Feature learning method, device, equipment, medium and program product

Abstract

According to the embodiment of the invention, a feature learning method and device, equipment, a medium and a program product are provided. The method comprises the steps that a first data set and a second data set are obtained, the first data set comprises first behavior data of a first group of objects on a target application and indicates whether the objects execute click behaviors and conversion behaviors on recommended content, and the second data set comprises second behavior data of a second group of objects on the target application and indicates whether the objects execute click behaviors and conversion behaviors on the recommended content; the instruction unit indicates whether an object performs a click behavior on recommended content; based on a union set of the first data set and the second data set, a plurality of training samples which respectively correspond to the plurality of objects and comprise attribute information of the objects and a plurality of labels are generated, and the first label and the second label respectively indicate whether the objects execute a click behavior and a conversion behavior on the recommended content or indicate a predetermined indicator; and training an object feature extraction model by utilizing a plurality of training samples, wherein the object feature extraction model is configured to extract feature embedding corresponding to the object based on attribute information of the input object.

Description

Methods, apparatus, devices, media, and program products for feature learning

[0001] The exemplary embodiments disclosed herein generally relate to the field of computer technology, and particularly to methods, apparatus, devices, computer-readable storage media, and computer program products for feature learning.

[0002] Feature embeddings are a crucial component of machine learning systems or models. A feature embedding is a vector with appropriate dimensions used to represent the original data. For example, in object prediction tasks, the input typically includes at least object-related data, and the output is the prediction result. Accurate prediction processes will produce similar prediction results for similar objects. For instance, when predicting whether a user will buy a product, the input is user-related data and product-related data, and the output is the probability of the user buying the product. When using a predictive model to perform probabilistic prediction tasks, it is desirable for the model to provide similar prediction probabilities for users with similar features. Therefore, representing user-related data with similar features using vectors with close proximity (i.e., feature embeddings) and using them as input makes the prediction results more consistent with expectations.

[0003] Summary of the Invention

[0004] In a first aspect of this disclosure, a method for feature learning is provided. The method includes: obtaining a first dataset and a second dataset, the first dataset including first behavioral data of a first group of objects on a target application, and the second dataset including second behavioral data of a second group of objects on the target application, wherein the first behavioral data indicates whether the objects perform click and conversion behaviors on recommended content, and the second behavioral data indicates whether the objects perform click behaviors on the recommended content; generating multiple training samples corresponding to multiple objects based on the union of the first and second datasets, each training sample including attribute information of the object and multiple labels, wherein a first label among the multiple labels indicates whether the object performs a click behavior on the recommended content, and a second label among the multiple labels indicates whether the object performs a conversion behavior or indicates a predetermined indicator on the recommended content; and training an object feature extraction model using the multiple training samples, the object feature extraction model being configured to extract feature embeddings corresponding to the objects based on the attribute information of the input objects.

[0005] In a second aspect of this disclosure, an apparatus for feature learning is provided. The apparatus includes: an acquisition module configured to acquire a first dataset and a second dataset, the first dataset including first behavioral data of a first group of objects on a target application, and the second dataset including second behavioral data of a second group of objects on the target application, the first behavioral data indicating whether an object performs a click or conversion action on recommended content, and the second behavioral data indicating whether an object performs a click action on recommended content; a generation module configured to generate multiple training samples corresponding to multiple objects based on the union of the first and second datasets, each training sample including attribute information of the object and multiple labels, a first label among the multiple labels indicating whether an object performs a click action on recommended content, and a second label among the multiple labels indicating whether an object performs a conversion action on recommended content or indicates a predetermined indicator; and a training module configured to train an object feature extraction model using the multiple training samples, the object feature extraction model being configured to extract feature embeddings corresponding to the object based on the attribute information of the input object.

[0006] In a third aspect of this disclosure, an electronic device is provided. The device includes at least one processing unit; and at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit. When executed by the at least one processing unit, the instructions cause the device to perform the method of the first aspect.

[0007] In a fourth aspect of this disclosure, a computer-readable storage medium is provided. The medium stores a computer program that, when executed by a processor, implements the method of the first aspect.

[0008] In a fifth aspect of this disclosure, a computer program product is provided. The computer program product is tangibly stored in a computer storage medium and includes computer-executable instructions that, when executed by a device, cause the device to perform the method of the first aspect.

[0009] It should be understood that the description in this section is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description.

[0010] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0011] Figure 1 shows a schematic diagram of an example environment in which embodiments of the present disclosure can be implemented;

[0012] Figure 2 illustrates a schematic diagram of the feature learning process based on traditional schemes;

[0013] Figure 3 shows a flowchart of a feature learning process according to some embodiments of the present disclosure;

[0014] Figure 4 illustrates a schematic diagram of a model training and application environment in which embodiments of the present disclosure can be implemented;

[0015] Figure 5 shows a flowchart of a feature learning process according to some embodiments of the present disclosure;

[0016] Figure 6 shows a block diagram of an apparatus for feature learning according to some embodiments of the present disclosure; and

[0017] Figure 7 shows a block diagram of an electronic device in which one or more embodiments of the present disclosure may be implemented.

[0018] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0019] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The term "some embodiments" should be understood as "at least some embodiments". Other explicit and implicit definitions may also be included below.

[0020] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0021] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure through appropriate means in accordance with relevant laws and regulations, and user authorization should be obtained.

[0022] For example, in response to receiving a user's active request, a prompt message is sent to the user to clearly inform the user that the requested operation will require the acquisition and use of the user's personal information, thereby enabling the user to choose whether to provide personal information to the software or hardware such as electronic devices, applications, servers or storage media that perform the operation of the technical solution disclosed herein, based on the prompt message.

[0023] As an optional but non-restrictive implementation, in response to a user's active request, a prompt message can be sent to the user, such as a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0024] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0025] As used in this paper, the term "model" refers to a model that learns the relationship between inputs and outputs from training data, enabling it to generate corresponding outputs for a given input after training. Model generation can be based on machine learning techniques. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs using multiple layers of processing units. A neural network model is an example of a deep learning-based model. In this paper, "model" may also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," and these terms are used interchangeably.

[0026] A neural network is a machine learning network based on deep learning. A neural network processes input and provides a corresponding output, typically consisting of an input layer, an output layer, and one or more hidden layers between the input and output layers. Neural networks used in deep learning applications often include many hidden layers, thus increasing the network's depth. The layers of a neural network are connected sequentially, so that the output of the previous layer is provided as the input to the next layer. The input layer receives the input to the neural network, while the output layer's output serves as the final output. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each node processing the input from the layer above.

[0027] Machine learning typically comprises three phases: training, testing, and application (also known as inference). In the training phase, a given model is trained using a large amount of training data, iteratively updating its parameter values ​​until the model can consistently generate inferences that meet the expected goals from the training data. Through training, the model can be considered to have learned the relationship between inputs and outputs (also known as the input-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing phase, test inputs are applied to the trained model to test whether it can provide the correct output, thus determining the model's performance. In the application phase, the model can be used to process actual inputs based on the trained parameter values ​​to determine the corresponding output.

[0028] Figure 1 illustrates a schematic diagram of an example environment 100 in which embodiments of the present disclosure can be implemented. In environment 100, a feature learning and application system 110 is configured to train and apply an object feature extraction model 112 (sometimes referred to as a UE model) and a prediction model 114 to perform predictions about objects in a predetermined prediction task. The object feature extraction model 112 is configured to extract feature embeddings corresponding to the input object based on the object's attribute information.

[0029] The task of predicting objects can be defined in different scenarios. These objects can include various entities in different scenarios, such as users, user groups, audiences, organizations, groups, and units.

[0030] For example, in a recommendation scenario, if the goal is to recommend content to users or an audience, the prediction task can include predicting the relevance of an object (i.e., a user or audience) to a recommended content item, in order to recommend content that better suits the audience's needs. A recommended content item refers to the content or resource to be recommended; examples can include applications, physical goods, virtual goods, articles, audio / video content, and so on. Example prediction models used to implement such prediction tasks can include object response prediction to predict the probability that an object will respond to a recommended content item. In some example applications, such prediction models can include conversion rate prediction (PVR) models to predict whether a user or audience will take a convertive action on a recommended content item, such as downloading an application, purchasing a product, reading an article, etc.

[0031] In addition to the examples of prediction tasks mentioned above, there can also be predictions about the object itself, as well as predictions about the object relative to other content or items, and so on.

[0032] In prediction tasks involving objects, to obtain more accurate prediction results, the input typically includes data related to the object. This data may consist of multiple fields, which can serve as attribute information describing different aspects of the object.

[0033] In the embodiments herein, it is assumed that the input to the prediction task of prediction model 114 includes data related to the object's behavior on the application. For example, the data required for the prediction task includes behavioral data of the object on the application, such as clicks, browsing, purchases, likes, comments, etc. Data of the object on different applications can characterize the object's features, thereby helping to perform predictions. Depending on the application needs, prediction model 114 can be constructed to include various types of model architectures, such as deep neural network (DNN) architectures. The embodiments of this disclosure do not limit the specific type and structure of prediction model 114.

[0034] As shown in Figure 1, objects 132-1, 132-2, ..., 132-N (collectively or individually referred to as object 132 for ease of discussion) can use one or more applications on terminal devices 130-1, 130-2, ..., 130-N (collectively or individually referred to as terminal device 130 for ease of discussion) through one or more terminal devices. Different applications in terminal device 130 can receive recommended content from corresponding service providers via network 105, such as service providers 120-1, 120-2, ..., 120-M (collectively or individually referred to as service provider 120 for ease of discussion). Service provider 120 or a third party can request the delivery of recommended content to the applications in terminal device 130.

[0035] In some embodiments, for recommended content requested by service provider 120 to be delivered to an application in terminal device 130, the conversion response can be completed through the same application or another application in terminal device 130. For example, the recommended content requested by service provider 120 to application A1 in terminal device 130-1 may be clicked by object 132-1 in application A1 and a conversion behavior such as purchase may be completed in application A1, or object 132-1 may click in application A1 and jump to application B1 to complete a conversion behavior such as purchase.

[0036] If the click and conversion actions of object 132-1 are completed in application A1 and application B1 respectively, the platform corresponding to application A1 cannot know the actions performed by object 132-1 in application B1. Therefore, for an application platform, this type of user can be referred to as an untraceable conversion user, or simply an untraceable user. For object 132, who clicks and completes the conversion action within the same application, it can be referred to as a traceable conversion user of that application platform, or simply a traceable user.

[0037] In environment 100, terminal device 130 can be any type of mobile terminal, fixed terminal, or portable terminal, including mobile phones, desktop computers, laptop computers, notebook computers, netbook computers, tablet computers, media computers, multimedia tablets, personal communication system (PCS) devices, personal navigation devices, personal digital assistants (PDAs), audio / video players, digital cameras / camcorders, positioning devices, television receivers, radio receivers, e-book devices, gaming devices, or any combination thereof, including accessories and peripherals of these devices or any combination thereof. In some embodiments, terminal device 130 may also support any type of user-facing interface (such as "wearable" circuitry). Model training and application system 110 can be implemented, for example, on various types of computing systems / servers capable of providing computing power, including but not limited to mainframes, edge computing nodes, computing devices in cloud environments, and so on.

[0038] It should be understood that the structure and function of the various elements in environment 100 are described for illustrative purposes only and do not imply any limitation on the scope of this disclosure.

[0039] Considering data security and privacy, the object-related behavioral data required for the training and use of prediction model 114 and object feature extraction model 112 are subject to various constraints, and the data is collected, stored, and used only after authorization. These constraints and authorizations involve multiple parties, including object 132, terminal device 130 running the application, and / or service provider 120.

[0040] Currently, the terminal device 130 running the application may require users to explicitly authorize the use of behavioral data from various applications. For example, if recommended content (e.g., advertisements) is served within the application on the terminal device, then user authorization is required for the application provider and service provider 120 to use the behavioral data. Otherwise, it will be impossible to collect and use user-granular behavioral data and corresponding conversion data.

[0041] Suppose that in a recommendation-related prediction task, the required input data includes user click behavior, download behavior, and purchase behavior of products recommended in the advertisement within the application.

[0042] With authorization, user-level behavioral data can be located using corresponding identifiers, enabling the differentiation of behavioral data related to different users. For example, if behavioral data is collected from application A1 on terminal device 130-1, this data can be associated with an identifier that identifies the terminal device and the identifier of application A1. The identifier for the terminal device could include, for example, an Identifier For Advertising (IDFA), which can be shared by different applications on the same device. In other words, for an application downloaded to different terminal devices, the IDFA can be used to locate the specific terminal device, which can then be considered to correspond to a particular user.

[0043] In prediction tasks, it's crucial to include behavioral data that differentiates different objects from the input, as these objects may exhibit distinct behavioral characteristics and should be associated with different prediction outcomes. For example, different users or audiences may be interested in different recommended content. Without object-level behavioral data, it becomes difficult to comprehensively describe the characteristics of each object.

[0044] In model processing, object-related data, i.e., behavioral data, is mapped into vector-like feature embeddings through an embedding mapping relationship. These feature embeddings are considered to represent the corresponding object. The object feature extraction model 112 in Figure 1 can extract the object's feature embeddings. Extracted feature embeddings have many uses, such as determining the similarity between objects by comparing their feature embeddings, and so on.

[0045] In traditional approaches, the model is trained separately for different behavioral data of different objects during training. Figure 2 illustrates a schematic diagram of the feature learning process 200 according to the traditional approach. Process 200 includes training the model separately for dataset 210 of trackable user behavior data and dataset 220 of untrackable user behavior data.

[0046] When training the object feature extraction model 112' using the dataset 210 of trackable users, the behavioral data related to click behavior is treated as a data stream and processed in its corresponding processing logic. The behavioral data related to conversion behavior (including behavioral data of actual conversion after click behavior and behavioral data of predicted conversion) is treated as a data stream and can be processed in the post-click behavior processing logic and PVR model 212.

[0047] When processing the dataset 210 of trackable users at box 211, the corresponding click behavior data can be labeled as [clicked, null] at box 213, where clicked indicates that a click behavior has been performed, and null indicates that the conversion behavior has not been tracked. When processing the corresponding data at box 212, the corresponding behavior data related to the conversion behavior can be labeled as [null, convert] at box 214, where null indicates that the click behavior has not been tracked, and convert indicates the actual conversion behavior and the predicted conversion behavior. Subsequently, the object feature extraction model 112' is trained using the labels at boxes 213 and 214 respectively.

[0048] When processing the dataset 220 containing untraceable users at box 221, the corresponding click behavior data can be labeled as [clicked, null] at box 223, where clicked indicates a click behavior has been performed, and null indicates that the conversion behavior was not tracked. When processing the dataset 222, the corresponding behavior data related to the conversion behavior can be labeled as [null, convert] at box 224, where null indicates a click behavior was not tracked, and post-click prediction indicates the predicted conversion behavior (because untraceable users did not actually perform any conversion behavior). Then, the object feature extraction model 112' is trained using the labels at boxes 223 and 224 respectively.

[0049] However, the feature learning process of such traditional methods has the following problems:

[0050] 1) There is a bias in feature learning for trackable and untrackable users: untrackable users may be underestimated because they have lower sending and cost. This is mainly due to the additional post-click signals (or behaviors) of trackable users.

[0051] 2) Overfitting risk: The model will be trained twice on the same request, once for clicks and once for conversions. The model can memorize data, but lacks generalization ability.

[0052] 3) The training efficiency is low, with an estimated waste of about 50% of the computation.

[0053] 4) Online and offline performance are inconsistent, and offline performance may be overestimated due to tag leakage.

[0054] Embodiments of this disclosure propose a feature learning scheme. According to various embodiments of this disclosure, a first dataset and a second dataset are obtained. The first dataset includes first behavioral data of a first group of objects on a target application, and the second dataset includes second behavioral data of a second group of objects on the target application. The first behavioral data indicates whether an object performs a click or conversion action on recommended content, and the second behavioral data indicates whether the object performs a click action on the recommended content. Then, based on the union of the first and second datasets, multiple training samples corresponding to multiple objects are generated. Each training sample includes attribute information of the object and multiple labels. A first label among the multiple labels indicates whether the object performs a click action on the recommended content, and a second label among the multiple labels indicates whether the object performs a conversion action on the recommended content or indicates a predetermined indicator. Accordingly, an object feature extraction model is trained using the multiple training samples. The object feature extraction model is configured to extract feature embeddings corresponding to the object based on the attribute information of the input object.

[0055] In this way, the scheme avoids feature learning bias between users whose conversion behavior is traceable and those whose conversion behavior is not by constructing multiple labels in the training samples corresponding to a single object using a unified strategy. Furthermore, by constructing training samples using the union of the first and second datasets, it is determined that only one training sample is constructed for each object to represent its click and conversion behavior towards a specific recommended content. This avoids the risk of overfitting caused by constructing multiple training samples for clicks and conversions for the same object separately. Moreover, the training efficiency of the object feature extraction model can be significantly improved by merging datasets and using multi-label training samples. In addition, since multiple labels can indicate whether each object clicked on the recommended content and whether it converted (or explicitly indicate that the conversion behavior is unknown), the problem of label leakage can be avoided, thus improving the performance of the trained model.

[0056] Some exemplary embodiments of this disclosure will now be described with reference to the accompanying drawings.

[0057] Figure 3 shows a flowchart of a process 300 for feature learning according to some embodiments of the present disclosure. For ease of discussion, it will be described in conjunction with Figure 1. Process 300 can be implemented at the feature learning and application system 110.

[0058] In process 300, the feature learning and application system 110 can obtain datasets 310 and 320. Dataset 310 may include first behavioral data of a first group of objects on the target application, indicating whether the objects performed click and conversion behaviors on recommended content. Such a first group of objects may, for example, include the aforementioned trackable users who can perform click and conversion behaviors on recommended content provided by service provider 120 within the target application. Click behaviors may, for example, include the action of the first group of objects clicking to view recommended content when browsing it in the target application. Conversion behaviors may, for example, include the actions of the first group of objects downloading, purchasing, etc., of the recommended content viewed after clicking. It should be understood that conversion behaviors can be set according to different application scenarios and are not limited here.

[0059] Dataset 320 may include second behavioral data of a second group of objects on the target application, indicating whether the objects performed a click action on recommended content. Such a second group of objects may, for example, include the aforementioned untraceable users who, as an example, can only perform click actions on recommended content provided by service provider 120 within the target application. For conversion actions performed by untraceable users on applications other than the target application, feature learning and application system 110 may not be able to obtain them without user authorization or under other circumstances. Furthermore, data obtainable on the target application can also be referred to as on-site data. Correspondingly, data not obtainable on the target application can also be referred to as off-site data.

[0060] It should be understood that in Figure 3, different datasets are distinguished only by the traceability of behavioral data. However, in reality, there can be multiple datasets collected and stored for behavioral data of traceable or untraceable users.

[0061] Furthermore, based on the union 330 of datasets 310 and 320, the feature learning and application system 110 can generate multiple training samples corresponding to multiple objects, each training sample including object attribute information and multiple labels. Based on such training samples, multi-label processing logic (340) can be executed. For example, the object attribute information may include information about whether the object belongs to a trackable user or an untrackable user. For example, data in the datasets can be added with identifiers about object attributes, such as relevant fields, so that the generated training samples contain object attribute information.

[0062] Based on the multi-label processing logic (340), the first label among multiple labels can indicate whether the object performs a click action on the recommended content. The second label among multiple labels can indicate whether the object performs a conversion action on the recommended content or indicate a predetermined indicator. In some embodiments, the predetermined indicator can indicate that the object's conversion action is untraceable. Such a predetermined indicator can be, for example, an indicator representing "empty," such as null. Exemplarily, an array can be used as the representation of the labels. Assuming that a training sample includes attribute information that the object belongs to a traceable user, the corresponding array of multiple labels can be represented, for example, as [click, trackable convert] at box 350. Here, click can represent the first label, indicating whether the object performs a click action on the recommended content. Trackable convert can represent the second label, indicating whether the object performs a conversion action on the recommended content. Assuming that another training sample includes attribute information that the object belongs to an untraceable user, the corresponding array of multiple labels can be represented, for example, as [click, null] at box 360. Here, null can represent the second label, indicating that the object's conversion action is untraceable.

[0063] In some embodiments, when generating multiple training samples corresponding to multiple objects based on the union of datasets 310 and 320, the feature learning and application system 110 can determine the behavior data of the first object from the first behavior data of dataset 310 and the second behavior data of dataset 320 based on the object identifier of the first object, and can generate the first training sample corresponding to the first object based on the behavior data of the first object. The first object can be one of the multiple objects.

[0064] Object identifiers may include, for example, the identifiers discussed above used to indicate object attributes (e.g., whether the object belongs to a trackable or untrackable user). When generating training samples, the corresponding attribute information can be determined based on the object identifiers. Object identifiers may also include identifiers used to indicate the object's actions in the target application (e.g., clicks, purchases, etc.). Such identifiers can be pre-added to the behavioral data of each dataset. The form of the identifiers may include, for example, expressions such as text and / or symbols. It should be understood that the form of object identifiers is not limited herein.

[0065] In some embodiments, when generating a first training sample corresponding to a first object based on the behavior data of the first object, the feature learning and application system 110 can determine a first label in the first training sample corresponding to the first object based on the behavior data of the first object indicating whether the first object performs a click behavior on the recommended content.

[0066] Since the first row of data in dataset 310 can indicate whether an object performed a click or conversion action on the recommended content, and the second row of data in dataset 320 can indicate whether an object performed a click action on the recommended content, the first label in the corresponding training sample can be determined based on the indication of whether an object performed a click action on the recommended content according to its behavioral data. For example, the specific value of the first label "click" in the above label array can be determined as 0 (which could indicate that no click action was performed) or 1 (which could indicate that a click action was performed). It should be understood that any other appropriate numerical value or symbol can also be used to represent whether the first object performed a click action or not.

[0067] In some embodiments, if the behavioral data of the first object includes an indication of whether the first object performs a conversion action on the recommended content, the feature learning and application system 110 can determine a second label in the first training sample based on the indication of whether the first object performs a conversion action on the recommended content.

[0068] Since the first behavioral data for the first group of objects (e.g., trackable users) contained in dataset 310 can indicate whether the object performed a conversion action on the recommended content, the second label in the corresponding training sample can be determined based on the object's behavioral data indicating whether the object performed a conversion action on the recommended content. For example, the specific value of the second label "trackable convert" in the above label array can be determined as 0 (which could indicate that no conversion action was performed) or 1 (which could indicate that a conversion action was performed). It should be understood that any other appropriate numerical value or symbol can also be used to represent whether the first object performed a conversion action or not.

[0069] In some embodiments, if the behavioral data of the first object does not include an indication of whether the first object performs a conversion action on the recommended content, the feature learning and application system 110 may determine a second label in the first training sample to indicate a predetermined indicator.

[0070] Since the second behavioral data for the second group of objects (e.g., untraceable users) included in dataset 320 can only indicate whether the object clicked on the recommended content, but not whether it converted to other content, the second label in the corresponding training sample can be directly determined to indicate a predefined indicator, such as null. It should be understood that any other suitable symbol or numerical value can also be used to represent the predefined indicator.

[0071] In the traditional feature learning scheme shown in Figure 2, the prediction of post-click behavior for trackable users is relatively accurate, while the prediction of post-click behavior for untrackable users is less accurate because untrackable users essentially have no post-click behavior. This results in almost no positive samples and only negative samples in the prediction of conversion behavior. However, for the post-click behavior of trackable users, i.e., conversion behavior, there are both predicted positive samples and original direct conversion positive samples. Therefore, there is an imbalance in the prediction of post-click behavior for trackable and untrackable users. Therefore, the embodiments of this disclosure do not employ a model for predicting post-click behavior.

[0072] Furthermore, the feature learning and application system 110 can use multiple training samples to train the object feature extraction model 112. The object feature extraction model 112 can be configured to extract the feature embeddings corresponding to the input object based on the attribute information of the input object. The feature embeddings of the object extracted by the object feature model can be used in various scenarios, such as determining the similarity between objects by comparing the feature embeddings of multiple objects.

[0073] By utilizing embodiments of this disclosure, multiple labels can be constructed using a unified strategy within the training samples corresponding to a single object, thus avoiding feature learning bias between users whose conversion behavior is traceable and those whose conversion behavior is not. Secondly, by constructing training samples using the union of datasets 310 and 320, it can be determined that only one training sample is constructed for each object to represent its click and conversion behavior towards a specific recommended content. This avoids the risk of overfitting caused by constructing multiple training samples for clicks and conversions for the same object separately. Furthermore, the training efficiency of the object feature extraction model can be significantly improved based on dataset merging and multi-label training samples. Additionally, since multiple labels can separately indicate whether each object clicked on the recommended content and whether it converted (or explicitly indicate that the conversion behavior is unknown), the problem of label leakage can be avoided, improving the performance of the trained model.

[0074] In some embodiments, the feature learning and application system 110 may input multiple test samples into the trained object feature extraction model 112 to obtain test feature embeddings corresponding to the multiple test samples. The multiple test samples may include a first set of test samples where the transformation behavior of the test object is traceable and a second set of test samples where the transformation behavior of the test object is not traceable. In other words, the attribute information of the test object in the test samples may be the same as the attribute information of the objects in the training dataset, thereby enabling further testing of the model's generalization ability and other performance.

[0075] In some embodiments, the number of parameters in the model-dense layer of the object feature extraction model 112 can be increased, thereby giving the new parameters better performance for predicting conversion rates. Specific test results can be found in Table 1 below.

[0076] Table 1

[0077] In Table 1 above, the multi-label processing logic baseline of this disclosure refers to the baseline model without performing model-dense layer optimization, while the multi-label processing logic optimization of this disclosure refers to the object feature extraction model after performing model-dense layer optimization. The performance metric AUC (Area Under Curve) is defined as the area under the ROC curve and the coordinate axis, and this area value will not be greater than 1, with AUC ranging between 0.5 and 1. The closer the AUC is to 1.0, the higher the realism of the detection method. Relative Information Gain (RIG) is a linear transformation of the logarithmic loss; the better the model prediction, the larger the RIG value should be.

[0078] In some embodiments, the feature learning and application system 110 can determine a first length distribution of the first set of feature embeddings corresponding to the first set of test samples. This allows for the calculation of the mean and standard deviation of the lengths of the feature embeddings of the first set of test objects, with the length distribution represented by the mean and standard deviation.

[0079] In some embodiments, the feature learning and application system 110 can determine a second length distribution of the second set of feature embeddings corresponding to the second set of test samples. Similarly, the mean and standard deviation of the lengths of the feature embeddings of the second set of test objects can be statistically analyzed.

[0080] In some embodiments, based on the difference between a first length distribution and a second length distribution, the feature learning and application system 110 can determine the test result of the object feature extraction model, indicating whether the object feature extraction model has passed the test. If the test result indicates that the object feature extraction model 112 has passed the test, then the feature extraction model 112 can be put into use. Conversely, if the test result indicates that the object feature extraction model 112 has failed the test, it means that the feature extraction model 112 has failed to achieve the intended effect and needs to be further trained.

[0081] In some embodiments, using a trained object feature extraction model 112, the feature learning and application system 110 can extract multiple target embedding representations of multiple target objects from the attribute information of multiple target objects, and then determine the similarity of the multiple target objects based on the similarity between the multiple target embedding representations. Thus, the object feature extraction model 112 in this disclosure embodiment can be used to evaluate the similarity of objects (e.g., users) in, for example, recommendation scenarios or other scenarios. Specifically, by extracting the feature embeddings of objects, it is possible to determine which users are similar users based on a comparison of the feature embeddings of multiple users.

[0082] In some embodiments, using a trained object feature extraction model 112, the feature learning and application system 110 can extract the target embedding representation of the target object from the attribute information of the target object, and then determine the probability of providing target recommended content to the target object based on the correlation between the target embedding representation and the embedding representation of the target recommended content. Thus, the object feature extraction model 112 in this embodiment can be further used in a recommendation scenario to evaluate whether to send target recommended content to a target user. Specifically, it can be determined whether to send recommended content to the user based on the correlation between the user's embedding representation and the embedding representation of the recommended content. If the determined probability is high, then the target recommended content will be sent to the user. The user can then choose whether to perform a click or conversion action on the target recommended content as needed.

[0083] Generally, the length of an embedding representation may be related to the number of samples used in training; that is, the more samples there are, the longer the mean length of the embedding representation may be. Tables 2 and 3 below show examples of the normal distribution of user embedding representations in conventional schemes and schemes of this disclosure, respectively.

[0084] Table 2 Normal distribution of user embedding representations in traditional schemes

[0085] In traditional schemes, when the sample size of trackable and untrackable users differs significantly, an exemplary test showed that the average length of the embedding representation of trackable users is 4.429184, while the average length of the embedding representation of untrackable users is 5.019679, with a difference of -0.5990495, or 13%.

[0086] Table 3 Normal distribution of user embedding representation in this disclosed scheme

[0087] In the scheme disclosed herein, even when the sample sizes of traceable and untraceable users differ significantly, exemplary testing shows that the average length of the embedding representation for traceable users is 2.8430356874783547, while the average length of the embedding representation for untraceable users is 2.915914931468604. The difference between the two is -0.07287924399024925, a difference of only -2.56%.

[0088] Therefore, the feature learning scheme of the embodiments of this disclosure can solve the problem of bias in predetermined results for trackable and untrackable users. The embodiments of this disclosure improve the efficiency of model training and achieve effective gains in overall model performance. Specifically, compared with traditional schemes, the scheme of this disclosure achieves better performance for model training on behavioral data of untrackable users.

[0089] Figure 4 illustrates a schematic diagram of a model training and application environment 400 in which embodiments of the present disclosure can be implemented. In environment 400 of Figure 4, the model is generally shown to involve different phases, including a training phase 402 and an application phase 406. A testing phase, not shown in the figure, may also be present after the training phase 402.

[0090] In training phase 402, model training system 410 is configured to train model 405 using training dataset 412. Model 405 may be, for example, the prediction model 114 and object feature extraction model 112 in Figure 1. At the start of training, model 405 may have initial parameter values. The training process involves updating the parameter values ​​of model 405 to desired values ​​based on the training data.

[0091] In the application phase 406, the obtained model 405, with trained parameter values, can be provided to the model application system 430 for use. In the application phase 406, model 405 can be used to process the corresponding target input 432 in the real-world scene and provide the corresponding target output 434. The model application system 430 can be configured to implement the feature learning and application system 110 of Figure 1.

[0092] In Figure 4, the model training system 410 and the model application system 430 can include any computing system with computing capabilities, such as various computing devices / systems, terminal devices, servers, etc. Terminal devices can involve any type of mobile terminal, fixed terminal, or portable terminal, including mobile phones, desktop computers, laptop computers, notebook computers, netbook computers, tablet computers, media computers, multimedia tablets, or any combination thereof, including accessories and peripherals of these devices or any combination thereof. Servers include, but are not limited to, mainframes, edge computing nodes, computing devices in cloud environments, etc.

[0093] It should be understood that the components and arrangements in environment 400 shown in Figure 4 are merely examples, and a computing system suitable for implementing the exemplary implementations described in this disclosure may include one or more different components, other components, and / or different arrangements. For example, although shown as separate, model training system 410 and model application system 430 may be integrated in the same system or device. Implementations of this disclosure are not limited in this respect.

[0094] Figure 5 shows a flowchart of a process 500 for feature learning according to some embodiments of the present disclosure. Process 500 may, for example, be implemented at the feature learning and application system 110 of Figure 1.

[0095] In box 510, a first dataset and a second dataset are obtained. The first dataset includes first behavioral data of the first group of objects on the target application, and the second dataset includes second behavioral data of the second group of objects on the target application. The first behavioral data indicates whether the objects perform click and conversion behaviors on the recommended content, and the second behavioral data indicates whether the objects perform click behaviors on the recommended content.

[0096] In box 520, based on the union of the first dataset and the second dataset, multiple training samples are generated corresponding to multiple objects respectively. Each training sample includes the object's attribute information and multiple labels. The first label among the multiple labels indicates whether the object performs a click action on the recommended content, and the second label among the multiple labels indicates whether the object performs a conversion action on the recommended content or indicates a predetermined indicator.

[0097] In box 530, an object feature extraction model is trained using multiple training samples. The object feature extraction model is configured to extract the feature embeddings corresponding to the object based on the attribute information of the input object.

[0098] In some embodiments, generating multiple training samples corresponding to multiple objects based on the union of the first dataset and the second dataset includes: determining the behavior data of the first object from the first behavior data of the first dataset and the second behavior data of the second dataset based on the object identifier of the first object; and generating the first training sample corresponding to the first object based on the behavior data of the first object.

[0099] In some embodiments, generating a first training sample corresponding to the first object based on the behavior data of the first object includes: determining a first label in the first training sample corresponding to the first object based on the behavior data of the first object indicating whether the first object performs a click behavior on the recommended content; if the behavior data of the first object includes an indication of whether the first object performs a conversion behavior on the recommended content, determining a second label in the first training sample based on the indication of whether the first object performs a conversion behavior on the recommended content; if the behavior data of the first object does not include an indication of whether the first object performs a conversion behavior on the recommended content, determining a second label in the first training sample to indicate a predetermined indicator.

[0100] In some embodiments, the predetermined indicator indicates that the transformation behavior of the object is untraceable.

[0101] In some embodiments, process 500 further includes: inputting multiple test samples into a trained object feature extraction model to obtain test feature embeddings corresponding to multiple test samples, the multiple test samples including a first set of test samples whose transformation behavior of the test object is traceable and a second set of test samples whose transformation behavior of the test object is not traceable; determining a first length distribution of the first set of feature embeddings corresponding to the first set of test samples; determining a second length distribution of the second set of feature embeddings corresponding to the second set of test samples; and determining a test result of the object feature extraction model based on the difference between the first length distribution and the second length distribution, the test result indicating whether the object feature extraction model passes the test.

[0102] In some embodiments, process 500 further includes: using a trained object feature extraction model to extract multiple target embedding representations of multiple target objects from the attribute information of multiple target objects respectively; and determining the similarity of multiple target objects based on the similarity between the multiple target embedding representations.

[0103] In some embodiments, process 500 further includes: using a trained object feature extraction model to extract target embedding representations of the target object from the attribute information of the target object; and determining the probability of providing target recommendation content to the target object based on the correlation between the target embedding representation and the embedding representation of the target recommendation content.

[0104] Figure 6 shows a schematic structural block diagram of an apparatus 600 for feature learning according to some embodiments of the present disclosure. The apparatus 600 may be implemented as or included in the feature learning and application system 110. The various modules / components in the apparatus 600 may be implemented by hardware, software, firmware, or any combination thereof.

[0105] As shown in the figure, the device 600 includes an acquisition module 610 configured to acquire a first dataset and a second dataset. The first dataset includes first behavioral data of a first group of objects on a target application, and the second dataset includes second behavioral data of a second group of objects on the target application. The first behavioral data indicates whether the objects perform click and conversion behaviors on the recommended content, and the second behavioral data indicates whether the objects perform click behaviors on the recommended content.

[0106] The device 600 also includes a generation module 620, configured to generate multiple training samples corresponding to multiple objects based on the union of the first dataset and the second dataset. Each training sample includes attribute information of the object and multiple labels. The first label among the multiple labels indicates whether the object performs a click action on the recommended content, and the second label among the multiple labels indicates whether the object performs a conversion action on the recommended content or indicates a predetermined indicator.

[0107] The device 600 also includes a training module 630 configured to train an object feature extraction model using multiple training samples. The object feature extraction model is configured to extract the feature embedding corresponding to the object based on the attribute information of the input object.

[0108] In some embodiments, the generation module 620 is further configured to determine the behavior data of the first object from the first behavior data of the first dataset and the second behavior data of the second dataset based on the object identifier of the first object; and to generate a first training sample corresponding to the first object based on the behavior data of the first object.

[0109] In some embodiments, the generation module 620 is further configured to: determine a first label in a first training sample corresponding to the first object based on the behavior data of the first object indicating whether the first object performs a click behavior on the recommended content; if the behavior data of the first object includes an indication of whether the first object performs a conversion behavior on the recommended content, determine a second label in the first training sample based on the indication of whether the first object performs a conversion behavior on the recommended content; if the behavior data of the first object does not include an indication of whether the first object performs a conversion behavior on the recommended content, determine a second label in the first training sample indicating a predetermined indicator.

[0110] In some embodiments, the predetermined indicator indicates that the transformation behavior of the object is untraceable.

[0111] In some embodiments, the apparatus 600 further includes a first extraction module configured to input multiple test samples into a trained object feature extraction model to obtain test feature embeddings corresponding to the multiple test samples, wherein the multiple test samples include a first set of test samples whose transformation behavior of the test object is traceable and a second set of test samples whose transformation behavior of the test object is not traceable; and to determine a first length distribution of the first set of feature embeddings corresponding to the first set of test samples; to determine a second length distribution of the second set of feature embeddings corresponding to the second set of test samples; and to determine the test result of the object feature extraction model based on the difference between the first length distribution and the second length distribution, wherein the test result indicates whether the object feature extraction model passes the test.

[0112] In some embodiments, the apparatus 600 further includes a second extraction module configured to use a trained object feature extraction model to extract multiple target embedding representations of multiple target objects from the attribute information of multiple target objects respectively; and to determine the similarity of multiple target objects based on the similarity between the multiple target embedding representations.

[0113] In some embodiments, the apparatus 600 further includes a third extraction module configured to use a trained object feature extraction model to extract the target embedding representation of the target object from the attribute information of the target object; and to determine the probability of providing the target recommendation content to the target object based on the correlation between the target embedding representation and the embedding representation of the target recommendation content.

[0114] Figure 7 illustrates a block diagram of an electronic device 700 in which one or more embodiments of the present disclosure may be implemented. It should be understood that the electronic device 700 shown in Figure 7 is merely exemplary and should not constitute any limitation on the functionality and scope of the embodiments described herein. The electronic device 700 shown in Figure 7 can be used to implement the feature learning and application system 110. The electronic device 700 may include or be implemented as the device 600 of Figure 6.

[0115] As shown in Figure 7, the electronic device 700 is in the form of a general-purpose computing device. Components of the electronic device 700 may include, but are not limited to, one or more processors or processing units 710, memory 720, storage devices 730, one or more communication units 740, one or more input devices 750, and one or more output devices 760. The processing unit 710 may be a physical or virtual processor and is capable of performing various processes according to programs stored in the memory 720. In a multiprocessor system, multiple processing units execute computer-executable instructions in parallel to improve the parallel processing capability of the electronic device 700.

[0116] Electronic device 700 typically includes multiple computer storage media. Such media can be any available media accessible to electronic device 700, including but not limited to volatile and non-volatile media, removable and non-removable media. Memory 720 can be volatile memory (e.g., registers, cache, random access memory (RAM)), non-volatile memory (e.g., read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory), or some combination thereof. Storage device 730 can be a removable or non-removable medium and can include machine-readable media, such as flash drives, disks, or any other media that can be used to store information and / or data (e.g., training data for training) and can be accessed within electronic device 700.

[0117] Electronic device 700 may further include additional removable / non-removable, volatile / non-volatile storage media. Although not shown in FIG. 7, disk drives for reading from or writing to removable, non-volatile disks (e.g., "floppy disks") and optical disk drives for reading from or writing to removable, non-volatile optical disks may be provided. In these cases, each drive may be connected to a bus (not shown) via one or more data media interfaces. Memory 720 may include computer program product 725 having one or more program modules configured to perform various methods or actions of various embodiments of the present disclosure.

[0118] The communication unit 740 enables communication with other electronic devices via a communication medium. Additionally, the functionality of the components of the electronic device 700 can be implemented using a single computing cluster or multiple computing machines capable of communicating via communication connections. Therefore, the electronic device 700 can operate in a networked environment using logical connections to one or more other servers, network personal computers (PCs), or another network node.

[0119] Input device 750 can be one or more input devices, such as a mouse, keyboard, trackball, etc. Output device 760 can be one or more output devices, such as a monitor, speaker, printer, etc. Electronic device 700 can also communicate with one or more external devices (not shown) via communication unit 740 as needed. These external devices include storage devices, display devices, etc., and can communicate with one or more devices that enable user interaction with electronic device 700, or with any device that enables electronic device 700 to communicate with one or more other electronic devices (e.g., network card, modem, etc.). Such communication can be performed via input / output (I / O) interface (not shown).

[0120] According to an exemplary implementation of this disclosure, a computer-readable storage medium is provided that stores computer-executable instructions thereon, wherein the computer-executable instructions are executed by a processor to implement the methods described above. According to an exemplary implementation of this disclosure, a computer program product is also provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, which are executed by a processor to implement the methods described above.

[0121] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, devices, and computer program products implemented according to this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0122] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0123] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0124] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0125] Various implementations of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed implementations. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described implementations. The terminology used herein is chosen to best explain the principles, practical applications, or improvements to technology in the market, or to enable others skilled in the art to understand the various implementations disclosed herein.

Claims

A feature learning method includes: Obtain a first dataset and a second dataset. The first dataset includes first behavioral data of a first group of objects on the target application. The second dataset includes second behavioral data of a second group of objects on the target application. The first behavioral data indicates whether the objects perform click and conversion behaviors on the recommended content. The second behavioral data indicates whether the objects perform click behaviors on the recommended content. Based on the union of the first dataset and the second dataset, multiple training samples are generated, each corresponding to multiple objects. Each training sample includes the object's attribute information and multiple labels. The first label among the multiple labels indicates whether the object performs a click action on the recommended content, and the second label among the multiple labels indicates whether the object performs a conversion action on the recommended content or indicates a predetermined indicator. The multiple training samples are then used to train an object feature extraction model, which is configured to extract the feature embedding corresponding to the object based on the attribute information of the input object. According to the method of claim 1, generating multiple training samples corresponding to multiple objects based on the union of the first dataset and the second dataset includes: Based on the object identifier of the first object, determine the behavior data of the first object from the first behavior data of the first dataset and the second behavior data of the second dataset; And based on the behavior data of the first object, generate the first training sample corresponding to the first object. According to the method of claim 2, generating the first training sample corresponding to the first object based on the behavior data of the first object includes: Based on the behavioral data of the first object, an indication is given as to whether the first object performs a click action on the recommended content, and the first label in the first training sample corresponding to the first object is determined; If the behavioral data of the first object includes an indication of whether the first object performs a conversion action on the recommended content, the second label in the first training sample is determined based on the indication of whether the first object performs a conversion action on the recommended content; If the behavioral data of the first object does not include an indication of whether the first object performs a conversion action on the recommended content, the second label in the first training sample is determined to indicate the predetermined indicator. According to the method of claim 1, the predetermined indicator indicates that the transformation behavior of the object is untraceable. The method according to claim 1 further includes: Multiple test samples are input into the trained object feature extraction model to obtain test feature embeddings corresponding to the multiple test samples, wherein the multiple test samples include a first set of test samples in which the transformation behavior of the test object is traceable and a second set of test samples in which the transformation behavior of the test object is not traceable; and a first length distribution of the first set of feature embeddings corresponding to the first set of test samples is determined. Determine the second length distribution of the second set of feature embeddings corresponding to the second set of test samples; based on the difference between the first length distribution and the second length distribution, determine the test result of the object feature extraction model, wherein the test result indicates whether the object feature extraction model passes the test. The method according to claim 1 further includes: Using the trained object feature extraction model, multiple target embedding representations of the multiple target objects are extracted from the attribute information of the multiple target objects respectively; And based on the similarity between the multiple target embedding representations, the similarity of the multiple target objects is determined. The method according to claim 1 further includes: Using the trained object feature extraction model, the target embedding representation of the target object is extracted from the attribute information of the target object; And based on the correlation between the target embedding representation and the target recommended content embedding representation, determine the probability of providing the target recommended content to the target object. An apparatus for feature learning, comprising: The acquisition module is configured to acquire a first dataset and a second dataset. The first dataset includes first behavioral data of a first group of objects on a target application, and the second dataset includes second behavioral data of a second group of objects on the target application. The first behavioral data indicates whether the objects perform click and conversion behaviors on the recommended content, and the second behavioral data indicates whether the objects perform click behaviors on the recommended content. The generation module is configured to generate multiple training samples corresponding to multiple objects based on the union of the first dataset and the second dataset. Each training sample includes the object's attribute information and multiple labels. The first label among the multiple labels indicates whether the object performs a click behavior on the recommended content, and the second label among the multiple labels indicates whether the object performs a conversion behavior on the recommended content or indicates a predetermined indicator. The training module is configured to train an object feature extraction model using the plurality of training samples, wherein the object feature extraction model is configured to extract the feature embedding corresponding to the object based on the attribute information of the input object. An electronic device, comprising: At least one processing unit; And at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions causing the device to perform the method according to any one of claims 1 to 7 when executed by the at least one processing unit. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method according to any one of claims 1 to 7. A computer program product, tangibly stored in a computer storage medium and comprising computer-executable instructions that, when executed by a device, cause the device to perform the method according to any one of claims 1 to 7.

Images