Information recommendation method and device, electronic equipment and storage medium

Abstract

The invention provides an information recommendation method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring a preset information base in any one of a plurality of domains, a search word input by a user and user information of the user; wherein the preset information base comprises at least two pieces of information to be retrieved; processing the user information, the retrieval word and each piece of to-be-retrieved information based on a preset cross-domain recommendation model, and obtaining a processing result; wherein the cross-domain recommendation model is obtained by training training samples of multiple domains by using an initial multi-target model, and one domain corresponds to one target in the initial multi-target model; and recommending information to the user based on the processing result. By means of the method, the retrieval performance in multiple fields can be improved at the same time, and the generalization ability is higher.

Description

Technical Field

[0001] The present disclosure relates to, but is not limited to, the field of artificial intelligence technology, and in particular to an information recommendation method and device, an electronic device, and a storage medium. Background Art

[0002] The recommendation system uses recommendation algorithms to generate a list of information that the user may be interested in based on the user's historical behavior, user interest preferences, etc. For example, if user A is a sports fan, then the information recommended to A is mostly sports-related information. In this case, it is difficult for user A to obtain other related information such as music or literature, and there is an information cocoon problem; in addition, due to the lack of historical information about new users or new materials to be recommended, it is difficult to provide recommendation services. This problem is generally called the cold start problem.

[0003] In order to solve the information cocoon or cold start problem in the recommendation system, the recommendation performance can be improved by sharing information and using cross-domain recommendation methods. However, the current cross-domain recommendation methods have the problem of limited scope of application. Summary of the invention

[0004] The present invention provides an information recommendation method and device, an electronic device, and a storage medium.

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

[0006] Acquire a preset information base under any domain of the multiple domains, a search term input by a user, and user information of the user; wherein the preset information base includes at least two pieces of information to be retrieved;

[0007] The user information, the search terms and each of the information to be searched are processed based on a preset cross-domain recommendation model to obtain a processing result; wherein the cross-domain recommendation model is obtained by training the training samples of the multiple domains using an initial multi-objective model, and one domain corresponds to one objective in the initial multi-objective model;

[0008] Recommend information to the user based on the processing result.

[0009] In some embodiments, the processing result includes the user's interest level in each of the to-be-retrieved information based on the search term;

[0010] The recommending information to the user based on the processing result includes:

[0011] The interest levels of each of the to-be-retrieved information are ranked, and information is recommended to the user based on the ranking results.

[0012] In some embodiments, the method further comprises:

[0013] Acquire training samples of the multiple domains, wherein one training sample includes sample user information in the domain, a sample search term, sample search information corresponding to the sample search term, and a label corresponding to the sample search information in the domain;

[0014] In response to the existence of a target training sample satisfying a preset information association condition in the training samples of different domains, information association processing is performed on the target training samples from different domains;

[0015] The initial multi-objective model is trained based on each training sample after information association processing to obtain the trained cross-domain recommendation model.

[0016] In some embodiments, in response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, performing information association processing on the target training samples from different domains includes:

[0017] In response to the existence of sample user information identifying the same sample user in training samples from different domains, the same sample user having the same sample search term in different domains, and target training samples having different sample retrieval information for the same sample search term in different domains, the sample user information of the target training samples from different domains is fused; and / or, the sample retrieval information corresponding to the same sample search term in the target training samples from different domains is fused.

[0018] In some embodiments, in response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, performing information association processing on the target training samples from different domains includes:

[0019] In response to the existence of sample user information identifying the same sample user in training samples from different domains, and the same sample user corresponding to target training samples with the same sample retrieval information in different domains, sample retrieval terms corresponding to the same sample retrieval information in target training samples from different domains are fused.

[0020] In some embodiments, in response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, performing information association processing on the target training samples from different domains includes:

[0021] In response to the existence of sample user information identifying the same sample user in training samples in different domains, but the preset target information in the sample user information is only included in target training samples in some domains, the preset target information is shared with the sample user information corresponding to the training samples associated with the same sample user in the domain that does not contain the preset target information.

[0022] In some embodiments, the initial multi-objective model is trained based on each training sample after information association processing to obtain the trained cross-domain recommendation model, including:

[0023] For each training sample, using the initial multi-objective model to determine a prediction result of the sample retrieval information in each training sample for a corresponding target;

[0024] For each training sample, based on the prediction result of each sample retrieval information for the corresponding target and the label of each sample retrieval information in the domain, calculate the loss value corresponding to each target;

[0025] Determining a total loss of the initial multi-objective model based on the loss value of each objective;

[0026] The parameters of the initial multi-objective model are adjusted based on the total loss to obtain the trained cross-domain recommendation model.

[0027] In some embodiments, the step of using the initial multi-objective model to determine a prediction result of the sample retrieval information in each training sample for a corresponding objective includes:

[0028] Using the feature extraction layer in the initial multi-objective model, feature extraction is performed on the sample user information, the sample search term, and the sample search information corresponding to the sample search term to obtain a first feature vector corresponding to each training sample; wherein the first feature vector includes a sample user feature vector, a sample search term feature vector, and a sample search information feature vector;

[0029] Using the feature normalization layer in the initial multi-objective model to batch normalize the sample user feature vector, the sample search term feature vector, and the sample search information feature vector to obtain a normalized second feature vector; wherein the feature normalization layer includes normalization parameters within the domain and normalization parameters between domains;

[0030] Determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second eigenvector using the initial multi-target model;

[0031] The adjusting the parameters of the initial multi-objective model based on the total loss to obtain the trained cross-domain recommendation model includes:

[0032] Parameters in the initial multi-objective model, including normalized parameters within the domain and normalized parameters between the domains, are adjusted based on the total loss to obtain the trained cross-domain recommendation model.

[0033] In some embodiments, the using the initial multi-objective model to determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second feature vector includes:

[0034] Using each expert layer in the initial multi-objective model to perform feature processing on the normalized second eigenvector to obtain a third eigenvector output by each expert layer;

[0035] Processing the normalized second eigenvector using the target-associated gating layer in the initial multi-target model to obtain the weights of each expert layer output by the gating layer;

[0036] Performing weighted fusion on the third feature vectors output by each of the expert layers based on the weights output by the target-associated gating layer to obtain a fourth feature vector;

[0037] Based on the weighted fused fourth eigenvector corresponding to each target and the target sub-network corresponding to each target in the initial multi-target model, a prediction result for each training sample output by each target sub-network for the corresponding target is obtained.

[0038] According to a second aspect of an embodiment of the present disclosure, there is provided an information recommendation device, including:

[0039] A first acquisition module is configured to acquire a preset information library under any domain of the multiple domains, a search term input by a user, and user information of the user; wherein the preset information library includes at least two pieces of information to be retrieved;

[0040] A processing module, configured to process the user information, the search terms and each of the information to be searched based on a preset cross-domain recommendation model and obtain a processing result; wherein the cross-domain recommendation model is obtained by training the training samples of the multiple domains using an initial multi-objective model, and one domain corresponds to one objective in the initial multi-objective model;

[0041] A recommendation module is configured to recommend information to the user based on the processing result.

[0042] In some embodiments, the processing result includes the user's interest level in each of the to-be-retrieved information based on the search term;

[0043] The recommendation module is further configured to sort the interest levels of each of the to-be-retrieved information, and recommend information to the user based on the sorting results.

[0044] In some embodiments, the apparatus further comprises:

[0045] A second acquisition module is configured to acquire training samples of the multiple domains, wherein one training sample includes sample user information in the domain, sample search terms, sample search information corresponding to the sample search terms, and a label corresponding to the sample search information in the domain;

[0046] an association module, configured to perform information association processing on the target training samples from different domains in response to the existence of target training samples that meet a preset information association condition in the training samples from different domains;

[0047] The training module is configured to train the initial multi-objective model based on each training sample after information association processing to obtain the trained cross-domain recommendation model.

[0048] In some embodiments, the association module is further configured to, in response to the existence of sample user information identifying the same sample user in training samples from different domains, and the same sample user corresponding to the same sample search term in different domains, and target training samples having different sample retrieval information for the same sample search term in different domains, fuse the sample user information of the target training samples from different domains; and / or fuse the sample retrieval information corresponding to the same sample search term in the target training samples from different domains.

[0049] In some embodiments, the association module is further configured to fuse the sample search terms corresponding to the same sample retrieval information in the target training samples from different domains in response to the presence of sample user information identifying the same sample user in training samples from different domains, and the same sample user corresponding to target training samples with the same sample retrieval information in different domains.

[0050] In some embodiments, the association module is further configured to, in response to the presence of sample user information identifying the same sample user in training samples in different domains, but target training samples whose preset target information is only included in some domains, share the preset target information with the sample user information corresponding to the training samples associated with the same sample user in the domain that does not contain the preset target information.

[0051] In some embodiments, the training module is further configured to determine, for each training sample, a prediction result of the sample retrieval information in each training sample for the corresponding target using the initial multi-target model; for each training sample, based on the prediction result of each sample retrieval information for the corresponding target and the label of each sample retrieval information in the domain, calculate the loss value corresponding to each target; based on the loss value of each target, determine the total loss of the initial multi-target model; based on the total loss, adjust the parameters of the initial multi-target model to obtain the trained cross-domain recommendation model.

[0052] In some embodiments, the training module is further configured to use the feature extraction layer in the initial multi-objective model to perform feature extraction on the sample user information, sample search terms, and sample retrieval information corresponding to the sample search terms, respectively, to obtain a first feature vector corresponding to each training sample; wherein the first feature vector includes a sample user feature vector, a sample search term feature vector, and a sample retrieval information feature vector; use the feature normalization layer in the initial multi-objective model to batch normalize the sample user feature vector, the sample search term feature vector, and the sample retrieval information feature vector to obtain a normalized second feature vector; wherein the feature normalization layer includes normalization parameters within the domain and normalization parameters between domains; use the initial multi-objective model to determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second feature vector; adjust the parameters in the initial multi-objective model including the normalization parameters within the domain and the normalization parameters between domains based on the total loss to obtain the trained cross-domain recommendation model.

[0053] In some embodiments, the training module is further configured to perform feature processing on the normalized second eigenvector using each expert layer in the initial multi-objective model to obtain a third eigenvector output by each expert layer; to process the normalized second eigenvector using a target-associated gating layer in the initial multi-objective model to obtain the weights of each expert layer output by the gating layer; to perform weighted fusion on the third eigenvectors output by each expert layer based on the weights output by the target-associated gating layer to obtain a fourth eigenvector; and to obtain a prediction result for the corresponding target for each training sample output by each target subnetwork based on the weighted fused fourth eigenvector corresponding to each target and a target subnetwork corresponding to each target in the initial multi-objective model.

[0054] According to a third aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0055] processor;

[0056] a memory for storing processor-executable instructions;

[0057] Wherein, the processor is configured to execute the method as described in the first aspect above.

[0058] According to a fourth aspect of an embodiment of the present disclosure, there is provided a storage medium, including:

[0059] When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method as described in the first aspect above.

[0060] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0061] In the embodiments of the present disclosure, since the cross-domain recommendation model is obtained by training samples of multiple domains using an initial multi-objective model, one domain corresponds to one objective, the essence of the multi-objective model is to learn information related to multiple fields through information sharing and improve the generalization ability of the model in multiple fields. Therefore, the cross-domain recommendation model obtained by training based on the multi-objective model is used for retrieval and recommendation in the embodiments of the present disclosure, which can achieve better recommendation effects in any domain. Compared with the cross-domain recommendation model that can only work for the target domain but is not helpful for the source domain, the solution of the embodiments of the present disclosure can simultaneously improve the retrieval performance in multiple fields and has stronger generalization ability.

[0062] 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

[0063] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0064] Figure 1 A schematic diagram of a cross-domain recommendation for EMCDR.

[0065] Figure 2 It is a flow chart of an information recommendation method shown in an embodiment of the present disclosure.

[0066] Figure 3 This is an example diagram of a search interface for a game center in an embodiment of the present disclosure.

[0067] Figure 4 This is an example diagram of a search interface of an application center in an embodiment of the present disclosure.

[0068] Figure 5 This is an example diagram of an MMOE model structure.

[0069] Figure 6 This is a detailed structural example diagram of the MMOE model.

[0070] Figure 7 The figure is a diagram of an information recommendation device according to an exemplary embodiment.

[0071] Figure 8 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0072] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. 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.

[0073] The terms "first / second / third" etc. involved in the present disclosure are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first / second / third" etc. can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein.

[0074] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present disclosure belongs. The terms used herein are only for the purpose of describing the present disclosure and are not intended to limit the present disclosure.

[0075] Related technologies When solving the cold start problem in the recommendation system, the first method is to actively solve the cold start problem by designing many decision-making strategies, such as pre-scoring new users or items and then giving them a higher recommendation weight, or using context bundling. This method requires strong domain prior knowledge and relies more on personal understanding of the cold start business, so it is difficult to generalize, and the strategy must be updated all the time to ensure the cold start effect. The second method is to use auxiliary information to help solve the problem of insufficient information in the cold start phase, so that the recommendation model can score correctly. This type of method mainly explores the use of various auxiliary information to improve the performance of cold start recommendations, such as using user attribute information to solve the user's cold start problem, using project attribute information to assist in solving the project cold start problem, using knowledge graphs to solve the cold start problem, and using auxiliary domain information to help solve the cold start problem in the target domain, that is, cross-domain recommendation (CDR) technology.

[0076] Taking the asymmetric cross-domain recommendation method EMCDR (Cross-Domain Recommendation: An Embedding and Mapping Approach) as an example, its recommendation target is users or items with little information in the target domain. For example, for a new user or item, its embedding cannot be found in the target domain, but it is more active in the source domain and its embedding can be found. At this time, the embedding of the target domain can be obtained through the bridge function (also called mapping function) between the target domain and the source domain before making a recommendation. Figure 1 A schematic diagram of a cross-domain recommendation for an EMCDR, such as Figure 1 As shown, it includes three steps: L1, generating embedding vectors of users or items in the source domain and target domain based on the rating matrix; L2, mapping using a bridge function, where the bridge function (mapping function) can be a linear mapping, such as multiplying a parameter matrix for mapping, or using a multi-layer perceptron (MLP) for mapping; L3, making cross-domain recommendations based on the vector of the mapped target domain. When making cross-domain recommendations based on the vector of the mapped target domain, the optimization goal is to minimize the distance between the embedding obtained by mapping and the embedding of the target domain obtained by pre-training, that is, the closer the two embeddings are, the better. The loss function is shown in the following formula (1):

[0077]

[0078] Where f(u) is the bridge function, u t i It is the user embedding of the target domain.

[0079] Generally, cross-domain recommendation technology works better because it directly uses the information of one domain, which is richer than the auxiliary information used by the first three. However, cross-domain recommendation technology usually uses the information of the source domain to initialize an embedding vector as the cold start embedding in the target domain. This will bring two problems: First, the model is not end-to-end, and adopts a segmented training method of training the embedding vector first and then making predictions. However, a good embedding training does not mean that the prediction will have good results, which means that it may not be possible to guarantee that the model can achieve the best prediction performance. The best effect achieved in a certain section may not be the best in the entire system; second, such a method can only work for the target domain, but is of no help to the source domain, and cannot improve the model prediction performance of both domains at the same time.

[0080] In this regard, the present disclosure provides an information recommendation method. Figure 2 is a flow chart of an information recommendation method shown in an embodiment of the present disclosure. Figure 2 As shown, the information recommendation method applied to the above electronic device includes the following steps:

[0081] S11, obtaining a preset information library under any domain of multiple domains, a search term input by a user, and user information of the user; wherein the preset information library includes at least two pieces of information to be searched;

[0082] S12, processing the user information, the search terms and each of the to-be-searched information based on a preset cross-domain recommendation model and obtaining a processing result; wherein the cross-domain recommendation model is obtained by training the training samples of the multiple domains using an initial multi-objective model, and one domain corresponds to one objective in the initial multi-objective model;

[0083] S13. Recommend information to the user based on the processing result.

[0084] In the embodiments of the present disclosure, the information recommendation method can be implemented by the terminal / server alone; or it can be implemented by the terminal and the server in collaboration. For example, the terminal alone undertakes the information recommendation method described below, or the terminal sends an information retrieval request carrying a search term to the server, and the server executes the information recommendation method based on the received information retrieval request, determines the target retrieval information corresponding to the search term, and recommends the target retrieval information to the terminal for display by the terminal.

[0085] The electronic device for information retrieval provided by the embodiments of the present disclosure may be various types of terminal devices or servers, wherein the server may be an independent physical server, or a server cluster or 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, and big data and artificial intelligence platforms; the terminal may be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, smart car connection, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited by the present disclosure.

[0086] In the disclosed embodiment, the information recommendation method is aimed at retrieval recommendation in the information retrieval scenario, and the cross-domain recommendation model used is a model trained for training samples from different retrieval scenarios (domains); different retrieval scenarios can be, for example, game retrieval scenarios in a game center and application retrieval scenarios in an application store; for another example, it can also be a product retrieval scenario in a shopping application and a retrieval scenario in a sharing application. It should be noted that the disclosed embodiment does not limit the retrieval scenarios targeted by the cross-domain recommendation model, nor the number of retrieval scenarios included.

[0087] In step S11, the information to be retrieved stored in the preset information library may be information in any domain targeted by the cross-domain recommendation model. For example, for the game search scenario of the game center, the information to be retrieved may be identification information such as the name of the game. The search term entered by the user is usually the search term entered by the user for the content of the information to be retrieved in the preset information library. For example, in the game search scenario, the search term may be a keyword of the game name, or an attribute identifier of the game (for example, a competitive game or a casual game). User information may include the user's account information (including account name, account level, etc.) in the search scenario corresponding to the preset information library, and may also include the user's identity attribute information such as age, gender and occupation, and related user portraits such as preference information.

[0088] In step S12, the electronic device uses a preset cross-domain recommendation model to extract features from the user information, search terms, and each piece of information to be retrieved in the preset information library and performs feature processing to obtain a processing result, and then recommends the information in the preset information library to the user based on the processing result in step S13. It should be noted that the processing result may include the user's interest level in each piece of information to be retrieved based on the search terms. At this time, the electronic device may recommend the information in the preset information library to the user based on the interest level. In addition, the processing result may also include the overall fitness of the user and the preset information library. The overall fitness can be comprehensively determined by the user's interest level in each piece of information to be retrieved. If the overall fitness representation of the user and the preset information library does not match, the electronic device may recommend information outside the preset information library to the customer.

[0089] In the disclosed embodiment, the cross-domain recommendation model is obtained by training the training samples of multiple domains using the initial multi-objective model, wherein the target in the multi-objective model refers to the domain (retrieval scenario) targeted by the cross-domain recommendation model, and one retrieval scenario corresponds to one target in the multi-objective model. The training samples used to train the cross-domain recommendation model are also samples from multiple domains, that is, the disclosed embodiment realizes model training by sharing information with training samples from multiple domains.

[0090] It should be noted that the multi-objective model is to learn multiple related objectives together. During the learning process, a shallow sharing is used to express the mutual sharing and mutual complementation of the learned field-related information, promote each other's learning, and improve the generalization effect. In the disclosed embodiment, one target corresponds to one retrieval scenario. Since the model trained based on the initial multi-objective model includes multiple retrieval scenarios, the obtained model is also a model for multiple domains, which can be called a cross-domain recommendation model. In the disclosed embodiment, the initial multi-objective model can be a hard parameter sharing (Hard Shared) model structure, a multi-gate mixture of experts (Multi-gate Mixture-of-Experts, MMOE) model structure, or a sub-network routing (Sub-Network-Routing, SNR) model structure, etc., which is not limited by the disclosed embodiment.

[0091] It can be understood that in the embodiments of the present disclosure, since the cross-domain recommendation model is obtained by training with an initial multi-objective model for training samples of multiple domains, one domain corresponds to one objective, and the essence of the multi-objective model is to learn information related to multiple fields through information sharing, and improve the generalization ability of the model in multiple fields. Therefore, the cross-domain recommendation model obtained by training based on the multi-objective model is used for retrieval and recommendation in the embodiments of the present disclosure, which can achieve better recommendation effects in any domain. Compared with the cross-domain recommendation model that can only work for the target domain but is not helpful for the source domain, the solution of the embodiments of the present disclosure can improve the retrieval performance in multiple fields at the same time, and has stronger generalization ability.

[0092] In some embodiments, the processing result includes the user's interest level in each of the to-be-retrieved information based on the search term;

[0093] The recommending information to the user based on the processing result includes:

[0094] The interest level of each of the to-be-retrieved information is ranked, and the information in the preset information database is recommended to the user according to the ranking result.

[0095] In this embodiment, the degree of interest can be characterized by the probability that the user will perform a predetermined behavior on the retrieved information. For example, for download recommendations for the information to be retrieved, the degree of interest is the probability that the user will download the retrieved information. The greater the probability that the user downloads the retrieved information, the higher the degree of interest of the user in the retrieved information. When recommending information in a preset information library to the user based on the sorting results, the retrieval information with a higher degree of interest can be recommended to the user. It should be noted that in the disclosed embodiment, the trained cross-domain recommendation model (multi-objective model) is a prediction of the probability that the user will perform a predetermined behavior after searching for the retrieved information based on the search terms, wherein the predetermined behavior is not limited to downloading behavior, for example, it can also be clicks, payments, comments, etc.

[0096] It is understandable that the method of sorting recommendations based on the degree of interest can recommend information that is more in line with the user's interests to the user, which can improve the accuracy of retrieval recommendations.

[0097] In some embodiments, the method further comprises:

[0098] Acquire training samples of the multiple domains, wherein one training sample includes sample user information in the domain, a sample search term, sample search information corresponding to the sample search term, and a label corresponding to the sample search information in the domain;

[0099] In response to the existence of a target training sample satisfying a preset information association condition in the training samples of different domains, information association processing is performed on the target training samples from different domains;

[0100] The initial multi-objective model is trained based on each training sample after information association processing to obtain the trained cross-domain recommendation model.

[0101] In the disclosed embodiment, the electronic device trains a multi-objective model based on training samples from multiple domains to obtain a cross-domain recommendation model for retrieval recommendation. The training samples can be samples in historical corresponding retrieval scenarios. A training sample includes sample user information in the domain, sample search terms, sample search information corresponding to the sample search terms, and labels corresponding to the sample search information in the domain. Taking the game search scenarios of the app store and the game center as an example, Figure 3 This is an example diagram of a search interface of a game center in an embodiment of the present disclosure. Figure 4 FIG. 1 is an example diagram of a search interface of an application center in an embodiment of the present disclosure. Figure 3 and Figure 4As shown in the figure, L4a is a sample search term in the game center, L4b is a sample search term in the application center, and the sample retrieval information displayed under the same search term is different for L5a and L5b. If the user downloads the sample retrieval information marked by L5b in the application center, the label corresponding to the sample retrieval information in the application center (domain) can be marked as a positive sample; if the user does not download the sample retrieval information marked by L5a in the game center, the label corresponding to the sample retrieval information in the game center (domain) can be marked as a negative sample.

[0102] It should be noted that, in some embodiments, the sample user information in the training sample may include information of all historical users, and the corresponding sample retrieval information may be information recommended by all historical users based on historical sample retrieval terms. During model training, the electronic device performing model training may extract identity attribute information of all historical users, such as learning the features of sample retrieval information of interest associated with sample retrieval terms by users with various identity attribute characteristics (topics of interest, presentation forms, etc.); during model application, the electronic device performing information recommendation may determine the identity of the user based on the user information of the current user, and make recommendations based on the features of sample retrieval information of interest associated with sample retrieval terms by users with various identity attribute characteristics that have been learned. Among them, the current user may be a new user in any domain, and this method may improve the user cold start problem.

[0103] In other embodiments, the sample user information in the training sample may also be only the user information of the current user to whom information recommendation is to be made, and the corresponding sample retrieval information may be the information recommended by the current user based on historical sample retrieval terms. During model training, the electronic device performing model training may be based on the sample retrieval terms and sample retrieval information of the current user, for example, learning the features of the sample retrieval information of interest to the current user based on the sample retrieval terms; during model application, the electronic device performing information recommendation may make recommendations based on the retrieval terms input by the current user and based on the learned features of the sample retrieval information of interest to the current user based on the sample retrieval terms. In this way, new sample retrieval information may be recommended to the current user to improve the cold start problem of the material.

[0104] In an embodiment of the present disclosure, after the electronic device obtains training samples from multiple domains, it further performs sample construction and associates the information of target training samples from different domains that meet preset information association conditions. The preset information association condition refers to an association condition for at least one of sample user information, sample search terms, and sample search information in training samples from different domains. Information association may include information fusion and / or sharing, wherein the information of target training samples in different domains after fusion is consistent, while the information of target training samples in different domains after sharing is not necessarily consistent.

[0105] In the related art, when making cross-domain recommendations, the sample information of a single domain is usually directly combined without further processing, that is, the data of different domains are not associated, and the cross-domain information is completely unused. Since the above information may overlap in the training samples of different domains, for example, two domains have the same users, the same search terms, or the same search information will be displayed, the method of simply merging training samples of different domains does not fully realize the sharing of information between domains, which is bound to limit the prediction performance of the multi-domain recommendation model. Therefore, the method of associating sample information between domains in the embodiment of the present disclosure can improve the prediction performance of the cross-domain recommendation model after training.

[0106] In some embodiments, in response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, performing information association processing on the target training samples from different domains includes:

[0107] In response to the existence of sample user information identifying the same sample user in training samples from different domains, the same sample user having the same sample search term in different domains, and target training samples having different sample retrieval information for the same sample search term in different domains, the sample user information of the target training samples from different domains is fused; and / or, the sample retrieval information corresponding to the same sample search term in the target training samples from different domains is fused.

[0108] In some embodiments, in response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, performing information association processing on the target training samples from different domains includes:

[0109] In response to the existence of sample user information identifying the same sample user in training samples from different domains, and the same sample user corresponding to target training samples with the same sample retrieval information in different domains, sample retrieval terms corresponding to the same sample retrieval information in target training samples from different domains are fused.

[0110] It should be noted that the target training samples that meet the above-mentioned preset information association conditions include the corresponding training samples of the same sample user in different domains. When performing information fusion, the corresponding information of each target training sample from different domains needs to be fused.

[0111] In the embodiments of the present disclosure, if the preset information association condition is a condition corresponding to the fusion of sample user information of different domains, in some cases, if a user has registered in different domains but has a user profile only in one domain, or the user profiles of the same user in different domains are different, then by fusion of the sample user information of the same sample user in different domains, more comprehensive sample user information can be obtained in each domain, thereby improving the accuracy of recommendations for each domain.

[0112] In the disclosed embodiment, the fusion is not simply performed based on whether the sample user information in different domains identifies the same sample user, but further determines whether the same sample user has the same sample search term in different domains, and whether the same sample search term corresponds to the same sample search information in different domains. When the same sample user has different sample search information based on the same sample search term in different domains, the sample user information of the target training samples from different domains is fused. Since there is usually more than one sample search information corresponding to a sample search term, when the same sample user has different sample search information based on the same sample search term in different domains, it indicates that the performance of the sample user in different domains may be different, and the reason may be different user profiles. Therefore, in this case, the fusion of sample user information from other domains can obtain a more comprehensive user representation, thereby improving the recommendation performance in each domain. At the same time, the method of directly fusion of sample user information of the same sample user can reduce the phenomenon of increasing the amount of data and affecting the training performance, such as when the sample search information corresponding to the same user based on the same sample search term in different domains is completely consistent.

[0113] For example, if the user portraits corresponding to user A in the app store and the game center are different, user A searches for game A in both the app store and the game center, and after searching for game A in the app store, not only game A but also game B is recommended, and user A downloads game B, while only game A is recommended in the game center. However, when training the task corresponding to the game center, if the user portrait in the app store is not associated, game B may not be recommended to user A when the game center makes a recommendation. Of course, the sample user information corresponding to user A in the app store also needs to be associated with the corresponding user portrait in the game center. It is understandable that the present disclosure can make recommendations more accurate and comprehensive by fusing sample user information of target training samples from different domains when the sample user information identifies the same sample user, and the same sample user has the same sample search terms in different domains, and the same sample search terms have different sample search information in different domains.

[0114] In addition, in the disclosed embodiment, the electronic device can also fuse the sample retrieval information of the target training samples from different domains to supplement the sample retrieval information in each domain, so as to improve the recommendation performance in each domain, when the same sample user has different sample retrieval information in different domains based on the same sample retrieval term. Still taking the above example as an example, by associating game B with the sample retrieval information corresponding to game A of user A in the game center, game B can also be recommended to user A when making recommendations in the game center.

[0115] In the embodiment of the present disclosure, when there are sample user information identifying the same sample user in the training samples of different domains, and the same sample user has target training samples with the same sample retrieval information in different domains, the sample search terms corresponding to the same sample retrieval information in the target training samples from different domains are merged. Usually, the search terms input by the user are usually keywords corresponding to the search target expected by the user, and the definition of keywords by the same user when searching for the same target at different times may also be random. Therefore, in the embodiment of the present disclosure, when the same sample user has target training samples with the same sample retrieval information in different domains, the sample search terms corresponding to the same sample retrieval information in the target training samples from different domains are merged, which can enrich the sample search term information in each domain and make the recommendation more accurate and comprehensive.

[0116] In some embodiments, in response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, performing information association processing on the target training samples from different domains includes:

[0117] In response to the existence of sample user information identifying the same sample user in training samples in different domains, but the preset target information in the sample user information is only included in target training samples in some domains, the preset target information is shared with the sample user information corresponding to the training samples associated with the same sample user in the domain that does not contain the preset target information.

[0118] As mentioned above, information association includes information sharing. In the embodiment of the present disclosure, the user age in the sample user information is used as the preset target information. If the game center has the age of user A, but the user A in the application store has no age information, and the user age is more important, the age of user A in the game center can be shared with the sample user information corresponding to the training sample associated with user A in the application store. It should be noted that the specific content of the preset target information is not limited in the embodiment of the present disclosure, and it can also be information such as gender, occupation, etc.

[0119] It is understandable that in the embodiments of the present disclosure, by sharing the preset target information, the training samples in the domain that does not contain the preset target information can also obtain more comprehensive and important sample user information, which is also beneficial to the accuracy of the recommendation.

[0120] In some embodiments, the initial multi-objective model is trained based on each training sample after information association processing to obtain the trained cross-domain recommendation model, including:

[0121] For each training sample, using the initial multi-objective model to determine a prediction result of the sample retrieval information in each training sample for a corresponding target;

[0122] For each training sample, based on the prediction result of each sample retrieval information for the corresponding target and the label of each sample retrieval information in the domain, calculate the loss value corresponding to each target;

[0123] Determining a total loss of the initial multi-objective model based on the loss value of each objective;

[0124] The parameters of the initial multi-objective model are adjusted based on the total loss to obtain the trained cross-domain recommendation model.

[0125] In the embodiment of the present disclosure, the loss value of each target is calculated based on the difference between the prediction result of the training sample for each target and the label. For example, the loss value corresponding to each target can be calculated based on logistic regression or mean square error. The specific implementation method of calculating the loss value of each target is not limited in the embodiment of the present disclosure.

[0126] It should be noted that the loss value of each target refers to the loss value of the task in each domain, which is relative to all training samples in each domain. The loss value of a single target can represent the average value of the difference between the prediction results and the label values ​​of all training samples in the corresponding domain. The smaller the loss value, the smaller the difference. Correspondingly, based on the loss value of each target, the smaller the total loss determined, the better the convergence degree of the model, that is, the accuracy of the model. Therefore, the embodiment of the present disclosure can adjust the parameters of the initial multi-target model based on the total loss to obtain a cross-domain recommendation model with better accuracy than the initial multi-task model.

[0127] In the embodiment of the present disclosure, when the total loss is determined based on the loss value of each target, the total loss can be determined based on a predetermined loss function, with the loss value of each target as an independent variable and the total loss as a dependent variable. For example, the sum of the loss values ​​of various tasks can be used as the total loss.

[0128] It can be understood that in the disclosed embodiment, the total loss is determined based on the loss value of each target, and the parameters in the initial multi-target model are adjusted based on the total loss, rather than adjusting the parameters based on a single target, so that the trained cross-domain recommendation model can balance the processing of different targets (domains), and the method of training multiple targets as a whole can enable the trained cross-domain recommendation model to obtain better retrieval recommendation effects in any domain. In addition, in the disclosed embodiment, the process of training the cross-domain recommendation model is coherent, and the input training samples are directly used to obtain the prediction results to adjust the total loss of the model to obtain the cross-domain recommendation model, that is, there is an independent and complete training stage. Compared with the aforementioned EMCDR segmented training method, the method of the disclosed embodiment is also more stable in improving the prediction performance.

[0129] In some embodiments, the step of using the initial multi-objective model to determine a prediction result of the sample retrieval information in each training sample for a corresponding objective includes:

[0130] Using the feature extraction layer in the initial multi-objective model, feature extraction is performed on the sample user information, the sample search term, and the sample search information corresponding to the sample search term to obtain a first feature vector corresponding to each training sample; wherein the first feature vector includes a sample user feature vector, a sample search term feature vector, and a sample search information feature vector;

[0131] Using the feature normalization layer in the initial multi-objective model to batch normalize the sample user feature vector, the sample search term feature vector, and the sample search information feature vector to obtain a normalized second feature vector; wherein the feature normalization layer includes normalization parameters within the domain and normalization parameters between domains;

[0132] Determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second eigenvector using the initial multi-target model;

[0133] The adjusting the parameters of the initial multi-objective model based on the total loss to obtain the trained cross-domain recommendation model includes:

[0134] Parameters in the initial multi-objective model, including normalized parameters within the domain and normalized parameters between the domains, are adjusted based on the total loss to obtain the trained cross-domain recommendation model.

[0135] In an embodiment of the present disclosure, when the electronic device uses the initial multi-objective model to determine the prediction result of the sample retrieval information in each training sample for the corresponding target, it first performs feature extraction to obtain a first feature vector for each training sample including a sample user feature vector corresponding to the sample user information, a sample retrieval term feature vector corresponding to the sample retrieval term, and a sample retrieval information feature vector corresponding to the sample retrieval information.

[0136] In the embodiment of the present disclosure, the first feature vector is an expression of the feature information of the current training sample. The feature extraction methods of different information in the training sample may be different, or the feature extraction methods of training samples from different domains may also be different. Correspondingly, the number level ranges of the sample user feature vector, the sample search term feature vector, and the sample search information feature vector may also be different. In order to speed up the model convergence speed, the initial multi-objective model implemented in the present disclosure also includes a feature normalization layer, which performs batch normalization (BN) processing.

[0137] Batch normalization assumes that all samples are independent and identically distributed (iid) and uses the shared statistics of all samples for normalization. The formula for batch normalization is shown below (2):

[0138] z=r i *(A)+b i (2)

[0139] Among them, ri and bi represent the normalization parameters in the i-th domain, A represents the standard normalization result of each training sample, and the standard normalization formula is shown as follows (3):

[0140]

[0141] Among them, ε is the mean of the feature vector of each training sample, σ is the variance of the feature vector, and A is the result of normalizing the feature vector x of a certain training sample.

[0142] Batch normalization can reduce the damage to the distribution characteristics of feature vectors by scaling and translating ri and bi. However, in multi-target modeling, samples are only independent and identically distributed within the domain. Using the same ri and bi confuses the distribution differences between domains, which is not conducive to more refined modeling. In this regard, the embodiment of the present disclosure also introduces inter-domain normalization parameters. Based on formula (2), the formula for batch normalization of the normalization layer of the embodiment of the present disclosure is as follows (4):

[0143] z=(r*r i )*(A)+(b+b i) (4)

[0144] Among them, r and b are the inter-domain normalization parameters.

[0145] It can be understood that during the training process of the embodiment of the present disclosure, when adjusting the parameters in the initial multi-objective model based on the total loss, the parameters including the normalization parameters within the domain and the normalization parameters between domains are also adjusted synchronously based on the above formula (4), so that when the model is trained, it not only retains the data distribution differences within the domain on the basis of a unified feature vector range, but also retains the data distribution differences between different domains, thereby facilitating subsequent more refined model building, thereby making the cross-domain recommendation model obtained through training more accurate.

[0146] In some embodiments, the using the initial multi-objective model to determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second feature vector includes:

[0147] Using each expert layer in the initial multi-objective model to perform feature processing on the normalized second eigenvector to obtain a third eigenvector output by each expert layer;

[0148] Processing the normalized second eigenvector using the target-associated gating layer in the initial multi-target model to obtain the weights of each expert layer output by the gating layer;

[0149] Performing weighted fusion on the third feature vectors output by each of the expert layers based on the weights output by the target-associated gating layer to obtain a fourth feature vector;

[0150] Based on the weighted fused fourth eigenvector corresponding to each target and the target sub-network corresponding to each target in the initial multi-target model, a prediction result for each training sample output by each target sub-network for the corresponding target is obtained.

[0151] In the disclosed embodiment, the first feature vector output by the feature extraction layer of the initial multi-objective model is a relatively rough feature. On this basis, the second feature vector normalized is not strong enough to express the training sample, and further feature optimization processing is required. In this regard, the initial multi-objective model in the disclosed embodiment also includes an expert layer and a gating layer, which are used for feature optimization processing.

[0152] In the disclosed embodiment, the initial multi-objective model may include multiple expert layers, one expert layer is equivalent to a multi-layer neural network, one expert layer outputs a feature vector, and K expert layers have a total of K feature vectors. The gating layer is a shallow neural network, and a weight parameter is assigned to each expert layer (the sum of all weights is 1), where the number of gating layers depends on the number of targets, and the input to the target sub-network is the weighted result of the features of multiple expert layers.

[0153] The above network structure is the MMOE model structure. Figure 5 This is an example diagram of the MMOE model structure, such as Figure 5 As shown, it includes an input layer, an expert layer (including expert 0, expert 1 and expert 2), a gating layer (including gating A and gating B) and a tower layer (including tower A and tower B corresponding to two targets respectively), wherein tower A and tower B are the target subnetworks of the embodiment of the present disclosure, one target subnetwork is used to execute the download recommendation task of the game center, and the other target subnetwork is used to execute the download recommendation task of the application center.

[0154] Figure 6 The detailed structure diagram of the MMOE model is shown in Figure 2. The number of weights output by the hidden units in the last fully connected layer of the gating network is the same as the number of experts, and the final output of the gating network will be normalized. Figure 6 As shown, the output of the last fully connected layer of the gating network, for example, after softmax normalization, acts on each expert. For example, the outputs L6, L7, and L8 (i.e., weights) of gating A in the figure act on experts 0, 1, and 2 respectively. In addition, it should be noted that the output of gating A acts on each hidden unit in the expert through a broadcast mechanism, such as the weight of L6 acts on the two hidden units of expert 0. The structure of gating B is similar to that of gating A and will not be described in detail here. It can be understood that in this network model, the role of the gating network is similar to that of the attention mechanism.

[0155] It can be understood that the embodiment of the present disclosure introduces an expert mechanism through a multi-gated hybrid expert model, so that more information can be extracted. In addition, the gating mechanism is introduced so that each target has a different weight for each expert to achieve a more flexible parameter sharing method, thereby improving the accuracy of the cross-domain recommendation model.

[0156] When information recommendation is performed based on the cross-domain recommendation model obtained by training according to the embodiment of the present disclosure, the retrieval recommendation of any one of the two scenarios of the game center and the application center is still taken as an example. For example, when a user enters a search term in the game center, the electronic device will extract the features corresponding to the user information of the user, the features corresponding to the search term, and the features corresponding to each piece of information to be retrieved in the game center based on the feature extraction layer of the cross-domain recommendation model, and then batch normalize all the above features based on the feature normalization layer of the cross-domain recommendation model. Subsequently, the batch normalized vectors are processed based on each expert layer of the cross-domain recommendation model, and the weight of each expert layer is determined based on the gating layer associated with the game center. The weights are then used to perform weighted fusion on the vectors processed by each expert layer, so as to finally obtain the user's interest level in each piece of information to be retrieved in the game center through the sub-network corresponding to the game center for recommendation.

[0157] In the same search scenario, the solution of the embodiment of the present disclosure is compared with the current mainstream cross-domain recommendation cold start model, and both achieve the best results, as shown in Table 1 below:

[0158] Model Mean absolute error Root mean square error TGT 4.2288 4.7920 CMF 1.5824 2.0887 DCDCSR 1.6731 2.0552 SSCDR 1.2243 1.5346 EMCDR 1.1843 1.4986 This method 1.0513* 1.4376*

[0159] Among them, the TGT model is a single-domain model that uses target domain data; CMF is a model that combines data from different domains into a single domain to act as auxiliary data; DCDCSR, SSCDR, and EMCDR are all models based on improving the embedding mapping of the target domain. In Table 1, the mean absolute error and root mean square error are used to characterize the difference between the prediction results of the trained model and the actual operation structure of the user during actual application. The smaller the mean absolute error and root mean square error, the closer to the true value, indicating that the performance of the model is better. As can be seen from Table 1, the method of the embodiment of the present disclosure achieves the best prediction effect. In addition, after the method in the embodiment of the present disclosure is applied to the game search recommendation scenarios of the application store and the game center, the average click-through rate, download rate, and payment rate of the game are greatly improved. However, it should be noted that the method of the embodiment of the present disclosure is not limited to the cross-domain recommendation model obtained based on the MMOE model training for retrieval recommendation. The cross-domain recommendation model obtained based on any multi-objective model training is applied to the retrieval recommendation of multiple domains. It is within the protection scope of the embodiment of the present disclosure.

[0160] Figure 7 FIG. 1 is a diagram showing an information recommendation device according to an exemplary embodiment. Figure 7 , the device comprises:

[0161] The first acquisition module 101 is configured to acquire a preset information library under any domain of multiple domains, a search term input by a user, and user information of the user; wherein the preset information library includes at least two pieces of information to be retrieved;

[0162] The processing module 102 is configured to process the user information, the search terms and each of the information to be searched based on a preset cross-domain recommendation model and obtain a processing result; wherein the cross-domain recommendation model is obtained by training the training samples of the multiple domains using an initial multi-objective model, and one domain corresponds to one objective in the initial multi-objective model;

[0163] The recommendation module 103 is configured to recommend information to the user based on the processing result.

[0164] In some embodiments, the processing result includes the user's interest level in each of the to-be-retrieved information based on the search term;

[0165] The recommendation module 103 is further configured to sort the interest levels of each of the to-be-retrieved information, and recommend information to the user according to the sorting result.

[0166] In some embodiments, the apparatus further comprises:

[0167] The second acquisition module 104 is configured to acquire training samples of the multiple domains, wherein one training sample includes sample user information in the domain, sample search terms, sample search information corresponding to the sample search terms, and a label corresponding to the sample search information in the domain;

[0168] The association module 105 is configured to perform information association processing on the target training samples from different domains in response to the existence of target training samples that meet the preset information association condition in the training samples from different domains;

[0169] The training module 106 is configured to train the initial multi-objective model based on each training sample after information association processing to obtain the trained cross-domain recommendation model.

[0170] In some embodiments, the association module 105 is further configured to, in response to the existence of target training samples in training samples from different domains, in which sample user information identifies the same sample user, the same sample user corresponds to the same sample search term in different domains, and the same sample search term has different sample retrieval information in different domains, fuse the sample user information of the target training samples from different domains; and / or fuse the sample retrieval information corresponding to the same sample search term in the target training samples from different domains.

[0171] In some embodiments, the association module 105 is further configured to, in response to the existence of sample user information identifying the same sample user in training samples from different domains, and the same sample user corresponding to target training samples with the same sample retrieval information in different domains, fuse the sample retrieval terms corresponding to the same sample retrieval information in the target training samples from different domains.

[0172] In some embodiments, the association module 105 is further configured to, in response to the presence of sample user information identifying the same sample user in training samples in different domains, but target training samples whose preset target information is only included in some domains, share the preset target information with the sample user information corresponding to the training samples associated with the same sample user in the domain that does not contain the preset target information.

[0173] In some embodiments, the training module 106 is further configured to determine, for each training sample, a prediction result of the sample retrieval information in each training sample for the corresponding target using the initial multi-target model; for each training sample, based on the prediction result of each sample retrieval information for the corresponding target and the label of each sample retrieval information in the domain, calculate the loss value corresponding to each target; based on the loss value of each target, determine the total loss of the initial multi-target model; based on the total loss, adjust the parameters of the initial multi-target model to obtain the trained cross-domain recommendation model.

[0174] In some embodiments, the training module 106 is further configured to use the feature extraction layer in the initial multi-objective model to perform feature extraction on the sample user information, sample search terms, and sample retrieval information corresponding to the sample search terms, respectively, to obtain a first feature vector corresponding to each training sample; wherein the first feature vector includes a sample user feature vector, a sample search term feature vector, and a sample retrieval information feature vector; use the feature normalization layer in the initial multi-objective model to batch normalize the sample user feature vector, the sample search term feature vector, and the sample retrieval information feature vector to obtain a normalized second feature vector; wherein the feature normalization layer includes normalization parameters within the domain and normalization parameters between domains; use the initial multi-objective model to determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second feature vector; adjust the parameters in the initial multi-objective model including the normalization parameters within the domain and the normalization parameters between domains based on the total loss to obtain the trained cross-domain recommendation model.

[0175] In some embodiments, the training module 106 is further configured to perform feature processing on the normalized second eigenvector using each expert layer in the initial multi-objective model to obtain a third eigenvector output by each expert layer; process the normalized second eigenvector using a target-associated gating layer in the initial multi-objective model to obtain the weights of each expert layer output by the gating layer; perform weighted fusion on the third eigenvectors output by each expert layer based on the weights output by the target-associated gating layer to obtain a fourth eigenvector; based on the weighted fused fourth eigenvector corresponding to each target and the target subnetwork corresponding to each target in the initial multi-objective model, obtain the prediction result of each training sample output by each target subnetwork for the corresponding target.

[0176] 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.

[0177] Figure 8 900 is a block diagram of an electronic device according to an exemplary embodiment. For example, the electronic device (device 900) may be a server for executing an information recommendation method.

[0178] Reference Figure 8 , the apparatus 900 includes a processing component 922, which further includes one or more processors, and a memory resource represented by a memory 932 for storing instructions, such as an application, that can be executed by the processing component 922. The application stored in the memory 932 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 922 is configured to execute the instructions to perform the above method.

[0179] The device 900 may also include a power supply component 926 configured to perform power management of the device 900, a wired or wireless network interface 950 configured to connect the device 900 to a network, and an input / output (I / O) interface 958. The device 900 may operate based on an operating system stored in the memory 932, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or the like.

[0180] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 932 including instructions, and the instructions can be executed by the processing component 922 of the device 900 to perform the above method. For example, the non-transitory 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.

[0181] A non-transitory computer-readable storage medium, when instructions in the storage medium are executed by a processor of an electronic device, enables the electronic device to perform an information recommendation method, the method comprising:

[0182] Acquire a preset information base under any domain of the multiple domains, a search term input by a user, and user information of the user; wherein the preset information base includes at least two pieces of information to be retrieved;

[0183] The user information, the search terms and each of the information to be searched are processed based on a preset cross-domain recommendation model to obtain a processing result; wherein the cross-domain recommendation model is obtained by training the training samples of the multiple domains using an initial multi-objective model, and one domain corresponds to one objective in the initial multi-objective model;

[0184] Recommend information to the user based on the processing result.

[0185] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure 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 description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims above.

[0186] 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. An information recommendation method, It is characterized in that The method comprises: Acquire a preset information base under any domain of the multiple domains, a search term input by a user, and user information of the user; wherein the preset information base includes at least two pieces of information to be retrieved; The user information, the search terms and each of the information to be searched are processed based on a preset cross-domain recommendation model to obtain a processing result; wherein the cross-domain recommendation model is obtained by training the training samples of the multiple domains using an initial multi-objective model, and one domain corresponds to one objective in the initial multi-objective model; Recommend information to the user based on the processing result.

2. The method according to claim 1, It is characterized in that The processing result includes the user's interest level in each of the to-be-searched information based on the search term; The recommending information to the user based on the processing result includes: The interest level of each of the to-be-retrieved information is ranked, and the information in the preset information database is recommended to the user according to the ranking result.

3. The method according to claim 1, It is characterized in that The method further comprises: Acquire training samples of the multiple domains, wherein one training sample includes sample user information in the domain, a sample search term, sample search information corresponding to the sample search term, and a label corresponding to the sample search information in the domain; In response to the existence of a target training sample satisfying a preset information association condition in the training samples of different domains, information association processing is performed on the target training samples from different domains; The initial multi-objective model is trained based on each training sample after information association processing to obtain the trained cross-domain recommendation model.

4. The method according to claim 3, It is characterized in that In response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, information association processing is performed on the target training samples from different domains, including: In response to the existence of sample user information identifying the same sample user in training samples from different domains, the same sample user having the same sample search term in different domains, and target training samples having different sample retrieval information for the same sample search term in different domains, the sample user information of the target training samples from different domains is fused; and / or, the sample retrieval information corresponding to the same sample search term in the target training samples from different domains is fused.

5. The method according to claim 3, It is characterized in that In response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, information association processing is performed on the target training samples from different domains, including: In response to the existence of sample user information identifying the same sample user in training samples from different domains, and the same sample user corresponding to target training samples with the same sample retrieval information in different domains, sample retrieval terms corresponding to the same sample retrieval information in target training samples from different domains are fused.

6. The method according to claim 3, It is characterized in that In response to the existence of target training samples satisfying a preset information association condition in the training samples from different domains, information association processing is performed on the target training samples from different domains, including: In response to the existence of sample user information identifying the same sample user in training samples in different domains, but the preset target information in the sample user information is only included in target training samples in some domains, the preset target information is shared with the sample user information corresponding to the training samples associated with the same sample user in the domain that does not contain the preset target information.

7. The method according to claim 3, It is characterized in that The initial multi-objective model is trained based on each training sample after information association processing to obtain the trained cross-domain recommendation model, including: For each training sample, using the initial multi-objective model to determine a prediction result of the sample retrieval information in each training sample for a corresponding target; For each training sample, based on the prediction result of each sample retrieval information for the corresponding target and the label of each sample retrieval information in the domain, calculate the loss value corresponding to each target; Determining a total loss of the initial multi-objective model based on the loss value of each objective; The parameters of the initial multi-objective model are adjusted based on the total loss to obtain the trained cross-domain recommendation model.

8. The method according to claim 7, It is characterized in that The method of using the initial multi-objective model to determine the prediction result of the sample retrieval information in each training sample for the corresponding target includes: Using the feature extraction layer in the initial multi-objective model, feature extraction is performed on the sample user information, the sample search term, and the sample search information corresponding to the sample search term to obtain a first feature vector corresponding to each training sample; wherein the first feature vector includes a sample user feature vector, a sample search term feature vector, and a sample search information feature vector; Using the feature normalization layer in the initial multi-objective model to batch normalize the sample user feature vector, the sample search term feature vector, and the sample search information feature vector to obtain a normalized second feature vector; wherein the feature normalization layer includes normalization parameters within the domain and normalization parameters between domains; Determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second eigenvector using the initial multi-target model; The adjusting the parameters of the initial multi-objective model based on the total loss to obtain the trained cross-domain recommendation model includes: Parameters in the initial multi-objective model, including normalized parameters within the domain and normalized parameters between the domains, are adjusted based on the total loss to obtain the trained cross-domain recommendation model.

9. The method according to claim 8, It is characterized in that The method of using the initial multi-objective model to determine the prediction result of the sample retrieval information in each training sample for the corresponding target based on the normalized second eigenvector includes: Using each expert layer in the initial multi-objective model to perform feature processing on the normalized second eigenvector to obtain a third eigenvector output by each expert layer; Processing the normalized second eigenvector using the target-associated gating layer in the initial multi-target model to obtain the weights of each expert layer output by the gating layer; Performing weighted fusion on the third feature vectors output by each of the expert layers based on the weights output by the target-associated gating layer to obtain a fourth feature vector; Based on the weighted fused fourth eigenvector corresponding to each target and the target sub-network corresponding to each target in the initial multi-target model, a prediction result for each training sample output by each target sub-network for the corresponding target is obtained.

10. An information recommendation device, It is characterized in that The device comprises: A first acquisition module is configured to acquire a preset information library under any domain of the multiple domains, a search term input by a user, and user information of the user; wherein the preset information library includes at least two pieces of information to be retrieved; A processing module, configured to process the user information, the search terms and each of the information to be searched based on a preset cross-domain recommendation model and obtain a processing result; wherein the cross-domain recommendation model is obtained by training the training samples of the multiple domains using an initial multi-objective model, and one domain corresponds to one objective in the initial multi-objective model; A recommendation module is configured to recommend information to the user based on the processing result.

11. An electronic device, It is characterized in that include: processor; a memory for storing processor-executable instructions; The processor is configured to execute the method as claimed in any one of claims 1 to 9.

12. A non-transitory computer-readable storage medium, It is characterized in that When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method as claimed in any one of claims 1 to 9.

Images