A method and system for estimating the effect of remarketing advertising based on multi-task learning

Abstract

The present invention discloses a remarketing advertising effect prediction method and system based on multi-task learning, which belongs to the field of online advertising marketing information. The present invention learns and predicts multiple user behaviors (clicks / browsing / add-to-cart / purchase) after the advertisement is displayed, considers multiple influencing factors at the same time in the display link, automatically learns the correlation between features and targets, and solves the remarketing advertising effect prediction problem with one model. Specifically, knowledge is learned from massive data through offline feature design, target determination, sample construction and model training, and feature extraction and model scoring are completed for requests online in real time. The knowledge trained from the offline model is applied to the production environment to realize real-time remarketing advertising effect prediction. The present invention uses one model to realize the prediction of multiple indicators, which reduces the complexity of traditional multi-model solutions and overcomes the problem of poor training effect caused by sparse data of purchase / add-to-cart and other behaviors.

Description

Technical Field

[0001] The present invention belongs to the field of online advertising marketing information technology, and relates to a remarketing advertising effect prediction method and system based on multi-task learning. Background Art

[0002] E-commerce websites often conduct secondary off-site reach with users who recently visited their websites and had purchase intent but didn't complete their purchases. This process requires remarketing advertising. Independent third-party advertising platforms help advertisers purchase traffic in the traffic market. Advertisers settle accounts with independent third-party advertising platforms based on a cost-per-sale (CPS) basis. This means that if a user clicks on an ad placed by the independent third-party advertising platform and enters the website to complete an order, the independent third-party advertising platform receives a percentage of the order amount. Therefore, independent third-party advertising platforms need to help advertisers evaluate the effectiveness of advertising on external media websites for users of the advertiser's website, specifically the probability that the user will click on the ad and complete the purchase on the website, in order to maximize the independent third-party advertising platform's profits.

[0003] like Figure 1 As shown in the figure, the conversion path of remarketing ads is as follows: user sees the remarketing ad on the media website (Impression) -> user clicks the remarketing ad (Click) -> user browses the product on the advertiser's website (ViewContent) -> user adds the product to the shopping cart on the advertiser's website (AddToCart) -> user completes payment on the advertiser's website (Purchase)

[0004] Estimating the probability of a user clicking an ad is CTR (Click-Through-Rate) estimation, and estimating the probability of a user completing a purchase after clicking an ad is CVR (Conversion Rate) estimation. Conventional approaches involve building two models to estimate click-through rate and conversion rate, respectively. Currently, the advertising industry is very mature in CTR / conversion rate estimation. The basic idea is to learn models based on historical ad exposure and conversion results, and then estimate traffic CTR and conversion rate online in real time. Common methods include linear LR models, nonlinear FM, FFM, and GBDT models, as well as deep learning models such as Wide&Deep and DeepFM. CTR estimation uses user clicks after an ad is displayed as positive samples and non-clicks as negative samples to learn the classification model. Conversion rate estimation uses user purchases after a click as positive samples and non-purchases as negative samples to learn the classification model.

[0005] For remarketing scenarios, traditional solutions cannot be used directly due to the following challenges:

[0006] 1) There is cross-site behavior in the remarketing scenario. User clicks occur on the media website, and user purchases occur on the advertiser's e-commerce website. User clicks are related to the media website, and whether the user purchases is related to the user's own purchasing intention. There is a conflict in the characterization of the feature level.

[0007] 2) There are many intermediate steps. Users need to click on the advertisement, browse the product details page, add the product to the shopping cart, and complete the payment to complete the entire purchase process. There is a gap in directly estimating whether the user will make a purchase.

[0008] 3) As the shopping process goes deeper, user behaviors will become less and less, and there will be problems such as insufficient samples and inability to accurately predict. Summary of the Invention

[0009] The purpose of the present invention is to solve the problems in the prior art and provide a method and system for predicting the effect of remarketing advertising based on multi-task learning.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] The present invention proposes a method for estimating the effectiveness of remarketing advertising based on multi-task learning, comprising the following steps:

[0012] The goals are whether the user clicks, browses the site, adds items to the shopping cart, and makes a purchase.

[0013] Advertisement-related features, user-related features, and context-related features are taken as a feature list;

[0014] Construct sample examples based on the target and feature list, and label the sample examples;

[0015] Train the MMoE model based on the labeled sample examples and learn the MMoE model parameters;

[0016] Based on the offline trained MMoE model, features are extracted online in real time to obtain scores for multiple targets, enabling real-time estimation of remarketing advertising effects.

[0017] Preferably, the advertisement-related features include the advertisement ID, the creative material ID, and the historical delivery effect of the advertisement;

[0018] Context-related features include the current day of the week, hour, ADX, media ID, ad slot type, ad slot size, and media’s historical delivery performance;

[0019] User-related characteristics include the user's geographic location (country / city), device information (mobile phone model / operating system / OS version / language) and user behavior statistics on the advertiser's website.

[0020] Preferably, whether the user clicks, whether the user browses the site, whether the user adds to the shopping cart, and whether the user purchases can be associated based on a unique tracking ID to achieve sample labeling.

[0021] Preferably, the MMoE model adopts a shared underlying structure. The Shared-Bottom network is located at the bottom of the MMoE model. Multiple tasks use the Shared-Bottom network to learn common features.

[0022] The Shared-Bottom network contains multiple expert networks that can learn different aspects of the sample characteristics; each task corresponds to a Tower Network to learn the unique features of the current task;

[0023] Multiple Gate gating networks are used to select subsets of expert networks according to different tasks.

[0024] Preferably, a Shared-Bottom structure is adopted to learn features and model parameters using sample examples when learning purchasing behavior.

[0025] Preferably, when the MMoE model is used to train the feature list, ClassWeight is added to use the inverse of the sample instance as the class weight.

[0026] Preferably, use the tf-serving component of the TensorFlow framework. tf-serving loads the trained MMoE model and combines the requests from Adx, the ad information in the ad library, and the user profiles in the DMP to complete feature extraction.

[0027] Based on the feature extraction results, we obtain the probability P(Click) of whether the user clicks, the probability P(ViewContent) of whether the user browses the site, the probability P(AddToCart) of whether the user adds the item to the shopping cart, and the probability P(Purchase) of whether the user purchases the item. Based on the probability P(Click) of whether the user clicks, the probability P(ViewContent) of whether the user browses the site, the probability P(AddToCart) of whether the user adds the item to the shopping cart, and the probability P(Purchase) of whether the user purchases the item, we obtain the scores of multiple targets. The scores of multiple targets are calculated as follows:

[0028] Score of multiple targets = P(Click) a *P(ViewContent) b *P(AddToCart) c *P(Purchase)d

[0029] Among them, a, b, c, and d are parameters. The scores of multiple targets are used to sort the candidate ads, and the ads with the highest scores are used as the currently displayed ads to achieve the estimation of remarketing advertising effects.

[0030] The present invention proposes a remarketing advertising effect prediction system based on multi-task learning, comprising:

[0031] A target acquisition module, which is used to determine whether the user clicks, whether the user browses the site, whether the user adds to the shopping cart, and whether the user purchases as a target;

[0032] A feature list acquisition module, wherein the feature list acquisition module is used to take advertisement-related features, user-related features, and context-related features as a feature list;

[0033] A sample example acquisition module, wherein the sample example acquisition module is used to construct a sample example according to the target feature and the feature list, and to label the sample example;

[0034] A model training module is used to train the MMoE model based on the labeled sample examples and learn the model parameters;

[0035] The online reasoning module is used to extract features online in real time based on the offline trained MMoE model, obtain scores of multiple targets, and achieve real-time remarketing advertising effect estimation.

[0036] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of a method for estimating the effect of remarketing advertising based on multi-task learning.

[0037] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for estimating the effect of remarketing advertising based on multi-task learning.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention proposes a remarketing advertising effect prediction method based on multi-task learning. It learns knowledge from massive historical delivery data through feature design, target determination, sample example construction and model training, completes feature extraction for requests in real time, obtains scores for multiple targets, and applies the knowledge trained from the offline model to the production environment to achieve real-time remarketing advertising effect prediction. In the feature design stage, the union of attributes related to each target will be used as a feature, and the relationship between each feature and target will be automatically learned in the model training stage. The present invention uses the MMoE model to realize the estimation of multiple indicators, which reduces the complexity of traditional multi-model solutions and overcomes the problem of poor training effect caused by sparse data on purchase / add-to-purchase behaviors.

[0040] Furthermore, the MMoE model can simultaneously learn multiple objectives. Through the expert network and gating network, all features related to each objective during the feature design phase can be input into the model, automatically learning the relationship between features and objectives. The Shared-Bottom architecture can learn common features across multiple tasks, while the design of multiple Towers can learn the unique characteristics of each task, enabling simultaneous training of multiple tasks.

[0041] Furthermore, through the Shared-Bottom structure, parameter learning is performed on common features, which alleviates the problem of poor feature learning results in the Purchase / AddToCart training tasks due to the small number of behaviors, because the common features have already been learned in the Shared-Bottom with the help of data from other tasks.

[0042] Furthermore, when calculating the loss, ClassWeight is taken into account, so that the MMoE model can balance the problem of inconsistent positive and negative sample ratios of the four targets. Ultimately, the training effects of each task are balanced, and there will be no situation where the performance on one task is very good and the performance on other tasks is very poor.

[0043] The present invention proposes a remarketing advertising effect prediction system based on multi-task learning. By dividing the prediction system into a target feature acquisition module, a feature list acquisition module, a sample example acquisition module, a model training module and an online reasoning module, the modular concept is adopted to make each module independent of each other, facilitating unified management of each module. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 A business process diagram for remarketing advertising based on existing technologies.

[0046] Figure 2 This is a flow chart of the remarketing advertising effect prediction method based on multi-task learning of the present invention.

[0047] Figure 3 This is a system architecture diagram of the present invention.

[0048] Figure 4 This is a structural diagram of the MMoE model in the prior art.

[0049] Figure 5 This is a diagram of the remarketing advertising effect prediction system based on multi-task learning in the present invention. DETAILED DESCRIPTION

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0051] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0052] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not require further definition or explanation in subsequent drawings.

[0053] In the description of the embodiments of the present invention, it should be noted that if the terms "upper," "lower," "horizontal," "inner," etc. appear, the orientation or positional relationship indicated is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the inventive product is typically placed when in use. These terms are merely for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0054] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0055] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0056] The present invention is described in further detail below with reference to the accompanying drawings:

[0057] The concept of this invention is still to train the model on large-scale advertising display and corresponding effect data, and evaluate the current advertising display in real time online. However, it does not adopt the traditional solution of modeling CTR and CVR separately. Instead, it uses a multi-task learning method to learn multiple objectives simultaneously in a single model. It learns four objectives simultaneously: 1) whether the user clicks, 2) whether the user has ViewContent behavior (i.e., browsing behavior within the site), 3) whether the user has AddToCart behavior, and 4) whether the user has Purchase behavior.

[0058] The present invention proposes a remarketing advertising effect prediction method based on multi-task learning, such as Figure 2 As shown, the following steps are included:

[0059] The goals are whether the user clicks, browses the site, adds items to the shopping cart, and makes a purchase.

[0060] Advertisement-related features, user-related features, and context-related features are taken as a feature list;

[0061] Construct sample examples based on the target and feature list, and label the sample examples;

[0062] Train the MMoE model based on the labeled sample examples and learn the model parameters;

[0063] Based on the offline trained MMoE model, features are extracted online in real time to obtain scores for multiple targets, enabling real-time estimation of remarketing advertising effects.

[0064] Specifically:

[0065] Step 1: Feature Design

[0066] A feature is a description of one dimension of an advertisement display. Since the present invention needs to estimate multiple effect indicators at the same time, it is necessary to design the factors that affect these goals into features.

[0067] Because the evaluation is about whether the current user will click on the current ad and make a purchase in the current context, the designed features mainly include three aspects: ad-related features, context-related features, and user-related features.

[0068] Ad-related features include the ad ID, creative material ID, historical ad delivery performance (for example, the ad's impressions / clicks / conversions / click-through rate / conversion rate / exposure-to-conversion rate in the last 3 / 7 / 14 days), etc.

[0069] Context-related features include the current day of the week, hour, ADX, media ID, ad slot type, ad slot size, media’s historical delivery performance (for example, the media’s impressions / clicks / conversions / click-through rate / conversion rate / exposure-to-conversion rate in the last 3 / 7 / 14 days), etc.

[0070] User-related features include the user's geographic location (country / city), device information (mobile phone model / operating system / operating system version / language), user behavior statistics on the advertiser's website (for example, the number of ViewContent / AddToCart / Purchase times this user has made in the last 3 / 7 / 14 days), etc.

[0071] In this feature system, some are related to whether the user clicks on the current ad: such as the ad slot type, ad slot size, and creative type; some are related to whether the user will make a purchase: for example, the user's behavior on the advertiser's website, if there has been AddToCart behavior recently, but no payment, then there is a high probability that the purchase will be completed; and some are related to multiple goals, such as the current day of the week, hour, etc.

[0072] Step 2: Sample generation

[0073] Sample generation is the process of constructing the training data to be output to the model. The original data is structured into pairs of <FeatureList, Optimization Target>. In this invention, because it is necessary to estimate multiple targets at the same time, the final constructed sample is as follows:

[0074] <Ad-related features, user-related features, contextual features, whether the user clicked (0 / 1), whether the user had a ViewContent action (0 / 1), whether the user had an AddToCart action (0 / 1), whether the user had a Purchase action (0 / 1)>

[0075] Clicks, viewContent, addToCart, and purchase actions can all be associated with a unique tracking ID to label a record. Offline learning is performed from massive amounts of historical advertising data. Ad impressions occur on media websites, while subsequent actions occur on advertiser websites. Using the same tracking ID, the behaviors at each stage are linked, extracted features are associated with targets, and training samples are generated based on historical advertising results.

[0076] Step 3: Model training

[0077] The model training uses Google's MMoE (Multi-gate Mixture-of-Experts) model for training, such as Figure 3 and Figure 4 As shown in the figure, the model adopts a shared underlying structure widely used in multi-task learning. The Shared-Bottom network is located at the bottom layer, and multiple tasks share this layer to learn some common features. The Shared-Bottom network contains multiple expert networks, which can learn different aspects of the sample characteristics; upwards, each task corresponds to a Tower Network to learn some unique features of the current task; the middle is connected by a Gate network, and a subset of the expert network is selected according to different tasks.

[0078] The MMoE model is implemented based on TensorFlow, using the Keras API to define the network structure, the loss calculation method for each task, the selection of the model optimizer, and the selection of evaluation metrics. Iterative learning is performed on the dataset over multiple epochs. At the end of each epoch, TruePositives, FalsePositives, TrueNegatives, FalseNegatives, BinaryAccuracy, Precision, Recall, and AUC are calculated on the training set and validation set, respectively. The learning effect of the model is evaluated in real time. An 8-card GPU server is used for training.

[0079] This paper proposes a remarketing advertising effect prediction method based on multi-task learning. It simultaneously learns four objectives: whether the user clicks, whether the user has ViewContent behavior (i.e., on-site browsing behavior), whether the user has AddToCart behavior, and whether the user has Purchase behavior. Because these four objectives are in a behavior funnel, the user behavior decreases as the funnel progresses. To solve this problem, this paper uses two solutions:

[0080] 1) Using a shared-bottom architecture, when learning Purchase behavior, we can leverage the Click / ViewContent / AddToCart samples to learn features and model parameters. This allows shared features to be used with samples from other tasks for parameter learning, addressing the issue of inaccurate parameter learning caused by insufficient positive samples and insufficient back-end behavior.

[0081] 2) Add ClassWeight and use the inverse of the number of samples as the class weight. When calculating the loss, take ClassWeight into account so that the model can optimize four objectives at the same time.

[0082] Step 4: Online Reasoning

[0083] Online inference uses the tf-serving component of the TensorFlow framework. tf-serving loads the trained model, combines online requests from Adx, advertising information in the advertising library, and user profiles in the DMP to complete feature extraction, and requests tf-serving with the features as parameters. The value of the feature vector entered in tf-serving is feedforward calculated on the MMoE network. The network parameters use the model parameters trained offline, and four probabilities can be obtained: the probability of whether the user clicks P(Click), the probability of whether the user has ViewContent behavior P(ViewContent), the probability of whether the user has AddToCart behavior P(AddToCart), and the probability of whether the user has Purchase behavior P(Purchase).

[0084] During the online sorting process, the four probabilities are exponentially weighted and multiplied together. The score of multiple targets = P(click) a *P(ViewContent) b *P(AddToCart) c *P(Purchase) d, where a, b, c, and d are parameters. The final scores are used to rank candidate ads based on the weights corresponding to the desired dimension. The highest-scoring ad is then displayed. In the validation set of the training data, the same formula used for online inference is used to calculate the final scores of all samples. The median is then used as the criterion for bidding. If the final score of the current request is greater than the threshold, the ad with the highest final score is used for bidding; if it is less than the threshold, the bid is abandoned.

[0085] This paper proposes a remarketing advertising effect prediction system based on multi-task learning. Figure 5 Shown, including:

[0086] A target acquisition module, wherein the target acquisition module is used to use whether the user clicks, whether the user browses the site, whether the user adds to the shopping cart, and whether the user purchases as target features;

[0087] A feature list acquisition module, wherein the feature list acquisition module is used to take advertisement-related features, user-related features, and context-related features as a feature list;

[0088] A sample example acquisition module, wherein the sample example acquisition module is used to construct a sample example according to the target feature and the feature list, and to label the sample example;

[0089] A model training module is used to train the MMoE model based on the labeled sample examples and learn the MMoE model parameters;

[0090] The online reasoning module is used to extract features online in real time based on the offline trained MMoE model, obtain scores of multiple targets, and achieve real-time remarketing advertising effect estimation.

[0091] An embodiment of the present invention provides a terminal device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of each of the aforementioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in each of the aforementioned apparatus embodiments are implemented.

[0092] The computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to accomplish the present invention.

[0093] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0094] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0095] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0096] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0097] The present invention proposes a remarketing advertising effect prediction method and system based on multi-task learning, which has the following advantages: 1) The features simultaneously include all target-related features, and the contribution of features to the targets is learned through the model; 2) The MMoE model is used to simultaneously estimate multiple targets, and the targets are progressive; 3) The Shared-Bottom structure allows the same features to be learned together, ensuring the robustness of the model, and adding Class Weight allows each target to be given equal attention during the optimization process.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A remarketing advertising effect prediction method based on multi-task learning, characterized by: The steps include: The goals are whether the user clicks, browses the site, adds items to the shopping cart, and makes a purchase. Advertisement-related features, user-related features, and context-related features are taken as a feature list; Construct sample examples based on the target and feature list, and label the sample examples; Train the MMoE model based on the labeled sample examples and learn the MMoE model parameters; Based on the offline trained MMoE model, features are extracted online in real time to obtain scores for multiple targets, enabling real-time remarketing advertising effect estimation. The MMoE model uses a shared underlying structure. The Shared-Bottom network is located at the bottom of the MMoE model. Multiple tasks use the Shared-Bottom network to learn common features. The Shared-Bottom network contains multiple expert networks that can learn different aspects of the sample characteristics; each task corresponds to a Tower Network to learn the unique features of the current task; Select subsets of expert networks based on different tasks through multiple Gate networks; Using the tf-serving component of the TensorFlow framework, tf-serving loads the trained MMoE model and combines requests from AdX, ad information in the ad library, and user profiles in the DMP to complete feature extraction. Get the probability P of whether the user clicks based on the feature extraction results Click , the probability P of whether the user has browsing behavior in the site ViewContent , the probability P of whether the user has added to the shopping cart behavior AddToCart And the probability P of whether the user has a purchase behavior Purchase , according to the probability P of whether the user clicks Click , the probability P of whether the user has browsing behavior in the site ViewContent , the probability P of whether the user has added to the shopping cart behavior AddToCart And the probability P of whether the user has a purchase behavior Purchase , get the scores of multiple targets, the scores of multiple targets are calculated as follows: Score of multiple targets = P Click a *P ViewContent b *P AddToCart c *P Purchase d Among them, a, b, c, and d are parameters. The scores of multiple targets are used to sort the candidate ads, and the ads with the highest scores are used as the currently displayed ads to achieve the estimation of remarketing advertising effects.

2. The method for estimating the effect of remarketing advertising based on multi-task learning according to claim 1, characterized in that: Ad-related features include ad ID, creative material ID, and historical ad delivery performance; Context-related features include the current day of the week, hour, ADX, media ID, ad slot type, ad slot size, and media’s historical delivery performance; User-related characteristics include the user's geographic location, including country and city; device information, including phone model, operating system, operating system version and language; and user behavior statistics on the advertiser's website.

3. The method for estimating the effect of remarketing advertising based on multi-task learning according to claim 2, characterized in that: Whether the user clicks, whether the user browses the site, whether the user adds to the shopping cart, and whether the user makes a purchase can be associated based on the unique tracking ID to achieve sample labeling.

4. The method for estimating the effect of remarketing advertising based on multi-task learning according to claim 1, characterized in that: Using the Shared-Bottom structure, when learning purchasing behavior, sample examples are used to learn features and model parameters.

5. The method for estimating the effect of remarketing advertising based on multi-task learning according to claim 1, characterized in that: When using the MMoE model to train the feature list, add ClassWeight and use the inverse of the sample size as the class weight.

6. A remarketing advertising effect prediction system based on multi-task learning, characterized by: include: A target acquisition module, which is used to determine whether the user clicks, whether the user browses the site, whether the user adds to the shopping cart, and whether the user purchases as a target; A feature list acquisition module, wherein the feature list acquisition module is used to take advertisement-related features, user-related features, and context-related features as a feature list; A sample example acquisition module, wherein the sample example acquisition module is used to construct a sample example according to the target feature and the feature list, and to label the sample example; A model training module is used to train the MMoE model based on the labeled sample examples and learn the model parameters; The online reasoning module is used to extract features online in real time based on the offline trained MMoE model, obtain scores of multiple targets, and achieve real-time remarketing advertising effect estimation.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the remarketing advertising effect prediction method based on multi-task learning as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting the effect of remarketing advertising based on multi-task learning according to any one of claims 1 to 5 are implemented.

Images