Method and device for estimating conversion rate based on time division

Abstract

The invention relates to a method and a device for estimating a conversion rate based on time periods. The method comprises the following steps: determining a delay duration of an advertisement exposure sample from an exposure moment to a current moment; according to the delay duration, time-division labels of all the conversion time periods are marked, determined labels in the time-division labels are used for indicating samples with the delay duration exceeding the corresponding conversion time periods, and uncertain labels in the time-division labels are used for indicating samples with the delay duration not exceeding the corresponding conversion time periods; inputting the advertisement exposure sample and the time-phased label into the initial model; according to a sample corresponding to the determined label, adjusting a parameter of a corresponding sub pre-estimation head, and adjusting a parameter of a main pre-estimation head when the label is determined in each conversion time period, so as to obtain a trained model; and inputting the advertisement exposure data into the trained model to obtain an accumulated conversion rate in a corresponding conversion period output by each sub-estimation head and an overall conversion rate output by the main estimation head. The conversion rate result output by the model is more accurate.

Description

Technical Field

[0001] This application relates to the field of advertising delivery technology, and in particular to a method and apparatus for estimating conversion rates based on time periods. Background Technology

[0002] In programmatic advertising scenarios, especially those with significant conversion delays, there is often a considerable time lag between a user's ad exposure and a conversion, with significant differences in conversion time distribution among different users. Currently, the industry commonly uses a single conversion prediction window for modeling, which only estimates the overall conversion rate of users without distinguishing the specific time when a conversion occurs.

[0003] In long-latency scenarios, due to the lag in the feedback of conversion data, the model's predicted results often do not match the actual conversion recovery, resulting in a significant reduction in prediction accuracy. For example, when traffic fluctuates, the predicted results for some periods may be too high, but because the corresponding conversion data has not yet been fed back in time, they are easily misjudged as prediction deviations, when in fact the prediction may be within a reasonable range, thus misleading the optimization of advertising strategies.

[0004] When modeling delayed data, incompletely collected conversion data results in incomplete labeling. Samples currently marked as unconverted may become converted in the future, posing a significant challenge to model training. However, waiting for all conversion data to be collected to obtain complete labeling would severely lag behind real-time traffic changes and user behavior dynamics, preventing timely model updates and impacting model accuracy. Summary of the Invention

[0005] This application provides a method and apparatus for predicting conversion rates based on time periods to address the problem of low model accuracy.

[0006] In a first aspect, this application provides a method for estimating conversion rates based on time-segmented periods, the method comprising:

[0007] Determine the delay time from the exposure time to the current time for the ad exposure sample;

[0008] Each conversion period is labeled with a time-segment tag based on the delay duration, wherein the definite tag in the time-segment tag is used to indicate samples whose delay duration exceeds the corresponding conversion period, and the indefinite tag in the time-segment tag is used to indicate samples whose delay duration does not exceed the corresponding conversion period;

[0009] The ad exposure samples and time-segment tags are input into the initial model, wherein the initial model includes a main prediction head and multiple sub-prediction heads corresponding to multiple conversion time periods;

[0010] The parameters of the corresponding sub-prediction head are adjusted according to the samples corresponding to the determined labels, and the parameters of the main prediction head are adjusted when the labels are determined in each conversion period to obtain the trained model. The samples corresponding to the uncertain labels do not participate in the model training.

[0011] By inputting the ad exposure data into the trained model, we can obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head.

[0012] Optionally, the time-segment labels for each conversion period based on the delay duration include:

[0013] Multiple conversion periods are defined, where each conversion period is the length of time from the sample exposure time to the preset cutoff time, and the length of time varies for different conversion periods;

[0014] If the delay time of the ad exposure sample exceeds the time length of the preset conversion period, then the preset conversion period is labeled with a determination tag;

[0015] Based on the conversion results of the ad exposure samples within the preset conversion period, the samples corresponding to the determined tags are labeled as positive or negative samples. The positive samples are used to indicate that the ad exposure samples have been converted within the preset conversion period, and the negative samples are used to indicate that the ad exposure samples have not been converted within the preset conversion period.

[0016] If the delay time of the ad exposure sample does not exceed the time length of the preset conversion period, then the preset conversion period is marked with an uncertain label.

[0017] Optionally, adjusting the parameters of the corresponding sub-prediction head based on the sample corresponding to the determined label, and adjusting the parameters of the main prediction head when the label is determined in each conversion period, includes:

[0018] The feature vector of the ad exposure sample is obtained through the underlying shared network of the initial model;

[0019] The feature vector is learned by sub-prediction heads in the corresponding transformation period, and the feature vector is learned by main prediction head in the entire time period.

[0020] The first difference between the predicted value of the sub-prediction head and the sample label value within the corresponding conversion period is calculated using a binary classification loss function, and the parameters of the sub-prediction head are adjusted based on the first difference. The predicted value of the sub-prediction head is the product of the task output of the sub-prediction head and the original output of the main prediction head.

[0021] When each conversion period has a defined label, the binary classification loss function is used to calculate the second difference between the predicted value of the master predictor and the sample label value throughout the entire period, and the parameters of the master predictor and the underlying shared network are adjusted based on the second difference.

[0022] Optionally, learning the transformation pattern of the feature vector in the corresponding transformation period through the prediction head includes:

[0023] Determine the positive or negative samples corresponding to each conversion period;

[0024] The prediction head is used to learn the characteristic patterns of positive samples transforming during the corresponding transformation period, and the prediction head is used to learn the characteristic patterns of negative samples not transforming during the corresponding transformation period.

[0025] Optionally, the method further includes:

[0026] When the delay of the ad exposure sample exceeds the deadline of the target conversion period, the label of the target conversion period is updated from an uncertain label to a certain label, and the sample corresponding to the certain label is used to perform incremental training of the corresponding sub-prediction head.

[0027] When all labels for each conversion period are updated to definite labels, the master predictor head is trained using samples corresponding to all conversion periods.

[0028] Optionally, the ad impression data is input into the trained model to obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head, including:

[0029] The ad exposure data is input into the trained model, and the feature vector is obtained by extracting features from the ad exposure data through a shared underlying network.

[0030] Based on the feature vector, the cumulative conversion rate within the corresponding conversion period is estimated by the sub-prediction head;

[0031] Based on the feature vector, the overall conversion rate is estimated by the main prediction head using samples from all conversion periods.

[0032] Optionally, after obtaining the cumulative conversion rate of each sub-prediction head output within the corresponding conversion period and the overall conversion rate output by the main prediction head, the method further includes:

[0033] If it is determined that the advertisement needs to achieve a successful conversion in a short period of time, then based on the estimated cumulative conversion rate of each segment, the advertisement with the highest cumulative conversion rate within the shortest conversion period will be prioritized for execution.

[0034] If it is determined that the advertisement needs to achieve successful conversions over a long period of time, then the advertising budget and targeting strategy should be adjusted based on the overall conversion rate of the main head and the cumulative conversion rate corresponding to the longest conversion period.

[0035] Secondly, this application provides an apparatus for estimating conversion rates based on time-segmented periods, the apparatus comprising:

[0036] The determination module is used to determine the delay time from the exposure time of the ad exposure sample to the current time;

[0037] The labeling module is used to label each conversion period with a time-segment label based on the delay duration. The definite label in the time-segment label is used to indicate samples whose delay duration exceeds the corresponding conversion period, and the indefinite label in the time-segment label is used to indicate samples whose delay duration does not exceed the corresponding conversion period.

[0038] The input module is used to input the ad exposure samples and time-segment labels into the initial model, wherein the initial model includes a main prediction head and multiple sub-prediction heads corresponding to multiple conversion time periods;

[0039] The adjustment module is used to adjust the parameters of the corresponding sub-prediction head according to the samples corresponding to the determined labels, and to adjust the parameters of the main prediction head when the labels are determined in each conversion period, so as to obtain the trained model. The samples corresponding to the uncertain labels do not participate in the model training.

[0040] The input / output module is used to input ad exposure data into the trained model to obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head.

[0041] Thirdly, this application provides an electronic device, comprising: at least one communication interface; at least one bus connected to the at least one communication interface; at least one processor connected to the at least one bus; and at least one memory connected to the at least one bus.

[0042] Fourthly, this application also provides a computer storage medium storing computer-executable instructions for executing the method for estimating conversion rates based on time periods as described in any of the preceding claims of this application.

[0043] Compared with the prior art, the technical solution provided in this application has the following advantages: It determines the delay duration of the ad exposure samples and then labels them according to the time period: samples with a delay duration exceeding the corresponding conversion time period are marked with a definite label, while those not exceeding are marked with an indefinite label. During model training, only samples with definite labels are used to avoid interference from samples with indefinite labels (conversion status is not yet determined), ensuring that each sub-prediction head learns a clear time-period conversion pattern. Subsequently, the samples and labels are input into an initial model with a multi-head structure. The sub-prediction heads adjust their parameters for the definite label samples of each time period, focusing on capturing the conversion features of the corresponding time period; the main prediction head is trained when all time periods have definite labels, focusing on the full-cycle conversion pattern. Finally, the trained model outputs the cumulative conversion rate by time period and the overall conversion rate. By selecting samples with definite labels, noisy data can be avoided, and the multi-head structure adapts to the time period and the full-cycle pattern respectively, making the conversion rate results output by the model more accurate. Attached Figure Description

[0044] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0047] Figure 1 A flowchart of a method for estimating conversion rates based on time periods is provided for embodiments of this application;

[0048] Figure 2 A schematic diagram of a model architecture provided for an embodiment of this application;

[0049] Figure 3 A schematic diagram of a device structure based on time-segmented conversion rate estimation provided in an embodiment of this application;

[0050] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] The following disclosure provides numerous different embodiments or examples for implementing various structures of this application. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the scope of this application. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0053] To address the issue of low model prediction accuracy mentioned in the background art, this application embodiment improves the model's robustness to traffic fluctuations and conversion delays by training with incompletely recovered conversion data. It adopts a multi-head design with a shared underlying network, where separate prediction heads and a main prediction head output corresponding results, and improves prediction accuracy through time-segmented calibration.

[0054] The embodiments of this application are applicable to high-latency conversion scenarios, such as advertising conversion delays and order placement delays.

[0055] The following will describe in detail a method for estimating conversion rates based on time periods, provided in the embodiments of this application, using a server as an example. Figure 1 As shown, the specific steps are as follows:

[0056] Step 101: Determine the delay time from the exposure time to the current time for the ad exposure sample;

[0057] Step 102: Label each conversion period with a time-segment label based on the delay duration. The definite label in the time-segment label is used to indicate samples whose delay duration exceeds the corresponding conversion period, and the indefinite label in the time-segment label is used to indicate samples whose delay duration does not exceed the corresponding conversion period.

[0058] Step 103: Input the ad exposure samples and time-segment tags into the initial model, wherein the initial model includes a main prediction head and multiple sub-prediction heads corresponding to multiple conversion time periods;

[0059] Step 104: Adjust the parameters of the corresponding sub-prediction head according to the samples corresponding to the determined labels, and adjust the parameters of the main prediction head when the labels are determined in each conversion period to obtain the trained model. Among them, the samples corresponding to the uncertain labels do not participate in the model training.

[0060] Step 105: Input the ad exposure data into the trained model to obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head.

[0061] First, the terms involved in the embodiments of this application will be explained, including the following content.

[0062] Ad exposure sample: refers to a single record of an ad being seen (exposed) by a user, including relevant information such as the time of exposure, user attributes, and ad characteristics.

[0063] Delay duration: The time elapsed from the moment the ad is displayed to the current moment, used to determine whether the conversion status of the sample can be determined in each conversion period.

[0064] Conversion period: Multiple preset time intervals (such as 6h, 12h, 18h, 24h after exposure, the next day, and the following 3 days) are used to subdivide the conversion situation in different time periods.

[0065] Define the label: When the delay time exceeds a certain conversion period, the label of that period is defined, and it can be further labeled as a positive sample (conversion occurred within that period) or a negative sample (no conversion occurred within that period).

[0066] Uncertain Labels: When the delay duration does not exceed a certain conversion period, the label for that period is an uncertain label, and the sample will not participate in the training of the corresponding prediction head.

[0067] Initial model: A deep learning model containing a main predictor head and multiple sub-predictors head, where each sub-predictor head corresponds one-to-one with a transformation period and shares the sample features extracted by the underlying network.

[0068] Prediction Header: A model component corresponding to a specific conversion period, used to output the predicted cumulative conversion rate for that period. Its parameters are adjusted based on the specific labeled samples for that period.

[0069] Main Prediction Header: A model component that outputs the overall conversion rate of the ad throughout the entire time period. Its parameters are adjusted only when all conversion periods have a fixed tag in order to learn the complete conversion pattern.

[0070] Cumulative conversion rate: The probability of a user converting within a certain conversion period (e.g., within 6 hours after exposure), output by the corresponding prediction header.

[0071] Overall Conversion Rate: The overall probability of a user converting within the entire period after ad exposure (covering all conversion periods), output by the main prediction head.

[0072] In step 101, the server extracts basic information about the ad exposure samples from the ad delivery log system, including the sample's unique identifier (such as exposure ID), exposure time (accurate to the second), and the current system time. The delay time is obtained by calculating the time difference between the current time and the exposure time; that is, the time elapsed from when the ad is seen by the user (exposure) to the current time. For example, if an ad is exposed by a user at 8:00 AM and the current time is 6:00 PM, the delay time is 10 hours. This step provides a time baseline for subsequent time-segmented tagging, ensuring that the tags accurately reflect the status of the sample in each conversion period, and is the foundation for achieving time-segmented modeling.

[0073] In step 102, the server is pre-configured with multiple conversion time periods, such as 6h, 12h, 18h, 24h, the next day, and the following 3 days. Each time period corresponds to the time interval from the exposure time to the preset end time. The core principle of this setting is based on the characteristics of user behavior in advertising conversion scenarios, which have significant delays and large differences in individual conversion times. Different users may convert within hours to days after exposure, and a single time window cannot accurately capture the conversion patterns at each stage.

[0074] For ad exposure samples, the server labels them by comparing their delay duration with the cutoff time of each time period. When the delay duration exceeds the cutoff time of a certain time period, the conversion status of that time period is clear (e.g., with a 10-hour delay, it is certain whether there will be a conversion within 6 hours and 12 hours), so it is labeled as "confirmed". When the delay duration does not exceed the cutoff time of a certain time period, the conversion status may still change (e.g., with a 10-hour delay, it is not yet known whether there will be a conversion after 18 hours), so it is labeled as "uncertain".

[0075] The purpose of this fine-grained annotation is to provide reliable data for model training at different stages. Samples with confirmed labels can be directly used for training the model in the corresponding time period, ensuring that the learning of conversion patterns in each time period is based on facts. Samples with uncertain labels are temporarily excluded from training to avoid introducing noise from undefined data. This approach fully utilizes the already defined conversion information while avoiding interference from uncertain data on model accuracy, laying a high-quality data foundation for subsequent time-segmented model training and improving adaptability to delayed conversion scenarios.

[0076] In step 103, the server calls a pre-built initial model, which includes a main prediction head and multiple sub-prediction heads corresponding to the conversion time periods (e.g., 6 sub-prediction heads corresponding to 6 time periods), and all prediction heads share the underlying feature extraction network (e.g., MLP or DeepFM). The server takes the feature data of the ad exposure samples (e.g., user age, ad type, exposure scenario, etc.) and the labeled time period labels as input and feeds them into the underlying network of the model: the underlying network first preprocesses the features (e.g., normalization, embedding), and then outputs high-dimensional feature vectors, which are sent to the main prediction head and each sub-prediction head respectively.

[0077] For example, after the feature vector of a sample is processed by the underlying network, it is simultaneously fed into the main prediction head and 6h and 12h sub-prediction heads. This step provides structured input for model training, ensuring the correlation between features and labels, and is a prerequisite for realizing multi-head parallel learning.

[0078] In step 104, the training of the sub-prediction head uses only samples (positive and negative samples) labeled with definite tags for the corresponding time period. The weight parameters of the sub-prediction head are adjusted based on the backpropagation algorithm, allowing it to learn the conversion patterns within that time period (e.g., the 6-hour sub-prediction head focuses on learning the conversion features from 0 to 6 hours). The core principle of using only samples labeled with definite tags for the corresponding time period is that the conversion status of the definite-tag samples is clear (e.g., whether a conversion occurred within 6 hours is unambiguous). Training based on this allows the sub-prediction head to accurately capture the conversion features within that time period (e.g., the correlation between user behavior preferences and conversion within 0-6 hours), avoiding noise interference introduced by uncertain-tag samples (where conversion status may change over time). This targeted training significantly improves the accuracy of conversion rate prediction for each time period.

[0079] For the master predictor, the sample is only used for training when all time periods of the sample have definite labels (i.e., the delay time exceeds the longest period, such as the last 3 days). Similarly, the difference between the predicted value and the actual conversion result of the whole cycle is calculated by the loss function, and the parameters of the master predictor are adjusted so that it learns the conversion pattern of the whole cycle. The parameter adjustment is only made when all time periods of the sample have definite labels because the master predictor needs to learn the conversion pattern of the whole cycle. Only when the conversion status of all time periods is clear (e.g., the delay exceeds the last 3 days, and the conversion situation of each time period is determined) can a complete full-cycle sample be formed, ensuring that the master predictor learns the true and unbiased long-term conversion trend.

[0080] Uncertain label samples do not participate in training, further ensuring the reliability of the training data. For example, samples with a 10-hour delay are only used to train the 6h and 12h sub-estimation heads, while samples with a delay exceeding 3 days serve all sub-estimation heads and the main estimation head. This mechanism of time-sharing focus + full-cycle calibration enables the sub-estimation heads to achieve refined time-period modeling and allows the main estimation head to grasp the overall conversion pattern, ultimately improving the accuracy and robustness of the overall conversion rate estimation.

[0081] In step 105, when new advertisement exposure data is generated, the server inputs its feature data (such as user behavior, advertisement attributes) into the trained model: each sub-estimation head outputs the cumulative conversion rate for the corresponding time period based on the feature vector and the optimized parameters (for example, the 6h sub-estimation head outputs the conversion probability from 0 to 6h, and the 12h sub-estimation head outputs the cumulative conversion probability from 0 to 12h); the main estimation head outputs the overall conversion rate, that is, the total conversion probability within the complete cycle after the advertisement exposure (such as the next 3 days), based on the feature vector and the full-cycle parameters.

[0082] In this application, the delay duration of the advertisement exposure sample is determined, and then sub-time period labels are marked according to the delay duration. Samples with a delay duration exceeding the corresponding conversion time period are marked as certain labels, and those not exceeding are uncertain labels. During model training, only certain label samples are used to avoid interference from uncertain label samples (whose conversion status is not yet determined) on the training accuracy, ensuring that each sub-estimation head learns clear time-period conversion patterns. Subsequently, the samples and labels are input into the initial model with a multi-head structure. The sub-estimation heads adjust the parameters for the certain label samples in each time period, focusing on capturing the conversion features of the corresponding time period; the main estimation head is trained when all time periods are certain labels, focusing on the full-cycle conversion pattern. Finally, the trained model outputs the sub-time period cumulative conversion rate and the overall conversion rate. By screening certain label samples, noise data can be avoided, and the multi-head structure is respectively adapted to the time period and full-cycle patterns, making the conversion rate results output by the model more accurate.

[0083] Figure 2 This is a schematic diagram of the multi-task learning model architecture of this application. It can be seen that the model includes SharedLayers (shared layer), and the shared layer can use the mmoe network or other common networks, such as underlying networks like MLP, DeepFM, DCN (deepcrossnet), masknet, etc. The shared layer extracts the feature vector of the advertisement exposure sample and delivers it to the main estimation head and each sub-estimation head respectively.

[0084] Each P(z < wi|y = 1, x) is a sub-task head, which predicts the conversion rate that z is less than the threshold wi under the condition of y = 1 (or other main task label assumptions) based on the shared layer features. P(y = 1|x) is the main task head, which directly predicts the conversion rate for the entire time period based on the shared layer features.

[0085] cross_entropy_loss: For each task branch (including the main task header and sub-task headers), the loss is calculated using labeled data to supervise model learning.

[0086] As an optional implementation, in step 101, the time-segment labels for each conversion period based on the delay duration include the following:

[0087] Step S11: Determine multiple conversion periods, where each conversion period is the length of time from the sample exposure time to the preset cutoff time, and the length of time for different conversion periods is different;

[0088] Step S12: If the delay time of the ad exposure sample exceeds the time length of the preset conversion period, then mark the preset conversion period with a confirmation label;

[0089] Step S13: Based on the conversion results of the ad exposure samples within the preset conversion period, mark the samples corresponding to the determined tags as positive samples or negative samples. Positive samples are used to indicate that the ad exposure samples have been converted within the preset conversion period, and negative samples are used to indicate that the ad exposure samples have not been converted within the preset conversion period.

[0090] Step S14: If the delay time of the ad exposure sample does not exceed the time length of the preset conversion period, then mark the preset conversion period with an uncertain label.

[0091] In step S11, to adapt to the significant time differences and delays in user conversion behavior in advertising scenarios (e.g., some users may convert within hours of exposure, while others may be delayed for several days), the server pre-sets multiple conversion time periods. Each time period is defined as the length of time from the sample exposure time to a preset deadline, and the duration of each time period is different (e.g., w1 = 6h, w2 = 12h, w3 = 24h, w4 = the next 3 days, etc.). The core principle of this design is to break down the entire conversion process into continuous time intervals, capture the conversion patterns of different stages through fine-grained time division, provide a basic framework for subsequent time-segmented modeling, and avoid the drawbacks of traditional single-window models that cannot distinguish between short-term and long-term conversions.

[0092] In step S12, for each advertisement exposure sample, the server calculates the latency duration z from the exposure to the current time, and compares it one by one with the cut-off times w1, w2, w3, w4 of each preset conversion period. If the latency duration z ≥ a certain period wi (e.g., z = 10h ≥ w1 = 6h, w2 = 12h), it is determined that the conversion result of this period has stabilized (will not change over time), and a definite label is assigned to it. The principle of this operation is based on the time boundary of the conversion behavior. After a specific duration, the possibility of user conversion no longer changes. Thus, samples with clear results can be screened out, providing a reliable data basis for model training and reducing the interference of uncertain data.

[0093] In step S13, for the period with a definite label assigned, the server further combines the actual conversion results (such as click, purchase, etc.) of the sample within the period 0 - wi, and labels the sample as a positive sample (conversion occurred) or a negative sample (no conversion occurred). Specifically, for training the above network, when annotating model data, in addition to regular feature generation, it is necessary to additionally annotate the label by period, which needs to be calculated based on the time from exposure to the current time. Taking the current latency as z and the target cut-off periods as w1, w2, w3, w4 for illustration, for the case of z ≥ wi, the label of the i-th head can be determined, so it is labeled as a determined sample. At the same time, it is labeled as a positive or negative sample according to the actual conversion situation within the time period 0 - wi. The principle of this process is to enable the model to learn the feature differences between conversion and non-conversion within each period (such as the correlation between user portraits, advertisement attributes and short-term conversion) through clear distinction between positive and negative samples, directly improving the accuracy of period-based prediction.

[0094] In step S14, if the sample latency duration z < a certain period wi (e.g., z = 10h < w3 = 24h), it indicates that this period has not ended, and the user may still convert within the remaining time (not converting currently does not mean not converting ultimately). Therefore, an uncertain label is assigned to this period. According to the annotation rule, for the case of z < wi, it means that the time represented by the i-th head has not arrived, and it is labeled as an uncertain sample of the wi-th head, and the sample is uniformly set as a negative sample. The principle of this design is to recognize the time delay of conversion. By temporarily storing uncertain samples and labeling them as negative samples (to avoid misjudgment), it prevents them from contaminating model training as pseudo-negative samples, ensuring that parameter updates are based on reliable annotation information.

[0095] In this application, through the double-layer annotation system of definite label / uncertain label + positive and negative samples, both the determined samples with z ≥ wi (stable results) are utilized, and the uncertain samples with z < wi (temporarily stored for verification) are processed, solving the problem of estimation deviation in traditional models caused by directly using incomplete conversion data.

[0096] As an optional implementation, step 104, adjusting the parameters of the corresponding sub-prediction head based on the sample corresponding to the determined label, and adjusting the parameters of the main prediction head when the label is determined in each conversion period, includes the following:

[0097] Step S21: Obtain the feature vector of the ad exposure sample through the underlying shared network of the initial model;

[0098] Step S22: Learn the transformation pattern of feature vectors in the corresponding transformation period through the sub-prediction head, and learn the transformation pattern of feature vectors in the whole period through the main prediction head;

[0099] Step S23: Calculate the first difference between the predicted value of the sub-predictor and the sample label value in the corresponding conversion period using a binary classification loss function, and adjust the parameters of the sub-predictor based on the first difference. The predicted value of the sub-predictor is the product of the task output of the sub-predictor and the original output of the main predictor.

[0100] Step S24: When each conversion period has a definite label, the second difference between the predicted value of the master predictor and the sample label value in the whole period is calculated using a binary classification loss function, and the parameters of the master predictor and the underlying shared network are adjusted based on the second difference.

[0101] In step S21, the underlying shared network of the initial model (such as MLP, DeepFM, etc.) is responsible for uniformly extracting and deeply processing the original features of the ad exposure samples (such as user age, ad type, exposure scenario, etc.) to generate high-dimensional feature vectors. Its core principle is based on a feature sharing mechanism, processing all sample features through the same network structure to ensure consistency and completeness of feature representations used by the sub-prediction heads and the main prediction head, avoiding learning biases caused by differences in feature extraction. This step lays the foundation for subsequent multi-task learning. By integrating scattered original features, the model can capture the deep correlation between user and ad interaction (such as the matching degree between user preferences and ad attributes), improving the explanatory power of features for conversion behavior.

[0102] In step S22, each sub-prediction head corresponds one-to-one with a conversion time period (e.g., 6h, 12h, etc.), focusing on learning the conversion patterns within the corresponding time period based on feature vectors. For example, the 6h sub-prediction head focuses on the immediate conversion characteristics within 6 hours after user exposure (e.g., short-term impulsive consumption signals), while the 12h sub-prediction head focuses on the conversion trend from 6 to 12 hours (e.g., decision-making characteristics after users learn more). The main prediction head, based on the same feature vector, learns the overall conversion patterns across all time periods (e.g., the following 3 days), integrating the correlations between features from different time periods (e.g., the evolution from short-term interest to long-term decision-making).

[0103] The principle behind this design is task decomposition and collaboration. By splitting tasks into time-segmented tasks, the model can specifically capture transformation patterns at different time granularities. By sharing feature vectors, the learning patterns of each head maintain inherent consistency. The effect is to avoid the drawback of a single model being unable to take into account both short-term and long-term transformations, allowing the sub-predictors and the main predictor to achieve accurate learning in their respective time dimensions.

[0104] In step S23, the predicted value of the sub-prediction head is the task output of the sub-prediction head multiplied by the original output of the main prediction head. The principle of this product structure is to combine the time period specificity and global correlation. The sub-task output reflects the conversion probability of the corresponding time period, and the main output reflects the overall trend of the conversion throughout the entire cycle. The multiplication of the two makes the cumulative conversion rate of the time period both fit the time period characteristics and is constrained by the global law.

[0105] The first difference between the predicted value and the corresponding sample label value (positive or negative) is calculated using a binary classification loss function (such as cross-entropy). Then, the parameters of the sub-prediction head are adjusted through the backpropagation algorithm. The effect is to make the sub-prediction head's predictions continuously approach the actual conversion results. For example, the 6-hour sub-prediction head, through loss adjustment, can more accurately distinguish the characteristics of samples that have been converted within 6 hours from those that have not been converted within 6 hours, significantly reducing the prediction error within the time period.

[0106] In step S24, when all conversion periods of the ad exposure sample have a definite label (i.e., the delay time exceeds the longest period, such as the last 3 days), the full-cycle conversion result is clear. At this time, the second difference between the predicted value of the main predictor (overall conversion rate for the entire period) and the label value of the sample for the entire period (whether it actually converted) is calculated using a binary classification loss function. Based on this difference, the model not only adjusts the parameters of the main predictor but also simultaneously optimizes the parameters of the underlying shared network. The principle is that the global learning of the main predictor relies on high-quality feature vectors, while the optimization of the underlying network parameters can further improve the accuracy of feature extraction, forming a closed-loop optimization of feature-prediction-parameter.

[0107] The main predictor learns from samples over a complete cycle, enabling it to more accurately capture long-term conversion patterns (such as the full-process features from user attention to final conversion); the synchronous optimization of the underlying network makes the feature vectors more aligned with the learning needs of each predictor, improving the overall feature representation capability of the model.

[0108] In this application, a collaborative mechanism of feature sharing, task splitting, loss adjustment, and closed-loop optimization is used to achieve multi-dimensional improvements in model performance. The underlying shared network ensures feature consistency and avoids learning bias in each head; the targeted learning of the sub-prediction heads and the main prediction head improves the prediction accuracy of conversion rates for different time periods and the entire cycle, respectively; the product structure of the predicted values ​​of the sub-prediction heads deeply correlates the results of different time periods with the global trend, reducing the one-sidedness of predictions for isolated time periods; and the coordinated optimization of the main prediction head and the underlying network enhances the adaptability of feature extraction while improving the global prediction capability.

[0109] As an optional implementation, step S22, learning the transformation rules of feature vectors in the corresponding transformation period by sub-prediction heads, includes: determining the positive or negative samples corresponding to each transformation period; learning the feature rules of positive samples transforming in the corresponding transformation period by sub-prediction heads, and learning the feature rules of negative samples not transforming in the corresponding transformation period by sub-prediction heads.

[0110] When learning the conversion patterns of feature vectors, the predictive head first identifies positive samples (samples that convert within the corresponding time period after exposure) and negative samples (samples that do not convert within the corresponding time period after exposure) based on the defined labels of each conversion time period. For example, for a 6-hour conversion time period, positive samples are those that complete conversion behaviors such as clicks and purchases within 6 hours after exposure, while negative samples are those that do not convert at all within 6 hours.

[0111] Subsequently, the prediction head performs targeted learning on the feature vectors of the two types of samples: for positive samples, the model focuses on extracting the patterns in their feature vectors that are related to conversion, such as high activity levels and historical click records of similar ads in user characteristics, or a combination of features such as large discounts and high matching degree of user interests in ad characteristics; for negative samples, the model focuses on learning the patterns in their feature vectors that are not related to conversion, such as low online time and no preference for ad categories in user characteristics, or a combination of features such as a single display format and mismatch between ad characteristics and user profile.

[0112] The core principle of this process is contrastive learning and pattern differentiation. By comparing the feature differences between positive and negative samples, the predictor can accurately distinguish the feature patterns of the two types of samples, thereby quickly identifying feature signals related to transformation in new samples during the inference stage and improving the prediction accuracy of the transformation probability for that period.

[0113] This application, through targeted learning of the characteristic patterns of positive and negative samples, enables the prediction head to capture the specific characteristics of conversion behavior in different time periods (such as the impulsive conversion characteristics in the 6-hour period and the rational decision-making characteristics in the 24-hour period). This avoids the crudeness of using a single pattern to explain conversions in all time periods, making the conversion rate predictions for each time period more consistent with real-world scenarios. For advertising scenarios with obvious delay characteristics (such as users needing multiple time periods to consider before converting), the prediction head's learning of the patterns of positive and negative samples in different time periods can effectively distinguish between samples that have not converted in the short term but may convert in the long term and samples that have been verified as not converting, reducing misjudgments caused by delayed conversions.

[0114] As an optional implementation, the method further includes the following:

[0115] Step S31: When the delay of the ad exposure sample exceeds the deadline of the target conversion period, update the label of the target conversion period from an uncertain label to a certain label, and use the sample corresponding to the certain label to perform incremental training of the corresponding sub-prediction head.

[0116] Step S32: When all labels for the conversion time periods are updated to definite labels, use the samples corresponding to all conversion time periods to train the master predictor.

[0117] In step S31, when the delay time of the ad exposure sample exceeds the deadline of a certain target conversion period (e.g., 12h), it means that the conversion status of that period is clear (and will no longer change over time). The server updates the label of that period from an uncertain label to a certain label and completes the labeling based on the actual conversion result of the sample in that period (if it occurs, it becomes a positive sample; if it does not occur, it becomes a negative sample).

[0118] To make full use of real-time data and ensure the timeliness of samples, it is necessary to relabel the samples every certain period of time as time goes by. Based on the latest delay time z', the samples are relabeled according to the above rules. For example, if a sample is initially delayed by 8 hours, the 12-hour period is an uncertain label; when the delay reaches 13 hours (z' = 13 hours), the 12-hour period is updated to a definite label after relabeling.

[0119] Based on the updated, definitive labeled samples, the sub-predictor head undergoes incremental training. Loss is calculated only for newly defined samples within that time period, and the sub-predictor head parameters are fine-tuned without retraining on all data. Its core principle is to dynamically track sample states. Through time-dimensional label updates, previously undefined samples are transformed into reliable training data, enabling the sub-predictor head to promptly capture newly determined conversion patterns (such as positive / negative sample features added within a 12-hour timeframe). This step avoids training lag caused by outdated sample states, allowing the sub-predictor head to continuously absorb the latest data and improve the prediction accuracy of conversion probabilities for the corresponding time period.

[0120] In step S32, when all the labels for the conversion time periods (e.g., 6h, 12h, 24h, etc.) of the ad exposure sample are updated to definite labels (i.e., the delay time exceeds the longest time period deadline, such as 3 days later), the full-cycle conversion data of the sample is complete. At this time, the server calls the definite label samples of all conversion time periods (covering positive / negative samples of each time period) to train the master predictor. By calculating the loss between the predicted value of the master predictor and the actual conversion results of the entire time period, the parameters of the master predictor and the underlying shared network are adjusted.

[0121] The principle behind this is that complete periodic data supports global learning. The main forecasting head needs to learn the overall conversion patterns across all time periods. Only when all time period labels are determined can the sample contain complete conversion information from exposure to the longest period, accurately reflecting the user's conversion behavior chain from short-term to long-term (such as the full-process characteristics of initial attention → in-depth understanding → final conversion). The effect of this step is to ensure that the parameter optimization of the main forecasting head is based on complete and unambiguous data, making its output overall conversion rate more consistent with the actual long-term conversion trend.

[0122] This application employs a collaborative mechanism of dynamic label updates, timed relabeling, and incremental training to achieve precise adaptation of the model to delayed conversion scenarios. Labels are updated periodically based on the latest delay duration z', ensuring that the sample status always reflects the current situation and avoiding training bias caused by using outdated labels. This resolves the contradiction between traditional models missing timeliness due to waiting for complete data and insufficient accuracy due to using incomplete data. The sub-prediction heads continuously absorb newly determined time-segment samples through incremental training, constantly optimizing the learning of conversion patterns for each time segment and improving short-term prediction accuracy. The main prediction head is trained after all time-segment data is complete, ensuring the accuracy of long-term conversion patterns. The two complement each other, providing short-term dynamic adaptation and long-term global calibration.

[0123] As an optional implementation, in step 105, the ad exposure data is input into the trained model to obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head, including the following:

[0124] Step S41: Input the ad exposure data into the trained model, and extract the feature vector by sharing the underlying network to extract features from the ad exposure data;

[0125] Step S42: Based on the feature vector, estimate the cumulative conversion rate within the corresponding conversion period using the sub-prediction head;

[0126] Step S43: Based on the feature vector, the overall conversion rate is estimated by the main prediction head based on samples from all conversion periods.

[0127] In step S41, the ad impression data is input into the trained model. The model first performs deep processing on this raw data through a shared underlying network (such as MLP, DeepFM, etc.): including feature cleaning, dimensionality transformation, cross-combination, etc., ultimately generating high-dimensional feature vectors. These feature vectors condense key information related to conversion behavior in the raw data (such as user interest in ads, ad matching degree, etc.). Its core principle is feature sharing and standardization. By extracting features through the same set of underlying networks, it ensures that the basic features used by the sub-prediction head and the main prediction head are consistent, avoiding prediction bias caused by differences in feature sources, and providing a unified and high-quality input foundation for subsequent multi-dimensional conversion rate prediction.

[0128] In step S42, based on the feature vector generated in step S41, each sub-prediction head (corresponding one-to-one with conversion time periods such as 6h, 12h, and 24h) combines its optimized parameters to predict the cumulative conversion rate for the corresponding time period. For example, the 6h sub-prediction head focuses on signals related to conversion within 6 hours after exposure in the feature vector (such as users' short-term click intention and immediate decision-making features), and outputs the cumulative conversion probability from 0 to 6 hours; the 12h sub-prediction head integrates the conversion patterns from 0 to 12 hours and outputs the cumulative conversion probability including the first 6 hours. The principle is time-specific modeling. Each sub-prediction head learns the positive and negative sample features of the corresponding time period, forming a unique understanding of the conversion patterns for that time period. It can accurately capture subtle patterns related to conversion in the feature vector for that time period, ensuring that the output cumulative conversion rate matches the actual conversion trend of that time period.

[0129] In step S43, the main prediction head, also based on the feature vector from step S41, combines the parameters obtained from full-cycle training and integrates the sample patterns of all conversion periods (such as the correlation between short-term impulsive conversions and long-term rational conversions) to output the overall conversion rate within the complete cycle after ad exposure (e.g., the last 3 days). The principle is global pattern integration. By learning from samples with defined labels in all time periods, the main prediction head has grasped the complete conversion chain characteristics from exposure to the longest period. It can extract signals reflecting long-term conversion potential from the feature vector (such as users' continued attention to the ad, the cumulative effect of brand preference, etc.), thus providing a conversion probability assessment covering the entire cycle.

[0130] In this application, the sub-prediction head focuses on the specific patterns of each time period, while the main prediction head focuses on the overall patterns of the entire cycle. The two work together based on the same feature vector, which not only ensures the accurate capture of the transformation details of each time period, but also enables a reliable judgment of the long-term transformation trend, thus resolving the contradiction between the one-sidedness of short-term prediction and the ambiguity of long-term prediction.

[0131] As an optional implementation method, after obtaining the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head, the method further includes the following: if it is determined that the advertisement needs to be successfully converted in a short period of time, then based on the cumulative conversion rate of the sub-prediction heads, the advertisement with the highest cumulative conversion rate within the shortest conversion period is prioritized for placement; if it is determined that the advertisement needs to be successfully converted in a long period of time, then the placement budget and targeting strategy are adjusted by combining the overall conversion rate of the main prediction head and the cumulative conversion rate corresponding to the longest conversion period.

[0132] When the goal of advertising is to achieve conversions in the short term (such as limited-time promotions, new product launches, etc., where clicks and purchases need to be increased rapidly within a short period), the system will prioritize ads with the highest cumulative conversion rate within the shortest conversion period (e.g., 6 hours after exposure) based on the cumulative conversion rate output by the sub-prediction header. For example, in a 6-hour flash sale event on an e-commerce platform, if ad A has a cumulative conversion rate of 15% and ad B has 8% within 6 hours, traffic will be prioritized for ad A to ensure maximum conversion efficiency in the short term.

[0133] When the advertising objective is long-term conversion (such as brand awareness cultivation, high-value customer retention, etc., which require continuous conversion over a longer period), the system will dynamically adjust the advertising strategy by combining the overall conversion rate output by the main forecast header (reflecting the full-cycle conversion potential) with the cumulative conversion rate corresponding to the longest conversion period (such as 3 days after exposure) (reflecting the actual progress of long-term conversion): If the overall conversion rate of an ad is 20% and the cumulative conversion rate over 3 days reaches 18%, it indicates that its long-term conversion stability is strong, and the advertising budget can be increased and targeted to high-potential user groups; if the overall conversion rate is high but the cumulative conversion rate over 3 days is low, the targeting strategy needs to be optimized (such as adjusting user profiles) or the tracking period needs to be extended to avoid wasting resources.

[0134] This application utilizes high-conversion-rate ads that focus on the shortest conversion period to accurately capture users' immediate decision-making intentions (such as impulsive consumption needs) within a limited time, maximizing short-term conversion results (such as clicks and immediate order volume). This is particularly suitable for time-sensitive marketing campaigns (such as limited-time discounts) and reduces cost losses caused by ineffective traffic. By combining the overall conversion rate with the cumulative conversion rate over the longest period, the long-term conversion potential and stability of the ads can be comprehensively evaluated, avoiding misjudgments caused by relying solely on short-term data (such as an ad having a low short-term conversion rate but high long-term potential).

[0135] This application provides an overall process for estimating conversion rates based on time periods, including the following steps.

[0136] I. Data preparation and preprocessing.

[0137] 1. Determine the conversion period: Preset multiple differentiated conversion periods, such as 6h, 12h, 18h, 24h after exposure, the next day and the following 3 days. Each period is the length of time from the exposure time to the corresponding preset end time.

[0138] 2. Calculate the delay duration: For each ad exposure sample, calculate the delay duration from the exposure time to the current time (i.e., the time elapsed after exposure).

[0139] 3. Label the time periods.

[0140] If the delay duration is greater than or equal to the deadline of a certain conversion period, that period is marked as a definite label, and based on whether the sample has been converted within that period, it is further labeled as a positive sample (converted) or a negative sample (not converted).

[0141] If the delay duration is less than the deadline of a certain conversion period, that period is labeled as uncertain, and the samples are temporarily set as negative samples.

[0142] Dynamic label updates: As time goes by, samples are relabeled at regular intervals based on the latest delay duration. Uncertain labels that meet the condition of delay duration ≥ deadline are updated to certain labels to ensure the timeliness of samples.

[0143] II. Model Construction and Training.

[0144] 1. Model structure: A shared underlying network + multi-head output architecture is adopted, including one main predictor and six sub-predictors corresponding to the conversion period. The underlying network can be MLP, DeepFM, etc., and the output of each sub-predictor is its own task output * the original output of the main predictor.

[0145] 2. Sub-prediction head training: Using only samples labeled with a definite label, the difference between the predicted value and the sample label is calculated through a binary classification loss function (such as cross-entropy). The corresponding sub-prediction head parameters are adjusted so that it learns the conversion pattern of that time period (e.g., the 6h head focuses on the conversion features from 0 to 6h).

[0146] 3. Master Predictor Training: Only when all conversion periods of the sample have definite labels, use full-period samples to calculate the loss and adjust the master predictor parameters to learn the conversion patterns throughout the entire cycle.

[0147] III. Model Reasoning and Application.

[0148] 1. Input ad exposure data: Input the newly generated ad exposure data (including user characteristics, ad attributes, etc.) into the trained model, and extract feature vectors through the shared underlying network.

[0149] 2. Output conversion rate results: The sub-prediction head outputs the cumulative conversion rate for the corresponding time period (such as the cumulative conversion probability within 6 hours and 12 hours); the main prediction head outputs the overall conversion rate for the entire cycle.

[0150] 3. Develop a deployment strategy.

[0151] Short-term conversion needs: Prioritize ads with the highest cumulative conversion rate for the shortest time period (e.g., 6 hours).

[0152] Long-term conversion needs: Adjust the campaign budget and targeting strategy by combining the overall conversion rate of the main estimated head with the cumulative conversion rate of the longest period (such as the last 3 days).

[0153] This application sets up multiple sub-prediction heads to output prediction results for different time periods separately, thereby providing more detailed conversion rates by time period. At the same time, it uses delayed conversion data with incomplete attribution to optimize the model, ensuring both data timeliness and accuracy, thus achieving a balance between the two.

[0154] Based on the same technical concept, this application provides a device for estimating conversion rates based on time periods, such as... Figure 3 As shown, the device includes:

[0155] The determination module 301 is used to determine the delay time from the exposure time to the current time of the advertisement exposure sample;

[0156] The labeling module 302 is used to label the time period of each conversion period according to the delay duration. The definite label in the time period label is used to indicate the sample whose delay duration exceeds the corresponding conversion period, and the indefinite label in the time period label is used to indicate the sample whose delay duration does not exceed the corresponding conversion period.

[0157] Input module 303 is used to input ad exposure samples and time-segment labels into the initial model, wherein the initial model includes a main prediction head and multiple sub-prediction heads corresponding to multiple conversion time periods;

[0158] The adjustment module 304 is used to adjust the parameters of the corresponding sub-predictor based on the samples corresponding to the determined labels, and to adjust the parameters of the main predictor when the labels are determined in each conversion period, so as to obtain the trained model. Samples corresponding to uncertain labels do not participate in model training.

[0159] The input / output module 305 is used to input advertising exposure data into the trained model to obtain the cumulative conversion rate of each sub-prediction head output within the corresponding conversion period and the overall conversion rate of the main prediction head output.

[0160] Optionally, the annotation module 302 is used for:

[0161] Multiple conversion periods are defined, where each conversion period is the length of time from the sample exposure time to the preset cutoff time, and the length of time varies for different conversion periods;

[0162] If the delay time of the ad exposure sample exceeds the time length of the preset conversion period, then the preset conversion period will be marked with a confirmation label;

[0163] Based on the conversion results of the ad exposure samples within the preset conversion period, the samples corresponding to the identified tags will be labeled as positive or negative samples. Positive samples indicate that the ad exposure samples have converted within the preset conversion period, while negative samples indicate that the ad exposure samples have not converted within the preset conversion period.

[0164] If the delay time of the ad exposure sample does not exceed the time length of the preset conversion period, then the preset conversion period will be marked with an uncertain label.

[0165] Optionally, the adjustment module 304 is used for:

[0166] The feature vectors of ad impression samples are obtained through the underlying shared network of the initial model;

[0167] The transformation rules of feature vectors in corresponding transformation periods are learned by using separate prediction heads, and the transformation rules of feature vectors in all time periods are learned by using the main prediction head.

[0168] The first difference between the predicted value of the sub-prediction head and the sample label value within the corresponding conversion period is calculated using a binary classification loss function, and the parameters of the sub-prediction head are adjusted based on the first difference. The predicted value of the sub-prediction head is the product of the task output of the sub-prediction head and the original output of the main prediction head.

[0169] When each conversion period has a definite label, a binary classification loss function is used to calculate the second difference between the predicted value of the master predictor and the sample label value throughout the period, and the parameters of the master predictor and the underlying shared network are adjusted based on the second difference.

[0170] Optionally, the adjustment module 304 is specifically used for:

[0171] Determine the positive or negative samples corresponding to each conversion period;

[0172] By using separate prediction heads, we learn the characteristic patterns of positive samples transforming within the corresponding transformation period, and by using separate prediction heads, we learn the characteristic patterns of negative samples not transforming within the corresponding transformation period.

[0173] Optionally, the device is also used for:

[0174] When the delay of the ad exposure sample exceeds the deadline of the target conversion period, the label of the target conversion period is updated from an uncertain label to a certain label, and the sample corresponding to the certain label is used for incremental training of the corresponding sub-prediction head.

[0175] When all labels for each conversion period are updated to definite labels, the master predictor head is trained using samples corresponding to all conversion periods.

[0176] Optionally, the input / output module 305 is used for:

[0177] The ad exposure data is input into the trained model, and the feature vector is obtained by extracting features from the ad exposure data through the shared underlying network.

[0178] Based on feature vectors, the cumulative conversion rate within the corresponding conversion period is estimated by using different prediction heads;

[0179] Based on feature vectors, the overall conversion rate is estimated by the main prediction head using samples from all conversion periods.

[0180] Optionally, the device is also used for:

[0181] If it is determined that the advertisement needs to achieve a successful conversion in a short period of time, then based on the estimated cumulative conversion rate of each segment, the advertisement with the highest cumulative conversion rate within the shortest conversion period will be prioritized for execution.

[0182] If it is determined that the advertisement needs to achieve successful conversions over a long period of time, then the advertising budget and targeting strategy should be adjusted based on the overall conversion rate of the main head and the cumulative conversion rate corresponding to the longest conversion period.

[0183] like Figure 4 As shown, this application provides an electronic device including a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404.

[0184] Memory 403 is used to store computer programs.

[0185] In one embodiment of this application, when the processor 401 executes the program stored in the memory 403, it implements the method based on time-segmented conversion rate estimation provided in any of the foregoing method embodiments.

[0186] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method based on time-segmented conversion rate estimation provided in any of the foregoing method embodiments.

[0187] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0188] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0189] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "one," "an," and "described" as used herein may also mean including the plural forms. The terms "comprising," "including," "containing," and "having" are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0190] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for predicting conversion rates based on time periods, characterized in that, The method includes: Determine the delay time from the exposure time to the current time for the ad exposure sample; Each conversion period is labeled with a time-segment tag based on the delay duration, wherein the definite tag in the time-segment tag is used to indicate samples whose delay duration exceeds the corresponding conversion period, and the indefinite tag in the time-segment tag is used to indicate samples whose delay duration does not exceed the corresponding conversion period; The ad exposure samples and time-segment tags are input into the initial model, wherein the initial model includes a main prediction head and multiple sub-prediction heads corresponding to multiple conversion time periods; The parameters of the corresponding sub-prediction head are adjusted according to the samples corresponding to the determined labels, and the parameters of the main prediction head are adjusted when the labels are determined in each conversion period to obtain the trained model. The samples corresponding to the uncertain labels do not participate in the model training. By inputting the ad exposure data into the trained model, we can obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head.

2. The method according to claim 1, characterized in that, The time-segment labels for each conversion period, based on the aforementioned delay duration, include: Multiple conversion periods are defined, where each conversion period is the length of time from the sample exposure time to the preset cutoff time, and the length of time varies for different conversion periods; If the delay time of the ad exposure sample exceeds the time length of the preset conversion period, then the preset conversion period is labeled with a determination tag; Based on the conversion results of the ad exposure samples within the preset conversion period, the samples corresponding to the determined tags are labeled as positive or negative samples. The positive samples are used to indicate that the ad exposure samples have been converted within the preset conversion period, and the negative samples are used to indicate that the ad exposure samples have not been converted within the preset conversion period. If the delay time of the ad exposure sample does not exceed the time length of the preset conversion period, then the preset conversion period is marked with an uncertain label.

3. The method according to claim 1, characterized in that, Adjusting the parameters of the corresponding sub-prediction head based on the sample corresponding to the determined label, and adjusting the parameters of the main prediction head when the label is determined in each conversion period, includes: The feature vector of the ad exposure sample is obtained through the underlying shared network of the initial model; The transformation pattern of the feature vector in the corresponding transformation period is learned by the sub-prediction head, and the transformation pattern of the feature vector in the whole time period is learned by the main prediction head. The first difference between the predicted value of the sub-prediction head and the sample label value within the corresponding conversion period is calculated using a binary classification loss function, and the parameters of the sub-prediction head are adjusted based on the first difference. The predicted value of the sub-prediction head is the product of the task output of the sub-prediction head and the original output of the main prediction head. When each conversion period has a defined label, the binary classification loss function is used to calculate the second difference between the predicted value of the master predictor and the sample label value throughout the entire period, and the parameters of the master predictor and the underlying shared network are adjusted based on the second difference.

4. The method according to claim 3, characterized in that, Learning the transformation patterns of the feature vectors during corresponding transformation periods by using a predictive head includes: Determine the positive or negative samples corresponding to each conversion period; The prediction head is used to learn the characteristic patterns of positive samples transforming during the corresponding transformation period, and the prediction head is used to learn the characteristic patterns of negative samples not transforming during the corresponding transformation period.

5. The method according to claim 1, characterized in that, The method further includes: When the delay of the ad exposure sample exceeds the deadline of the target conversion period, the label of the target conversion period is updated from an uncertain label to a certain label, and the sample corresponding to the certain label is used to perform incremental training of the corresponding sub-prediction head. When all labels for each conversion period are updated to definite labels, the master predictor head is trained using samples corresponding to all conversion periods.

6. The method according to claim 1, characterized in that, By inputting ad impression data into the trained model, we obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate output by the main prediction head, including: The ad exposure data is input into the trained model, and the feature vector is obtained by extracting features from the ad exposure data through a shared underlying network. Based on the feature vector, the cumulative conversion rate within the corresponding conversion period is estimated by the sub-prediction head; Based on the feature vector, the overall conversion rate is estimated by the main prediction head using samples from all conversion periods.

7. The method according to claim 1, characterized in that, After obtaining the cumulative conversion rate of each sub-prediction head output within the corresponding conversion period and the overall conversion rate output by the main prediction head, the method further includes: If it is determined that the advertisement needs to achieve a successful conversion in a short period of time, then based on the estimated cumulative conversion rate of each segment, the advertisement with the highest cumulative conversion rate within the shortest conversion period will be prioritized for execution. If it is determined that the advertisement needs to achieve successful conversions over a long period of time, then the advertising budget and targeting strategy should be adjusted based on the overall conversion rate of the main head and the cumulative conversion rate corresponding to the longest conversion period.

8. A device for predicting conversion rates based on time periods, characterized in that, The device includes: The determination module is used to determine the delay time from the exposure time of the ad exposure sample to the current time; The labeling module is used to label each conversion period with a time-segment label based on the delay duration. The definite label in the time-segment label is used to indicate samples whose delay duration exceeds the corresponding conversion period, and the indefinite label in the time-segment label is used to indicate samples whose delay duration does not exceed the corresponding conversion period. The input module is used to input the ad exposure samples and time-segment labels into the initial model, wherein the initial model includes a main prediction head and multiple sub-prediction heads corresponding to multiple conversion time periods; The adjustment module is used to adjust the parameters of the corresponding sub-prediction head according to the samples corresponding to the determined labels, and to adjust the parameters of the main prediction head when the labels are determined in each conversion period, so as to obtain the trained model. The samples corresponding to the uncertain labels do not participate in the model training. The input / output module is used to input ad exposure data into the trained model to obtain the cumulative conversion rate of each sub-prediction head within the corresponding conversion period and the overall conversion rate of the main prediction head.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images