Method, device, electronic device and storage medium for determining recommended object

Abstract

The present disclosure relates to a method, device, electronic device and storage medium for determining a recommended object, the method comprising: in response to an object recommendation request of a target user account, obtaining attribute information, current scene information and a historical interaction information sequence of the target user account; fusing the current scene information with the attribute information and the historical interaction information sequence respectively to obtain a first fusion feature and a second fusion feature; determining the degree of similarity corresponding to each candidate recommended object based on the first fusion feature and the second fusion feature; determining the target candidate recommended object based on the degree of similarity corresponding to each candidate recommended object. The present disclosure can accurately recall without a recall model trained separately based on the current scene, and thus does not require a recall model corresponding to each scene to be trained separately for different scenes, thereby improving the recall efficiency and further improving the efficiency of determining the recommended object.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a method, device, electronic device and storage medium for determining a recommendation object. Background Art

[0002] The recommendation process of the recommendation system generally includes three stages: recall, coarse sorting and fine sorting. Recall, as the first stage of the entire system, mainly learns the feature representation of user accounts and candidate recommendation objects, and based on the feature representation, quickly locates and recalls the candidate recommendation objects that the user account may be interested in from the candidate recommendation object pool. These candidate recommendation objects that may be of interest will be further subjected to coarse sorting and fine sorting to determine the candidate recommendation objects that are finally recommended to the user account.

[0003] In actual applications, the data of a user account may involve multiple scenarios (for example, the main version and the express version of the same application are different scenarios). In related technologies, the recommendation system is based on a recall model trained separately for each scenario when recalling candidate recommendation objects (that is, a deep semantic matching model, also known as a dual-tower model). As a result, it is unable to quickly capture the user's interest preferences in different scenarios, which reduces the recall efficiency and, in turn, the recommendation efficiency. Summary of the invention

[0004] The present disclosure provides a method, device, electronic device and storage medium for determining a recommendation object, so as to at least solve the problem that the related art cannot quickly capture the user's interest preferences in different scenarios, thereby reducing the recall efficiency and recommendation efficiency. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a method for determining a recommendation object is provided, comprising:

[0006] In response to an object recommendation request of a target user account, acquiring attribute information, current scene information, and a historical interaction information sequence of the target user account;

[0007] Fusing the current scene information with the attribute information and the historical interaction information sequence respectively to obtain a first fused feature and a second fused feature;

[0008] Determining the similarity level corresponding to each candidate recommendation object according to the first fusion feature and the second fusion feature;

[0009] According to the degree of similarity corresponding to each of the candidate recommendation objects, a target candidate recommendation object corresponding to the object recommendation request is determined.

[0010] In an exemplary implementation, determining the similarity level corresponding to each candidate recommendation object according to the first fusion feature and the second fusion feature includes:

[0011] Determining a first feature representation of the target user account according to the first fused feature and the second fused feature;

[0012] Determining a second feature representation of each of the candidate recommendation objects according to the recommendation object information of each of the candidate recommendation objects;

[0013] Determine the degree of similarity between the first feature representation and the second feature representation of each of the candidate recommendation objects, and obtain the degree of similarity corresponding to each of the candidate recommendation objects.

[0014] In an exemplary embodiment, the fusing the current scene information with the attribute information and the historical interaction information sequence to obtain a first fused feature and a second fused feature includes:

[0015] Inputting the current scene information, the attribute information and the historical interaction information sequence into a first feature representation network of a recommendation object determination model, encoding the current scene information, the attribute information and the historical interaction information sequence respectively through an encoding network of the first feature representation network to obtain a current scene encoding, an attribute encoding and a historical interaction encoding sequence;

[0016] Inputting the current scene code and the attribute code into a first feature fusion network of the first feature representation network, performing feature extraction on the current scene code and the attribute code respectively through the first feature fusion network, and determining the first fusion feature according to the product of the features extracted from the feature extraction;

[0017] The current scene code and the historical interaction code sequence are input into the second feature fusion network of the first feature representation network, and the second feature fusion network is used to extract features from the current scene code and the historical interaction code sequence respectively, and the second fusion feature is determined according to the product of the features extracted.

[0018] In an exemplary embodiment, the first feature fusion network includes a plurality of cascaded first feature cross units, and the second feature fusion network is used to extract features from the current scene code and the historical interaction code sequence, and the second fusion feature is determined according to the product of the features extracted, including:

[0019] Inputting the current scene code and the attribute code into a first-level first feature cross unit, performing feature extraction on the current scene code and the attribute code respectively through the first-level first feature cross unit, and taking the product of the features obtained by the feature extraction as the output feature of the first-level first feature cross unit;

[0020] For the first feature intersection unit of the current level among the remaining first feature intersection units, the output feature of the first feature intersection unit of the previous level and the current scene code are input to the first feature intersection unit of the current level, and the output feature and the current scene code are respectively extracted by the first feature intersection unit of the current level, and the product of the features obtained by the feature extraction is used as the output feature of the first feature intersection unit of the current level;

[0021] The output feature of the last level of the first feature intersection unit in the cascaded multiple first feature intersection units is determined as the first fusion feature.

[0022] In an exemplary embodiment, the second feature fusion network includes a second feature intersection unit and a first self-attention network; the second feature fusion network is used to extract features of the current scene code and the historical interaction code sequence, and the second fusion feature is determined according to the product of the features extracted, including:

[0023] For each historical interaction code in the historical interaction code sequence, the historical interaction code and the current scene code are input into the second feature cross unit, and the second feature cross unit performs feature extraction on the historical interaction code and the current scene code respectively, and uses the product of the features obtained by the feature extraction as the fusion feature corresponding to the historical interaction code, so as to obtain a fusion feature sequence corresponding to the historical interaction code sequence;

[0024] Inputting the fused feature sequence into the first self-attention network, and determining the weight corresponding to each fused feature in the fused feature sequence based on the self-attention mechanism;

[0025] A weighted sum is performed on each of the fused features and the corresponding weights to obtain the second fused feature.

[0026] In an exemplary implementation, determining a first feature representation of the target user account according to the first fused feature and the second fused feature includes:

[0027] Inputting the current scene code into the scene feature extraction network of the first feature representation network to obtain current scene features;

[0028] The first fused feature, the second fused feature, and the current scene feature are concatenated to obtain a first feature representation of the target user account.

[0029] In an exemplary embodiment, determining the second feature representation of each candidate recommendation object according to the recommendation object information of each candidate recommendation object includes:

[0030] Acquire recommended object information of the candidate recommended object; the recommended object information includes recommended object attribute information and a sub-object sequence, and the sub-object sequence includes a plurality of sub-objects associated with the candidate recommended object;

[0031] Inputting the recommended object attribute information and the sub-object sequence into a second feature representation network of the recommended object determination model, and encoding the recommended object attribute information and the sub-object sequence through an encoding network of the second feature representation network to obtain a recommended object code and a sub-object code sequence;

[0032] Inputting the recommended object code into a recommended object feature extraction network of the second feature representation network to obtain recommended object features;

[0033] Inputting the sub-object encoding sequence into the second self-attention network of the second feature representation network, determining the weight corresponding to each sub-object encoding in the sub-object encoding sequence based on the self-attention mechanism; performing weighted summation on each sub-object encoding and the corresponding weight to obtain a sub-object feature;

[0034] The recommended object feature and the sub-object feature are concatenated to obtain a second feature representation of the candidate recommended object.

[0035] In an exemplary embodiment, the method further comprises:

[0036] Acquire training samples and label information corresponding to the training samples; the training samples include sample attribute information, sample scenario information, and sample historical interaction information sequence corresponding to the sample user account, and recommended object information corresponding to the sample recommended object; the label information indicates whether the sample user account is interested in the sample recommended object;

[0037] Inputting the sample attribute information, sample scene information and sample historical interaction information sequence into an initial first feature representation network of a preset neural network model to obtain a predicted first feature representation output by the initial first feature representation network;

[0038] Inputting the recommended object information into the initial second feature representation network of the preset neural network model to obtain the predicted second feature representation output by the initial second feature representation network;

[0039] Determining a similarity between the predicted first feature representation and the predicted second feature representation to obtain a predicted similarity;

[0040] A loss value is determined according to the difference between the predicted similarity and the label information, and model parameters of the preset neural network model are adjusted based on the loss value. The adjusted preset neural network model is continuously iteratively trained until a preset training end condition is met, thereby obtaining the recommended object determination model.

[0041] According to a second aspect of an embodiment of the present disclosure, there is provided a device for determining a recommendation object, including:

[0042] A first acquisition unit is configured to execute in response to an object recommendation request of a target user account, and acquire attribute information, current scene information, and a historical interaction information sequence of the target user account;

[0043] a fusion unit configured to fuse the current scene information with the attribute information and the historical interaction information sequence respectively to obtain a first fusion feature and a second fusion feature;

[0044] a similarity determination unit, configured to determine the similarity corresponding to each candidate recommendation object according to the first fusion feature and the second fusion feature;

[0045] The target recommendation object determination unit is configured to determine the target candidate recommendation object corresponding to the object recommendation request according to the similarity degree corresponding to each candidate recommendation object.

[0046] In an exemplary embodiment, the similarity determination unit includes:

[0047] A first feature representation determining unit is configured to determine a first feature representation of the target user account according to the first fused feature and the second fused feature;

[0048] A second feature representation determining unit is configured to determine a second feature representation of each candidate recommendation object according to the recommendation object information of each candidate recommendation object;

[0049] The similarity determination subunit is configured to determine the similarity between the first feature representation and the second feature representation of each candidate recommendation object, and obtain the similarity corresponding to each candidate recommendation object.

[0050] In an exemplary embodiment, the fusion unit comprises:

[0051] A first encoding unit is configured to execute inputting the current scene information, the attribute information and the historical interaction information sequence into a first feature representation network of a recommendation object determination model, and encoding the current scene information, the attribute information and the historical interaction information sequence through an encoding network of the first feature representation network to obtain a current scene code, an attribute code and a historical interaction code sequence;

[0052] A first feature fusion unit is configured to execute a first feature fusion network that inputs the current scene code and the attribute code into the first feature representation network, extract features from the current scene code and the attribute code respectively through the first feature fusion network, and determine the first fusion feature according to the product of features obtained by the feature extraction;

[0053] The second feature fusion unit is configured to execute a second feature fusion network that inputs the current scene code and the historical interaction code sequence into the first feature representation network, extract features from the current scene code and the historical interaction code sequence respectively through the second feature fusion network, and determine the second fusion feature according to the product of the features obtained by the feature extraction.

[0054] In an exemplary embodiment, the first feature fusion network includes a plurality of first feature crossover units connected in cascade;

[0055] The first feature fusion unit is specifically configured to execute:

[0056] Inputting the current scene code and the attribute code into a first-level first feature cross unit, performing feature extraction on the current scene code and the attribute code respectively through the first-level first feature cross unit, and taking the product of the features obtained by the feature extraction as the output feature of the first-level first feature cross unit;

[0057] For the first feature intersection unit of the current level among the remaining first feature intersection units, the output feature of the first feature intersection unit of the previous level and the current scene code are input to the first feature intersection unit of the current level, and the output feature and the current scene code are respectively extracted by the first feature intersection unit of the current level, and the product of the features obtained by the feature extraction is used as the output feature of the first feature intersection unit of the current level;

[0058] The output feature of the last level of the first feature intersection unit in the cascaded multiple first feature intersection units is determined as the first fusion feature.

[0059] In an exemplary embodiment, the second feature fusion network includes a second feature intersection unit and a first self-attention network;

[0060] The second feature fusion unit is specifically configured to execute:

[0061] For each historical interaction code in the historical interaction code sequence, the historical interaction code and the current scene code are input into the second feature cross unit, and the second feature cross unit performs feature extraction on the historical interaction code and the current scene code respectively, and uses the product of the features obtained by the feature extraction as the fusion feature corresponding to the historical interaction code, so as to obtain a fusion feature sequence corresponding to the historical interaction code sequence;

[0062] The fused feature sequence is input into the first self-attention network, and the weight corresponding to each fused feature in the fused feature sequence is determined based on the self-attention mechanism; and each fused feature and the corresponding weight are weighted summed to obtain the second fused feature.

[0063] In an exemplary embodiment, the first feature representation unit includes:

[0064] a current scene feature extraction unit, configured to execute a scene feature extraction network that inputs the current scene code into the first feature representation network to obtain a current scene feature;

[0065] The first splicing unit is configured to splice the first fusion feature, the second fusion feature and the current scene feature to obtain a first feature representation of the target user account.

[0066] In an exemplary embodiment, the second feature representation unit includes:

[0067] A second acquisition unit is configured to acquire recommended object information of the candidate recommended object; the recommended object information includes recommended object attribute information and a sub-object sequence, and the sub-object sequence includes a plurality of sub-objects associated with the candidate recommended object;

[0068] A second encoding unit is configured to input the recommended object attribute information and the sub-object sequence into a second feature representation network of the recommended object determination model, and encode the recommended object attribute information and the sub-object sequence through an encoding network of the second feature representation network to obtain a recommended object code and a sub-object code sequence;

[0069] a recommended object feature extraction unit, configured to execute a recommended object feature extraction network that inputs the recommended object encoding into the second feature representation network to obtain recommended object features;

[0070] The sub-object feature determination unit is configured to execute the second self-attention network of the second feature representation network by inputting the sub-object code sequence into the second feature representation network, determine the weight corresponding to each sub-object code in the sub-object code sequence based on the self-attention mechanism; perform weighted summation on each sub-object code and the corresponding weight to obtain the sub-object feature;

[0071] The second concatenation unit is configured to concatenate the recommended object feature and the sub-object feature to obtain a second feature representation of the candidate recommended object.

[0072] In an exemplary embodiment, the apparatus further includes a model training unit; the model training unit is configured to execute:

[0073] Acquire training samples and label information corresponding to the training samples; the training samples include sample attribute information, sample scenario information, and sample historical interaction information sequence corresponding to the sample user account, and recommended object information corresponding to the sample recommended object; the label information indicates whether the sample user account is interested in the sample recommended object;

[0074] Inputting the sample attribute information, sample scene information and sample historical interaction information sequence into an initial first feature representation network of a preset neural network model to obtain a predicted first feature representation output by the initial first feature representation network;

[0075] Inputting the recommended object information into the initial second feature representation network of the preset neural network model to obtain the predicted second feature representation output by the initial second feature representation network;

[0076] Determining a similarity between the predicted first feature representation and the predicted second feature representation to obtain a predicted similarity;

[0077] A loss value is determined according to the difference between the predicted similarity and the label information, and model parameters of the preset neural network model are adjusted based on the loss value. The adjusted preset neural network model is continuously iteratively trained until a preset training end condition is met, thereby obtaining the recommended object determination model.

[0078] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0079] processor;

[0080] a memory for storing instructions executable by the processor;

[0081] The processor is configured to execute the instructions to implement the recommended object determination method of the first aspect.

[0082] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method for determining a recommended object of the first aspect.

[0083] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, including a computer program / instruction, which, when executed by a processor, implements the method for determining a recommended object according to the first aspect.

[0084] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0085] When a target user account initiates an object recommendation request, the attribute information, current scene information and historical interaction information sequence of the target user account are obtained, and the current scene information is fused with the attribute information and the historical interaction information sequence to obtain a first fusion feature and a second fusion feature, and then the similarity corresponding to each candidate recommendation object is determined based on the first fusion feature and the second fusion feature, and the target candidate recommendation object is determined based on the similarity, so that the current scene information is integrated into the first fusion feature and the second fusion feature, and accurate recall can be achieved without the need for a recall model trained separately based on the current scene. There is no need to train a recall model corresponding to each scene separately for different scenes, which improves the recall efficiency and thereby improves the efficiency of determining the recommended objects.

[0086] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0088] Figure 1 is a schematic diagram of an application environment of a method for determining a recommendation object according to an exemplary embodiment;

[0089] Figure 2 is a flow chart of a method for determining a recommendation object according to an exemplary embodiment;

[0090] Figure 3 is a schematic diagram of a model structure of a recommendation object determination model according to an exemplary embodiment;

[0091] Figure 4 is a schematic diagram of a model structure of a feature cross unit according to an exemplary embodiment;

[0092] Figure 5is a flowchart showing a method of determining a first feature representation of a target user account according to a first fused feature and a second fused feature according to an exemplary embodiment;

[0093] Figure 6 is a flowchart of determining a second feature representation of a candidate recommendation object according to an exemplary embodiment;

[0094] Figure 7 is a structural block diagram of a device for determining a recommended object according to an exemplary embodiment;

[0095] Figure 8 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0096] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0097] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0098] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0099] In actual applications, for the same application, user interactions will occur in multiple different scenarios. For example, the interactions in the main version application and the interactions in the express version application are interactions in different scenarios. The interactions in different page display forms (such as single-column display and multi-column display) of the content aggregation page are also interactions in different scenarios. The interactions in different presentation forms of the content in the content aggregation page (such as static picture form and dynamic video stream form) are also interactions in different scenarios, etc. There are differences in the data and interaction information of user accounts between different scenarios. In the related art, when recalling candidate recommendation objects, the recommendation system is based on the recall model trained separately for each scenario (such as the dual-tower model), which cannot capture the user's interest preferences in different scenarios, reducing the recall efficiency and thus reducing the recommendation efficiency.

[0100] In view of this, an embodiment of the present disclosure provides a method for determining a recommended object. When a target user account initiates an object recommendation request, the method obtains attribute information, current scene information and a historical interaction information sequence of the target user account, and fuses the current scene information with the attribute information and the historical interaction information sequence to obtain a first fusion feature and a second fusion feature, respectively. Then, based on the first fusion feature and the second fusion feature, the degree of similarity corresponding to each candidate recommended object is determined, and based on the degree of similarity, the target candidate recommended object is determined, thereby integrating the current scene information into the first fusion feature and the second fusion feature, and accurately recalling the object without the need for a recall model trained separately based on the current scene. Furthermore, there is no need to train a recall model corresponding to each scene separately for different scenes, thereby improving the recall efficiency and thereby improving the efficiency of determining the recommended object.

[0101] See also Figure 1 , which is a schematic diagram of an application environment of a method for determining a recommendation object according to an exemplary embodiment. The application environment may include a terminal 110 and a recommendation system 120.

[0102] The terminal 110 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc., but is not limited thereto. The terminal 110 may be installed with client software such as an application (Application, referred to as App) that provides a recommended object display function. The application may be an independent application or a subprogram in the application. Exemplarily, the application may include a game application, a live broadcast application, etc. The user of the terminal 110 may log in to the application through a pre-registered user account.

[0103] The recommendation system 120 may be a server that provides recommendation services for applications in the terminal 110. The recommendation object is associated with the specific business of the application. For example, for a live broadcast application, the recommendation object may be live broadcast, that is, the recommendation system may recommend live broadcast to the terminal 110. The recommendation system 120 and the terminal 110 may be connected via a wired network or a wireless network.

[0104] Specifically, Figure 1As shown, the recommendation system 120 may include a candidate recommendation object pool 121, a recall server 122, a rough ranking server 123, and a fine ranking server 124. The recall server 122 may receive an object recommendation request sent by the terminal 110, and in response to the object recommendation request, obtain the attribute information, current scene information, and historical interaction information sequence of the target user account corresponding to the terminal 110, fuse the current scene information with the attribute information and the historical interaction information sequence to obtain a first fusion feature and a second fusion feature, and then obtain a first feature representation of the target user account based on the first fusion feature, the second fusion feature, and the current scene information, and determine the degree of similarity between the first feature representation and the second feature representation of each candidate recommendation object in the candidate recommendation object pool 121, based on the first fusion feature, the second fusion feature, and the current scene information. The target candidate recommendation object is selected from the candidate recommendation object pool as the recall result of the corresponding object recommendation request according to the degree of similarity, and the target candidate recommendation object is sent to the rough ranking server 123, and the rough ranking server 123 further screens the target candidate recommendation object in combination with the rough ranking strategy to obtain a rough ranking result, and the rough ranking result is sent by the rough ranking server 123 to the fine ranking server 124, and the fine ranking server 124 further screens the rough ranking result in combination with the fine ranking strategy to obtain the final candidate recommendation object to be recommended (i.e., the fine ranking result), and returns the candidate recommendation object to be recommended to the terminal 110, which is displayed by the terminal 110.

[0105] In a specific application scenario, the recall server 122 may store a trained recommendation object determination model, and call the recommendation object determination model to recall the target candidate recommendation object from the candidate recommendation object pool when performing a recall in response to an object recommendation request.

[0106] It should be noted that the server mentioned in the embodiments of the present disclosure can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, as well as big data and artificial intelligence platforms.

[0107] Figure 2 is a flowchart of a method for determining a recommended object according to an exemplary embodiment. Figure 2 As shown, this method is applied to Figure 1 Taking the recall server in as an example, the following steps are included.

[0108] In step S21, in response to an object recommendation request of a target user account, attribute information, current scene information, and a historical interaction information sequence of the target user account are obtained.

[0109] Among them, the target user account is the user account that initiates the object recommendation request. The attribute information of the target user account may include portrait information such as name, age, gender, and other information that can characterize the attributes of the user account. The current scene information of the target user account may include context information and partial portrait information. The context information may include the time information and location information of initiating the object recommendation request, the page information that triggers the object recommendation request (which may include page identification, page display form, presentation form of content in the page, etc.), the version information of the application, etc., which can characterize the virtual environment in which the target user account is located when the object recommendation request is initiated; partial portrait information may be partial portrait information extracted from the attribute information of the target user account, and the specific amount of extracted data can be set according to actual experience.

[0110] The historical interaction information sequence of the target user account may include the interaction information of the target user account in a preset time period before the current time. Taking live broadcast recommendation as an example, the interaction information in the historical interaction information sequence may include relevant information of the live broadcast room that the target user account has entered in the preset time period before the current time, such as the live broadcast room identifier, anchor identifier, etc. The preset time period may be set according to actual needs, such as one week or one month, etc.

[0111] In step S23, the current scene information is fused with the attribute information and the historical interaction information sequence respectively to obtain a first fused feature and a second fused feature.

[0112] Specifically, the current scene information and the attribute information are fused to obtain a first fusion feature, which can characterize the interest preference of the target user account in the current scene; the current scene information and the historical interaction information sequence are fused to obtain a second fusion feature, which can characterize the behavior preference of the target user account in the current scene.

[0113] In step S25, the similarity level corresponding to each candidate recommendation object is determined according to the first fusion feature and the second fusion feature.

[0114] The similarity level corresponding to the candidate recommendation object can represent the interest level of the target user account in the candidate recommendation object.

[0115] In a specific implementation, the similarity can be obtained by calculating the similarity between the first feature representation corresponding to the target user account and the second feature representation corresponding to the candidate recommendation object. The similarity can be cosine similarity or other metrics that can characterize the similarity between two vectors. The specific characterization method of the similarity is not limited in the embodiment of the present disclosure. Based on this, the above step S25 can include the following steps when implemented:

[0116] Determining a first feature representation of the target user account according to the first fused feature and the second fused feature;

[0117] Determining a second feature representation of each of the candidate recommendation objects according to the recommendation object information of each of the candidate recommendation objects;

[0118] Determine the degree of similarity between the first feature representation and the second feature representation of each of the candidate recommendation objects, and obtain the degree of similarity corresponding to each of the candidate recommendation objects.

[0119] Specifically, the first fused feature and the second fused feature may be concatenated to obtain a first feature representation of the target user account.

[0120] The second feature representation of the candidate recommendation object is obtained according to the recommendation object information of the candidate recommendation object. The recommendation object information may include recommendation object attribute information and a sub-object sequence, and the sub-object sequence includes multiple sub-objects associated with the candidate recommendation object. Taking live broadcast recommendation as an example, the multiple sub-objects associated with the candidate live broadcast may be live broadcast items (such as clothing, food, etc.) associated with the candidate live broadcast, and the recommendation object attribute information may include anchor information, live broadcast item information (such as live broadcast item name, category, etc.) and provider information of the live broadcast item.

[0121] In the above embodiment, the feature representation of the target user account is determined by the first fusion feature and the second fusion feature, so that the feature representation of the user account corresponds to the current scene, and the similarity between the feature representation of the target user account and the feature representation of the candidate recommendation objects is used to accurately characterize the similarity corresponding to each candidate recommendation object, thereby improving the accuracy of the similarity, which is beneficial to improving the accuracy of determining the target recommendation object.

[0122] In step S27, the target candidate recommended object corresponding to the object recommendation request is determined according to the degree of similarity corresponding to each of the candidate recommended objects.

[0123] In a specific implementation, candidate recommendation objects with a similarity greater than a preset threshold may be determined as target candidate recommendation objects recalled in response to the object recommendation request. The preset threshold may be set based on actual experience, and the present disclosure does not specifically limit this.

[0124] In another specific implementation, the candidate recommendation objects may be sorted according to the degree of similarity, and a preset number of the largest candidate recommendation objects may be selected according to the sorting result as the target candidate recommendation objects to be recalled in response to the object recommendation request. For example, the candidate recommendation objects may be sorted in descending order according to the degree of similarity, and then a preset number of the candidate recommendation objects ranked first may be selected as the target candidate recommendation objects. The preset number may be set according to actual needs, and the embodiments of the present disclosure do not specifically limit this.

[0125] In the above embodiment, the current scene information is respectively integrated into the attribute information and historical interaction information of the user account, which fully captures the interest preferences and behavioral preferences of the user account in the current scene, so that accurate recall can be achieved without the need for a recall model trained separately based on the current scene. Furthermore, there is no need to train recall models corresponding to different scenes separately, which improves the recall efficiency and further improves the efficiency of determining the recommended objects.

[0126] The method for determining a recommended object in the embodiment of the present disclosure can be implemented based on a trained recommended object determination model. The model structure of the recommended object determination model is as follows: Figure 3 The method may include: a first feature representation network corresponding to a user account, a second feature representation network corresponding to a candidate recommendation object, and a matching layer. The first feature representation network is used to implement a feature representation of the user account to obtain a first feature representation, the second feature representation network is used to implement a feature representation of the recommendation object to obtain a second feature representation, and the matching layer is used to determine the similarity between the first feature representation and the second feature representation.

[0127] The first feature representation network includes an encoding network (i.e., an embedding layer), a first feature fusion network, a second feature fusion network, and a scene feature extraction network. The encoding network is used to encode the input to convert it into an embedding vector. The scene feature extraction network is used to extract features from scene information to obtain scene features. The scene feature extraction network can be a shallow deep neural network (DNN), such as a layer of a DNN network.

[0128] The first feature fusion network is implemented based on Co-Action Unit (feature cross unit). Co-Action means that multiple features are related to each other and jointly affect the final output. The specific structure of Co-Action Unit is as follows: Figure 4As shown, it includes two deep neural networks (DNNs) and a cross layer. In order to distinguish the expressions, the feature cross unit in the first feature fusion network is hereinafter referred to as the first feature cross unit. The inputs of the two DNNs in the first feature cross unit correspond to the attribute information and scene information of the user account respectively, and the cross layer outputs the feature cross result based on the element-wise multiplication according to the outputs of the two DNNs.

[0129] The second feature fusion network is also implemented based on the Co-Action Unit (feature cross unit). Its structure can be seen in Figure 4 In order to distinguish the expressions, the feature cross unit in the second feature fusion network is hereinafter referred to as the second feature cross unit. The inputs of the two DNNs in the second feature cross unit correspond to the historical interaction information and the scene information in the historical interaction information sequence respectively, and the cross layer outputs the feature cross result based on the element-wise multiplication according to the outputs of the two DNNs.

[0130] Based on this, in an exemplary implementation, the above step S23 may include:

[0131] Inputting the current scene information, the attribute information and the historical interaction information sequence into a first feature representation network of a recommendation object determination model, encoding the current scene information, the attribute information and the historical interaction information sequence respectively through an encoding network of the first feature representation network to obtain a current scene encoding, an attribute encoding and a historical interaction encoding sequence;

[0132] Inputting the current scene code and the attribute code into a first feature fusion network of the first feature representation network, performing feature extraction on the current scene code and the attribute code respectively through the first feature fusion network, and determining the first fusion feature according to the product of the features extracted from the feature extraction;

[0133] The current scene code and the historical interaction code sequence are input into the second feature fusion network of the first feature representation network, and the second feature fusion network is used to extract features from the current scene code and the historical interaction code sequence respectively, and the second fusion feature is determined according to the product of the features extracted.

[0134] In the above embodiment, the first feature fusion network of the first feature representation network in the recommendation object determination model realizes the intersection of the attributes of the current scene and the user account, and the second feature fusion network realizes the intersection of the current scene and the user interaction, which is conducive to fully integrating the current scene into the feature representation of the user account.

[0135] In an exemplary embodiment, in order to improve the effect of feature fusion, the first feature fusion network may include a plurality of cascaded feature cross units (in order to distinguish the expressions, the feature cross unit in the first feature fusion network is hereinafter referred to as the first feature cross unit), for example, Figure 3 As shown in , it can include three cascaded first feature cross units. For the first-level first feature cross unit, its input is the attribute code of the user account obtained after encoding processing and the current scene code, and for the remaining first feature cross units, its input is the output feature of the first feature cross unit of the previous level and the current scene code, and the output feature of the last level first feature cross unit is the first fusion feature output by the first feature fusion network.

[0136] Based on this, the above-mentioned extracting features from the current scene code and the attribute code respectively through the first feature fusion network, and determining the first fusion feature according to the product of the features obtained by the feature extraction, when implemented, may include:

[0137] Inputting the current scene code and the attribute code into a first-level first feature cross unit, performing feature extraction on the current scene code and the attribute code respectively through the first-level first feature cross unit, and taking the product of the features obtained by the feature extraction as the output feature of the first-level first feature cross unit;

[0138] For the first feature intersection unit of the current level among the remaining first feature intersection units, the output feature of the first feature intersection unit of the previous level and the current scene code are input to the first feature intersection unit of the current level, and the output feature and the current scene code are respectively extracted by the first feature intersection unit of the current level, and the product of the features obtained by the feature extraction is used as the output feature of the first feature intersection unit of the current level;

[0139] The output feature of the last level of the first feature intersection unit in the cascaded multiple first feature intersection units is determined as the first fusion feature.

[0140] The above embodiment, through the cascade of multiple feature cross-units, fuses the attributes of the user account with the current scene at each level, thereby achieving deep fusion of the attributes of the user account and the current scene, strengthening the current scene information in the feature representation of the user account, and fully capturing the interest preferences of the user account in the current scene, so that the first fused feature is more matched with the current scene, and the recommended objects subsequently recalled based on the first feature representation are more in line with the user's interests in the current scene, thereby improving the accuracy of the recall.

[0141] In order to obtain a second fusion feature containing more information based on the historical interaction information sequence and the current scene information, such as Figure 3 As shown, the second feature fusion network includes a second feature intersection unit and a first self-attention network, wherein the second feature intersection unit is used to realize the intersection of each historical interaction information with the current scene information, and the first self-attention network determines the weight of each historical interaction based on the self-attention mechanism. In order to enhance the ability of the first self-attention network to capture the differences between different scene features, the embodiment of the present disclosure will obtain a fused feature sequence based on the second feature intersection unit as the input of the first self-attention network to better combine the current scene to determine the weight.

[0142] Based on this, in an exemplary embodiment, the above-mentioned extracting features from the current scene code and the historical interaction code sequence respectively through the second feature fusion network, and determining the second fusion feature according to the product of the features obtained by the feature extraction, may include:

[0143] For each historical interaction code in the historical interaction code sequence, the historical interaction code and the current scene code are input into the second feature cross unit, and the second feature cross unit performs feature extraction on the historical interaction code and the current scene code respectively, and uses the product of the features obtained by the feature extraction as the fusion feature corresponding to the historical interaction code, so as to obtain a fusion feature sequence corresponding to the historical interaction code sequence;

[0144] Inputting the fused feature sequence into the first self-attention network, and determining the weight corresponding to each fused feature in the fused feature sequence based on the self-attention mechanism;

[0145] A weighted sum is performed on each of the fused features and the corresponding weights to obtain the second fused feature.

[0146] In the above embodiment, the weights corresponding to each historical interaction are learned through the self-attention mechanism in combination with the current scene information, and then a second fusion feature is obtained by weighted summation based on the weights and the corresponding historical interactions and the fusion features of the current scene. While retaining more user interest information, the interest of historical interactions in the current scene is fully captured, making the second fusion feature more matching and more accurate with the current scene. The recommended objects subsequently recalled based on the first feature representation are more in line with the interests in the current scene, thereby improving the accuracy of the recall.

[0147] In order to enhance the scene information in the first feature representation, in an exemplary embodiment, Figure 5 As shown, in the above embodiment, when determining the first feature representation of the target user account according to the first fusion feature and the second fusion feature, it may include:

[0148] In step S51, the current scene code is input into the scene feature extraction network of the first feature representation network to obtain the current scene feature;

[0149] In step S53, the first fused feature, the second fused feature and the current scene feature are concatenated to obtain a first feature representation of the target user account.

[0150] Based on the above recommendation object determination model, it can be understood that the embodiment of the present disclosure also includes a process for determining the second feature representation of the candidate recommendation object, such as Figure 6 As shown, the above-mentioned determining the second feature representation of each candidate recommendation object according to the recommendation object information of each candidate recommendation object may include:

[0151] In step 61, the recommended object information of the candidate recommended object is obtained; the recommended object information includes recommended object attribute information and a sub-object sequence, and the sub-object sequence includes a plurality of sub-objects associated with the candidate recommended object.

[0152] In step 63, the recommended object attribute information and the sub-object sequence are input into the second feature representation network of the recommended object determination model, and the recommended object attribute information and the sub-object sequence are respectively encoded by the encoding network of the second feature representation network to obtain a recommended object code and a sub-object code sequence;

[0153] In step 65, the recommended object code is input into the recommended object feature extraction network of the second feature representation network to obtain the recommended object feature;

[0154] Among them, the recommended object feature extraction network can be a multi-layer DNN network, such as Figure 3 In the 3-layer DNN network shown in the figure, the recommended object encoding is subjected to feature extraction by the multi-layer DNN network to obtain the recommended object features.

[0155] In step 67, the sub-object code sequence is input into the second self-attention network of the second feature representation network, and the weight corresponding to each sub-object code in the sub-object code sequence is determined based on the self-attention mechanism; each sub-object code and the corresponding weight are weighted summed to obtain the sub-object feature;

[0156] In step 69, the recommended object features and the sub-object features are concatenated to obtain a second feature representation of the candidate recommended object.

[0157] In the above embodiment, the weight of each sub-object is learned through the self-attention mechanism, and then the sub-object features are obtained by weighted summation based on the weight and the corresponding sub-object encoding, so that the sub-object features can retain updated information, thereby improving the accuracy of the second feature representation of the candidate recommended object, which is conducive to improving the accuracy of recall.

[0158] It can be understood that the implementation of the present disclosure may also include a training process for a recommendation object determination model. Specifically, the training of the recommendation object determination model may include the following steps:

[0159] (1) Obtain training samples and label information corresponding to the training samples.

[0160] Among them, the training sample includes sample attribute information, sample scene information and sample historical interaction information sequence corresponding to the sample user account, and recommended object information corresponding to the sample recommended object. The label information represents whether the sample user account is interested in the sample recommended object. Specifically, the label information can be determined based on the interaction information of the sample user account with the sample recommended object. Taking live broadcast recommendation as an example, for example, if the sample user account enters the sample live broadcast, it is considered to be interested in the sample live broadcast, and the corresponding label information can be set to "1". Conversely, if the sample user account does not enter the sample live broadcast, it is considered to be not interested in the sample live broadcast, and the corresponding label information can be set to "0".

[0161] (2) Inputting the sample attribute information, sample scene information and sample historical interaction information sequence into an initial first feature representation network of a preset neural network model to obtain a predicted first feature representation output by the initial first feature representation network.

[0162] (3) Inputting the recommended object information into the initial second feature representation network of the preset neural network model to obtain the predicted second feature representation output by the initial second feature representation network.

[0163] (4) Determine the similarity between the predicted first feature representation and the predicted second feature representation to obtain a predicted similarity.

[0164] Among them, the model structure of the preset neural network model is Figure 3 The model structure shown in , the detailed processing process of sample attribute information, sample scene information and sample historical interaction information sequence in the initial first feature representation network can be found in the description of the related processing of input data by the aforementioned first feature representation network in the embodiment of the present disclosure, and will not be repeated here.

[0165] The detailed processing process of the recommended object information in the initial second feature representation network can be found in the description of the related processing of the input data by the aforementioned second feature representation network in the embodiment of the present disclosure, which will not be repeated here.

[0166] The predicted similarity may be cosine similarity or other metrics that can characterize the similarity between two vectors.

[0167] (5) Determine a loss value according to the difference between the predicted similarity and the label information, and adjust the model parameters of the preset neural network model based on the loss value, and continue to iteratively train the adjusted preset neural network model until the preset training end condition is met, thereby obtaining the recommendation object determination model.

[0168] The loss value may be obtained based on a preset loss function, which may be, but is not limited to, a cross entropy loss function. When adjusting the model parameters of the preset neural network model based on the loss value, gradient descent may be used for back propagation. The preset training end condition may be that the difference between two loss values ​​reaches a preset difference, or the number of iterations reaches a preset number of iterations. Both the preset difference and the preset number of iterations may be set based on actual experience.

[0169] In the above-mentioned embodiment, the scene information is respectively integrated with the attribute information of the user account and the historical interaction information sequence, so that the model can learn the user's interest preferences in different scenarios well, and then the recommendation object determination model obtained based on the training can determine the user's behavior preferences and interest preferences in different scenarios, thereby eliminating the need to train the recall model separately for each scenario, which is conducive to improving the recall efficiency.

[0170] Figure 7 is a structural block diagram of a device for determining a recommended object according to an exemplary embodiment. Figure 7 As shown, the recommended object determination device 700 includes:

[0171] A first acquisition unit 710 is configured to execute, in response to an object recommendation request of a target user account, acquire attribute information, current scene information, and a historical interaction information sequence of the target user account;

[0172] A fusion unit 720 is configured to fuse the current scene information with the attribute information and the historical interaction information sequence to obtain a first fusion feature and a second fusion feature;

[0173] A similarity determination unit 730 is configured to determine the similarity corresponding to each candidate recommendation object according to the first fusion feature and the second fusion feature;

[0174] The target recommendation object determination unit 740 is configured to determine the target candidate recommendation object corresponding to the object recommendation request according to the similarity degree corresponding to each of the candidate recommendation objects.

[0175] In an exemplary implementation, the similarity determination unit 730 includes:

[0176] A first feature representation determining unit is configured to determine a first feature representation of the target user account according to the first fused feature and the second fused feature;

[0177] A second feature representation determining unit is configured to determine a second feature representation of each candidate recommendation object according to the recommendation object information of each candidate recommendation object;

[0178] The similarity determination subunit is configured to determine the similarity between the first feature representation and the second feature representation of each candidate recommendation object, and obtain the similarity corresponding to each candidate recommendation object.

[0179] In an exemplary embodiment, the fusion unit comprises:

[0180] A first encoding unit is configured to execute inputting the current scene information, the attribute information and the historical interaction information sequence into a first feature representation network of a recommendation object determination model, and encoding the current scene information, the attribute information and the historical interaction information sequence through an encoding network of the first feature representation network to obtain a current scene code, an attribute code and a historical interaction code sequence;

[0181] A first feature fusion unit is configured to execute a first feature fusion network that inputs the current scene code and the attribute code into the first feature representation network, extract features from the current scene code and the attribute code respectively through the first feature fusion network, and determine the first fusion feature according to the product of features obtained by the feature extraction;

[0182] The second feature fusion unit is configured to execute a second feature fusion network that inputs the current scene code and the historical interaction code sequence into the first feature representation network, extract features from the current scene code and the historical interaction code sequence respectively through the second feature fusion network, and determine the second fusion feature according to the product of the features obtained by the feature extraction.

[0183] In an exemplary embodiment, the first feature fusion network includes a plurality of first feature crossover units connected in cascade;

[0184] The first feature fusion unit is specifically configured to execute:

[0185] Inputting the current scene code and the attribute code into a first-level first feature cross unit, performing feature extraction on the current scene code and the attribute code respectively through the first-level first feature cross unit, and taking the product of the features obtained by the feature extraction as the output feature of the first-level first feature cross unit;

[0186] For the first feature intersection unit of the current level among the remaining first feature intersection units, the output feature of the first feature intersection unit of the previous level and the current scene code are input to the first feature intersection unit of the current level, and the output feature and the current scene code are respectively extracted by the first feature intersection unit of the current level, and the product of the features obtained by the feature extraction is used as the output feature of the first feature intersection unit of the current level;

[0187] The output feature of the last level of the first feature intersection unit in the cascaded multiple first feature intersection units is determined as the first fusion feature.

[0188] In an exemplary embodiment, the second feature fusion network includes a second feature intersection unit and a first self-attention network;

[0189] The second feature fusion unit is specifically configured to execute:

[0190] For each historical interaction code in the historical interaction code sequence, the historical interaction code and the current scene code are input into the second feature cross unit, and the second feature cross unit performs feature extraction on the historical interaction code and the current scene code respectively, and uses the product of the features obtained by the feature extraction as the fusion feature corresponding to the historical interaction code, so as to obtain a fusion feature sequence corresponding to the historical interaction code sequence;

[0191] The fused feature sequence is input into the first self-attention network, and the weight corresponding to each fused feature in the fused feature sequence is determined based on the self-attention mechanism; and each fused feature and the corresponding weight are weighted summed to obtain the second fused feature.

[0192] In an exemplary embodiment, the first feature representation unit includes:

[0193] a current scene feature extraction unit, configured to execute a scene feature extraction network that inputs the current scene code into the first feature representation network to obtain a current scene feature;

[0194] The first splicing unit is configured to splice the first fusion feature, the second fusion feature and the current scene feature to obtain a first feature representation of the target user account.

[0195] In an exemplary embodiment, the second feature representation unit includes:

[0196] A second acquisition unit is configured to acquire recommended object information of a candidate recommended object; the recommended object information includes recommended object attribute information and a sub-object sequence, and the sub-object sequence includes a plurality of sub-objects associated with the candidate recommended object;

[0197] A second encoding unit is configured to input the recommended object attribute information and the sub-object sequence into a second feature representation network of the recommended object determination model, and encode the recommended object attribute information and the sub-object sequence through an encoding network of the second feature representation network to obtain a recommended object code and a sub-object code sequence;

[0198] a recommended object feature extraction unit, configured to execute a recommended object feature extraction network that inputs the recommended object encoding into the second feature representation network to obtain recommended object features;

[0199] The sub-object feature determination unit is configured to execute the second self-attention network of the second feature representation network by inputting the sub-object code sequence into the second feature representation network, determine the weight corresponding to each sub-object code in the sub-object code sequence based on the self-attention mechanism; perform weighted summation on each sub-object code and the corresponding weight to obtain the sub-object feature;

[0200] The second concatenation unit is configured to concatenate the recommended object feature and the sub-object feature to obtain a second feature representation of the candidate recommended object.

[0201] In an exemplary embodiment, the apparatus further includes a model training unit; the model training unit is configured to execute:

[0202] Acquire training samples and label information corresponding to the training samples; the training samples include sample attribute information, sample scenario information, and sample historical interaction information sequence corresponding to the sample user account, and recommended object information corresponding to the sample recommended object; the label information indicates whether the sample user account is interested in the sample recommended object;

[0203] Inputting the sample attribute information, sample scene information and sample historical interaction information sequence into an initial first feature representation network of a preset neural network model to obtain a predicted first feature representation output by the initial first feature representation network;

[0204] Inputting the recommended object information into the initial second feature representation network of the preset neural network model to obtain the predicted second feature representation output by the initial second feature representation network;

[0205] Determining a similarity between the predicted first feature representation and the predicted second feature representation to obtain a predicted similarity;

[0206] A loss value is determined according to the difference between the predicted similarity and the label information, and model parameters of the preset neural network model are adjusted based on the loss value. The adjusted preset neural network model is continuously iteratively trained until a preset training end condition is met, thereby obtaining the recommended object determination model.

[0207] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0208] In an exemplary embodiment, an electronic device is also provided, including a processor; a memory for storing instructions executable by the processor; wherein, when the processor is configured to execute the instructions stored in the memory, the method for determining a recommended object provided in any of the above embodiments is implemented.

[0209] The electronic device may be a terminal, a server or a similar computing device. For example, the electronic device is a server. Figure 8 is a block diagram of an electronic device for a method for determining a recommendation object according to an exemplary embodiment. Figure 8 As shown, the server 800 may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPU) 810 (the processor 810 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 830 for storing data, and one or more storage media 820 (such as one or more mass storage devices) for storing application programs 823 or data 822. Among them, the memory 830 and the storage medium 820 can be short-term storage or permanent storage. The program stored in the storage medium 820 may include one or more modules, each of which may include a series of instruction operations on the server. Furthermore, the central processing unit 810 can be configured to communicate with the storage medium 820 and execute a series of instruction operations in the storage medium 820 on the server 800. The server 800 may also include one or more power supplies 860, one or more wired or wireless network interfaces 850, one or more input and output interfaces 840, and / or one or more operating systems 821, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0210] The input / output interface 840 may be used to receive or send data via a network. The specific example of the network may include a wireless network provided by a communication provider of the server 800. In one example, the input / output interface 840 includes a network adapter (Network Interface Controller, NIC), which may be connected to other network devices via a base station so as to communicate with the Internet. In one example, the input / output interface 840 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0211] It can be understood by those skilled in the art that Figure 8 The structure shown is for illustration only and does not limit the structure of the above electronic device. Figure 8 More or fewer components as shown, or with Figure 8 Different configurations shown.

[0212] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 830 including instructions, and the instructions can be executed by the processor 810 of the device 800 to complete the above method. Alternatively, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0213] In an exemplary embodiment, a computer program product is also provided, including a computer program / instruction, and when the computer program / instruction is executed by a processor, the method for determining a recommended object provided in any of the above embodiments is implemented.

[0214] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0215] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for determining a recommendation object, characterized in that: include: In response to an object recommendation request of a target user account, acquiring attribute information, current scene information, and a historical interaction information sequence of the target user account; Inputting the current scene information, the attribute information and the historical interaction information sequence into a first feature representation network of a recommendation object determination model, encoding the current scene information, the attribute information and the historical interaction information sequence respectively through an encoding network of the first feature representation network to obtain a current scene encoding, an attribute encoding and a historical interaction encoding sequence; Inputting the current scene code and the attribute code into a first feature fusion network of the first feature representation network, performing feature extraction on the current scene code and the attribute code respectively through the first feature fusion network, and determining a first fusion feature according to the product of the features extracted from the feature extraction; the first feature fusion network includes a plurality of cascaded first feature crossover units; Inputting the current scene code and the historical interaction code sequence into a second feature fusion network of the first feature representation network, performing feature extraction on the current scene code and the historical interaction code sequence respectively through the second feature fusion network, and determining a second fusion feature according to the product of the features extracted from the feature extraction; the second feature fusion network includes a second feature intersection unit and a first self-attention network; Determining the similarity level corresponding to each candidate recommendation object according to the first fusion feature and the second fusion feature; According to the degree of similarity corresponding to each of the candidate recommendation objects, a target candidate recommendation object corresponding to the object recommendation request is determined.

2. The method for determining a recommended object according to claim 1, characterized in that: The determining, according to the first fusion feature and the second fusion feature, the degree of similarity corresponding to each candidate recommendation object includes: Determining a first feature representation of the target user account according to the first fused feature and the second fused feature; Determining a second feature representation of each of the candidate recommendation objects according to the recommendation object information of each of the candidate recommendation objects; Determine the degree of similarity between the first feature representation and the second feature representation of each of the candidate recommendation objects, and obtain the degree of similarity corresponding to each of the candidate recommendation objects.

3. The method for determining a recommended object according to claim 1, characterized in that: The extracting features of the current scene code and the attribute code respectively through the first feature fusion network, and determining the first fusion feature according to the product of the features extracted from the features, includes: Inputting the current scene code and the attribute code into a first-level first feature cross unit, performing feature extraction on the current scene code and the attribute code respectively through the first-level first feature cross unit, and taking the product of the features obtained by the feature extraction as the output feature of the first-level first feature cross unit; For the first feature intersection unit of the current level among the remaining first feature intersection units, the output feature of the first feature intersection unit of the previous level and the current scene code are input to the first feature intersection unit of the current level, and the output feature and the current scene code are respectively extracted by the first feature intersection unit of the current level, and the product of the features obtained by the feature extraction is used as the output feature of the first feature intersection unit of the current level; The output feature of the last level of the first feature intersection unit in the cascaded multiple first feature intersection units is determined as the first fusion feature.

4. The method for determining a recommended object according to claim 1, characterized in that: The extracting features of the current scene code and the historical interaction code sequence respectively through the second feature fusion network, and determining the second fusion feature according to the product of the features extracted from the features, includes: For each historical interaction code in the historical interaction code sequence, the historical interaction code and the current scene code are input into the second feature cross unit, and the second feature cross unit performs feature extraction on the historical interaction code and the current scene code respectively, and uses the product of the features obtained by the feature extraction as the fusion feature corresponding to the historical interaction code, so as to obtain a fusion feature sequence corresponding to the historical interaction code sequence; Inputting the fused feature sequence into the first self-attention network, and determining the weight corresponding to each fused feature in the fused feature sequence based on the self-attention mechanism; A weighted sum is performed on each of the fused features and the corresponding weights to obtain the second fused feature.

5. The method for determining a recommended object according to claim 2, characterized in that: The determining, according to the first fused feature and the second fused feature, a first feature representation of the target user account includes: Inputting the current scene code into the scene feature extraction network of the first feature representation network to obtain current scene features; The first fused feature, the second fused feature, and the current scene feature are concatenated to obtain a first feature representation of the target user account.

6. The method for determining a recommended object according to claim 5, characterized in that: The determining, according to the recommendation object information of each of the candidate recommendation objects, a second feature representation of each of the candidate recommendation objects comprises: Acquire recommended object information of the candidate recommended object; the recommended object information includes recommended object attribute information and a sub-object sequence, and the sub-object sequence includes a plurality of sub-objects associated with the candidate recommended object; Inputting the recommended object attribute information and the sub-object sequence into a second feature representation network of the recommended object determination model, and encoding the recommended object attribute information and the sub-object sequence through an encoding network of the second feature representation network to obtain a recommended object code and a sub-object code sequence; Inputting the recommended object code into a recommended object feature extraction network of the second feature representation network to obtain recommended object features; Inputting the sub-object encoding sequence into the second self-attention network of the second feature representation network, determining the weight corresponding to each sub-object encoding in the sub-object encoding sequence based on the self-attention mechanism; performing weighted summation on each sub-object encoding and the corresponding weight to obtain a sub-object feature; The recommended object feature and the sub-object feature are concatenated to obtain a second feature representation of the candidate recommended object.

7. The method for determining a recommended object according to claim 6, characterized in that: The method further comprises: Acquire training samples and label information corresponding to the training samples; the training samples include sample attribute information, sample scenario information, and sample historical interaction information sequence corresponding to the sample user account, and recommended object information corresponding to the sample recommended object; the label information indicates whether the sample user account is interested in the sample recommended object; Inputting the sample attribute information, sample scene information and sample historical interaction information sequence into an initial first feature representation network of a preset neural network model to obtain a predicted first feature representation output by the initial first feature representation network; Inputting the recommended object information into the initial second feature representation network of the preset neural network model to obtain the predicted second feature representation output by the initial second feature representation network; Determining a similarity between the predicted first feature representation and the predicted second feature representation to obtain a predicted similarity; A loss value is determined according to the difference between the predicted similarity and the label information, and model parameters of the preset neural network model are adjusted based on the loss value. The adjusted preset neural network model is continuously iteratively trained until a preset training end condition is met, thereby obtaining the recommended object determination model.

8. A device for determining a recommended object, characterized in that: include: A first acquisition unit is configured to execute in response to an object recommendation request of a target user account, and acquire attribute information, current scene information, and a historical interaction information sequence of the target user account; A first encoding unit is configured to execute inputting the current scene information, the attribute information and the historical interaction information sequence into a first feature representation network of a recommendation object determination model, and encoding the current scene information, the attribute information and the historical interaction information sequence through an encoding network of the first feature representation network to obtain a current scene code, an attribute code and a historical interaction code sequence; a first feature fusion unit configured to execute a first feature fusion network that inputs the current scene code and the attribute code into the first feature representation network, extract features from the current scene code and the attribute code respectively through the first feature fusion network, and determine a first fusion feature according to the product of features obtained by the feature extraction; the first feature fusion network includes a plurality of cascaded first feature intersection units; a second feature fusion unit configured to execute a second feature fusion network that inputs the current scene code and the historical interaction code sequence into the first feature representation network, extracts features from the current scene code and the historical interaction code sequence respectively through the second feature fusion network, and determines a second fusion feature according to the product of features extracted from the features; the second feature fusion network includes a second feature intersection unit and a first self-attention network; a similarity determination unit, configured to determine the similarity corresponding to each candidate recommendation object according to the first fusion feature and the second fusion feature; The target recommendation object determination unit is configured to determine the target candidate recommendation object corresponding to the object recommendation request according to the similarity degree corresponding to each candidate recommendation object.

9. The device for determining a recommended object according to claim 8, characterized in that: The similarity determination unit comprises: A first feature representation determining unit is configured to determine a first feature representation of the target user account according to the first fused feature and the second fused feature; A second feature representation determining unit is configured to determine a second feature representation of each candidate recommendation object according to the recommendation object information of each candidate recommendation object; The similarity determination subunit is configured to determine the similarity between the first feature representation and the second feature representation of each candidate recommendation object, and obtain the similarity corresponding to each candidate recommendation object.

10. The device for determining a recommended object according to claim 8, characterized in that: The first feature fusion unit is specifically configured to execute: Inputting the current scene code and the attribute code into a first-level first feature cross unit, performing feature extraction on the current scene code and the attribute code respectively through the first-level first feature cross unit, and taking the product of the features obtained by the feature extraction as the output feature of the first-level first feature cross unit; For the first feature intersection unit of the current level among the remaining first feature intersection units, the output feature of the first feature intersection unit of the previous level and the current scene code are input to the first feature intersection unit of the current level, and the output feature and the current scene code are respectively extracted by the first feature intersection unit of the current level, and the product of the features obtained by the feature extraction is used as the output feature of the first feature intersection unit of the current level; The output feature of the last level of the first feature intersection unit in the cascaded multiple first feature intersection units is determined as the first fusion feature.

11. The device for determining a recommended object according to claim 8, characterized in that: The second feature fusion unit is specifically configured to execute: For each historical interaction code in the historical interaction code sequence, the historical interaction code and the current scene code are input into the second feature cross unit, and the second feature cross unit performs feature extraction on the historical interaction code and the current scene code respectively, and uses the product of the features obtained by the feature extraction as the fusion feature corresponding to the historical interaction code, so as to obtain a fusion feature sequence corresponding to the historical interaction code sequence; Inputting the fused feature sequence into the first self-attention network, and determining the weight corresponding to each fused feature in the fused feature sequence based on the self-attention mechanism; A weighted sum is performed on each of the fused features and the corresponding weights to obtain the second fused feature.

12. The device for determining a recommended object according to claim 9, characterized in that: The first feature representation unit comprises: a current scene feature extraction unit, configured to execute a scene feature extraction network that inputs the current scene code into the first feature representation network to obtain a current scene feature; The first splicing unit is configured to splice the first fusion feature, the second fusion feature and the current scene feature to obtain a first feature representation of the target user account.

13. The device for determining a recommended object according to claim 12, characterized in that: The second feature representation unit comprises: A second acquisition unit is configured to acquire recommended object information of the candidate recommended object; the recommended object information includes recommended object attribute information and a sub-object sequence, and the sub-object sequence includes a plurality of sub-objects associated with the candidate recommended object; A second encoding unit is configured to input the recommended object attribute information and the sub-object sequence into a second feature representation network of the recommended object determination model, and encode the recommended object attribute information and the sub-object sequence through an encoding network of the second feature representation network to obtain a recommended object code and a sub-object code sequence; a recommended object feature extraction unit, configured to execute a recommended object feature extraction network that inputs the recommended object encoding into the second feature representation network to obtain recommended object features; a sub-object feature determination unit configured to execute inputting the sub-object encoding sequence into a second self-attention network of the second feature representation network, determining a weight corresponding to each sub-object encoding in the sub-object encoding sequence based on a self-attention mechanism; and performing a weighted summation on each sub-object encoding and the corresponding weight to obtain a sub-object feature; The second concatenation unit is configured to concatenate the recommended object feature and the sub-object feature to obtain a second feature representation of the candidate recommended object.

14. The device for determining a recommended object according to claim 13, characterized in that: The device also includes a model training unit; the model training unit is configured to perform: Acquire training samples and label information corresponding to the training samples; the training samples include sample attribute information, sample scenario information, and sample historical interaction information sequence corresponding to the sample user account, and recommended object information corresponding to the sample recommended object; the label information indicates whether the sample user account is interested in the sample recommended object; Inputting the sample attribute information, sample scene information and sample historical interaction information sequence into an initial first feature representation network of a preset neural network model to obtain a predicted first feature representation output by the initial first feature representation network; Inputting the recommended object information into the initial second feature representation network of the preset neural network model to obtain the predicted second feature representation output by the initial second feature representation network; Determining a similarity between the predicted first feature representation and the predicted second feature representation to obtain a predicted similarity; A loss value is determined according to the difference between the predicted similarity and the label information, and model parameters of the preset neural network model are adjusted based on the loss value. The adjusted preset neural network model is continuously iteratively trained until a preset training end condition is met, thereby obtaining the recommended object determination model.

15. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the recommended object determination method as described in any one of claims 1 to 7. 16 . A computer-readable storage medium, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method for determining a recommended object as claimed in any one of claims 1 to 7.

17. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the method for determining a recommended object according to any one of claims 1 to 7 is implemented.

Images