Method for object recommendation, and device, storage medium and program product

Abstract

According to the embodiments of the present disclosure, provided are a method for object recommendation, and a device, a storage medium and a program product. The method comprises: obtaining behavior information related to a behavior sequence which is executed by a target user on objects, and scenario information related to a target scenario used for recommending a target object; using each of a plurality of first feature generation layers, which respectively correspond to a plurality of behaviors, in a trained recommendation model to respectively determine, on the basis of the scenario information, processing parameter values specific to the behaviors, and on the basis of the processing parameter values specific to the behaviors and a shared processing parameter value among the plurality of behaviors, extracting behavior feature representations of the behaviors from the behavior information related to the behavior sequence; using a second feature generation layer in the recommendation model to extract an object feature representation from the scenario information and object information of the target object; and at least on the basis of the respective behavior feature representations of the plurality of behaviors and the object feature representation, determining a prediction result of the target user executing a specified behavior on the target object.

Description

Method, device, storage medium, and program product for object recommendation Technical Field

[0001] Example embodiments of the present disclosure generally relate to the field of computer technology, and more particularly, to a method for object recommendation, an apparatus for object recommendation, an electronic device, a computer-readable storage medium, and a computer program product. Background Art

[0002] With the development of digital content and the widespread use of the Internet, recommendation systems are configured to perform prediction tasks for object recommendations, which plays an important role in alleviating information overload and enhancing user experience. In a prediction task, the input of the prediction task usually includes at least data related to the object, and the output is the prediction result. An accurate prediction process will give similar prediction results for similar objects. For example, when predicting whether a user will buy a certain product, the input of the prediction task is user-related data and product-related data, and the output is the probability that the user will perform a certain behavior on the product. When using a prediction model to perform a probabilistic prediction task, it is expected that the prediction model can give similar prediction probabilities for users with similar features. In this way, representing the data related to users with similar features with vectors of similar distances (i.e., feature embedding) and using them as input can make the prediction results more in line with expectations.

[0003] Summary of the Invention

[0004] In a first aspect of the present disclosure, a method for object recommendation is provided. The method includes: obtaining behavior information related to a sequence of behaviors performed by a target user on an object and scene information related to a target scene for recommending the target object, wherein the behavior sequence includes multiple behaviors; utilizing each of multiple first feature generation layers corresponding to the multiple behaviors in a trained recommendation model to determine, based on scene information, a processing parameter value specific to the behavior, and extracting a behavior feature representation of the behavior from the behavior information related to the behavior sequence based on the behavior-specific processing parameter value and a shared processing parameter value between the multiple behaviors; utilizing a second feature generation layer in the recommendation model to extract an object feature representation from the scene information and object information of the target object; and determining a prediction result of the target user performing a specified behavior on the target object based on at least the behavior feature representation and the object feature representation of each of the multiple behaviors.

[0005] In a second aspect of the present disclosure, a device for object recommendation is provided. The device includes: an information acquisition module configured to obtain behavior information related to a sequence of behaviors performed by a target user on an object and scene information related to a target scene for recommending the target object, wherein the behavior sequence includes multiple behaviors; a first feature generation module configured to use each of multiple first feature generation layers corresponding to the multiple behaviors in a trained recommendation model to determine, based on scene information, a behavior-specific processing parameter value, and extract a behavior feature representation of the behavior from the behavior information related to the behavior sequence based on the behavior-specific processing parameter value and the shared processing parameter value between the multiple behaviors; a second feature generation module configured to use the second feature generation layer in the recommendation model to extract an object feature representation from the scene information and the object information of the target object; and a prediction determination module configured to determine a prediction result of the target user performing a specified behavior on the target object based on at least the behavior feature representation and the object feature representation of each of the multiple behaviors.

[0006] In a third aspect of the present 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 the present disclosure, a computer-readable storage medium is provided, wherein a computer program is stored on the medium, and when the computer program is executed by a processor, the method of the first aspect is implemented.

[0008] In a fifth aspect of the present disclosure, a computer program product is provided, which includes a computer program, and when the computer program is executed by a processor, the method of the first aspect is implemented.

[0009] It should be understood that the content described in this section is not intended to limit the key features or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:

[0011] FIG1 shows a schematic diagram of an example environment in which embodiments of the present disclosure can be implemented;

[0012] FIG2 shows a schematic structural block diagram of a recommendation model according to some embodiments of the present disclosure;

[0013] FIG3 shows a schematic structural block diagram of a feature generation layer in a recommendation model according to some embodiments of the present disclosure;

[0014] FIG4A shows a schematic structural block diagram of another feature generation layer in a recommendation model according to some embodiments of the present disclosure;

[0015] FIG4B shows a schematic diagram of query feature generation based on grouping according to some embodiments of the present disclosure;

[0016] FIG5 shows a schematic diagram of a model training and application environment in which embodiments of the present disclosure can be implemented;

[0017] FIG6 shows a flowchart of a process for object recommendation according to some embodiments of the present disclosure;

[0018] FIG7 shows a block diagram of an apparatus for object recommendation according to some embodiments of the present disclosure; and

[0019] FIG8 illustrates a block diagram of an electronic device in which one or more embodiments of the present disclosure may be implemented. DETAILED DESCRIPTION

[0020] The following describes embodiments of the present disclosure in more detail with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0021] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, i.e., "including but not limited to". The term "based on" should be understood as "based at least in part 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 be included below.

[0022] It is understandable that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) must comply with the requirements of relevant laws, regulations and relevant provisions.

[0023] It is understandable that before using the technical solutions disclosed in the various embodiments of this disclosure, the type, scope of use, usage scenarios, etc. of the personal information involved in this disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0024] For example, in response to receiving a user's active request, a prompt message is sent to the user to clearly remind the user that the operation requested to be performed will require obtaining and using the user's personal information, so that the user can independently choose whether to provide personal information to the electronic device, application, server or storage medium and other software or hardware that performs the operation of the technical solution of the present disclosure based on the prompt message.

[0025] As an optional but non-limiting implementation, in response to receiving a user's active request, a prompt message may be sent to the user, for example, in the form of a pop-up window, in which the prompt message may be presented in text form. Furthermore, the pop-up window may also include a selection control for the user to select "agree" or "disagree" to provide personal information to the electronic device.

[0026] It is understandable that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0027] As used herein, the term "model" can learn the association between corresponding inputs and outputs from training data, so that after training is completed, corresponding outputs can be generated for given inputs. The generation of the model can be based on machine learning technology. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs by using multiple layers of processing units. A neural network model is an example of a model based on deep learning. In this article, "model" may also be referred to as "machine learning model", "learning model", "machine learning network" or "learning network", and these terms are used interchangeably in this article.

[0028] A "neural network" is a machine learning network based on deep learning. A neural network is capable of processing inputs and providing corresponding outputs. It typically includes 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 typically include many hidden layers, thereby increasing the depth of the network. The layers of a neural network are connected in sequence so that the output of the previous layer is provided as input to the next layer, where the input layer receives the input of the neural network and the output of the output layer serves as the final output of the neural network. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each of which processes the input from the previous layer.

[0029] Generally speaking, machine learning can be roughly divided into three stages, namely the training stage, the testing stage, and the application stage (also known as the inference stage). In the training stage, a given model can be trained using a large amount of training data, and the parameter values ​​are continuously updated iteratively until the model can obtain consistent inferences that meet the expected goals from the training data. Through training, the model can be considered to be able to learn the association between input and output (also known as input-to-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing stage, the test input is applied to the trained model to test whether the model can provide the correct output, thereby determining the performance of the model. In the application stage, the model can be used to process the actual input based on the parameter values ​​obtained through training to determine the corresponding output.

[0030] FIG1 shows a schematic diagram of an example environment 100 in which an embodiment of the present disclosure can be implemented. One or more content providers can use a recommendation system 120 to manage content to be on a content distribution platform 110. One or more terminal devices 130-1, 130-2, 130-3, etc. (for ease of discussion, collectively referred to as or individually referred to as terminal devices 130) are associated with the content distribution platform 110 and can, for example, access various types of content provided on the content distribution platform 110 based on corresponding users 132-1, 132-2, 132-3, etc. (for ease of discussion, collectively referred to as or individually referred to as users 132). As an example, the content distribution platform 110 can be an application, a website, a webpage, and other accessible platforms. The terminal device 130 can be installed with an application for accessing the content distribution platform 110, or can access the content distribution platform 110 in an appropriate manner.

[0031] The recommendation system 120 can be configured to deliver one or more specific recommended content items related to one or more objects to the audience based on the corresponding recommendation strategy (e.g., provide or present on the terminal device 130). One or more recommended objects 142-1, 142-2, ... 142-M (for ease of discussion, collectively or individually referred to as recommended objects 142) in the recommendation database 140.

[0032] In some embodiments, the recommendation system 120 can use a recommendation model to perform tasks related to object recommendations. For example, if it is desired to recommend content to a user or an audience group, the prediction task of the recommendation model may include predicting the likelihood of the user performing a specific behavior on the object to be recommended to determine whether to recommend the object to the user. In the recommendation scenario, the specific behavior to be predicted may consider click behavior, viewing behavior, download behavior (for downloadable applications), interactive behavior (like, favorite, comment, etc.), add to shopping cart behavior, purchase behavior, etc. The output of the recommendation model can be a binary classification result to predict whether a specific behavior occurs or not, or it can be a probability value within a continuous value range to indicate the probability of occurrence of a specific behavior.

[0033] In prediction tasks involving object recommendations, in order to obtain more accurate prediction results, the input of the prediction task usually includes information about the user and object to be recommended. This information may include multiple fields describing different attributes or aspects of the user and object.

[0034] The prediction task of the recommendation model may include predicting that a user performs a specific behavior on an object (for example, a click behavior). The information required for the prediction task may include the user's behavior data on the application, such as clicks, browsing, purchases, likes, comments, and other behavior data. The user's data on different applications can characterize the characteristics of the object, thereby helping to perform predictions. Depending on the application needs, the recommendation model can be constructed to include various types of model architectures, including deep neural network (DNN) architectures. The embodiments of the present disclosure do not limit the specific type and structure of the recommendation model.

[0035] In the environment 100, the terminal device 110 can be any type of mobile terminal, fixed terminal or portable terminal, including a mobile phone, a desktop computer, a laptop computer, a notebook computer, a netbook computer, a tablet computer, a media computer, a multimedia tablet, a personal communication system (PCS) device, a personal navigation device, a personal digital assistant (PDA), an audio / video player, a digital camera / camcorder, a positioning device, a television receiver, a radio broadcast receiver, an e-book device, a gaming device or any combination thereof, including accessories and peripherals of these devices or any combination thereof. In some embodiments, the terminal device 110 can also support any type of interface for the user (such as a "wearable" circuit, etc.). The recommendation system 120 can, for example, be various types of computing systems / servers that can provide computing capabilities, including but not limited to mainframes, edge computing nodes, computing devices in cloud environments, and the like.

[0036] It should be understood that the structure and function of the various elements in the environment 100 are described for illustrative purposes only and do not imply any limitation on the scope of the present disclosure.

[0037] Feature representation is a crucial component of machine learning systems or model processing. A feature representation is a vector of appropriate dimensions that represents raw data. Distinguishing behavioral data from different objects is crucial in prediction input, as different objects may exhibit distinct behavioral characteristics and should be associated with different predictions. For example, different users or audiences may be interested in different recommended content. Without object-level behavioral data, it becomes difficult to fully describe user and object characteristics, resulting in suboptimal prediction results.

[0038] In different scenarios of object recommendation (also called different domains), users often show different behavior patterns and user needs. For example, object recommendation may include recommending objects in information flow, recommending objects on shopping pages, recommending objects on live broadcast pages, recommending objects in different types of applications, and so on. Some solutions will train different recommendation models for different recommendation scenarios (for example, the information flow scenario, live broadcast scenario, or mall scenario of an application). Some other solutions may consider mixing training data from different scenarios to train machine learning models so that the model can be used in different scenarios. However, the scenario-level model cannot take into account the commonalities between different scenarios, while the general recommendation model finds it difficult to take into account the differences between different scenarios.

[0039] To minimize the differences between different scenarios, these solutions usually emphasize the use of complex model structures to adapt to different input data distributions, but this ignores the inherent understanding of user behavior sequences from a multi-domain perspective.

[0040] Many recommendation models utilize the attention mechanism, which can be written as a query-key-value paradigm and expressed as g(Repcand,Repseq)×Repseq, where Repcand represents the feature representation of the object to be recommended (to predict the user's click or conversion probability for the object), Repseq represents the feature representation of the behavior sequence related to the user, and g represents the calculation of the attention weight between any pair of feature representations.

[0041] Current recommendation systems generally adopt a shared-base embedding layer model, where the embedding representations of objects and user behaviors are shared, without considering the differences between objects and users belonging to different scenarios. Furthermore, objects are used as shared query features with different key / value pairs. This fails to consider the differences between multiple objects in a user's sequence, which often occurs across multiple scenarios. This poses potential challenges to current multi-scenario recommendations in two ways:

[0042] Representations of user behavior sequences lack differentiation. Multi-domain representations of the same objects appearing in different users' behavior sequences are essential for multi-domain recommendations. For example, new users may prefer to watch videos with high likes, while long-time users may prefer to watch videos created by creators in their followlist. Therefore, the appeal of the same video to different users may differ, while its feature representation is shared across users, hindering the recommendation system's ability to identify the needs of different users.

[0043] For different users, or different items in the same user's behavior sequence, recommended objects have different influences and functions due to multi-domain influences. For example, a user's shopping behavior on a shopping page may be primarily influenced by the price of the item, while for content-oriented users, the influence of the video creator on their shopping behavior in the video streaming scenario is more significant. Therefore, the same object should be processed differently to adapt to the different historical behavior items in the user's behavior sequence.

[0044] In the embodiments of this disclosure, an improved object recommendation scheme is proposed. In this embodiment, a behavior-specific feature generation layer within a behavior sequence comprehensively enhances the differentiated capabilities in modeling the behavior sequences perceived by the target object across different scenarios. This scheme can adaptively determine the user's interest in a specific object in different scenarios and accurately predict the likelihood of the user performing a specified behavior on the object, which facilitates more accurate service recommendations for the user.

[0045] Some example embodiments of the present disclosure will be described below with reference to the accompanying drawings.

[0046] FIG2 shows a schematic block diagram of a recommendation model 200 according to some embodiments of the present disclosure. The recommendation model 200 can be used, for example, by the recommendation system 120 of FIG1 , for example, to help the recommendation system 120 rank multiple objects to be recommended to determine the objects to be recommended to a particular user or a particular user group.

[0047] The model output of recommendation model 200 is prediction result 252, which can indicate the probability that a specific user will perform a specified action on a specific object. In other words, recommendation model 200 is trained based on specified actions. In the context of online recommendations, specified actions can include, for example, clicking on an object, viewing it, downloading it (for downloadable applications), interacting with it (likes, favorites, comments, etc.), adding it to a shopping cart, purchasing it, and so on. The actions predicted by recommendation model 200 can be customized based on the specific situation.

[0048] The model inputs of the recommendation model 200 include at least behavior information 204 related to the sequence of behaviors performed by the target user on an object, scenario information 206 related to the target scenario for recommending the target object, and object information 208 of the target object. The "target user" here refers to the user for whom the recommendation model 200 is used to predict which object to recommend. The "target object" here refers to the object currently being predicted in the set of candidate objects to be recommended.

[0049] Behavior information 204 relates to the target user's behavior sequence, including multiple behaviors. Examples of behavior sequences may include, but are not limited to, viewing behavior, clicking behavior, interactive behavior, adding to a shopping cart behavior, purchasing behavior, and the like. In some embodiments, these behaviors may be a series of behaviors performed on the same object, or may be a series of behaviors performed on different objects. For example, a behavior sequence may include relevant information about viewing object A, relevant information about clicking on object A, relevant information about adding object A to a shopping cart, relevant information about purchasing object A, and the like. For another example, a behavior sequence may include relevant information about viewing object A, relevant information about clicking on object A, relevant information about clicking on object B, relevant information about viewing object B, collecting relevant information about object B, and the like. Behavior information 204 includes relevant information corresponding to each behavior, which may include, but is not limited to, an indication of the object to which the behavior is directed, an indication of the behavior, contextual information regarding the occurrence of the behavior, and the like. The specific definition and requirements of the behavior sequence, as well as the specific content contained in the behavior information, can be determined based on actual applications.

[0050] In this article, scene refers to the domain or range used to recommend objects. The division criteria or division granularity of the recommendation scene can be configured according to actual conditions. For example, for a specific application that supports object recommendation, multiple scenes (multiple domains) can be divided, including information flow scenes, live broadcast scenes, shopping scenes, and so on. Of course, different scenes may involve different applications or different platforms. For example, different applications can be determined as different scenes. The embodiments of the present disclosure do not limit the scope of the scene. The scene information 206 related to the target scene may include feature information of one or more categories related to the scene, such as the category and description of the scene, etc.

[0051] Object information 208 may include descriptive information of the target object, including object category, object creator, and so on.

[0052] In some embodiments, the model input of the recommendation model 200 may also include user information 202 related to the target user. User information 202 may be user-side information and may include user context information. In some embodiments, the model input of the recommendation model 200 may also include other relevant information, such as cross-feature information between the user and the object, context information related to the object recommendation, and so on.

[0053] During the processing of the recommendation model 200 , various types of information input to the model can be directly converted into embedded representations via the embedding layer 210 of the recommendation model 200 .

[0054] The prediction process of the recommendation model 200 can be generally expressed as follows:

[0055] Where E() indicates the embedding function of the embedding layer 210, which directly converts the original model input into an embedded representation. In the embedding layer 210, the continuous feature information can first be bucketed and embedded into a dense vector representation; E U Indicates user information 202, E I Indicates object information 208, E b Indicates behavior information 204, E D Indicates scene information 206, and E O It should be understood that in addition to the above four types of information, the model input of the recommendation model 200 may also include other types of information, and the embodiments of the present disclosure are not limited in this respect. In the above formula (1), f represents the overall processing process of the recommendation model 200, The prediction result 252 of the recommendation model 200 is shown.

[0056] An example structure of the recommendation model 200 is shown in FIG2 . The recommendation model 200 includes a feature extraction portion 220 configured to extract feature representations from the embedded features of the partial information generated by the embedding layer 210. The feature extraction portion 220 includes at least a plurality of (T) feature generation layers 224-1, 224-2, 224-3, ..., 224-T (collectively referred to as or referred to as feature generation layers 224, T is an integer greater than 1). In an embodiment of the present disclosure, the number (T) of feature generation layers 224 is equal to the number (T) of behaviors included in the behavior sequence in the behavior information 204. That is, for each behavior in the behavior sequence, a corresponding feature generation layer 224 may be configured.

[0057] Many current recommendation models simply embed raw behavior information into a dense vector representation. This approach gives each specific behavior a unified embedding representation, which is shared across different scenarios and user behavior sequences, ignoring the impact of differences between scenarios.

[0058] In an embodiment of the present disclosure, the behavior embedding representation 214 corresponding to each behavior and the scene embedding representation 216 corresponding to the target scene are input into the feature generation layer 224 to generate the behavior feature representation 225 corresponding to each behavior. The scene information 206 may include explicit scene indications and implicit scene indications. The explicit scene indication can distinguish which scene the current model input belongs to. For example, the value range of the explicit scene indication may be [0, 1, 2], where the three values ​​in this range respectively indicate the information flow scene, live broadcast scene, and shopping scene used for object recommendation in the same application. Of course, it should be understood that if more possible scenes are involved, the value range of the explicit scene indication can be defined to be larger. The implicit scene indication can be descriptive information related to the scene included in the scene information 206, such as the user's activity in a certain aspect in the target scene, etc. The implicit scene indication can be formulated through feature engineering, and the statistical features included can be information that is believed to better capture and distinguish different scenes.

[0059] Based on the scene information 206 and the behavior information 204, the feature generation layer 224 is configured to adaptively generate behavior feature representations corresponding to each behavior in the behavior sequence under the guidance of the scene information 206. This allows different behavior feature representations to be extracted when the same behavior for the same object appears in the behavior sequences of different users in different scenes due to the differences between the users in the multiple scenes. The feature generation layer 224 corresponding to each behavior is configured to determine the behavior feature representation of the behavior based on the information portion of the behavior information 204 corresponding to the behavior and also based on the scene information 206.

[0060] To support model processing, both the behavior information 204 and the scene information 206 are converted into corresponding embedding representations by the embedding layer 210. The behavior information portions of each of the T behaviors in the behavior sequence are respectively converted into behavior embedding representations 214-1, 214-2, 214-3, ..., 214-T (collectively referred to as or referred to as behavior embedding representations 214). The scene information 206 is converted into a scene embedding representation 216. Therefore, the feature generation layers 224-1, 224-2, 224-3, ..., 224-T respectively receive the behavior embedding representations 214-1, 214-2, 214-3, ..., 214-T and the scene embedding representation 216 of the corresponding behaviors, and respectively output the behavior feature representations 225-1, 225-2, 225-3, ..., 225-T (collectively referred to as or referred to as behavior feature representations 225) corresponding to the T behaviors.

[0061] The feature generation layer 224 corresponding to each behavior is configured to determine a behavior-specific processing parameter value based on the scene information related to the target scene for recommending the target object (i.e., based on the scene embedding representation 216). The feature generation layer 224 corresponding to each behavior is then configured to extract a behavior feature representation 225 of the behavior from the behavior information related to the behavior sequence (i.e., based on the behavior embedding representation 214 corresponding to each behavior) based on the behavior-specific processing parameter value and the shared processing parameter value between multiple behaviors.

[0062] The recommendation model 220 also includes a feature generation layer 226. The feature generation layer 226 is configured to extract object feature representations from the scene information 206 and the object information 208 of the target object. To support model processing, the scene information 206 and the object information 208 are converted into corresponding embedded representations by the embedding layer 210, including the scene embedding representation 216 and the object embedding representation 218. Therefore, the input of the feature generation layer 226 includes the scene embedding representation 216 and the object embedding representation 218.

[0063] In some embodiments, the feature generation layer 226 is configured to process the scene embedding representation 216 and the object embedding representation 218 to generate a shared object feature representation. In some embodiments, in order to adapt to the different feature information exhibited by the same object in different scenes, the feature generation layer 226 is configured to capture different aspects of each behavior in the behavior sequence and generate multiple adaptive feature representations, that is, multiple object feature representations for multiple behaviors. These object feature representations can reflect the different impacts on different behaviors in the behavior sequence. As shown in Figure 2, the feature generation layer 226 will generate multiple object feature representations 228-1, 228-2, 228-3, ..., 228-T (collectively referred to as or referred to as object feature representation 228) for multiple behaviors.

[0064] Based on the behavior feature representation and object feature representation, the subsequent processing layers of the recommendation model 220 determine the predicted outcome of the target user performing a specified behavior on the target object based on at least the behavior feature representation and object feature representation of each of the multiple behaviors. The subsequent processing layers of the recommendation model 220 may include an attention layer 230, a cascade layer 240, and a multi-layer perceptron (MLP) layer 250. The processing mechanisms of these subsequent layers will be described in detail later.

[0065] Next, the specific structure of the feature generation layer 224 will be first introduced, and then the specific structure of the feature generation layer 226 in the embodiment of generating multiple object feature representations 228 will be introduced.

[0066] Figure 3 shows a schematic structural block diagram of the feature generation layer 224 in the recommendation model 200 according to some embodiments of the present disclosure. As previously mentioned, for each behavior in the behavior sequence, a corresponding feature generation layer 224 can be configured to generate a behavior feature representation corresponding to the behavior.

[0067] The feature generation layer 224 is configured to determine behavior-specific weights and behavior-specific bias values ​​based on the scene information 206 (specifically, based on the scene embedding representation 216 corresponding to the scene information 206). The feature generation layer 224 further generates weights for the behaviors based on the behavior-specific weights and the shared weights between multiple behaviors, and extracts behavior feature representations of the behaviors from the behavior information 204 (specifically, based on the behavior embedding representation 214 corresponding to the behavior information 204) based on the generated weights and behavior-specific bias values.

[0068] Assume that the behavior embedding representation 214 is represented as Where T and d S The number of behaviors in the behavior sequence and the embedding dimension of the behavior embedding representation 214 corresponding to each behavior are respectively identified. The scene embedding representation 216 is represented as where d D Represents the embedding dimension of the scene embedding representation 216. In each feature generation layer 224, the weight generation process includes private weights (i.e., behavior-specific weights) and shared weight parts, which can capture the individuality and commonality of multiple scenes respectively. For the private weight part, as shown in Figure 3, the feature generation layer 224 includes a weight generation and shaping layer 310, which is configured to generate behavior-specific weights 312 from the scene embedding representation 216. In one example, the weight generation and shaping layer 310 can be implemented based on a multi-layer perceptron (MLP), for example, it can include two layers of MLP. In some embodiments, the generation process of the behavior-specific weights 312 can be expressed as follows:

[0069] in And d h Denotes the hidden layer dimension. Note that the MLP-based weight generation and shaping layer 310 can not only improve the expressiveness of the model, but also significantly reduce the model parameters and computational cost, because d S The dimensions of d are usually very large (tens of dimensions), and h The dimension is much smaller than d S , that is, d h <<(d S ×d S ).

[0070] Next, share weight 314 Also applied to the behavior-specific weights 312. Shared weights 314 It is also a learnable parameter of the recommendation model 200, but its value is shared among multiple behaviors of the user group. In order to simultaneously learn the sharing between scenarios and the individuality of each scenario in the behavior feature processing, the weight combiner 330 in the feature generation 224 is configured to generate a weight 316 for each behavior based on the behavior-specific weight 312 and the shared weight 314 among multiple behaviors. The generation process of the weight 316 can be expressed as follows: generated =η*(W shared ⊙W private ) (3)

[0071] Where ⊙ represents the element-wise product. Due to the introduction of the Sigmoid function, W private The value range is from [0,1], and the scaling parameter η can be used to further amplify W private The value range of . Once you, the calculation process of formula (3) can be regarded as the shared weight W shared Adaptive scaling process.

[0072] In addition, the feature generation layer 224 also includes a bias generation layer 320, which is configured to generate behavior-specific bias values ​​322 from the scene embedding representation 216. The processing of the bias generation layer 320 can also be similar to the weight generation and shaping layer 310. Similar to the weight generation process, the bias 322b generated The generation process can be determined by:

[0073] in and

[0074] The generated weights 316 are applied to the behavior embedding representation 214 via the multiplier 340, and the weighted result is combined with the behavior-specific bias value 322 to obtain the behavior feature representation 225. S-personalized The generation process of E can be expressed as follows: S-personalized =E S Reshape(W generated ) T +b generated (5)

[0075] The Reshape operator is to convert the weight 316W into a one-dimensional (1D) vector. generated Reshape into a two-dimensional (2D) matrix, the dimension of which is d S ×d S .

[0076] In addition to modeling the features of the behavior sequence, another feature generation layer 226 of the recommendation model 200 is used to model the multi-scenario features of the target object. Generally, feature modeling for the target object can also include two aspects. On the one hand, similar to the feature representation of the behavior sequence, the feature representation of the target object itself is also obtained through a shared embedding layer 210, and the embedding representation is shared for different users and different scenarios. On the other hand, it is considered that the target object may play different roles in different user actions in different scenarios. For example, the user's follow list on the shopping page reflects the user's shopping interests, while the follow list in the information stream reflects the user's demand for pushed content. Therefore, the different needs of users should be captured through more differentiated object feature representations.

[0077] Therefore, the feature generation layer 226 also generates multiple object feature representations for each of the multiple behaviors in the behavior sequence. Figure 4A shows a schematic block diagram of the feature generation layer 226 in the recommendation model 200 according to some embodiments of the present disclosure. As previously described, in considering the differentiated representation of the behavior sequence, the feature generation layer 226, guided by the scene information, generates an object feature representation 228 corresponding to each behavior in the behavior sequence.

[0078] The feature generation layer 226 is configured to extract multiple intermediate feature representations 422 from the scene information (specifically, based on the scene embedding representation 216 corresponding to the scene information 206) and the object information (specifically, based on the object embedding representation 218 corresponding to the object information 208). The shared feature representation 412 is then extracted from the object information. The feature generation layer 226 is further configured to generate multiple object feature representations 228-1, 228-2, ..., 228-T for multiple behaviors based on the multiple intermediate feature representations 422 and the shared feature representation 412.

[0079] Assume that the scene embedding representation 216 is The object embedding representation 218 is where d D and d Q are the embedding dimensions of the scene embedding representation and the object embedding representation, respectively. The feature generation layer 226 includes a cascade layer 410, which is configured to cascade the scene embedding representation 216 and the object embedding representation 218 to obtain a cascaded embedding representation. The feature generation layer 226 also includes a feature generation and shaping layer 410, which is configured to generate multiple intermediate feature representations 422 based on the cascaded embedding representation. The generation process of multiple intermediate feature representations 422 can be expressed as follows:

[0080] in refers to the concatenation of the scene embedding representation 216 and the object embedding representation 218, and Note that due to the embedding dimension d of the scene embedding representation 216 D There may usually be dozens of hidden layers, and the hidden layer dimension d h <<(T×d Q ), then the computational overhead required for the generation of the intermediate feature representation 422 will not be too large and can be controlled within O(d h Td Q ) such a level.

[0081] In some embodiments, the number of the plurality of intermediate feature representations 422 may be equal to the number T of behaviors in the behavior sequence. In some embodiments, the number of the plurality of intermediate feature representations 422 is less than the number T of behaviors in the behavior sequence.

[0082] In an example embodiment where the number of multiple intermediate feature representations 422 is less than the number T of behaviors in the behavior sequence, considering that adjacent behaviors of a user usually occur in the same scenario, the T behaviors in the behavior sequence can be divided into G behavior groups in sequence (G is less than T, and G can be an integer greater than 1). Each behavior group includes at least two adjacent behaviors in the behavior sequence or a single behavior. Then, for the G behavior groups, the feature generation and shaping layer 410 generates G intermediate feature representations 422 respectively. In this way, each intermediate feature representation is specific to a behavior group.

[0083] For each of the G behavior groups, an object feature representation 228 for at least two adjacent behaviors or a single behavior in the behavior group is generated based on the intermediate feature representation 422 and the shared feature representation 412 corresponding to the behavior group. In other words, the object feature representation 228 corresponding to the behaviors in the same behavior group is repeated. In this case, Can be repeated to the dimension of the intermediate feature representation for T behaviors The computational cost can be obtained from O(d h Td Q ) is further reduced to O(d h Gd Q ), because G is smaller than T.

[0084] As shown in Figure 4B, mapping relationship 450 indicates that for the same target object, feature generation layer 226 needs to generate T object feature representations 228-1, 228-2, 228-3, ..., 228-T, each corresponding to a behavior in the target user's behavior sequence. Mapping relationship 452 indicates that after the behavior grouping, feature generation layer 226 actually needs to generate G different object feature representations, namely object feature representations 228-1, ..., 228-G. For multiple behaviors in the same behavior group, their object feature representations 228 are the same. For example, if adjacent behaviors 1, 2, and 3 are grouped into behavior group 1, then feature generation layer 226 actually only needs to calculate and generate the same object feature representation 228-1 for these three behaviors. The generation process for object feature representations corresponding to other behavior groups is similar. Of course, it should be understood that Figure 4B only provides an example of behavior grouping. In actual applications, T behaviors can be divided into any number of behavior groups to calculate object feature representations based on actual needs and computational overhead requirements.

[0085] Feature generation layer 226 embeds the object into representation 218E T Determined as shared feature representation 412, namely E Q-shared =tile(E T ), where tile refers to the repetition operator, E T Repeat T times, that is Corresponding to the T intermediate feature representations 422 (or the T intermediate feature representations 422 obtained by repetition within the group from the G intermediate feature representations 422), T object feature representations 228 for the T behaviors are generated based on the shared feature representation 412 and the multiple intermediate feature representations 422. Then, the final generated object feature representation 228 can be determined by the combiner 430 in a residual manner, that is: E Q-personalized =Reshape(E Q-private +E Q-shared ) (7)

[0086] The Reshape operator is to convert the one-dimensional (1D) vector E Q-personalized Reshape into a two-dimensional (2D) matrix, the dimension of which is expressed as T×d Q .

[0087] Referring back to FIG. 2 , in the plurality of behavior characteristics representations 225 and multiple object feature representations 228 Based on this, the attention layer 230 in the recommendation model 200 can apply an appropriate attention mechanism to process these feature representations to further determine the degree of relevance between the target user and the target object in a specific scenario, thereby predicting whether the user is likely to perform a specified behavior on the object. The attention layer 230 may include attention blocks 232-1, 232-2, 232-3, ..., 232-T (collectively or individually referred to as attention blocks 232) for T behaviors.

[0088] By utilizing multiple attention blocks 232 corresponding to multiple behaviors, and applying an attention mechanism to the behavior feature representation 225 corresponding to each behavior and the object feature representation 228 for each behavior, multiple weighted feature representations for multiple behaviors are generated, and based at least on the multiple weighted feature representations, a prediction result 452 of the target user performing a specified behavior on the target object is determined.

[0089] In some embodiments, each attention block 232 can apply an attention mechanism based on a Transformer (sometimes also referred to as a "transformer") structure. The input to the Transformer is defined as the query feature, key feature, and value feature input to the Transformer. The processing of the Transformer can be understood as calculating a self-attention weight matrix using the query feature Q and the key feature K, and using the self-attention weight matrix to perform a weighted summation on the value feature V.

[0090] In some embodiments, in each attention block 232 , the object feature representation 228 is determined as the query feature Q, and the behavior feature representation 225 is determined as the key feature K and the value feature V.

[0091] The attention block 232 calculates the attention weight for each behavior and aggregates the behavior sequence under the guidance of the query feature Q. It can be understood that the attention block 232 can adopt various model structures based on the attention mechanism. Taking the multi-head target attention mechanism as an example, for each head, the object feature representation 228 and the behavior feature representation 225 are first transformed as follows: Q = E Q-personalized W Q (8)

[0092] K=E S-personalized W K (9)

[0093] V=E S-personalized W V (10)

[0094] in W K and d A Refers to the dimension size in the target attention mechanism.

[0095] The attention weight z′[t] in the t-th query-key pair is It can be obtained in the following ways:

[0096] Where 1≤t≤T (11)

[0097] After obtaining the attention weights z′[t], a Softmax-based operation may be performed to normalize the resulting weights and aggregate the weighted feature representations based on the behavioral feature representations 225 (i.e., the value features V), which may be expressed as:

[0098] z=softmax(z′), (12)

[0099] Where z represents the weighted feature representation for each behavior, and s is the weighted feature representation after aggregation via the aggregator 234.

[0100] The feature representation s obtained after sequence modeling can be cascaded with other feature embedding representations in the cascade layer 240 of the recommendation model 200, including the object embedding representation 218E corresponding to the object information. I , and the user embedding representation 212E corresponding to the user information of the user U , and possibly other information E O The corresponding embedding representation. The cascaded embedding representation can be determined as:

[0101] The concatenated embedding representations are then processed through one or more MLP layers 250 in the recommendation model 200 to determine a prediction result 252 .

[0102] The above describes the processing of the recommendation model 200. During the training process, the training data of the recommendation model 200 may include model input samples (which include information required for various model inputs) and the labels y of the actual model outputs. Based on the known labels y, the training process of the recommendation model 200 can be expressed as determining the training loss via a binary cross entropy function, and minimizing the training loss or achieving the desired target by continuously updating the model parameters. The loss function can be expressed as follows:

[0103] It should be understood that the embodiments of the present disclosure do not limit the specific training algorithm and training strategy of the recommendation model 200.

[0104] According to the embodiments of the present disclosure, the object recommendation process can fully exploit the differentiated modeling capabilities of current recommendation systems based on attention mechanisms, providing more accurate and adaptive prediction capabilities in multi-scenario tasks. Specifically, the feature generation layer specific to the behavior sequence in the recommendation model is configured with a shared-private structure to learn feature representations for different scenarios in the user behavior sequence, aiming to generate differentiated representations for sequence behavior, that is, the same object appearing in different user sequences has different representations. In another feature generation layer of the recommendation model, the recommended object generates different scene-object representations based on different behaviors in the behavior sequence, which further enhances the differentiated modeling capabilities. These feature generation layers can capture multi-scenario features from the perspective of user behavior sequences and recommended objects, improving the multi-scenario training efficiency of the current attention mechanism. In addition, by using scene-related information as input to the feature generation structure, the influence of multiple scenarios is fully injected into the modeling of feature representations, thereby enhancing the understanding of user recommendation needs. The recommendation model proposed in the embodiments of the present disclosure can be embedded in any appropriate system or platform to determine or assist in determining recommendation predictions for users.

[0105] Figure 5 shows a schematic diagram of a model training and application environment 500 in which embodiments of the present disclosure can be implemented. In the environment 500 of Figure 5 , the model is generally shown to involve different stages, including a training stage 502 and an application stage 506. After the training stage is completed, there may also be a testing stage, which is not shown in the figure.

[0106] In the training phase 502, the model training system 510 is configured to perform training of the model 505 using the training data set 512. The model 505 may be, for example, the recommendation model 200 in FIG2. At the beginning of the training, the model may have initial parameter values. The training process is to update the parameter values ​​of the model 505 to the desired values ​​based on the training data.

[0107] In the application phase 506, the obtained model 505 has trained parameter values ​​and can be provided to the model application system 530 for use. In the application phase 506, the model 505 can be used to process the corresponding target input 532 in the actual scenario and provide the corresponding target output 534. The model application system 530 can be configured to implement the recommendation system 120 of Figure 1.

[0108] In Figure 5 , the model training system 510 and the model application system 530 may include any computing system with computing capabilities, such as various computing devices / systems, terminal devices, servers, etc. Terminal devices may include any type of mobile, fixed, 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, mainframe computers, edge computing nodes, computing devices in cloud environments, etc.

[0109] It should be understood that the components and arrangements in the environment 500 shown in FIG5 are merely examples, and a computing system suitable for implementing the exemplary implementations described in the present disclosure may include one or more different components, other components, and / or different arrangements. For example, although shown as separate, the model training system 510 and the model application system 530 may be integrated into the same system or device. Implementations of the present disclosure are not limited in this respect.

[0110] FIG6 shows a flowchart of a process 600 for object recommendation according to some embodiments of the present disclosure. The process 600 may be implemented, for example, at the recommendation system 120 of FIG1 .

[0111] In block 610 , the recommendation system 120 obtains behavior information related to a behavior sequence performed by a target user on an object and scenario information related to a target scenario for recommending the target object, the behavior sequence including a plurality of behaviors.

[0112] In box 620, the recommendation system 120 uses each of the multiple first feature generation layers corresponding to the multiple behaviors in the trained recommendation model to determine the behavior-specific processing parameter value based on the scenario information, and extracts the behavioral feature representation of the behavior from the behavior information related to the behavior sequence based on the behavior-specific processing parameter value and the shared processing parameter value between the multiple behaviors.

[0113] At block 630 , the recommendation system 120 utilizes a second feature generation layer in the recommendation model to extract object feature representations from the scene information and the object information of the target object.

[0114] At block 640 , the recommendation system 120 determines a predicted result of the target user performing the specified behavior on the target object based on at least the behavior feature representations and the object feature representations of each of the plurality of behaviors.

[0115] In some embodiments, determining a behavior-specific processing parameter value based on the context information includes: for each of a plurality of behaviors, determining a behavior-specific weight and a behavior-specific bias value based on the context information using a corresponding first feature generation layer. In some embodiments, extracting a behavior feature representation of a behavior from behavior information associated with a sequence of behaviors includes: generating a weight for the behavior based on the behavior-specific weight and a shared weight among the plurality of behaviors; and extracting a behavior feature representation of the behavior from the behavior information based on the generated weight and the behavior-specific bias value.

[0116] In some embodiments, extracting object feature representations from scene information and object information includes: extracting multiple intermediate feature representations from the scene information and object information; extracting shared feature representations from the object information; and generating multiple object feature representations for multiple behaviors based on the multiple intermediate feature representations and the shared feature representations.

[0117] In some embodiments, the number of the multiple intermediate feature representations is less than the number of behaviors in the behavior sequence. In some embodiments, generating multiple object feature representations for multiple behaviors includes: sequentially dividing the multiple behaviors in the behavior sequence into multiple behavior groups, each intermediate feature representation being specific to a behavior group, and each behavior group including at least two adjacent behaviors in the behavior sequence or including a single behavior; for each behavior group in the multiple behavior groups, generating an object feature representation for at least two adjacent behaviors or a single behavior in the behavior group based on the intermediate feature representation and shared feature representation corresponding to the behavior group.

[0118] In some embodiments, the object feature representation includes multiple object feature representations for multiple behaviors. In some embodiments, determining a prediction result of a target user performing a specified behavior on a target object includes: utilizing multiple attention blocks corresponding to the multiple behaviors in a recommendation model, applying an attention mechanism to the behavior feature representation corresponding to each behavior and the object feature representation for each behavior, to generate multiple weighted feature representations for the multiple behaviors; and determining a prediction result of the target user performing the specified behavior on the target object based at least on the multiple weighted feature representations.

[0119] In some embodiments, the plurality of attention blocks include a transformer-based attention block, and wherein in each attention block, the object feature representation is determined as a query feature, and the behavior feature representation is determined as a key feature and a value feature.

[0120] In some embodiments, determining the predicted result of the target user performing the specified behavior on the target object also includes: determining the predicted result of the target user performing the specified behavior on the target object based on the object embedding representation corresponding to the object information and the user embedding representation corresponding to the user information of the target user.

[0121] FIG7 shows a block diagram of an apparatus 700 for object recommendation according to some embodiments of the present disclosure. Apparatus 700 may be implemented as or included in the recommendation system 120 of FIG1 . Each module / component in apparatus 700 may be implemented by hardware, software, firmware, or any combination thereof.

[0122] As shown in the figure, the device 700 includes an information acquisition module 710, which is configured to obtain behavior information related to a sequence of behaviors performed by a target user on an object and scenario information related to a target scenario for recommending the target object, wherein the behavior sequence includes multiple behaviors. The device 700 also includes a first feature generation module 720, which is configured to use each of multiple first feature generation layers corresponding to the multiple behaviors in the trained recommendation model to determine a behavior-specific processing parameter value based on the scenario information, and extract a behavior feature representation of the behavior from the behavior information related to the behavior sequence based on the behavior-specific processing parameter value and the shared processing parameter value between the multiple behaviors.

[0123] The apparatus 700 further includes a second feature generation module 730 configured to extract an object feature representation from the scene information and the object information of the target object using the second feature generation layer of the recommendation model. The apparatus 700 further includes a prediction determination module 740 configured to determine a prediction result of the target user performing a specified behavior on the target object based on at least the behavior feature representation and the object feature representation of each of the plurality of behaviors.

[0124] In some embodiments, the first feature generation module 720 is further configured to: for each of the multiple behaviors, determine, using the corresponding first feature generation layer, a behavior-specific weight and a behavior-specific bias value based on the scene information. In some embodiments, extracting a behavior feature representation of a behavior from behavior information associated with a behavior sequence includes: generating a weight for the behavior based on the behavior-specific weight and a shared weight among the multiple behaviors; and extracting a behavior feature representation of the behavior from the behavior information based on the generated weight and the behavior-specific bias value.

[0125] In some embodiments, the second feature generation module 730 is further configured to: extract multiple intermediate feature representations from the scene information and the object information; extract shared feature representations from the object information; and generate multiple object feature representations for multiple behaviors based on the multiple intermediate feature representations and the shared feature representations.

[0126] In some embodiments, the number of the plurality of intermediate feature representations is less than the number of behaviors in the behavior sequence. In some embodiments, the second feature generation module 730 is further configured to: sequentially divide the plurality of behaviors in the behavior sequence into a plurality of behavior groups, wherein each intermediate feature representation is specific to a behavior group, and each behavior group includes at least two adjacent behaviors in the behavior sequence or includes a single behavior; for each behavior group in the plurality of behavior groups, based on the intermediate feature representation and the shared feature representation corresponding to the behavior group, generate an object feature representation for at least two adjacent behaviors or a single behavior in the behavior group.

[0127] In some embodiments, the object feature representation includes multiple object feature representations for multiple behaviors. In some embodiments, the prediction result module 740 is further configured to: utilize multiple attention blocks corresponding to the multiple behaviors in the recommendation model, apply an attention mechanism to the behavior feature representation corresponding to each behavior and the object feature representation for each behavior, generate multiple weighted feature representations for the multiple behaviors; and determine, based at least on the multiple weighted feature representations, a prediction result of the target user performing the specified behavior on the target object.

[0128] In some embodiments, the plurality of attention blocks include a transformer-based attention block, and wherein in each attention block, the object feature representation is determined as a query feature, and the behavior feature representation is determined as a key feature and a value feature.

[0129] In some embodiments, the prediction result module 740 is further configured to: determine a prediction result of the target user performing a specified behavior on the target object based on the object embedding representation corresponding to the object information and the user embedding representation corresponding to the user information of the target user.

[0130] FIG8 illustrates a block diagram of an electronic device 800 in which one or more embodiments of the present disclosure may be implemented. It should be understood that the electronic device 800 shown in FIG8 is merely exemplary and should not be construed as limiting the functionality and scope of the embodiments described herein. The electronic device 800 shown in FIG8 may be used to implement the recommendation 110 of FIG1 . The electronic device 800 may include or be implemented as the apparatus 500 of FIG5 .

[0131] As shown in FIG8 , electronic device 800 is in the form of a general-purpose computing device. Components of electronic device 800 may include, but are not limited to, one or more processors or processing units 810, memory 820, storage device 830, one or more communication units 840, one or more input devices 850, and one or more output devices 860. Processing unit 810 may be a real or virtual processor and is capable of performing various processes according to programs stored in memory 820. In a multi-processor system, multiple processing units execute computer-executable instructions in parallel to enhance the parallel processing capabilities of electronic device 800.

[0132] The electronic device 800 typically includes a plurality of computer storage media. Such media can be any available media accessible to the electronic device 800, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 820 can be a volatile memory (e.g., registers, cache, random access memory (RAM)), a non-volatile memory (e.g., read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory), or some combination thereof. The storage device 830 can be a removable or non-removable medium and can include a machine-readable medium, such as a flash drive, a disk, or any other medium that can be used to store information and / or data (e.g., training data for training) and can be accessed within the electronic device 800.

[0133] The electronic device 800 may further include additional removable / non-removable, volatile / non-volatile storage media. Although not shown in FIG8 , a disk drive for reading from or writing to a removable, non-volatile disk (e.g., a “floppy disk”) and an optical drive for reading from or writing to a removable, non-volatile optical disk may be provided. In these cases, each drive may be connected to a bus (not shown) by one or more data media interfaces. The memory 820 may include a computer program product 825 having one or more program modules configured to perform various methods or actions of various embodiments of the present disclosure.

[0134] The communication unit 840 enables communication with other electronic devices via a communication medium. Additionally, the functions of the components of the electronic device 800 can be implemented in a single computing cluster or multiple computing machines that can communicate via a communication connection. Thus, the electronic device 800 can operate in a networked environment using a logical connection with one or more other servers, a network personal computer (PC), or another network node.

[0135] The input device 850 may be one or more input devices, such as a mouse, keyboard, or trackball. The output device 860 may be one or more output devices, such as a display, a speaker, or a printer. The electronic device 800 may also communicate with one or more external devices (not shown) via the communication unit 840 as needed, such as a storage device, a display device, or the like, with one or more devices that allow a user to interact with the electronic device 800, or with any device that allows the electronic device 800 to communicate with one or more other electronic devices (e.g., a network card, a modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).

[0136] According to an exemplary implementation of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is also provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, and the computer-executable instructions are executed by a processor to implement the method described above.

[0137] Various aspects of the present disclosure are described herein with reference to flowcharts and / or block diagrams of methods, apparatuses, devices, and computer program products implemented according to the present disclosure. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer-readable program instructions.

[0138] 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 device, thereby producing a machine, such that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device is generated that implements the functions / actions specified in one or more blocks in the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium, where these instructions cause the computer, programmable data processing device, and / or other device to operate in a specific manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing various aspects of the functions / actions specified in one or more blocks in the flowchart and / or block diagram.

[0139] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0140] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the systems, methods and computer program products according to multiple implementations of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a part for a module, program segment or instruction, and a part for a module, program segment or instruction comprises one or more executable instructions for realizing the logical function of the specification. In some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be realized by a special hardware-based system that performs the function or action of the specification, or can be realized by a combination of special hardware and computer instructions.

[0141] While various implementations of the present disclosure have been described above, the foregoing description is intended to be illustrative, not exhaustive, and not 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 selected to best explain the principles of the implementations, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the various implementations disclosed herein.

Claims

1. A method for object recommendation, comprising: Obtaining behavior information related to a behavior sequence performed by a target user on an object and scene information related to a target scene for recommending the target object, wherein the behavior sequence includes a plurality of behaviors; Using each of the first feature generation layers in the trained recommendation model corresponding to the plurality of behaviors, Based on the scenario information, respectively determining processing parameter values ​​specific to the behavior, and extracting a behavior feature representation of the behavior from the behavior information associated with the behavior sequence based on the processing parameter value specific to the behavior and the shared processing parameter value among the plurality of behaviors; Extracting object feature representation from the scene information and the object information of the target object using a second feature generation layer in the recommendation model; as well as A prediction result of the target user performing a specified behavior on the target object is determined based at least on the behavior feature representations of each of the multiple behaviors and the object feature representation.

2. The method of claim 1 , wherein determining a processing parameter value specific to the behavior based on the context information comprises: For each of the plurality of behaviors, using the corresponding first feature generation layer, determining a weight specific to the behavior and a bias value specific to the behavior based on the scene information; And wherein extracting the behavior feature representation of the behavior from the behavior information related to the behavior sequence comprises: generating weights for the behaviors based on the weights specific to the behaviors and the shared weights among the plurality of behaviors; as well as Based on the generated weights and a bias value specific to the behavior, a behavior feature representation of the behavior is extracted from the behavior information.

3. The method according to claim 1, wherein extracting the object feature representation from the scene information and the object information comprises: extracting a plurality of intermediate feature representations from the scene information and the object information; extracting a shared feature representation from the object information; as well as Based on the multiple intermediate feature representations and the shared feature representation, a plurality of object feature representations for the multiple behaviors are generated.

4. The method according to claim 3, wherein the number of the plurality of intermediate feature representations is less than the number of behaviors in the behavior sequence; and wherein generating a plurality of object feature representations for the plurality of behaviors comprises: Sequentially dividing the plurality of behaviors in the behavior sequence into a plurality of behavior groups, each intermediate feature representation being specific to a behavior group, and each behavior group including at least two adjacent behaviors in the behavior sequence or including a single behavior; For each behavior group in the multiple behavior groups, an object feature representation for at least two adjacent behaviors or a single behavior in the behavior group is generated based on the intermediate feature representation corresponding to the behavior group and the shared feature representation.

5. The method according to claim 1, wherein the object feature representation comprises a plurality of object feature representations for the plurality of behaviors, and wherein determining a prediction result that the target user performs a specified behavior on the target object comprises: Using a plurality of attention blocks in the recommendation model corresponding to the plurality of behaviors respectively, by applying an attention mechanism to the behavior feature representation corresponding to each of the behaviors and the object feature representation for each of the behaviors, a plurality of weighted feature representations for the plurality of behaviors are generated; as well as Based at least on the multiple weighted feature representations, a prediction result of the target user performing a specified behavior on the target object is determined.

6. The method of claim 5, wherein the plurality of attention blocks comprises a transformer-based attention block, and wherein in each attention block, the object feature representation is determined as a query feature, and the behavior feature representation is determined as a key feature and a value feature.

7. The method according to claim 1, wherein determining the prediction result of the target user performing a specified behavior on the target object further comprises: A prediction result of the target user performing a specified behavior on the target object is also determined based on the object embedding representation corresponding to the object information and the user embedding representation corresponding to the user information of the target user.

8. A device for object recommendation, comprising: an information obtaining module configured to obtain behavior information related to a behavior sequence performed by a target user on an object and scene information related to a target scene for recommending the target object, wherein the behavior sequence includes a plurality of behaviors; A first feature generation module is configured to use each first feature generation layer of a plurality of first feature generation layers respectively corresponding to the plurality of behaviors in the trained recommendation model, Based on the scenario information, respectively determining processing parameter values ​​specific to the behavior, and extracting a behavior feature representation of the behavior from the behavior information associated with the behavior sequence based on the processing parameter value specific to the behavior and the shared processing parameter value among the plurality of behaviors; A second feature generation module is configured to extract object feature representation from the scene information and the object information of the target object using a second feature generation layer in the recommendation model; as well as The prediction determination module is configured to determine a prediction result of the target user performing a specified behavior on the target object based at least on the behavior feature representations of each of the multiple behaviors and the object feature representation.

9. An electronic device, comprising: at least one processing unit; as well as At least one memory, the at least one memory being 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.

10. A computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the method according to any one of claims 1 to 7 when executed by a processor.

11. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images