Virtual space item recommendation model training method, item recommendation method and device

Abstract

The application discloses a kind of virtual space article recommendation model training method, article recommendation method and device, wherein method includes: the training sample corresponding to the user of the overlap in reality domain and virtual domain is constructed;Real domain user feature is input to inter-domain mapping model, and the real domain user vector is obtained;The user vector is obtained by fusing the virtual domain user vector corresponding to real domain user vector and virtual domain user feature, as the input of user tower in double-tower recommendation model;The article vector corresponding to virtual domain article feature is used as the input of article tower in double-tower recommendation model;Inter-domain mapping model and double-tower recommendation model are jointly trained, and the article recommendation model is obtained.By the above-mentioned mode, the mapping mode of cross-domain model is used, the situation that reality domain feature cannot be directly applied to virtual domain because of data security and other problems is solved, the user feature in virtual domain is effectively supplemented, and inter-domain mapping model and double-tower recommendation model joint training and recommendation are realized.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a method for training a virtual space item recommendation model, an item recommendation method, and an apparatus. Background Technology

[0002] In real-world shopping malls, there are generally two ways to guide customers. One is through electronic screens placed on each floor, where customers can search for and find the products they want. The second is by asking staff at the service desk and obtaining the information they need through conversation.

[0003] Virtual malls aim to provide users in the metaverse with an immersive shopping experience. For example, many e-commerce companies, both domestically and internationally, can build virtual shops on their websites and integrate them with real-world shopping mall layouts to create virtual malls. Users can experience a virtual mall from the comfort of their homes using VR glasses or similar devices, just like a real-world mall, but with a much wider variety of goods. However, the existing methods mentioned above cannot meet the needs of virtual malls with a greater number of shops and products in the virtual world. Especially when users browse virtual malls without a clear purpose, it means that the "digital intelligence" system needs to have personalized recommendation capabilities.

[0004] Traditional single-domain recommendation systems use user profile features as input to recommend products that the user may be interested in. However, in virtual shopping scenarios, when users have little or no user profile features in the virtual domain, the recommendation results of traditional single-domain recommendation systems are not accurate enough.

[0005] The real-world domain and the virtual domain differ in the following ways: 1. User characteristics in the real-world domain and the virtual domain are not necessarily the same. For example, a user might be male in the real-world domain but female in the virtual domain, and the virtual domain may have more open feature types than the real-world domain; 2. Virtual domain models lack data, and due to data security and authorization issues, it is impossible to directly obtain the characteristics of real-world users during training. Therefore, a new solution is needed to address these challenges. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide a method, apparatus and device for training a virtual space item recommendation model that overcomes or at least partially solves the above problems.

[0007] According to one aspect of the present invention, a method for training an item recommendation model in a virtual space is provided, comprising:

[0008] Construct training samples corresponding to users overlapping in the real domain and virtual domain; the training samples include: real domain user features, virtual domain user features, virtual domain item features, and user-item behavior association data;

[0009] The real-domain user features are input into the inter-domain mapping model to obtain the real-domain user vector;

[0010] The user vector is obtained by fusing the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features, and is used as the input to the user tower in the dual-tower recommendation model; the item vector corresponding to the virtual-domain item features is used as the input to the item tower in the dual-tower recommendation model.

[0011] By using the user-item behavior association data, the inter-domain mapping model and the dual-tower recommendation model are jointly trained to obtain the item recommendation model.

[0012] Optionally, the real-world user characteristics include: real-world user attribute characteristics and / or real-world user preference characteristics and / or device type accessing the virtual space;

[0013] The virtual domain user characteristics include: virtual domain user attribute characteristics and / or virtual domain user preference characteristics;

[0014] The virtual domain item characteristics include: category and / or price and / or quality and / or sales volume and / or rating and / or appearance;

[0015] The user and item behavior association data includes: user and item purchase and / or attention behavior data generated in the virtual domain.

[0016] Optionally, the inter-domain mapping model includes a fully connected layer network model;

[0017] The dual-tower recommendation model includes a first neural network layer, which is used to perform non-linear mapping on the input vector to obtain a dense user vector.

[0018] The dual-tower recommendation model includes a second neural network layer, which performs a non-linear mapping on the input vector to obtain a dense array of item vectors.

[0019] Optionally, fusing the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features to obtain the user vector includes any one of the following operations:

[0020] The real-domain user vector and the virtual-domain user vector are concatenated together to form a user vector;

[0021] The user vector is obtained by multiplying the real-domain user vector and the virtual-domain user vector by two variables whose sum is 1, and then summing them.

[0022] Optionally, the step of jointly training the inter-domain mapping model and the dual-tower recommendation model using the user-item behavior association data to obtain the item recommendation model includes:

[0023] Calculate the similarity between the user vector and the item vector;

[0024] A loss function is constructed or selected based on the similarity;

[0025] By comparing the user and item behavior association data, the training samples are input into the inter-domain mapping model and the dual-tower recommendation model for joint training, and the model parameters are updated according to the backpropagation algorithm to reduce the loss value of the loss function. This process is repeated until the loss value of the loss function no longer decreases.

[0026] Based on the final similarity score, items are recommended to the user, resulting in a trained item recommendation model.

[0027] Optionally, after obtaining the item recommendation model, the method includes:

[0028] The trained inter-domain mapping model is distributed to the devices connected to the virtual space, and the trained dual-tower recommendation model is saved to the virtual domain recommendation system.

[0029] The item vectors are saved to the item vector index library.

[0030] According to another aspect of the present invention, a method for recommending items in a virtual space is provided, comprising:

[0031] Obtain the real-domain user profile characteristics and virtual-domain user profile characteristics of the target users;

[0032] Based on the recommendation request, the real-domain user profile features are input into the trained inter-domain mapping model to obtain the real-domain user profile vector;

[0033] The user profile vector is obtained using the trained dual-tower recommendation model, and the inner product between the user profile vector and each item vector is calculated. Based on the calculation results, the item to be recommended is found and recommended to the target user. The inter-domain mapping model and the dual-tower recommendation model are obtained according to the above-mentioned virtual space item recommendation model training method.

[0034] According to another aspect of the present invention, a training apparatus for a virtual space item recommendation model is provided, comprising:

[0035] The training sample construction module is suitable for constructing training samples corresponding to users overlapping in the real domain and virtual domain; the training samples include: real domain user features, virtual domain user features, virtual domain item features, and behavioral association data between users and items;

[0036] The user vector acquisition module is adapted to input the real-domain user features into the inter-domain mapping model to obtain real-domain user vectors.

[0037] The recommendation model construction module fuses the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features to obtain a user vector, which serves as the input to the user tower in the dual-tower recommendation model; and uses the item vector corresponding to the virtual-domain item features as the input to the item tower in the dual-tower recommendation model.

[0038] The recommendation model training module is adapted to use the user-item behavior association data to jointly train the inter-domain mapping model and the dual-tower recommendation model to obtain the item recommendation model.

[0039] According to another aspect of the present invention, an electronic device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0040] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the virtual space item recommendation model training method, and / or to perform the operation corresponding to the virtual space item recommendation method.

[0041] According to another aspect of the present invention, a computer storage medium is provided, the storage medium storing at least one executable instruction, the executable instruction causing a processor to perform operations corresponding to the above-described virtual space item recommendation model training method, and / or to perform operations corresponding to the above-described virtual space item recommendation method.

[0042] According to the virtual space item recommendation model training method and recommendation method of the present invention, existing real-world user features can be mapped to the virtual domain, thereby realizing the joint training of cross-domain model and dual-tower recommendation model, and introducing real-world user features into the virtual domain as a supplement to user behavior features; and by using the mapping method of cross-domain model, the problem that real-world features cannot be directly used in the virtual domain due to data security and other considerations is overcome.

[0043] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0044] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0045] Figure 1 A flowchart of a virtual space item recommendation model training method provided in an embodiment of the present invention is shown;

[0046] Figure 2 This diagram illustrates the overall flowchart of the virtual space item recommendation model training and recommendation provided in one embodiment of the present invention.

[0047] Figure 3 A flowchart illustrating the training steps of an item recommendation model according to an embodiment of the present invention is shown.

[0048] Figure 4 This diagram illustrates a structural schematic of a training sample provided in one embodiment of the present invention.

[0049] Figure 5 A detailed flowchart illustrating a similarity-based recommendation process provided by an embodiment of the present invention is shown.

[0050] Figure 6 A schematic diagram of the structure of an inter-domain mapping model provided in an embodiment of the present invention is shown;

[0051] Figure 7 The diagram illustrates the process flow of two models provided in one embodiment of the present invention;

[0052] Figure 8 A flowchart illustrating a virtual space item recommendation method according to an embodiment of the present invention is shown.

[0053] Figure 9 This diagram illustrates the data flow in a virtual space item recommendation system according to an embodiment of the present invention.

[0054] Figure 10 This diagram illustrates the structure of a virtual space item recommendation model training device according to an embodiment of the present invention.

[0055] Figure 11 A schematic diagram of the structure of an electronic device provided in one embodiment of the present invention is shown. Detailed Implementation

[0056] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0057] Figure 1 A flowchart illustrating an embodiment of the virtual space item recommendation model training method of the present invention is shown, which is applied to electronic devices. The aforementioned items include goods, auction items, exhibits, and other items that may be appreciated in a virtual space such as a metaverse. The electronic devices include various servers, laptops, PCs, tablets, and other smart terminal devices. Figure 1 As shown, the method includes the following steps:

[0058] Step 110: Construct training samples corresponding to users overlapping in the real domain and virtual domain; the training samples include: real domain user features, virtual domain user features, virtual domain item features, and user-item behavior association data.

[0059] The training samples are samples that exist in both the real domain and the virtual domain. Of course, there may be cases where the amount of data in the real domain is large and the amount of data in the virtual domain is small. The sample data preferably includes real domain user characteristic data, virtual domain user characteristic data, virtual domain items related to the user, and user behavior association data with items, such as whether they have purchased or followed them.

[0060] Step 120: Input the real-domain user features into the inter-domain mapping model to obtain the real-domain user vector.

[0061] In order to obtain a complete user profile, especially in the early stages of building a virtual space, applying user characteristic data from the real domain to the virtual domain is a necessary supplement to the virtual domain user profile, thereby making item recommendations more accurate.

[0062] Step 130: Fuse the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features to obtain a user vector, which is used as the input to the user tower in the dual-tower recommendation model; use the item vector corresponding to the virtual-domain item features as the input to the item tower in the dual-tower recommendation model.

[0063] It should be noted that user features and item features generally require preprocessing such as normalization. This step mainly sets the composition of the dual-tower recommendation model and the input data. The user vectors and item vectors used for fusion can be text vectors or numerical vectors converted from text vectors. Numerical vector extraction preferably uses an embedding layer, which can be set either within the dual-tower recommendation model or outside of it.

[0064] Step 140: Using the user and item behavior association data, jointly train the inter-domain mapping model and the dual-tower recommendation model to obtain the item recommendation model.

[0065] Among these, user and item behavior correlation data are mainly used for intervention or optimization during training, or as a basis for judgment in backpropagation. To achieve the purpose of recommendation, the item recommendation model preferably uses the pyramid method to solve for the similarity between user vectors and item vectors, and makes recommendations based on the magnitude of the similarity. The greater the similarity, the more positively correlated the two are, and the more worthy of recommendation. The similarity can be determined by any one of the following: inner product, cosine similarity, Euclidean distance, cosine distance, and Mahalanobis distance.

[0066] Furthermore, combined with Figure 2 and Figure 3 As shown, this invention employs a "joint model training and separate usage" method. First, two models are trained end-to-end by merging them together. The trained models are then distributed separately to the user's VR glasses and the backend recommendation system for joint recommendations. After the models are built and trained, when a user enters a virtual space using their wearable device, items of interest are recommended to the user based on the aforementioned models.

[0067] In summary, by utilizing the scheme disclosed in this embodiment, the problem that real-world domain features cannot be directly applied to the virtual domain due to data security and other issues can be solved by using the mapping method of the cross-domain model. This effectively supplements the user features in the virtual domain and enables end-to-end joint training of the inter-domain mapping model and the dual-tower recommendation model.

[0068] In one or more embodiments, the real-world user characteristics include: real-world user attribute characteristics and / or real-world user preference characteristics and / or device type accessing the virtual space;

[0069] The virtual domain user characteristics include: virtual domain user attribute characteristics and / or virtual domain user preference characteristics;

[0070] The virtual domain item characteristics include: category and / or price and / or quality and / or sales volume and / or rating and / or appearance;

[0071] The user and item behavior association data includes: user and item purchase and / or attention behavior data generated in the virtual domain.

[0072] User attribute features include gender, age, occupation, address, and wealth; user preference features include user preference information or user interest information. During specific training, the number and type of the above features can be selected according to the actual situation.

[0073] Combination Figure 4 and Figure 5 As shown, each training sample includes at least the four features mentioned above. Of course, during model training, the number and composition of the features can vary, and each feature has a different function, namely, generating user vectors, generating item vectors, and supervising model training.

[0074] In one or more embodiments, such as Figure 6 As shown, the inter-domain mapping model includes a fully connected layer network model.

[0075] Neural networks are typically multi-layered, with each layer consisting of multiple neurons. Each neuron can be understood as a parameter to be trained in the neural network model. The input to the model is a vector structure, and through the information transmission of the multi-layered neural network, the output is also a vector structure. Neural networks possess rich information representation and knowledge learning capabilities, enabling them to effectively learn the various implicit rules for transforming real-domain vectors into virtual-domain vectors. As an inter-domain mapping model, they can obtain more information about real-domain user characteristics.

[0076] The dual-tower recommendation model includes a user tower comprising a first neural network layer, which performs a non-linear mapping on the input vector to obtain a dense user vector; and an item tower comprising a second neural network layer, which performs a non-linear mapping on the input vector to obtain a dense item vector.

[0077] Specifically, the dual-tower recommendation model transforms the features of users (such as gender, age, and interests) and items (category, appearance, etc.) into vectors through model parameters. These vectors are represented by two "vector towers" (the "user tower" and the "item tower"). Each "vector tower" is a multi-layer neural network that non-linearly transforms the user and item features into dense vector representations. The similarity between the two vectors is then calculated by performing an inner product at the "top of the tower."

[0078] In one or more embodiments, the step 130 of fusing the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features to obtain the user vector includes any one of the following operations:

[0079] The user vector is formed by concatenating the real-domain user vector and the virtual-domain user vector; or...

[0080] The user vector is obtained by multiplying the real-domain user vector and the virtual-domain user vector by two variables whose sum is 1, and then summing them.

[0081] As can be seen from this embodiment, the user vector can be obtained by concatenating the real domain user vector and the virtual domain user vector, or it can be fused by simple addition or other methods.

[0082] In one or more embodiments, step 140, which involves jointly training the inter-domain mapping model and the dual-tower recommendation model using the user-item behavior association data to obtain the item recommendation model, includes:

[0083] Calculate the similarity between the user vector and the item vector;

[0084] A loss function is constructed or selected based on the similarity;

[0085] By comparing the user and item behavior association data, the training samples are input into the inter-domain mapping model and the dual-tower recommendation model for joint training, and the model parameters are updated according to the backpropagation algorithm to reduce the loss value of the loss function. This process is repeated until the loss value of the loss function no longer decreases.

[0086] After the model parameters are determined, items are recommended to the user based on the final similarity score, and the trained item recommendation model is obtained.

[0087] It should be noted that the backpropagation algorithm is a commonly used and effective algorithm for training neural network models. Its main steps include: (1) inputting training samples into the input layer of the neural network, passing through intermediate layers, and finally reaching the output layer and outputting the result. This is the forward propagation process; (2) since there is an error between the output result of the neural network and the actual result, the error between the estimated value and the actual value is calculated, or the error is determined by using a loss function, and the error is propagated backward from the output layer to the intermediate layer until it reaches the input layer; (3) during the backpropagation process, the values ​​of various parameters are adjusted according to the error; the above process is iterated continuously until convergence.

[0088] Combination Figure 5As shown, the training objective of the dual-tower model is to train and adjust the model parameters in the "vector tower" using sample data such as users' "purchases" and "likes" of items, so that the similarity of items that users are interested in increases in the inner product of the "tower top", and the similarity of items that they are not interested in decreases in the "tower top". The specific training process also includes a convergence process through a loss function. Preferably, an optimizer can be used to optimize the training process.

[0089] In one or more embodiments, combined with Figure 7 As shown, after obtaining the item recommendation model, the method includes:

[0090] The trained inter-domain mapping model is distributed to devices connected to the virtual space, and the trained dual-tower recommendation model is saved to the virtual domain recommendation system, thus realizing the separate use of the models.

[0091] To improve the efficiency of recommendations, the item vectors can also be saved to an item vector index library for later use.

[0092] It is important to emphasize that in the above embodiments, during model training, iterative training is performed using samples constructed from features and user behaviors towards items (purchases, clicks, etc.). With positive samples, the model learns and adjusts its parameters, increasing the similarity between the calculated user vector and item vector, and vice versa. During training, user-item behavior association data from the training samples can be used to verify the accuracy of recommendations based on similarity.

[0093] In a preferred embodiment, the complete steps of model training include:

[0094] 1) Randomly initialize the training parameters of the inter-domain mapping model and the dual-tower recommendation model;

[0095] 2) Input the real-domain vectors of overlapping users into the inter-domain mapping model, and output the real-domain user vectors through the neural network;

[0096] 3) The user's features in the virtual domain are converted into virtual domain vectors through a neural network layer;

[0097] 4) Average and sum the user's real-domain vector and virtual-domain vector to obtain the final user vector, which serves as the input to the "user tower" of the dual-tower recommendation model;

[0098] 5) Similarly, the "item tower" is used to convert item features into item vectors, and the inner product of the user vector and the item vector is calculated as the loss function for model optimization.

[0099] 6) The model begins training. The process involves using the backpropagation algorithm of the neural network to update the parameters to be trained, reducing the loss of the loss function (i.e., the item vector and the user vector are closer), and iterating repeatedly until the loss function no longer decreases, and the model optimization is complete.

[0100] See Figure 8 As shown, another aspect of the present invention also provides an embodiment of a virtual space item recommendation method, the item recommendation method comprising:

[0101] Step 810: Obtain the real-domain user profile features and virtual-domain user profile features of the target user;

[0102] Step 820: Based on the recommendation request, input the real-domain user profile features into the trained inter-domain mapping model to obtain the real-domain user profile vector;

[0103] Step 830: Obtain the user profile vector using the trained dual-tower recommendation model, calculate the inner product between the user profile vector and each item vector, find the item to be recommended based on the calculation result, and recommend the item to be recommended to the target user; wherein, the inter-domain mapping model and the dual-tower recommendation model are obtained according to any of the above-described virtual space item recommendation model training method embodiments.

[0104] Specifically, in combination Figure 9 As shown, before a new user enters the virtual metaverse wearing a VR device, the VR device prompts the user to authorize the acquisition of their real-world profile features. At this time, the dual-tower model of the real-world domain built into the VR device transforms the user's real-world profile features into a vector of the user's real-world domain.

[0105] When a user initiates a personalized recommendation request, the recommendation system's backend receives the user's feature vector from the real-world domain and simultaneously receives the user's feature vector from the virtual domain. These two vectors are combined, and the system searches a pre-defined vector retrieval database for items with similar vector distances to recommend. When recommending items to a user, the system first converts the user's features into user vectors using the user tower parameters of a dual-tower model, and then retrieves similar items from the vector retrieval database for recommendation.

[0106] Therefore, when recommending items to users, the user features can be converted into user vectors through the user tower of the dual-tower model, and then similar items can be retrieved from the vector retrieval library for recommendation.

[0107] Figure 10 A schematic diagram of an embodiment of the virtual space item recommendation model training device of the present invention is shown. Figure 10 As shown, the device 1000 includes:

[0108] The training sample construction module 1010 is adapted to construct training samples corresponding to users overlapping in the real domain and the virtual domain; the training samples include: real domain user features, virtual domain user features, virtual domain item features, and user and item behavior association data.

[0109] User vector acquisition module 1020 is adapted to input the real-domain user features into the inter-domain mapping model to obtain real-domain user vectors;

[0110] The recommendation model construction module 1030 fuses the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features to obtain a user vector, which is used as the input to the user tower in the dual-tower recommendation model; and uses the item vector corresponding to the virtual-domain item features as the input to the item tower in the dual-tower recommendation model.

[0111] The recommendation model training module 1040 is adapted to use the user and item behavior association data to jointly train the inter-domain mapping model and the dual-tower recommendation model to obtain the item recommendation model.

[0112] In one embodiment, the real-world user characteristics include: real-world user attribute characteristics and / or real-world user preference characteristics and / or device type accessing the virtual space;

[0113] The virtual domain user characteristics include: virtual domain user attribute characteristics and / or virtual domain user preference characteristics;

[0114] The virtual domain item characteristics include: category and / or price and / or quality and / or sales volume and / or rating and / or appearance;

[0115] The user and item behavior association data includes: user and item purchase and / or attention behavior data generated in the virtual domain.

[0116] In one embodiment, the inter-domain mapping model includes a fully connected layer network model;

[0117] The dual-tower recommendation model includes a first neural network layer, which is used to perform non-linear mapping on the input vector to obtain a dense user vector.

[0118] The dual-tower recommendation model includes a second neural network layer, which performs a non-linear mapping on the input vector to obtain a dense array of item vectors.

[0119] In one embodiment, the recommendation model building module 1030 is further adapted to:

[0120] The real-domain user vector and the virtual-domain user vector are concatenated together to form a user vector;

[0121] The user vector is obtained by multiplying the real-domain user vector and the virtual-domain user vector by two variables whose sum is 1, and then summing them.

[0122] In one embodiment, the recommendation model training module 1040 is further adapted to:

[0123] Calculate the similarity between the user vector and the item vector;

[0124] A loss function is constructed or selected based on the similarity;

[0125] By comparing the user and item behavior association data, the training samples are input into the inter-domain mapping model and the dual-tower recommendation model for joint training. The model parameters are updated according to the backpropagation algorithm to reduce the loss value of the loss function. This process is repeated until the loss value of the loss function no longer decreases, thus obtaining the trained item recommendation model.

[0126] In one embodiment, the apparatus further includes a model configuration module 1050, which is adapted to:

[0127] The trained inter-domain mapping model is distributed to the devices connected to the virtual space, and the trained dual-tower recommendation model is saved to the virtual domain recommendation system.

[0128] The item vectors are saved to the item vector index library.

[0129] The beneficial effects of this invention include:

[0130] The above embodiments utilize a combination of a dual-tower recommendation model and an inter-domain mapping model to establish a connection between the virtual and real domains. Through this connection, personalized recommendations are made in the virtual domain for users who are active in the real domain but inactive in the virtual domain. Compared to traditional recommendation models, this invention offers at least the following two advantages:

[0131] 1. Jointly train the cross-domain model and the dual-tower model recommendation model to introduce the features of the user's real domain into the virtual domain as a supplement to the user's behavioral features;

[0132] 2. By utilizing the mapping method of cross-domain models, the situation where real-domain features cannot be directly used in the virtual domain due to data security and other considerations can be resolved.

[0133] This invention provides a non-volatile computer storage medium storing at least one executable instruction that can execute the virtual space item recommendation model training method or the virtual space item recommendation method in any of the above method embodiments.

[0134] Figure 11The diagram shows a structural schematic of an embodiment of the electronic device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the electronic device.

[0135] like Figure 11 As shown, the electronic device may include: a processor 1102, a communications interface 1104, a memory 1106, and a communications bus 108.

[0136] The processor 1102, communication interface 1104, and memory 1106 communicate with each other via communication bus 1108. Communication interface 1104 is used to communicate with other network elements such as clients or other servers. The processor 1102 executes program 1110, specifically performing the relevant steps in the above-described method for training a virtual space item recommendation model for electronic devices or the embodiment of the virtual space item recommendation method.

[0137] Specifically, program 1110 may include program code that includes computer operation instructions.

[0138] Processor 1102 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The electronic device may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0139] Memory 1106 is used to store program 1110. Memory 1106 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0140] Specifically, program 1110 can be used to cause processor 1102 to execute the above-mentioned virtual space item recommendation model training method or the operation corresponding to the virtual space item recommendation method embodiment.

[0141] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the content of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0142] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0143] Similarly, it should be understood that, in order to simplify the invention and aid in understanding one or more of the various inventive aspects, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim. Rather, as reflected in the following claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into this detailed description, wherein each claim itself is a separate embodiment of the invention.

[0144] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0145] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0146] The various component embodiments of the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of the present invention. The present invention can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such programs implementing the present invention can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0147] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A method for training an item recommendation model in a virtual space, characterized in that, include: Construct training samples corresponding to users whose real-world and virtual domains overlap; The training samples include: real-world user characteristics, virtual-world user characteristics, virtual-world item characteristics, and user-item behavior association data; The real-domain user features are input into the inter-domain mapping model to obtain the real-domain user vector; the inter-domain mapping model includes a fully connected layer network model. The user vector is obtained by fusing the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features, and is used as the input to the user tower in the dual-tower recommendation model; the item vector corresponding to the virtual-domain item features is used as the input to the item tower in the dual-tower recommendation model; the user tower in the dual-tower recommendation model includes a first neural network layer, which is used to perform non-linear mapping on the input vector to obtain a dense user vector; the item tower in the dual-tower recommendation model includes a second neural network layer, which is used to perform non-linear mapping on the input vector to obtain a dense item vector. Using the user-item behavior association data, the inter-domain mapping model and the dual-tower recommendation model are jointly trained to obtain an item recommendation model. Specifically, the similarity between the user vector and the item vector is calculated; a loss function is constructed or selected based on the similarity; the training samples are input into the inter-domain mapping model and the dual-tower recommendation model for joint training, using the user-item behavior association data, and the model parameters are updated according to the backpropagation algorithm to reduce the loss value of the loss function. This process is iterated repeatedly until the loss value of the loss function no longer decreases, resulting in a fully trained item recommendation model.

2. The method according to claim 1, characterized in that, The real-domain user characteristics include: real-domain user attribute characteristics and / or real-domain user preference characteristics and / or device type accessing the virtual space; The virtual domain user characteristics include: virtual domain user attribute characteristics and / or virtual domain user preference characteristics; The virtual domain item characteristics include: category and / or price and / or quality and / or sales volume and / or rating and / or appearance; The user and item behavior association data includes: user and item purchase and / or attention behavior data generated in the virtual domain.

3. The method according to claim 1, characterized in that, The process of fusing the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features to obtain the user vector includes any one of the following operations: The real-domain user vector and the virtual-domain user vector are concatenated together to form a user vector; The user vector is obtained by multiplying the real-domain user vector and the virtual-domain user vector by two variables whose sum is 1, and then summing them.

4. The method according to claim 1, characterized in that, After obtaining the item recommendation model, the method includes: The trained inter-domain mapping model is distributed to the devices connected to the virtual space, and the trained dual-tower recommendation model is saved to the virtual domain recommendation system. The item vectors are saved to the item vector index library.

5. A method for recommending items in a virtual space, characterized in that, include: Obtain the real-domain user profile characteristics and virtual-domain user profile characteristics of the target users; Based on the recommendation request, the real-domain user profile features are input into the trained inter-domain mapping model to obtain the real-domain user profile vector; The user profile vector is obtained using the trained dual-tower recommendation model, and the inner product between the user profile vector and each item vector is calculated. Based on the calculation result, the item to be recommended is found and recommended to the target user. The inter-domain mapping model and the dual-tower recommendation model are obtained by the virtual space item recommendation model training method according to any one of claims 1-4.

6. A training device for a virtual space item recommendation model, characterized in that, include: The training sample construction module is suitable for constructing training samples corresponding to users that overlap in the real domain and the virtual domain. The training samples include: real-world user characteristics, virtual-world user characteristics, virtual-world item characteristics, and user-item behavior association data; The user vector acquisition module is adapted to input the real-domain user features into the inter-domain mapping model to obtain real-domain user vectors; the inter-domain mapping model includes a fully connected layer network model. The recommendation model construction module fuses the real-domain user vector and the virtual-domain user vector corresponding to the virtual-domain user features to obtain a user vector, which serves as the input to the user tower in the dual-tower recommendation model; it also uses the item vector corresponding to the virtual-domain item features as the input to the item tower in the dual-tower recommendation model. The user tower in the dual-tower recommendation model includes a first neural network layer, which performs a non-linear mapping on the input vector to obtain a dense user vector; the item tower in the dual-tower recommendation model includes a second neural network layer, which performs a non-linear mapping on the input vector to obtain a dense item vector. The recommendation model training module is adapted to jointly train the inter-domain mapping model and the dual-tower recommendation model using the user-item behavior association data to obtain an item recommendation model. Specifically, it calculates the similarity between the user vector and the item vector; constructs or selects a loss function based on the similarity; inputs the training samples into the inter-domain mapping model and the dual-tower recommendation model for joint training, using the user-item behavior association data, and updates the model parameters according to the backpropagation algorithm to reduce the loss value of the loss function. This process is iterated repeatedly until the loss value of the loss function no longer decreases, resulting in a trained item recommendation model.

7. An electronic device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the virtual space item recommendation model training method as described in any one of claims 1-4, and / or to perform the operation corresponding to the virtual space item recommendation method as described in claim 5.

8. A computer storage medium storing at least one executable instruction, the executable instruction causing a processor to perform an operation corresponding to the virtual space item recommendation model training method as described in any one of claims 1-4, and / or to perform an operation corresponding to the virtual space item recommendation method as described in claim 5.

Images