan, and it is determined to be usable after downgrading; when any of the following cases are true, it is marked as a level four evidence level: the first case: there is no matching binning key in the calibration coefficient knowledge base; the second case: the matched calibration coefficient is determined to be severely expired after downgrading; when outputting the results, the generation timestamp of the calibration coefficient is also output.

[0138] Furthermore, the method also includes a lifecycle maintenance step for the calibration coefficient knowledge base: for each calibration coefficient stored in the calibration coefficient knowledge base Set an effective lifespan; periodically scan the knowledge base, and when a certain calibration coefficient is detected... When the storage time exceeds the effective lifetime, a downgrade process is performed, and the corresponding evidence level is lowered; the evidence level status of a specific bucket is monitored, and when the evidence level is detected to be lower than the preset level threshold, a new incremental parameter correction training request is automatically generated.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An attention-weighted attribution method for advertising effectiveness enhancement, characterized in that, include: Based on the family identifier dimension, user behavior logs from different terminals are aggregated; based on the user behavior logs, multi-dimensional interaction features associated with user exposure behavior during the ad exposure process are collected; the multi-dimensional interaction features are vectorized to generate attention feature vectors. Attention feature vectors are analyzed based on logistic regression models to calculate the attention score for a single ad exposure. The attention score is used as a weight and input into the multi-touchpoint attribution MTA model to calculate the preliminary attribution contribution value of each ad touchpoint. Trigger the incremental parameter correction training process: perform sparse sampling on the specified advertising campaign to construct a calibration sample set, and obtain the real incremental conversion value of the calibration sample set in a specific bucket; The process of triggering incremental parameter correction training and constructing a calibration sample set includes: dividing the target traffic into training set traffic and baseline set traffic based on a preset sampling strategy; executing advertising delivery logic for the training set traffic and airdrop delivery logic for the baseline set traffic to form a homogeneous calibration sample set; and calculating the difference between the conversion rate in the training set traffic and the natural conversion rate in the baseline set traffic to obtain the true incremental conversion value. The actual incremental conversion value is used as the anchor point; for the sample users in the calibration sample set, the predicted incremental conversion value is calculated based on the multi-touchpoint attribution MTA model; the calibration coefficient is calculated based on the actual incremental conversion value and the predicted incremental conversion value; and the calibration coefficient is associated with the multi-dimensional bucketing strategy and stored as a correction parameter in the calibration coefficient knowledge base; the dimensions based on the multi-dimensional bucketing strategy include at least one of the following: ad placement type, marketing objective, audience tag, and industry category; The calibration coefficient is applied to correct the initial attribution contribution value, and the incremental ad return of the specified advertising campaign and the incremental reach of the training set traffic are output with evidence level and timeliness label. The incremental reach is calculated as follows: obtain the total number of conversions in the training set traffic in the calibration sample set; calculate the organic conversion rate based on the baseline traffic; calculate the number of organic conversions in the training set traffic, where the number of organic conversions is equal to the total number of users in the training set traffic multiplied by the organic conversion rate; subtract the number of organic conversions from the total number of conversions to obtain the incremental reach.

2. The attribution method for enhanced advertising effectiveness based on attention weighting according to claim 1, characterized in that: The multi-dimensional interactive features include at least one of the following: visual duration feature, audio-visual status feature, screen ratio feature, and active interaction feature; the visual duration feature is the ratio of the actual playback time of the advertisement on the screen to the total duration of the advertisement; the audio-visual status feature indicates whether the device is in a silent state and whether the device is in a picture-in-picture playback state. The screen occupancy feature is the proportion of the display screen area occupied by the advertising content; The active interaction feature represents the active operation behavior generated by the user during the exposure period and within a preset time window after the exposure. The active operation behavior includes at least one of voice search, cross-screen interaction, fast forward operation, rewind operation, and pause operation.

3. The attribution method for enhanced advertising effectiveness based on attention weighting according to claim 1, characterized in that: The construction process of the logistic regression model includes: defining positive samples as events in which a preset conversion behavior occurs on a mobile device associated with a family identifier within a preset time window after an ad exposure; defining negative samples as exposure events in which the preset conversion behavior does not occur; training the model using a supervised learning classification algorithm to learn the functional relationship between the attention feature vector and the probability of the preset conversion behavior; and using the probability of occurrence output by the model as the attention score for a single ad exposure.

4. The attribution method for enhanced advertising effectiveness based on attention weighting according to claim 1, characterized in that: The step of using attention scores as weights and inputting them into the multi-touchpoint attribution (MTA) model to calculate the preliminary attribution contribution value of each ad touchpoint includes: constructing the MTA model using the Shapley value algorithm; defining the conversion path value, which is the aggregate value of the attention scores of all touchpoints on the path; and reconstructing the feature function in the Shapley value algorithm. ,in For channel alliance, Defined as all containing only The sum of path values ​​of the conversion paths in the middle channel; based on the reconstructed feature function The marginal contribution of each touchpoint is calculated, and the marginal contribution is positively correlated with the attention score.

5. The attribution method for enhanced advertising effectiveness based on attention weighting according to claim 1, characterized in that: The generation logic for the evidence level and timeliness markers includes: obtaining the generation time of the calibration coefficient matched by the current attribution request, and calculating the interval between the generation time and the current time; when the attribution request triggers the incremental parameter correction training process for the advertising campaign, it is marked as Level 1 evidence; when the attribution request does not trigger the training process, and simultaneously meets the three conditions of the interval being less than a first preset time threshold, the calibration coefficient being within its effective lifespan, and the binning dimension of the calibration coefficient being completely matched with the advertising campaign attributes, it is marked as Level 2 evidence; when any of the following conditions are met, it is marked as Level 3 evidence: First condition: the binning dimension of the calibration coefficient is... The advertising campaign attributes only partially match; the second case: the binning dimension of the calibration coefficient completely matches the advertising campaign attributes, but the interval duration is greater than the first preset time threshold, and it is determined to be usable after downgrading; the third case: the binning dimension of the calibration coefficient completely matches the advertising campaign attributes, but the calibration coefficient exceeds its valid lifespan, and it is determined to be usable after downgrading; when any of the following cases are true, it is marked as a level four evidence level: the first case: there is no matching binning key in the calibration coefficient knowledge base; the second case: the matched calibration coefficient is determined to be severely expired after downgrading; when outputting the results, the generation timestamp of the calibration coefficient is also output.

6. The attribution method for enhanced advertising effectiveness based on attention weighting according to claim 1, characterized in that: The method also includes a lifecycle maintenance step for the calibration coefficient knowledge base: for each calibration coefficient stored in the calibration coefficient knowledge base Set an effective lifespan; periodically scan the knowledge base, and when a certain calibration coefficient is detected... When the storage time exceeds the effective lifetime, a downgrade process is performed, and the corresponding evidence level is lowered; the evidence level status of a specific bucket is monitored, and when the evidence level is detected to be lower than the preset level threshold, a new incremental parameter correction training request is automatically generated.

Images