Inter-domain transition recommendation system based on user interests and preferences

Abstract

To provide an inter-domain shift recommendation system and an inter-domain migration recommendation system on the basis of interest or preference of a user.SOLUTION: The inter-domain shift recommendation system includes a user data acquisition module and a recommendation module. The user data acquisition module acquires user account information, action habit information, an application scenario, and a history action record in a target domain indicating used by the application scenario and formed according to a specific type of media resource of the application scenario, and acquires the account information, action habit information, and history action record of a user. the recommendation module recommends a media resource in the target domain to the user on the basis of the account information and the action habit information of the user and the history action record in the target domain.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to the field of media resource recommendation, in particular cross-domain recommendation based on user interests and preferences. transition Concerning recommendation systems. [Background technology]

[0002] With the rapid development of the Internet, media resources Recommendation is used in all aspects of people's daily lives. For example, in reading apps, User Recommend books that interest you You can, A long and short video application can recommend long and short videos or short videos that are of interest to a user, and an audio application can recommend storytelling, audiobooks, radio dramas, etc. to a user, but these applications have no data interaction with each other, and cannot optimize their own recommendation capabilities based on available information in other applications. Summary of the Invention [Problem to be solved by the invention]

[0003] The present invention is intended to solve at least one of the above technical problems. [Means for solving the problem]

[0004] Inter-domain communication based on user interests transition The recommendation system includes a user data acquisition module 101 and a recommendation module 102 . The user data acquisition module 101 acquires user account information, behavioral habit information, and Application scenario and, Application scenario Used by Application scenario and a media resource space formed by a particular type of media resource corresponding to the Target Domain Obtain historical behavior records in. In this application, for a user from each domain, the media resources in that domain that the user is interested in can be predicted based on its account information, behavioral habit information, and historical behavior records in that domain, and the domain is the above-mentioned target domain. The target domain in this application is: Transition recommendation It may be any of the various domains supported by the system.

[0005] In fact, the embodiment of the present application can recommend different types of media resources to a user, although there are various applications. However, for the same user, there is a commonality in hobbies, for example, the user likes history. Videos, audios, and books are likely to be historical, so making full use of this common information can help predict the user's preferences more accurately. This is also the reason that the embodiment of the present invention can recommend different types of media resources to a user based on the user's interest preferences. Ku Between domains transition One of the reasons for building a system is to transition Through this, we can provide cross-domain information relevant to individual recommendation systems for individual applications. transition A comprehensive recommendation system for transition This allows for cross-domain transition The system can provide recommendation services corresponding to each domain it supports, and can realize the fusion of common information from different domains and highly accurate recommendations for each domain. Between the domains in this application transitionThe domains involved in the above include at least one media resource space formed by a video-type media resource played in a video playback scene, where the specific type in a video playback scene refers to the type video; a media resource space formed by an audio-type media resource played in an audio playback scene, where the specific type in an audio playback scene refers to the type audio; and a media resource space formed by a text-type media resource played in a book browsing scene, where the specific type in a book browsing scene refers to the type text. That is, a personalized recommendation system corresponding to applications related to these domains is developed across domains. transition By doing so, transition The system can be obtained. The user data acquisition module 101 acquires user account information, behavioral habit information, and cross-domain information that can be used in the recommendation scene. transition The domains supported by the system (target D main) In By entering historical behavior records, transition System Recommendation The results can be obtained by targeting D To users in the main Recommendation Media used R It may be the source.

[0006] The recommendation module 102 recommends media resources in a target domain to the user based on the user's account information, behavioral habit information, and historical behavior records in the target domain, where the recommendation module is obtained by transitioning common association preference information in at least two domains. The recommendation module 102 is transition Inter-domain Recommended The system core component is used, among other things, to perform the following actions: The user's account information and history are used to identify a specific type of media resource for the user in the target domain. preference Media resources in the target domain that characterize the information preference Predict information. Predicting user behavior instruction preference information that characterizes the user's behavioral tendency in a target domain based on the user's account information, behavioral habit information, and static attribute information; The recommended media resources are predicted from the media resource preference information of the target domain and the user behavior instruction preference information. Specifically, the account information is identifier, The attribution includes at least one of the following: account affiliation, account activity frequency, and account level. The static attribute information includes at least one of age information, gender information, educational background information, regional information, and profile information of the user to which the account belongs, and age information, gender information, educational background information, regional information, and profile information of the user to which the account belongs. The behavioral habit information includes at least one of whether a user belonging to an account has a certain interest in the long term and whether he / she likes to consume the same media resource repeatedly. Historical behavior information: The historical behavior of the consumption of media resources of the user to which the account belongs.

[0007] The recommendation module described herein performs the following steps as shown in FIG. training The neural network is constructed from: Step S101. Obtain sample data corresponding to at least one batch of sample users, including account information, static attribute information, behavioral habit information, historical behavior information of at least one sample domain, and feedback behavior information of the sample users. In the embodiment of the present specification, training is batch training However, for each batch trainingYou may use step S102. Of course, the sample user's account information, static attribute information, and behavioral habit information included in the sample data are as described above, but the sample domain is between The historical behavior information is the behavior of the sample user in the domain supported by the recommendation system, for example, in a video application, the video viewing behavior, which includes what videos were viewed, how much, whether comments, "likes" or ratings were made, etc. The feedback behavior information specification is based on the sample user's account information, static attribute information, behavioral habit information, and historical behavior information in at least one sample domain, such as whether clicked, whether viewed, whether commented, whether expressed likes or dislikes, etc. speculation Consumption information of media resources recommended by the recommendation system.

[0008] Step S102. For each user's sample data for each domain, perform the following operations: The sample domain is input into the first network for distributed activation processing, and fusion feature parameters corresponding to each of the second networks are obtained, and the neural network includes second networks corresponding to at least two domains that are used to output common associative preference information of the media resources of the corresponding domains themselves, and the neural network obtains fusion feature parameters corresponding to each of the second networks.

[0009] In the embodiment of the present invention, although the domains are different, the interests shown by users in different domains can have a common association expression that matches. For example, in a domain related to a reading application, a user is likely to like both history books and biographies, and in a domain related to a video application, a user is likely to like both historical movies and biographical movies. This is one of the construction foundations of the inventive idea of ​​the embodiment of the present application. ofThe association information is recorded in the recommendation system of each domain, but usually the recommendation system of each domain is built with a neural network, and this information is recorded in the parameter matrix of the neural network. That is, the second network can come from the recommendation system of the subset corresponding to each domain, and the cross-domain association with the second network can be transition By training a recommendation system, the second network can be used to transition For example, a reading app Recommendation System, video app Recommendation System, audio app Recommendation Among these recommendation systems, of Association preference The network in which the information is recorded is then fed into the neural network above. transition The second network in this neural network is itself a cross-domain network. transition This information is then used to record the domain-specific information. transition In order to be applied to the recommendation system, it needs to be iteratively optimized according to the neural network training process. Therefore, the second network can be considered to be linked to the neural network by inter-domain transition from the recommendation system corresponding to the corresponding domain. Recommendation The system is of course maintained by the corresponding software development team and can be used directly after obtaining a license, and this application does not describe how to build it.

[0010] Between the first network and each second network, an activation layer sequence is provided, the second network having an activation layer sequence, the activation layer sequence including at least two activation layers connected in series; Any second network Regarding The above Example Domain is input to the first network to characterize a feature property, Example Domainand applying the feature to an activation layer close to the first network in a corresponding activation layer sequence of the second network. Example Domain For any activation layer sequence that inputs a feature and outputs a corresponding fused feature parameter by an activation layer of the activation layer sequence close to the second network, the activation functions of adjacent activation layers in the activation layer sequence do not match. In the embodiment of the present application, the structure of the first network is not limited, and it is sufficient if feature extraction is possible. However, there are many neural network structures for feature extraction, and for example, partial structures in CNN or DNN can be selected, so please refer to the prior art and no limitation is made here. This specification does not limit the number of activation layers in a series of activation layers, which is preferably two or three. The activation layer sequence is essentially: Example Domain The aim of the present invention is to map the characteristics to a constant that can act as the weight of the information output by the second network. The embodiment of the present invention does not limit the activation function of the activation layer, and the inter-domain transition, which is the fusion feature parameter, can be obtained by at least two consecutive activations, as long as the activation functions of adjacent activation layers in the activation layer sequence do not match. Recommendation It is possible to obtain constants that better suit the system. An embodiment of the present invention is input to a first network. Example Domain Not limited to the expression method of Example Domain The identification information of the corresponding

[0011] The fusion feature parameters corresponding to each of the second networks are respectively associated with the input second networks. Enter The information output from each of the second networks is fused to generate a common model based on the sample domain. of Association preference Get information. The parameters in each second network are used to record the common association priority information of the corresponding domain of the second network. After inputting the fusion feature parameters into the second network, the common association priority information of the domain of the second network is recorded. of Association preference The effect of information reduction is achieved. That is, if the domain corresponding to the second network and the sample domain are close to each other, the important thing is to take it into consideration when making cross-domain recommendations. If they are far away, the comparison is not important and is not taken into consideration when making cross-domain recommendations. Thus, the common association based on the sample domain is reduced. preference By obtaining information, realizing the fusion of information across domains and adapting to specific sample domains, common associations in the sample domains can be obtained. preference The technical effect of the information is obtained. The above Example Domain The historical behavior information for is input to a third network for embedding feature extraction, Example Domain We obtain the corresponding historical behavior features. In the embodiments of the present application, the structure of the third network is not limited, and it is sufficient to perform embedded feature extraction. However, there are a great number of neural network structures for embedded feature extraction, and for example, a substructure in CNN or DNN can be selected. Therefore, the prior art will be used as a reference and no limitations will be imposed here.

[0012] Corresponding historical behavior of the sample domain Features and the common association preference information of the sample domain are input together into a fourth network for encoding to obtain corresponding media resource preference information of the sample user in the sample domain. In the embodiment of the present invention, the structure of the fourth network is not limited, and it is sufficient to encode it. However, there are a great number of neural network structures to be encoded, and for example, a partial structure in a transformer can be selected and used. Therefore, please refer to the prior art and no restrictions are imposed here. The purpose of encoding is to find historical behavior that corresponds to the sample domain. Featuresand the common associated preference information of the sample domain to obtain the corresponding media resource preference information of the sample user in the sample domain.

[0013] The account information, static attribute information and behavioral habit information of the sample user are input to a fifth network to perform feature extraction based on fusion, and behavioral instruction preference information corresponding to the sample user is obtained. In the embodiments of the present application, the structure of the fifth network is not limited, and it is sufficient to perform feature extraction based on fusion. For example, the fifth network can be obtained by using an organic combination of a convolution layer, a pooling layer, a max pooling layer, etc., and the construction method thereof is not limited, and those skilled in the art can design it according to actual circumstances.

[0014] (6) Input the media resource preference information corresponding to the sample domain and the behavior instruction preference information corresponding to the sample user into a sixth network to obtain the interested resources personalized by the sample user under the sample domain. In the embodiments of the present application, the structure of the sixth network is not limited, and it is sufficient to perform feature extraction based on fusion. For example, the sixth network can be obtained by using an organic combination of a convolution layer, a pooling layer, a max pooling layer, etc., and the construction method thereof is not limited, and those skilled in the art can design it according to actual situations. (7) The loss generated by the neural network is determined based on the resource of interest and the feedback behavior information in the sample data for the resource of interest. The embodiments of the present invention are not limited to the method of measuring the loss, and a method of determining the loss based on predicted resources and feedback information from those resources is commonly seen in recommendation neural network training, but this will not be mentioned. (8) Adjust parameters of the first network, the second network, the third network, the fourth network, the fifth network, and the sixth network according to the loss feedback.

[0015] Before the neural network training process, steps (1) to (7) are all forward feedback steps, and step (8) is a feedback adjustment step. The forward feedback, reverse adjustment, and reverse adjustment adjustment methods, as well as the conditions for ending neural network training, are all existing technology and can be found in other materials, and no further reference will be made here. In an embodiment of the present invention, the neural network is trained using multiple batch Since the neural network is huge, it is difficult to control the convergence speed and accuracy. In contrast, in the embodiment of this application, the convergence is divided into two consecutive batch During training, we perform neural network pruning, i.e., we perform two steps to optimize the neural network: pruning and training. batch They propose that by repeating the process over and over again, it is possible to speed up convergence, simplify the neural network while still ensuring sufficient accuracy, and improve the training speed.

[0016] Specifically, the at least one batch of sample data corresponding to the sample user includes at least two batches of sample data corresponding to the sample user; training The method is as follows: Target Batch Based on the sample data, training and obtaining an intermediate state neural network based on a difference between the accuracy information of the intermediate state neural network and a preset accuracy information target value. Pruning do Pruning After getting the neural network, Pruning Later neural networks training The second Target Batch This is based on sample data.

[0017] First target batch sample and second target batch The sample of is any two consecutive batchcan be understood as a sample of two consecutive batch In training samples, the accuracy of the intermediate state of the neural network is used to determine the Pruning By carrying out the following, Pruning The amount of the later neural network can be reduced, the training speed can be increased, and the training speed and rapid convergence can be achieved. In addition, the target value of the accuracy information in this embodiment can be set according to the actual situation. The accuracy information can be obtained by using the neural network accuracy information evaluation software to calculate the accuracy information of the neural network. Pruning There are several methods and rules for cutting, and the specific cutting method can be implemented based on the prior art. The embodiments of this application are not limited to such overlap cutting, batch It should be noted that the reference adjustment training method based on the above is an inventive idea proposed in the embodiments of the present application to make the neural network converge quickly and ensure accuracy.

[0018] The embodiment of the present invention is based on the user's interests and preferences. Ku Between domains transition In this scenario, we provide a recommendation system that supports individual applications. Recommended Inter-domain transition Between domains via transition Integrated Recommended On the stem transition This allows cross-domain transition The system can provide recommendation services corresponding to each domain it supports, and can realize the fusion of common information from different domains and highly accurate recommendations for each domain. In another aspect, an embodiment of the present invention provides an electronic device including at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the at least one processor executes the instructions stored in the memory to perform the above-mentioned cross-domain navigation based on a user's interest preferences. transitionRealize a recommendation system. On the other hand, the embodiment of the present invention, when executed by a processor, performs the above-mentioned inter-domain search based on a user's interest preferences. transition A computer program product is provided that includes a computer program or instructions for implementing the recommendation system.

[0019] The above are merely preferred embodiments of the present application, and do not limit the embodiments of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application. Effect of the Invention

[0020] In the present invention, the user data acquisition module acquires user account information, behavioral habit information, and Application scenario and, Application scenario Used by Application scenario and a media resource space formed corresponding to a particular type of media resource. In the target domain The user data acquisition module is a module for acquiring a user's account information, behavioral habit information, and historical behavior records, and the user data acquisition module is a module for acquiring a user's account information, behavioral habit information, and historical behavior records. Recommendation The module provides the user with media resources in the target domain based on the user's account information, behavioral habit information, and historical behavior records in the target domain. Recommendation Here, Recommendation A module is obtained by transferring common associative preference information in at least two domains. [Brief description of the drawings]

[0021] [Figure 1] FIG. 1 is a virtual structural diagram of an inter-domain transition recommendation system based on a user's interests and preferences provided by an embodiment. [Diagram 2]FIG. 2 is a flow diagram of a neural network training method provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0022] FIELD OF THE DISCLOSURE The present invention relates to the field of media resource recommendation, and in particular to cross-domain media resource recommendation based on user interest preferences. transition Concerning recommendation systems. EXAMPLES

[0023] Inter-domain communication based on user interests Transition recommendation The system includes a user data acquisition module and a recommendation module. The user data acquisition module 101 acquires user account information, behavioral habit information, and Application scenario and, Application scenario Used by Application scenario and a media resource space formed corresponding to a particular type of media resource. In the target domain This is a module for obtaining historical behavior records. The recommendation module 102 recommends media resources in the target domain to the user based on the user's account information, behavioral habit information, and history behavior records in the target domain, where: Recommendation The module 102 obtains common associated preference information in at least two domains by transitioning the common associated preference information in at least two domains. The recommendation module 102 is transition Inter-domain Recommendation It is the system core component and is used to perform, among other things: The user's account information and history are used to identify a specific type of media resource for the user in the target domain. preference Media resources in the target domain that characterize the information preference Predict information. Preference information indicating a user's behavior that characterizes the user's behavioral tendency in the target domain is predicted based on the user's account information, behavioral habit information, and static attribute information. Recommendation The media resources to be played are predicted based on the media resource preference information of the target domain and the user behavior instruction preference information.

[0024] As shown in FIG. 2, the recommendation module 102 performs the following steps: training The neural network is constructed from: Step S101. Obtain sample data corresponding to at least one batch of sample users, including the sample user's account information, static attribute information, behavioral habit information, historical behavior information of at least one sample domain, and feedback behavior information; and obtain the sample data corresponding to at least one batch of sample users, including the sample user's account information, static attribute information, behavioral habit information, historical behavior information of at least one sample domain, and feedback behavior information. Step S102. For each user's sample data for each domain, perform the following operations: The sample domain is input into the first network for distributed activation processing, and fusion feature parameters corresponding to each of the second networks are obtained, and the neural network includes second networks corresponding to at least two domains that are used to output common associative preference information of the media resources of the corresponding domains themselves, and the neural network obtains fusion feature parameters corresponding to each of the second networks. The fusion feature parameters corresponding to each of the second networks are respectively associated with the input second networks, and the information output from each of the second networks is fused to generate a common feature based on the sample domain. of Association preference Get information. The above Example Domain The historical behavior information for is input to a third network for embedding feature extraction, Example DomainWe obtain the corresponding historical behavior features. Corresponding historical behavior of the sample domain Features and the common association preference information of the sample domain are input together into a fourth network for encoding to obtain corresponding media resource preference information of the sample user in the sample domain. The account information, static attribute information and behavioral habit information of the sample user are input to a fifth network to perform feature extraction based on fusion, and behavioral instruction preference information corresponding to the sample user is obtained. Media resources corresponding to sample domains preference Information and sample user actions are preference The information is input into a sixth network to show obtaining personalized interest resources for a sample user under a sample domain. Neural Networks The loss resulting from is determined based on feedback behavior information of the resource of interest and sample data for the resource of interest. Parameters of the first network, the second network, the third network, the fourth network, the fifth network, and the sixth network are adjusted according to the loss feedback.

[0025] In one embodiment, the second network is from a recommendation system corresponding to the corresponding domain, and ... transition is linked to the neural network by In one embodiment, an activation layer sequence is provided between the first network and each second network, the second network having an activation layer sequence, the activation layer sequence including at least two activation layers connected in series, Any second network About , the above Example Domain is input to the first network to characterize a feature property, Example Domain and applying the feature to an activation layer close to the first network in a corresponding activation layer sequence of the second network. Example DomainA feature is input and a corresponding fused feature parameter is output by an activation layer of an activation layer sequence similar to the second network. For any activation layer sequence, the activation functions of adjacent activation layers in the activation layer sequence do not match.

[0026] In one embodiment, the at least one sample domain comprises at least one of the following sample domains: A media resource space formed by the video-type media resources played in a video playback scene; A media resource space formed by the audio-type media resources played in the audio playback scene; A media resource space formed by text-type media resources played in a book browsing scene.

[0027] In one embodiment, the at least one batch of sample data corresponding to the sample user includes at least two batches of sample data corresponding to the sample user: training The method further comprises the steps of: 1. Target Batch The neural network is trained based on the sample data to obtain an intermediate state neural network. Based on the difference between the accuracy information of the intermediate state neural network and a preset accuracy information target value, the intermediate state neural network is Pruning do Pruning After that we get a neural network. Pruning Later neural networks training The second Target Batch This is based on sample data.

[0028] In another aspect, embodiments of the present invention provide a method for determining whether a user's interests are satisfied based on the user's interests. Ku Between domains transitionTo implement the recommendation system, a computer readable storage medium is provided having stored thereon at least one instruction or at least one program that is loaded and executed by a processor. In another aspect, an embodiment of the present invention provides an electronic device including at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the at least one processor executes the instructions stored in the memory to perform the above-mentioned cross-domain navigation based on a user's interest preferences. transition Realize a recommendation system.

[0029] On the other hand, the embodiment of the present invention, when executed by a processor, performs the above-mentioned inter-domain search based on a user's interest preferences. transition A computer program product is provided that includes a computer program or instructions for implementing the recommendation system. The embodiment of the present invention is based on the user's interests and preferences. Ku Between domains transition In this scenario, we provide a recommendation system that supports individual applications. Recommended Inter-domain transition Between domains via transition Integrated Recommended On the stem transition This allows for cross-domain transition The system can provide recommendation services corresponding to each domain it supports, and can realize the fusion of common information from different domains and highly accurate recommendations for each domain.

Claims

1. An inter-domain transition recommendation system based on a user's interests and preferences, comprising: a user data acquisition module; and a recommendation module; The user data acquisition module includes: User account information, Behavioral habit information, A historical behavior record in a target domain indicating an application scenario and a media resource space used by the application scenario and formed corresponding to a specific type of media resource of the application scenario; This is a module for obtaining The recommendation module is a module for recommending media resources in the target domain to the user according to the user's account information, the behavioral habit information, and a history behavior record in the target domain; The recommendation module is obtained by transitioning common association preference information in at least two domains; The recommendation module: predicting media resource preference information in the target domain from the user's account information and the historical behavior record, the media resource preference information characterizing the user's preference information for a particular type of media resource in the target domain; predicting user behavior instruction preference information that characterizes the user's behavioral tendency in the target domain based on the user's account information, behavioral habit information, and static attribute information; Predicting recommended media resources according to the media resource preference information of the target domain and the user's behavior instruction preference information; The inter-domain transition recommendation system according to the present invention is characterized by executing the above steps.

2. The recommendation module is constructed from a neural network, and the neural network comprises: Obtaining at least one batch of sample data corresponding to the sample user, the sample data including account information, static attribute information, behavioral habit information, and historical behavior information and feedback behavior information of at least one sample domain of the sample user; For each sample domain, for each sample user, The sample domain is input into a first network for distributed activation processing, and a fusion feature parameter corresponding to each of the second networks is obtained, and the neural network includes at least two second networks corresponding to the at least two domains, which are used to output common association preference information of the media resource of the corresponding domain itself; The fusion feature parameters corresponding to the second networks are input to the second networks, respectively, and the information output from the second networks is fused to obtain common associated preference information based on the sample domain; inputting historical behavior information for the sample domain into a third network for embedding feature extraction to obtain historical behavior features corresponding to the sample domain; Inputting the historical behavior features corresponding to the sample domain and the common association preference information of the sample domain together into a fourth network for encoding to obtain corresponding media resource preference information of the sample user in the sample domain; The account information, static attribute information, and behavioral habit information of the sample user are input to a fifth network to perform feature extraction based on fusion, and behavioral instruction preference information corresponding to the sample user is obtained; Input the media resource preference information corresponding to the sample domain and the behavior instruction preference information corresponding to the sample user into a sixth network to obtain personalized interest resources of the sample user in the sample domain; determining a loss resulting from the neural network based on the resource of interest and feedback behavioral information of the sample data with respect to the resource of interest; Adjusting parameters of the first network, the second network, the third network, the fourth network, the fifth network, and the sixth network according to the loss feedback. Performing an operation; and They are trained using a training method that includes 2. The system of claim 1 .

3. The account information includes at least one of an account identifier, an account affiliation, an account activity frequency, and an account level; The static attribute information includes at least one of age information, gender information, educational background information, region information, and profile information of the user to which the account belongs, The behavioral habit information includes at least one of whether the user to which the account belongs has a certain interest in the long term and whether the user likes to consume the same media resource repeatedly; The historical behavior information is a historical behavior of a user to which the account belongs, the user consuming media resources.

3. The system of claim 2.

4. The second network comes from a recommendation system corresponding to the corresponding domain and is linked to the neural network by an inter-domain transition.

4. The system of claim 3.

5. Between the first network and each of the second networks, an activation layer sequence is provided, the activation layer sequence including at least two activation layers connected in series, for each second network, the sample domain is input to the first network to characterize a feature property, a sample domain feature is obtained, the sample domain feature is input to an activation layer of the corresponding activation layer sequence of the second network close to the first network, and a corresponding fused feature parameter is output by an activation layer of the activation layer sequence close to the second network; For any given activation layer sequence, the activation functions of adjacent activation layers in the activation layer sequence do not match.

5. The system of claim 4.

6. The at least one sample domain is a sample domain a media resource space formed by the video type media resources played in the video playback scene; a media resource space formed by the audio-type media resources played in the audio playback scene; The media resource space formed by the text-type media resources played in the book browsing scene, 6. The system of claim 5, further comprising at least one of:

7. the at least one batch of sample data corresponding to the sample user includes at least two batches of sample data corresponding to the sample user; The training method further comprises: Training a neural network based on the sample data of a first target batch to obtain an intermediate state neural network; pruning the intermediate state neural network based on a difference between accuracy information of the intermediate state neural network and a preset accuracy information target value to obtain a pruned neural network; training the pruned neural network based on a second target batch of sample data; 7. The system of claim 6, further comprising:

8. At least one instruction or at least one program is stored, which is loaded and executed by a processor to realize the inter-domain transition recommendation system based on a user's interest preference according to any one of claims 1 to 7. A computer-readable storage medium comprising:

9. 1. An electronic device comprising: at least one processor; and a memory communicatively coupled to the at least one processor, The memory stores instructions executable by the at least one processor, and the at least one processor executes the instructions stored in the memory to realize the inter-domain transition recommendation system based on a user's interests and preferences according to any one of claims 1 to 7.

1. An electronic device comprising:

Images