Information pushing method and apparatus, computer device, storage medium, and computer program product

Abstract

The present application relates to an information pushing method and apparatus, a computer device, a storage medium, and a computer program product. The method comprises: acquiring an interactive item identifier sequence, wherein the interactive item identifier sequence comprises identifiers of interactive items of a user; acquiring an interactive item feature sequence and an interactive behavior feature sequence on the basis of the interactive item identifier sequence; fusing the interactive item feature sequence and the interactive behavior feature sequence to obtain a fused feature sequence, and extracting autocorrelation information of fused features in the fused feature sequence to obtain an autocorrelation feature sequence; and identifying, from various pieces of preset scenario feature enhancement information, at least two pieces of scenario feature enhancement information corresponding to the autocorrelation feature sequence, and performing feature enhancement on the autocorrelation feature sequence on the basis of the at least two pieces of scenario feature enhancement information to obtain an enhanced feature sequence, wherein the enhanced feature sequence is used for implementing item information pushing. By using this method, the accuracy of information pushing can be improved.

Description

Information push method, device, computer equipment, storage medium and computer program product

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on March 25, 2024, with application number 2024103417477 and application name “Information Push Method, Device, Computer Equipment and Storage Medium”, all contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of artificial intelligence technology, and in particular to an information push method, apparatus, computer equipment, storage medium, and computer program product. Background Art

[0003] With the development of artificial intelligence, information push technology has emerged. This technology can extract user representations and then determine whether to push information by calculating the similarity between the user representation and the representation of the information to be pushed. Currently, when extracting user representations, the user's historical behavior information is usually used to extract the user's embedded representation. However, the user's embedded representation extracted from historical behavior information is less accurate, which can easily reduce the accuracy of information push and waste push resources. Summary of the Invention

[0004] Based on this, it is necessary to provide an information push method, device, computer equipment, computer-readable storage medium and computer program product that can improve the accuracy of information push and thus save push resources in response to the above technical problems.

[0005] In one aspect, the present application provides an information push method. The method comprises:

[0006] Obtaining an interactive item identification sequence, where the interactive item identification sequence includes the identifications of each interactive item of the user;

[0007] Obtaining an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identifier sequence. The interactive item feature sequence is obtained by extracting features from the item information corresponding to each interactive item identifier. The interactive behavior feature sequence is obtained by extracting features from the interactive behavior information corresponding to each interactive item identifier. The interactive behavior information is obtained from the interactive scene to which the interactive item identifier belongs.

[0008] The interactive item feature sequence and the interactive behavior feature sequence are fused to obtain a fused feature sequence, and the autocorrelation information of the fused features in the fused feature sequence is extracted to obtain an autocorrelation feature sequence;

[0009] At least two scene feature enhancement information corresponding to the autocorrelation feature sequence are identified from each preset scene feature enhancement information, and the autocorrelation feature sequence is feature enhanced based on the at least two scene feature enhancement information to obtain an enhanced feature sequence; the enhanced feature sequence is used to predict the interaction degree corresponding to each preset candidate item identification, and item information is pushed based on the interaction degree corresponding to each preset candidate item identification.

[0010] In a second aspect, the present application also provides an information push device. The device includes:

[0011] An identification sequence acquisition module is used to acquire an interactive item identification sequence, where the interactive item identification sequence includes the identifications of each interactive item of the user;

[0012] A feature sequence acquisition module is used to acquire an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identifier sequence. The interactive item feature sequence is obtained by extracting features from the item information corresponding to each interactive item identifier. The interactive behavior feature sequence is obtained by extracting features from the interactive behavior information corresponding to each interactive item identifier. The interactive behavior information is obtained from the interactive scene to which the interactive item identifier belongs.

[0013] A feature extraction module is used to fuse the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, and extract the autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence;

[0014] The push module is used to identify at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and to enhance the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence; the enhanced feature sequence is used to predict the interaction degree corresponding to each preset candidate item identification, and to push item information based on the interaction degree corresponding to each preset candidate item identification.

[0015] On the other hand, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned information push method when executing the computer program.

[0016] On the other hand, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above-mentioned information push method when executed by a processor.

[0017] On the other hand, the present application also provides a computer program product, which includes a computer program that implements the steps of the above-mentioned information push method when executed by a processor.

[0018] The above-mentioned information push method, apparatus, computer device, storage medium, and computer program product obtain an interactive item identification sequence, which includes each user's interactive item identification; fuse the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence; and extract the autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence; identify at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and enhance the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence. That is, by using multiple scene feature enhancement information to enhance the autocorrelation feature sequence, it is possible to better fit the item information and the interactive behavior information of the interactive scene, thereby improving the accuracy of the obtained enhanced feature sequence. The enhanced feature sequence can then be used to predict the interaction level corresponding to each preset candidate item identification, and item information is pushed based on the interaction level corresponding to each preset candidate item identification. That is, by using the highly accurate enhanced feature sequence for prediction and information push, the accuracy of the information push is improved, and the push of item information with low interaction level can be reduced, thereby saving push resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0020] FIG1 is a diagram illustrating an application environment of an information push method according to an embodiment;

[0021] FIG2 is a schematic diagram of a flow chart of an information push method in one embodiment;

[0022] FIG3 is a schematic diagram of a network architecture of a large language model in a specific embodiment;

[0023] FIG4 is a schematic diagram of a process for obtaining an enhanced feature sequence in one embodiment;

[0024] FIG5 is a schematic diagram of a process for obtaining an enhanced feature sequence in another embodiment;

[0025] FIG6 is a schematic diagram of a network architecture for obtaining an enhanced feature sequence in a specific embodiment;

[0026] FIG7 is a schematic diagram of a network architecture of a sparse hybrid expert network in a specific embodiment;

[0027] FIG8 is a diagram of a network architecture for obtaining a depth-enhanced feature sequence in a specific embodiment;

[0028] FIG9 is a schematic diagram of a model architecture of an information push model in a specific embodiment;

[0029] FIG10 is a schematic diagram showing the principle of an information push method in a specific embodiment;

[0030] FIG11 is a schematic flow chart of an information push method in a specific embodiment;

[0031] FIG12 is a structural block diagram of an information push method and apparatus according to an embodiment;

[0032] FIG13 is a diagram showing the internal structure of a computer device according to one embodiment;

[0033] FIG14 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0035] The information push method provided in the embodiments of the present application can be applied in the application environment shown in FIG1 . In this embodiment, terminal 102 communicates with server 104 via a network. A data storage system can store data that server 104 needs to process. The data storage system can be set up separately, integrated with server 104, or placed in the cloud or on other servers. The server 104 may receive an information push request sent by the user's terminal 102 and obtain an interactive item identification sequence based on the information push request. The interactive item identification sequence includes each interactive item identification of the user. The server 104 obtains an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identification sequence. The interactive item feature sequence is obtained by extracting features from item information corresponding to each interactive item identification. The interactive behavior feature sequence is obtained by extracting features from interactive behavior information corresponding to each interactive item identification. The interactive behavior information is obtained from the interactive scene to which the interactive item identification belongs. The server 104 fuses the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, and extracts autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence. The server 104 identifies at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and performs feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence. The server uses the enhanced feature sequence to predict the degree of interaction corresponding to each preset candidate item identification, and pushes the item information to the user's terminal 102 based on the degree of interaction corresponding to each preset candidate item identification. The terminals may include, but are not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices may include smart watches, smart bracelets, head-mounted devices, etc. The server may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services. The terminals and servers may be connected directly or indirectly via wired or wireless communication, which is not limited in this application.

[0036] In one embodiment, as shown in FIG2 , a method for pushing information is provided. This method is described using the server in FIG1 as an example. It is understood that the method can also be applied to a terminal, or to a system including a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0037] S202: Acquire an interactive item identification sequence, where the interactive item identification sequence includes identifications of each interactive item of the user.

[0038] Interactive item identifiers are used to uniquely identify interactive items. Interactive items are items with which users interact. Interactive behavior refers to interactions between a user and the item information of an item. Such interactions can include various user operations on the item information, including but not limited to browsing item information, clicking on item information, commenting on item information, and forwarding item information. The interactive item identifier sequence is a sequence of the user's interactive item identifiers. For example, it can be a sequence of the IDs (identity documents) of the user's interactive items. The order of the interactive item identifier sequence can be determined based on the order in which users interact with the item information, with the identifiers of interactive items that interacted first appearing first in the sequence and those that interacted later appearing last. The order of the interactive item identifier sequence can also be determined based on the number of interactions between the user and the item information, with the identifiers of interactive items with more frequent interactions appearing earlier in the sequence and those with fewer frequent interactions appearing later in the sequence. The order of the interactive item identifier sequence can also be manually set.

[0039] Specifically, the server can directly obtain the interactive item identification sequence from a database. The server can also obtain the interactive item identification sequence from a service provider providing information push services. The server can also obtain the interactive item identification sequence uploaded by a terminal. The server can also obtain the interactive item identification sequence from a service provider providing data services. The server can also obtain each user's interactive item identification and then sort the interactive item identifications to obtain the interactive item identification sequence. For example, the server can obtain each user's interactive item identification and the interaction time points corresponding to each interactive item identification from a database, and then sort the interactive item identifications according to the order of the interaction time points to obtain the interactive item identification sequence. For another example, the server can obtain each user's interactive item identification and the importance of each interactive item identification from a database, and then sort the interactive item identifications according to the order of importance to obtain the interactive item identification sequence. For another example, the server can obtain each user's interactive item identification and the number of interactions corresponding to each interactive item identification from a database, and then sort the interactive item identifications according to the order of the number of interactions to obtain the interactive item identification sequence.

[0040] S204, obtaining an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identification sequence. The interactive item feature sequence is obtained by extracting features from the item information corresponding to each interactive item identification. The interactive behavior feature sequence is obtained by extracting features from the interactive behavior information corresponding to each interactive item identification. The interactive behavior information is obtained from the interactive scene to which the interactive item identification belongs.

[0041] The interactive item feature sequence is a sequence of features of each user's interactive items. Interactive item features are used to characterize the user's interactive items and are obtained by extracting features from the item information corresponding to the interactive item identifier. Item information refers to information describing the interactive item corresponding to the interactive item identifier. Item information can be multimodal, for example, including at least one of images, text, and audio. The interactive behavior feature sequence is a sequence of features of each user's interactive behavior. Interactive behavior features are used to characterize the user's interactive behavior with the interactive items and are obtained by extracting features from the interactive behavior information corresponding to the interactive item identifier. Interactive behavior information describes the user's interactive behavior with the interactive items and can be textual information. For example, interactive behavior information can be text describing a click on the item information of an interactive item, i.e., the interactive behavior information can be "clicked on item A." The interactive scenario refers to the specific scenario used to display the item information of the interactive item, for example, a client scenario, a web scenario, a mobile scenario, and so on. In different interactive scenarios, the same interactive behavior can interact with interactive items in different ways. For example, in scenario 1, the interactive behavior of clicking on the interactive item can be performed by long pressing the corresponding button, while in scenario 2, the interactive behavior of clicking on the interactive item can be performed by sliding the corresponding button. The same interactive item can obtain different interactive behavior information in different interactive scenarios. For example, the interactive behavior information of item A can be "item A clicked in scenario 1" and "item A clicked in scenario 2". The sequence order of the interactive item feature sequence and the sequence order of the interactive behavior feature sequence are consistent with the sequence order of the interactive item identification sequence. For example, the first interactive item feature in the interactive item feature sequence corresponds to the first interactive item identifier in the interactive item identification sequence, and the first interactive behavior feature in the interactive behavior feature sequence corresponds to the first interactive item identifier.

[0042] Specifically, the server can obtain the corresponding interactive item features based on each interactive item identifier in the interactive item identifier sequence to obtain an interactive item feature sequence. The server can pre-acquire the item information of all items and then perform feature extraction on the item information of all items to obtain the item features of all items, which can then be stored. The server can then search the item features in the database for the corresponding item features based on the interactive item identifiers in the interactive item identifier sequence to obtain the interactive item features for each interactive item identifier. The server can also obtain the item information for each interactive item identifier from the database and then perform feature extraction on the item information for each interactive item identifier to obtain the interactive item features for each interactive item identifier. The server can also obtain the interactive item features for each interactive item identifier sent by the service provider providing the information push service.

[0043] The server can simultaneously obtain the interaction behavior features corresponding to each interactive item identifier to obtain an interaction behavior feature sequence. Specifically, the server can collect user interaction behavior information regarding the item information associated with the interactive item identifier from the interaction scenario to which the interactive item identifier belongs, and then perform feature extraction on the interaction behavior information associated with each interactive item identifier to obtain the interaction behavior features associated with each interactive item identifier. Alternatively, the server can obtain the interaction behavior features associated with each interactive item identifier from a service provider providing information push services. Alternatively, the server can search for the interaction behavior features associated with each interactive item identifier in a database.

[0044] S206 , fusing the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, and extracting autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence.

[0045] The fused feature sequence includes the fused features of each interactive item identifier. The fused feature is obtained by fusing the interactive item features and interactive behavior features of the same interactive item identifier. The autocorrelation feature sequence includes each autocorrelation feature. The autocorrelation feature is obtained by extracting the autocorrelation information of the fused feature. This autocorrelation information is used to represent the correlation between different positions in the fused feature sequence. In other words, the autocorrelation feature contains the relevant semantic information between different positions in the fused feature.

[0046] Specifically, the server fuses each interactive item feature in the interactive item feature sequence with the corresponding interactive behavior feature in the interactive behavior feature sequence to obtain a fused feature sequence. This fusion can be performed by linear addition, weighted sum, or vector operations on feature vectors, such as calculating vector sums or vector products. The server then extracts autocorrelation information of the fused features in the fused feature sequence using autocorrelation extraction parameters to obtain an autocorrelation feature sequence. The autocorrelation extraction parameters can be pre-set or trained. For example, the autocorrelation extraction parameters can be neural network parameters obtained after completing neural network training.

[0047] S208: Identify at least two pieces of scene feature enhancement information corresponding to the autocorrelation feature sequence from each piece of preset scene feature enhancement information, and perform feature enhancement on the autocorrelation feature sequence based on the at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence. The enhanced feature sequence is used to predict the interaction level corresponding to each preset candidate item identifier, and item information is pushed based on the interaction level corresponding to each preset candidate item identifier.

[0048] Among them, the preset scene feature enhancement information refers to the feature enhancement information of the pre-set interactive scene, and the predicted scene feature enhancement information refers to the information for feature enhancement of the autocorrelation features in the autocorrelation feature sequence, which may include feature enhancement parameters. The feature enhancement parameters may be obtained by pre-training, determined by testing, or network parameters in a neural network. The scene feature enhancement information can be used to enhance the semantic information of the corresponding interactive scene contained in the autocorrelation feature sequence, and different scene feature enhancement information can be used to enhance the semantic expression of different scenes in the autocorrelation feature sequence. The enhanced feature sequence includes various enhanced features, and the enhanced feature sequence is used to characterize the corresponding user. The enhanced feature sequence can be used as an embedded representation of the user for subsequent task processing, for example, prediction of information push, classification, and so on.

[0049] Specifically, the server calculates the degree of correlation between each preset scene feature enhancement information and the autocorrelation feature sequence, and then selects at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information based on the degree of correlation between each preset scene feature enhancement information and the autocorrelation feature sequence. The selected at least two scene feature enhancement information are the scene feature enhancement information that is most correlated with the autocorrelation feature sequence. The at least two scene feature enhancement information are then used to enhance the autocorrelation feature sequence. This can be done by using each selected scene feature enhancement information to enhance the autocorrelation feature sequence to obtain each enhanced feature sequence. All enhanced feature sequences are then fused to obtain an enhanced feature sequence for the user corresponding to the input sequence. That is, feature enhancement is performed using scene feature enhancement information of the interactive scene related to the input. This can improve the accuracy of the enhanced feature sequence obtained, thereby improving the accuracy of the user's embedded representation. It also avoids using all scene feature enhancement information to enhance the autocorrelation feature sequence, which can reduce computational complexity while ensuring accuracy. The server can then save the enhanced feature sequence as a representation of the user, and can then use the enhanced feature sequence for subsequent tasks, such as using the user's enhanced feature sequence to predict information push. In one embodiment, after the user's interactive behavior information and the user's interactive item identification are updated, the saved enhanced feature sequence can be updated using the updated user's interactive item identifications to obtain an updated enhanced feature sequence, thereby ensuring the accuracy of the enhanced feature sequence and facilitating subsequent tasks.

[0050] Preset candidate item identifiers refer to pre-set item identifiers that need to be screened. These preset candidate item identifiers can be the identifiers of all items or a subset of items, for example, identifiers of items with which the user has not interacted. The interactivity level indicates the likelihood that a user will interact with the item information associated with the preset candidate item identifiers. The higher the interactivity level, the more likely the user is to interact with the corresponding item information.

[0051] Specifically, the server can use the user's enhanced feature sequence to push item information. That is, the enhanced feature sequence is used to predict each preset candidate item identifier, obtain the interaction level corresponding to each preset candidate item identifier, and then use the interaction level corresponding to each preset candidate item identifier to filter each preset candidate item identifier. The filtering can be performed according to a pre-set filtering quantity. For example, the server can filter the preset candidate item identifier with the highest interaction level and push the item information of the preset candidate item identifier with the highest interaction level to the user's terminal. Alternatively, the server can filter the preset candidate item identifiers ranked by the top three interaction levels and push the item information of the top three preset candidate item identifiers to the user's terminal. The user's terminal receives the item information pushed by the server and displays it, and the user can interact with the pushed item information.

[0052] The above-mentioned information push method obtains an interactive item identification sequence, which includes each user's interactive item identification; fuses the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence; and extracts the autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence; identifies at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and performs feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence. Specifically, by using multiple scene feature enhancement information to enhance the autocorrelation feature sequence, the feature can better fit the item information and interactive behavior information of different interactive scenarios, thereby improving the accuracy of the obtained enhanced feature sequence, and thus improving the accuracy of the obtained user representation. The enhanced feature sequence can then be used to predict the interaction level corresponding to each preset candidate item identification, and item information is pushed based on the interaction level corresponding to each preset candidate item identification. Specifically, by using the highly accurate enhanced feature sequence for prediction and information push, the accuracy of information push is improved, and the push of item information with low interaction levels can be reduced, thereby saving push resources.

[0053] In one embodiment, S202, acquiring an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identification sequence includes:

[0054] The method comprises searching for interactive item features corresponding to each interactive item identifier in each pre-extracted interactive item feature to obtain an interactive item feature sequence; obtaining at least two interactive behavior information corresponding to each interactive item identifier from at least two interactive scenes to which each interactive item identifier belongs; extracting embedded representations of at least two interactive behavior information for each interactive item identifier to obtain at least two behavior representations, fusing the at least two behavior representations to obtain an interactive behavior feature for each interactive item identifier, and obtaining an interactive behavior feature sequence based on the interactive behavior feature of each interactive item identifier.

[0055] Pre-extracted interactive item features refer to pre-extracted features of item information associated with interactive item identifiers. Behavior representations refer to embedded representations of interactive behavior information. Different interaction scenarios generate different interactive behavior information, resulting in different behavioral features.

[0056] Specifically, the server's database stores pre-extracted interactive item features corresponding to each interactive item identifier. The server can then use each interactive item identifier in the interactive item identifier sequence to search for the interactive item features corresponding to each interactive item identifier among the pre-extracted interactive item features, and sort them according to the order of the interactive item identifier sequence to obtain an interactive item feature sequence. Specifically, the server searches the database for identical interactive item identifiers and uses the item features corresponding to the identical interactive item identifiers as the interactive item features. The server can then obtain at least two pieces of interactive behavior information corresponding to each interactive item identifier from at least two interactive scenarios to which each interactive item identifier belongs. Scenario identifiers for multiple interactive scenarios to which each interactive item identifier belongs can be found, and then, based on the scenario identifiers, the server searches the database for interactive behavior information generated by the interactive item identifier in the interactive scenario, thereby obtaining interactive behavior information for multiple different interactive scenarios associated with each interactive item identifier.

[0057] At this point, the server uses an embedded representation extraction algorithm to extract the embedded representation of each interactive behavior information corresponding to each interactive item identifier, obtaining the corresponding behavior representation for each interactive behavior information. The embedded representation extraction algorithm can be a neural network algorithm, a vectorization algorithm, or other algorithms. Finally, the server fuses all behavior representations of the same interactive item identifier across different interaction scenarios and calculates the sum of all behavior representations to obtain the interactive behavior characteristics of the interactive item identifier. The server traverses and calculates all interactive item identifiers to obtain the interactive behavior characteristics of all interactive item identifiers and sorts them according to the order of the interactive item identifier sequence to obtain the interactive behavior feature sequence.

[0058] In some embodiments, the server stores pre-extracted item features. These item features can be obtained by extracting the item information of all items. Once the user's interactive item identifiers are obtained, the server can search for item features that match the interactive item identifiers to obtain the interactive item features corresponding to the respective interactive item identifiers.

[0059] In the above embodiment, an interactive item feature sequence is obtained by searching for the interactive item features corresponding to each interactive item identifier within each pre-extracted interactive item feature, eliminating the need for real-time extraction and improving the efficiency of obtaining the interactive item feature sequence. Then, at least two pieces of interactive behavior information are obtained for each interactive item identifier. For each interactive item identifier, embedded representations of the at least two pieces of interactive behavior information are extracted to obtain at least two behavior representations, which are then fused to obtain an interactive behavior feature. By extracting interactive behavior information for the same item in multiple different interaction scenarios, the interactive behavior feature of the same item is extracted. This ensures that the interactive behavior feature includes semantic information about the same item in different interaction scenarios, improving the accuracy of the item's interactive behavior feature and, in turn, the accuracy of the obtained interactive behavior feature sequence.

[0060] In one embodiment, the server can obtain the item information of each interactive item identifier in the interactive item identifier sequence and extract the embedded representation of the item information to obtain an interactive item feature sequence. Simultaneously, the server can obtain the interactive behavior information of each interactive item identifier from the interactive scene to which each interactive item identifier belongs and extract the embedded representation of the interactive behavior information to obtain an interactive behavior feature sequence. This allows for real-time extraction of the embedded representation of the item information of each interactive item identifier, eliminating the need to pre-extract item features for all items, thereby conserving server computing resources.

[0061] In one embodiment, the server can obtain the modal information of each item corresponding to each interactive item identifier in the interactive item identifier sequence, then extract the embedded representation of each item modal information to obtain the respective modal embedding features, and fuse the respective modal encoding features to obtain an interactive item feature sequence. For example, the server can extract the embedded representation of the text information in the modal information corresponding to each interactive item identifier to obtain a text embedding feature sequence, and extract the embedded representation of the image information in the modal information corresponding to each interactive item identifier to obtain an image embedding feature sequence. Finally, the text embedding feature sequence and the image embedding feature sequence are fused to obtain an interactive item feature sequence. This means that by extracting the embedded representations of information from different modalities and then fusing them, the accuracy of the obtained interactive item features is improved.

[0062] In one embodiment, the server can obtain a target fusion feature sequence based on the interactive item identifier sequence. The target fusion feature sequence is obtained by extracting features from the target item information of each interactive item identifier. The target item information is obtained by splicing the item information of each interactive item identifier with the interactive behavior information of each interactive item identifier. For example, the server can splice the text information in the item information with the interactive behavior information to obtain spliced ​​text information, and then use the spliced ​​text information and item information of other modalities as the target item information. The server then uses the target fusion feature sequence as the fusion feature sequence for subsequent enhanced feature sequence extraction to obtain an enhanced feature sequence. That is, by splicing the interactive behavior information and item information to obtain the target item information, and then using the target item information for feature extraction, the efficiency of obtaining the fusion feature sequence can be improved.

[0063] In a specific embodiment, a large language model can be used to extract features from the item information of each interactive item identifier to obtain an interactive item feature sequence. Figure 3 shows a schematic diagram of the network architecture of the large language model. Specifically, the server obtains the item information of each interactive item identifier. This item information is multimodal and includes an item image and item text. The item text may include the item's name, description, category, and so on. The item image and item text are then input into the large language model to obtain a sequence of item text tokens (the basic units of the large language model, typically words, punctuation marks, or other symbols, which can also be used to represent contextual information in the text). The item text can be divided into these basic unit tokens to obtain a token sequence. This text token sequence is then input into a text embedding layer to extract a text embedding representation, obtaining a text embedding feature sequence. Simultaneously, the item image is encoded using an image encoder in the large language model to obtain an embedding representation of the image token. This image token embedding representation is then linearly projected into a space of the same dimensionality as the text token embedding representation, obtaining an image embedding feature sequence. The server then fuses the image embedding feature sequence and the text embedding feature sequence, inputs them into the transformer (a neural network model based on the self-attention mechanism, used to process sequence data) decoding layer, and decodes them through multiple transformer decoding networks in the transformer decoding layer, that is, generates text tokens through the transformer layer, and obtains the output interactive item feature sequence. Among them, the large language model can be LLaVa (a multimodal pre-training model that achieves cross-modal understanding and generation), MiniGPT (enhanced visual language understanding and advanced large language model), BLIP-2 (a multimodal Transformer model), etc. The large language model can fuse image features and text features to generate text based on image information, thereby improving the accuracy of the obtained interactive item feature sequence.

[0064] In one embodiment, S206, fusing the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, includes the following steps:

[0065] The interactive item features in the interactive item feature sequence are fused with the corresponding interactive behavior features in the interactive behavior feature sequence to obtain a fused feature sequence.

[0066] Specifically, the server can calculate the sum of the interactive item features in the interactive item feature sequence and the corresponding interactive behavior features in the interactive behavior feature sequence to obtain a fused feature sequence, wherein the interactive item features and the corresponding interactive behavior features are features of the same item, that is, the correspondence is determined based on the item to which they belong, and the interactive item features of an item do not correspond to the interactive behavior features of other items. The server can also obtain the pre-set importance corresponding to each interactive item identifier, and then use the importance to fuse the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, that is, calculate the feature sum of the interactive item features and interactive behavior features of the same interactive item identifier, and use the importance to weight the feature sum to obtain the fused feature of the interactive item identifier. All interactive item identifiers are traversed to obtain a fused feature sequence.

[0067] In one embodiment, the server may calculate the linear sum of each feature element in the interactive item feature sequence and the interactive behavior feature sequence according to the same position to obtain a fused feature sequence.

[0068] In one embodiment, the server may input the interactive item feature sequence and the interactive behavior feature into a trained neural network for feature fusion to perform feature fusion and obtain an output fused feature sequence.

[0069] In the above embodiment, the interactive item features in the interactive item feature sequence are fused with the corresponding interactive behavior features in the interactive behavior feature sequence to obtain a fused feature sequence. The fused feature contains not only item information but also interactive behavior information of the interactive scene, thereby improving the accuracy of the obtained fused feature sequence.

[0070] In one embodiment, S206, i.e., extracting the autocorrelation information of the fused features in the fused feature sequence to obtain the autocorrelation feature sequence, includes the following steps:

[0071] The fused feature sequence is linearly transformed to obtain the target query sequence, target key sequence and target value sequence; the correlation between the target query sequence and the target key sequence is calculated to obtain the correlation sequence, and the target value sequence is transformed based on the correlation sequence to obtain the autocorrelation feature sequence.

[0072] Specifically, the server obtains linear transformation parameters, which may include parameters for transforming to obtain a target query sequence, parameters for transforming to obtain a target key sequence, and parameters for transforming to obtain a target value sequence. The linear transformation parameters are pre-set or pre-trained, for example, trained through a self-attention neural network. The fused feature sequence is then linearly transformed using the linear transformation parameters. Specifically, the fused feature sequence is linearly transformed using the parameters for transforming to obtain the target query sequence, the fused feature sequence is linearly transformed using the parameters for transforming to obtain the target key sequence, and the fused feature sequence is linearly transformed using the parameters for transforming to obtain the target value sequence. The server then calculates the correlation between the target query sequence and the target key sequence using a similarity algorithm to obtain a correlation sequence. The similarity algorithm may be a distance similarity algorithm, a cosine similarity algorithm, a dot product operation, or the like. Finally, the server may normalize the correlation sequence to obtain a normalized correlation sequence, and then use the normalized correlation sequence to weight the target value sequence to obtain an autocorrelation feature sequence. That is, the autocorrelation information of the fused features in the fused feature sequence can be extracted through the self-attention mechanism to obtain the autocorrelation feature sequence.

[0073] In the above embodiment, by linearly transforming the fused feature sequence to obtain a target query sequence, a target key sequence and a target value sequence, and calculating the correlation between the target query sequence and the target key sequence to obtain a correlation sequence, the dependency relationship between different positions in the fused feature sequence can be fully extracted, and then the target value sequence is transformed based on the correlation sequence to obtain an autocorrelation feature sequence, thereby improving the accuracy of the obtained autocorrelation feature sequence.

[0074] In one embodiment, as shown in FIG4 , S208, identifying at least two pieces of scene feature enhancement information corresponding to the autocorrelation feature sequence from each piece of preset scene feature enhancement information, and performing feature enhancement on the autocorrelation feature sequence based on the at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence, includes:

[0075] S402: performing information selection calculation based on the autocorrelation feature sequence to obtain the selection degree corresponding to each preset scene feature enhancement information.

[0076] The selection degree is used to characterize the correlation between the autocorrelation feature sequence and the preset scene feature enhancement information. The more correlated the autocorrelation feature sequence is with the preset scene feature enhancement information, the higher the corresponding selection degree is.

[0077] Specifically, the server may perform a fully connected operation on the autocorrelation feature sequence to obtain the degree of selection corresponding to each preset scene feature enhancement information. When performing the fully connected operation, the operation may be performed using pre-trained fully connected operation parameters for calculating the degree of selection, or using pre-set fully connected operation parameters. In one embodiment, the server may input the autocorrelation feature sequence into a fully connected neural network for selecting degree calculation to calculate the degree of selection corresponding to each preset scene feature enhancement information. The fully connected neural network may be pre-trained for selecting degree calculation.

[0078] S404, screening each preset scene feature enhancement information based on the selection degree corresponding to each preset scene feature enhancement information to obtain at least two scene feature enhancement information corresponding to the autocorrelation feature sequence;

[0079] S406 , performing feature enhancement on the autocorrelation feature sequence based on at least two pieces of scene feature enhancement information to obtain at least two current enhanced feature sequences.

[0080] The current enhanced feature sequence is obtained by performing feature enhancement on the autocorrelation feature sequence using the current scene feature enhancement information. Different current enhanced feature sequences are obtained by performing feature enhancement on the autocorrelation feature sequence using different scene feature enhancement information.

[0081] Specifically, the server screens each preset scene feature enhancement information according to the selection degree corresponding to each preset scene feature enhancement information, and can obtain a pre-set number to be screened, which is at least two. Then, the scene feature enhancement information is selected in order from large to small according to the selection degree, so as to obtain at least two scene feature enhancement information that are most relevant to the autocorrelation feature sequence. At this time, the server uses each scene feature enhancement information to perform feature enhancement on the autocorrelation feature sequence, wherein the feature enhancement can be to use the feature enhancement parameters in the scene feature enhancement information to perform weighted calculation on the autocorrelation feature sequence, that is, to calculate the product of the feature increase parameter and the autocorrelation feature in the autocorrelation feature sequence. In some embodiments, feature enhancement can also be to calculate the sum of the feature enhancement parameters in the scene feature enhancement information and the autocorrelation feature sequence. In some embodiments, feature enhancement can also be to use the scene feature enhancement information to perform a nonlinear transformation on the autocorrelation feature sequence. In some embodiments, feature enhancement can also be to use the attention mechanism to weight the autocorrelation feature sequence through the scene feature enhancement information to enhance the key information in the autocorrelation feature sequence, thereby obtaining the current enhanced feature sequence corresponding to each scene feature enhancement information.

[0082] In some embodiments, a dynamic routing mechanism can be used to select scene feature enhancement information from various preset scene feature enhancement information. This dynamic routing mechanism is a technology that adaptively selects paths within a neural network. It allows the data propagation path in the network to dynamically change based on the characteristics of the input data. This can make the network more efficient, that is, it can only activate the neurons most relevant to the current task, thereby reducing unnecessary calculations and parameter usage.

[0083] S408: Fusing at least two current enhanced feature sequences according to the selection degrees corresponding to at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence.

[0084] Specifically, the server calculates the weighted sum of all current enhanced feature sequences and corresponding selection degrees to obtain the enhanced feature sequence, that is, calculates the product of each current enhanced feature sequence and the corresponding selection degree, and then calculates the sum of all products to obtain the enhanced feature sequence.

[0085] In the above embodiment, at least two scene feature enhancement information are obtained by calculating the selection degree, and then the autocorrelation feature sequence is enhanced and then fused according to the selection degree to obtain an enhanced feature sequence. That is, the scene feature enhancement information most relevant to the autocorrelation feature sequence is used for feature enhancement, and then the enhanced features are weightedly fused according to the selection degree, thereby improving the accuracy of the obtained enhanced feature sequence.

[0086] In one embodiment, as shown in FIG5 , S408 , at least two current enhanced feature sequences are fused according to the selection degrees corresponding to at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence, including:

[0087] S502, obtaining information quantities of at least two scene feature enhancement information, and updating the selection degrees corresponding to the respective preset scene feature enhancement information based on the information quantities to obtain the update degrees corresponding to the respective preset scene feature enhancement information.

[0088] The "number of information" refers to the number of scene feature enhancement information selected, which may be a preset number of scene feature enhancement information to be selected. The "update degree" refers to the updated selection degree, which is used to indicate the likelihood of the corresponding scene feature enhancement information being selected during feature enhancement.

[0089] Specifically, the server obtains the pre-set number of scene feature enhancement information to be screened. The server can also count the number of scene feature enhancement information that has been screened. Then, according to the number of information, the selection degree corresponding to each preset scene feature enhancement information is updated. The server can divide the selection degree corresponding to each preset scene feature enhancement information into different types of selection degrees according to the number of information, and then update the different types of selection degrees according to different types of pre-set update rules. The different types of selection degrees may include selection degrees that need to be updated and selection degrees that need to remain unchanged. The selection degrees corresponding to at least two scene feature enhancement information obtained by screening can be used as the selection degrees that remain unchanged, and the selection degrees corresponding to the remaining scene feature enhancement information can be used as the selection degrees that need to be updated. Then, the selection degrees that need to be updated are updated. For example, they can be weighted according to a pre-set update weight, or the selection degrees that need to be updated can be updated to a pre-set target value, while the selection degrees that remain unchanged are kept unchanged, thereby obtaining the update degree corresponding to each preset scene feature enhancement information.

[0090] S504 , normalizing the update degrees corresponding to each preset scene feature enhancement information to obtain the target selection degrees corresponding to each preset scene feature enhancement information.

[0091] Among them, the target selection degree refers to the selection degree obtained by normalizing the update degree of each preset scene feature enhancement information. The sum of the target selection degrees of all preset scene feature enhancement information is the target value, which can be 1 or other values.

[0092] Specifically, the server can use a normalization algorithm to normalize the update degree corresponding to each preset scene feature enhancement information. The normalization algorithm can be a minimum-maximum normalization algorithm, a scaling normalization algorithm, a mean-variance normalization algorithm, etc., to obtain the target selection degree of each preset scene feature enhancement information.

[0093] S506, determining target selection degrees corresponding to at least two pieces of scene feature enhancement information from the target selection degrees corresponding to the preset scene feature enhancement information;

[0094] S508 : Fusing at least two current enhanced feature sequences according to target selection degrees corresponding to at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence.

[0095] Specifically, the server determines the target selection degrees corresponding to at least two scene feature enhancement information obtained by screening from the target selection degrees corresponding to each preset scene feature enhancement information, and then calculates the weighted sum of at least two enhanced feature sequences according to the target selection degrees corresponding to the at least two scene feature enhancement information to obtain an enhanced feature sequence.

[0096] In one embodiment, the server can also obtain the selection degrees corresponding to the at least two scene feature enhancement information obtained through screening, and then directly normalize the selection degrees corresponding to the at least two scene feature enhancement information to obtain the target selection degrees corresponding to the at least two scene feature enhancement information.

[0097] In the above embodiment, the selection degrees of all scene feature enhancement information are updated and normalized, and then the target selection degree of the scene feature enhancement information obtained by screening is determined, that is, the accuracy of the obtained target selection degree is improved through unified updating and normalization, and then the target selection degree is used to fuse at least two current enhancement feature sequences to obtain an enhanced feature sequence, thereby improving the accuracy of the obtained enhancement feature sequence.

[0098] In one embodiment, S502, based on the amount of information, the selection degree corresponding to each preset scene feature enhancement information is updated to obtain the update degree corresponding to each preset scene feature enhancement information, including the steps of:

[0099] Based on the amount of information, each first selection degree and each second selection degree are determined from the selection degrees corresponding to each preset scene feature enhancement information, each first selection degree is greater than each second selection degree, and the number of each first selection degree is the same as the amount of information; each first selection degree remains unchanged, and each second selection degree is updated to the preset target value to obtain the update degree corresponding to each preset scene feature enhancement information.

[0100] Each first selection degree refers to the selection degree obtained by filtering all selection degrees from large to small to obtain the same number of selection degrees as the amount of information. Each second selection degree refers to the selection degree other than each first selection degree, that is, the selection degree remaining after filtering the selection degrees of information from large to small. The preset target value refers to a pre-set target value used to filter the preset scene feature enhancement information with a lower selection degree to ensure that the filtered preset scene feature enhancement information is used when performing feature fusion. The preset target value can be negative infinity.

[0101] Specifically, the server selects the selection levels corresponding to each preset scene feature enhancement information in descending order based on the number of pieces of information, obtaining first selection levels. The number of these first selection levels is the same as the number of pieces of information. The remaining selection levels are then used as second selection levels. The server then maintains the first selection levels unchanged and updates all second selection levels to the preset target values. At this point, the updated level for each preset scene feature enhancement information is obtained.

[0102] In a specific embodiment, when updating the selection degree corresponding to each preset scene feature enhancement information, the following formula (1) may be used for updating.

[0103] Among them, top k refers to the first k in the sorting. k can be the number of information, and k can be pre-set to 2. i It refers to the degree of selection of the feature enhancement information of the i-th preset scene. -∞ means negative infinity. TopK(L) i It refers to the selection degree after the i-th update. Then, the update degree can be normalized using the following formula (2).

[0104] G(x)=Softmax(TopK(xW g )) Formula (2)

[0105] Where x refers to the autocorrelation characteristic sequence. g Refers to the parameters of information selection calculation, for example, it can be the full connection operation parameter. By calculating the product of the autocorrelation feature sequence and the parameters of information selection calculation, the selection degree of each preset scene feature enhancement information is obtained, and then the TopK(xW g ), the TopK(xW g ) refers to the update degree of the feature enhancement information of the first k preset scenes in the ranking. Softmax refers to the normalization function. G(x) refers to the target selection degree obtained by normalizing the update degree of the feature enhancement information of the first k preset scenes in the ranking.

[0106] In the above embodiment, each first selection degree is kept unchanged, and each second selection degree is updated to a preset target value to obtain the update degree corresponding to each preset scene feature enhancement information. Then, during normalization, the unselected preset scene feature enhancement information can be filtered to avoid using the unselected preset scene feature enhancement information for feature enhancement, thereby improving the accuracy of feature enhancement.

[0107] In one embodiment, S508, that is, fusing at least two current enhanced feature sequences according to the target selection degrees corresponding to at least two scene feature enhancement information to obtain an enhanced feature sequence, includes:

[0108] For each scene feature enhancement information, the corresponding current enhanced feature sequence is weighted according to the target selection degree to obtain each weighted feature sequence; each weighted feature sequence is fused to obtain an enhanced feature sequence.

[0109] Specifically, the server calculates the product of the target selection degree of the screened scene feature enhancement information and the current enhanced feature sequence of the scene feature enhancement information to obtain the weighted feature sequence corresponding to the scene feature enhancement information, and traverses all the screened scene feature enhancement information to obtain each weighted feature sequence. Finally, the server calculates the sum of all weighted feature sequences, which can be achieved by linearly adding the feature elements at the same position in all weighted feature sequences, or by calculating the product of the weighted feature sequences, which can be achieved by multiplying the feature elements at the same position in all weighted feature sequences to obtain the enhanced feature sequence.

[0110] In a specific embodiment, the enhanced feature sequence can be calculated using the formula (3) shown below.

[0111] Among them, G(x) i It refers to the target selection degree of the i-th scene feature enhancement information, E i (x) refers to the i-th current enhanced feature sequence, and n refers to the number of information of at least two scene feature enhancement information.

[0112] In the above embodiment, the corresponding current enhanced feature sequence is weighted by the target selection degree to obtain each weighted feature sequence, and then the each weighted feature sequence is fused to obtain the enhanced feature sequence, that is, weighting by the target selection degree and then fusion is performed, thereby improving the accuracy of the obtained enhanced feature sequence.

[0113] In a specific embodiment, as shown in FIG6 , a schematic diagram of a network architecture for obtaining an enhanced feature sequence is provided. Specifically, the server inputs the fused feature sequence into the decoding network of the transformer. The decoding network of the transformer is obtained by replacing the feedforward neural network after the multi-head self-attention network with a sparse hybrid expert network, that is, the decoding network of the transformer includes a multi-head self-attention network with a mask mechanism and a sparse hybrid expert network. The decoding network of the transformer also includes a residual connection (Residual Connection) and a layer normalization network not shown in the figure. The server performs multi-head self-attention calculation on the fused feature sequence through the multi-head self-attention network with a mask mechanism to capture the dependency between different positions in the input sequence, thereby obtaining an autocorrelation feature sequence. Residual connection is then performed through the residual connection (Residual Connection) and the layer normalization network, and the result of the residual connection is normalized, thereby improving the performance of the network. The autocorrelation feature sequence is feature enhanced by the sparse hybrid expert network to obtain an enhanced feature sequence. The network architecture diagram of the sparse hybrid expert network can be shown in Figure 7. The sparse hybrid expert network includes a routing network and various expert networks. The various expert networks are used to characterize the corresponding preset scene feature enhancement information. The network architecture of the expert network is the same as the network architecture of the feedforward neural network in the transformer. Each expert network can focus on processing the subtasks it is good at, so that different expert networks can handle the autoregressive problems in the interactive scenarios they are good at. The server then inputs the autocorrelation feature sequence into the sparse hybrid expert network. First, the routing network is used to determine the expert network that needs to be input into the autocorrelation feature sequence. For example, there can be a total of 8 expert networks in the sparse hybrid expert network, and only the top k (topk) expert networks need to be activated. The k can be 2. That is, the server inputs the autocorrelation feature sequence into the routing network, and the selection probability of the 8 expert networks is calculated by the routing network. The routing network can be obtained through L=xW g Calculate the selection probability, W gis the network parameter of the routing network, which can be a fully connected neural network. Then, based on the selection probabilities of the 8 expert networks, the 2 largest selection probabilities and the corresponding expert networks are determined. Then, the selection probabilities of the 8 expert networks are updated using formula (1), that is, the 2 largest selection probabilities are kept unchanged and the other 6 selection probabilities are set to negative infinity. Then, the updated selection probabilities are normalized using formula (2), so that the updated probability of the input autocorrelation feature sequence entering the topk expert network can be determined, ensuring that the probability of entering the non-topk expert network is 0. Then, the autocorrelation feature sequence is input into the expert networks with the 2 largest selection probabilities for feature enhancement, and the current enhanced feature sequence output by the 2 expert networks is obtained. Then, the updated probabilities of the 2 expert networks and the current enhanced feature sequence output by the 2 expert networks are used to calculate the enhanced feature sequence through formula (3), that is, only 2 expert networks can be activated for feature enhancement, which reduces the computational complexity and improves the computational efficiency.

[0114] In one embodiment, after S208, that is, after the autocorrelation feature sequence is enhanced based on at least two scene feature enhancement information to obtain an enhanced feature sequence, the method further includes the following steps:

[0115] The enhanced feature sequence is used as the fused feature sequence, and the autocorrelation information of the fused features in the fused feature sequence is extracted and returned to obtain the autocorrelation feature sequence. The step of obtaining the autocorrelation feature sequence is iteratively executed until the depth enhancement completion condition is met, and a depth enhanced feature sequence is obtained; the depth enhanced feature sequence is used to predict the depth interaction degree corresponding to each preset candidate item identifier, and item information is pushed based on the depth interaction degree corresponding to each preset candidate item identifier.

[0116] The "deep enhancement completion condition" refers to the condition for terminating the loop iteration, and can be a condition for extracting a deep enhancement feature sequence, including but not limited to the number of iterations reaching a set maximum. A deep enhancement feature sequence refers to an enhanced feature sequence obtained after multiple iterations, capable of extracting deep patterns and associations, so that the resulting deep enhancement feature sequence contains deep semantic information. The "deep interaction level" refers to the degree of user interaction with the item information of the preset candidate item identifiers obtained after multiple iterations.

[0117] Specifically, the server can perform multiple iterations to extract depth information from the fused feature sequence to obtain a depth-enhanced feature sequence. The parameters used in each iteration for autocorrelation feature extraction and feature enhancement can be the same or different. That is, the server can use the enhanced feature sequence as the fused feature sequence, and return the steps of extracting the autocorrelation information of the fused features in the fused feature sequence to obtain the autocorrelation feature sequence. Each time the enhanced feature sequence is obtained, it is determined whether the depth enhancement completion condition is met. When the depth enhancement completion condition is not met, the obtained enhanced feature sequence is used as the starting input for the next iteration. When the depth enhancement completion condition is met, the obtained enhanced feature sequence is used as the final extracted depth-enhanced feature sequence, and the depth-enhanced feature sequence is used as the representation of the user. The server can then use the depth-enhanced feature sequence to process subsequent tasks, such as pushing item information, recalling item information, sorting item information, and so on.

[0118] In one embodiment, as shown in FIG8 , a network architecture diagram for obtaining a deep enhanced feature sequence is provided. The network architecture includes a decoding network of n transformers. The network architecture of the decoding network of the n transformers can be the network architecture shown in FIG6 . The network parameters of the decoding networks of the n transformers are different and pre-trained. n can be set in advance based on experience, for example, it can be set to 6. Then, the server inputs the fused feature sequence into the decoding network of the first transformer to extract the output enhanced feature sequence, and uses the enhanced feature sequence as the input of the decoding network of the second transformer. The output enhanced feature sequence is extracted by the decoding network of the second transformer and input into the decoding network of the next transformer. Then, until the output enhanced feature sequence is extracted by the decoding network of the last transformer, a deep enhanced feature sequence is obtained. By extracting enhanced features through the decoding networks of multiple transformers, the accuracy of the obtained enhanced feature sequence is improved.

[0119] In the above embodiment, the depth enhancement feature sequence is obtained through iterative execution, which can deeply extract relevant information in the input sequence, improve the accuracy of the depth enhancement feature sequence, and then use the depth enhancement feature sequence to push information, thereby improving the accuracy of information push.

[0120] In one embodiment, after S208, that is, after identifying at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and performing feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence, the following steps are further included:

[0121] Based on each preset candidate item identifier, the enhanced feature sequence is linearly transformed to obtain the linear transformation features corresponding to each preset candidate item identifier. The linear transformation features corresponding to each preset candidate item identifier are then mapped to an interaction degree to obtain the interaction degree corresponding to each preset candidate item identifier. Based on the interaction degree corresponding to each preset candidate item identifier, each preset candidate item identifier is screened to obtain the target item identifier. The item information of the target item identifier is pushed.

[0122] The linear transformation feature is the feature obtained by linearly transforming the enhanced feature sequence. The target item identifier is the candidate item identifier corresponding to the maximum interaction degree obtained by the final screening.

[0123] Specifically, the server obtains the linear transformation parameters corresponding to each preset candidate item identifier and uses the linear transformation parameters to linearly transform the enhanced feature sequence to obtain a linear transformation feature sequence. In some embodiments, all preset candidate item identifiers may also correspond to common linear transformation parameters, and the enhanced feature sequence can then be linearly transformed using the common linear transformation parameters. The linear transformation feature sequence is then normalized and mapped using a normalization function to obtain the interaction level corresponding to each preset candidate item identifier, which can be represented as a probability. The preset candidate item identifiers are then screened based on the interaction level to obtain the target item identifier corresponding to the maximum interaction level. Finally, the server pushes the item information of the target item identifier to the user's terminal.

[0124] In one embodiment, the server can use a pre-configured neural network output layer to predict the degree of interaction. Specifically, an enhanced feature sequence is input into the neural network output layer. The enhanced feature sequence is then output and calculated using the neural network output layer's linear transformation network and normalization function to obtain the degree of interaction corresponding to each preset candidate item identifier. The candidate item identifier corresponding to the highest degree of interaction is then selected as the target item identifier. Alternatively, multiple candidate item identifiers ranked top can be selected as the item identifiers to be pushed. Finally, the item information associated with the target item identifier is pushed to the user's terminal. Specifically, the output of the last transformer decoding network is mapped to a vector space with a dimension equal to the number of candidate item identifiers through a linear transformation. A softmax function converts each element in this vector space into a probability value representing the probability of the next candidate item identifier to be pushed. The target item identifier is then filtered from the candidate item identifiers based on the probabilities of all candidate item identifiers. The item information associated with the target item identifier is then obtained and pushed to the user's terminal.

[0125] In the above embodiment, the interaction level corresponding to each preset candidate item identifier is calculated using an enhanced feature sequence. The candidate item identifier with the highest interaction level is then screened to obtain the target item identifier. The item information of the target item identifier with the highest interaction level is then pushed to the user terminal. This avoids pushing information that the user is not interested in, thereby improving the accuracy of information push.

[0126] In a real-time embodiment, the information push method further includes the steps of:

[0127] The interactive item identification sequence is input into the information push model, which includes a feature extraction network, an autocorrelation feature extraction network, a feature enhancement network and an information push network; the interactive item feature sequence and the interactive behavior feature sequence corresponding to the interactive item identification sequence are obtained through the feature extraction network, and the interactive item feature sequence and the interactive behavior feature sequence are fused to obtain a fused feature sequence; the autocorrelation information of the fused features in the fused feature sequence is extracted through the autocorrelation feature extraction network to obtain an autocorrelation feature sequence; at least two scene feature enhancement information corresponding to the autocorrelation feature sequence are identified from each preset scene feature enhancement information through the feature enhancement network, and the autocorrelation feature sequence is feature enhanced based on the at least two scene feature enhancement information to obtain an enhanced feature sequence; the information push network predicts the degree of interaction corresponding to each preset candidate item identification based on the enhanced feature sequence, and pushes item information based on the degree of interaction corresponding to each preset candidate item identification.

[0128] Among them, the information push model refers to a neural network model used for information push, and the information push model can be trained using a historical interaction item identification sequence, where the historical interaction item identification sequence includes the user's various historical interaction item identifications. The feature extraction network refers to a neural network that performs feature extraction, and can be a neural network that extracts embedded representations, or a vectorized neural network, etc. The autocorrelation feature extraction network refers to a neural network that performs autocorrelation features, and can be a self-attention network. The feature enhancement network refers to a neural network that performs feature enhancement, and can be a sparse mixed expert network. The information push network refers to a neural network that performs information push, and can be the output layer of the neural network.

[0129] Specifically, the interactive item identification sequence is input into an information push model, which can implement the steps of the information push method described in any of the aforementioned embodiments. The information push model includes a feature extraction network, an autocorrelation feature extraction network, a feature enhancement network, and an information push network. The feature extraction network can implement the steps of the method for obtaining a fused feature sequence described in any of the aforementioned embodiments. The autocorrelation feature extraction network can implement the steps of the method for obtaining an autocorrelation feature sequence described in any of the aforementioned embodiments. The feature enhancement network can implement the steps of the method for obtaining an enhanced feature sequence described in any of the aforementioned embodiments. The information push network can implement the steps of the method for obtaining the degree of interaction corresponding to each preset candidate item identification in any of the aforementioned embodiments. That is, the server can obtain the degree of interaction corresponding to each preset candidate item identification through the information push model, and then push item information based on the degree of interaction corresponding to each preset candidate item identification.

[0130] In one embodiment, when the server obtains the interactive item identifiers and the length of the sequence formed by these identifiers is less than a preset sequence length, it obtains a padding mask and uses the interactive item identifiers and the padding mask to generate an interactive item identifier sequence of the preset length. In this embodiment, by inputting the interactive item identifier sequence into the information push model to obtain the interactivity level corresponding to each preset candidate item identifier, item information is pushed based on the interactivity level corresponding to each preset candidate item identifier. This, in other words, by using a pre-trained information push model to push information, can improve information push efficiency.

[0131] In one embodiment, the training of the information push model includes the following steps:

[0132] Obtain a historical interactive item identification sequence, and determine a training interactive item identification sequence and an item identification training label based on the historical interactive item identification sequence; input the training interactive item identification sequence into the initial information push model to obtain the training interaction degree corresponding to each training candidate item identification; perform loss calculation based on the training interaction degree and item identification training label corresponding to each training candidate item identification to obtain loss information; train the initial information push model based on the loss information, and obtain the information push model when the training completion condition is met.

[0133] Among them, the historical interactive item identification sequence includes the user's various historical interactive item identifications, and the historical interactive item identification refers to the item identification with which the user has historically interacted. The training interactive item identification sequence refers to the historical interactive item identification sequence used during training. The item identification training label refers to the label of the item identification used during training, including the item push label and the item non-pushed label. The initial information push model refers to the information push model for initializing model parameters. The initial information push model can be established using a neural network, which can be a convolutional neural network, a recurrent neural network, a feedforward neural network, etc. The loss information is used to characterize the error between the degree of training interaction and the item identification training label. The training completion condition refers to the condition for the end of training, including but not limited to the number of iterations reaching the maximum number of iterations, the model parameters no longer changing, or the loss information of the model being less than the preset loss threshold. The training candidate item identification is used to uniquely identify the candidate item identification used in the training process, and can be the same as or different from the preset candidate item identification used in the application process.

[0134] Specifically, the server can retrieve a historical interaction item identifier sequence from a database, then select a training interaction item identifier sequence from the historical interaction item identifier sequence, and determine the next historical interaction item identifier in the training interaction item identifier sequence as the item identifier training label. For example, if there are m historical interaction item identifiers in the historical interaction item identifier sequence, the server can use the first q-1 (q less than m) historical interaction item identifiers in the historical interaction item identifier sequence as the training interaction item identifier sequence. The server then determines the item identifier training label based on the qth historical interaction item identifier, setting the qth historical interaction item identifier as the next item to be pushed label, and setting the remaining historical interaction item identifiers as the item not pushed label. The server then inputs the training interaction item identifier sequence into the initial information push model, performs prediction using the initialized model parameters in the initial information push model, and outputs the training interaction degree corresponding to each preset candidate item identifier. A pre-set loss function is then used to calculate the error between the training interaction degree corresponding to each preset candidate item identifier and the item identifier training label to obtain loss information. The loss function can be a cross-entropy loss function or a logarithmic loss function. The initial information push model is then reversely updated using a gradient descent algorithm to obtain an updated information push model. At this point, the server determines whether the training completion conditions have been met. If not, the server uses the updated information push model as the initial information push model and iterates back to the initial information push model steps until the training completion conditions are met. The server then uses the last updated information push model as the final trained information push model. At this point, the server can deploy and use the trained information push model.

[0135] In a specific embodiment, the objective function of server training can be a negative log-likelihood loss. Specifically, the objective of the log-likelihood loss function is to maximize the probability of the user interacting with the next item under the condition of predicting the next item. The log-likelihood loss function can be shown in the following formula (4).

[0136] Among them, LOSS refers to the negative log-likelihood loss. ＜t It refers to the sequence of historical interaction item identifiers from the 1st to the t-1th position. The item sequence X of length n = (x1…x n ) t Refers to the historical interaction item identifier at the t-th position. P(x t |x <t ) means that at a given x <t After the historical interaction item identification sequence, predict x tDuring the training process, we use optimization algorithms such as stochastic gradient descent to minimize the negative logarithmic loss function and update the parameters of the initial information push model, so that the information push model can better predict the next item identifier that the user will interact with.

[0137] In the above embodiment, the initial information push model is trained by using the training interaction item identification sequence and the item identification training label, and when the training completion condition is met, the information push model is obtained, thereby improving the accuracy of information push by the information push model.

[0138] In a specific embodiment, as shown in FIG9 , a schematic diagram of the model architecture of an information push model is provided. The information push model includes an input layer, an embedding layer, a transformer decoding layer, and an output layer. Specifically, the server inputs the interactive item identification sequence into the information push model through the output layer. The information push model then extracts the interactive item features and interactive behavior features corresponding to the interactive item identification sequence through the embedding layer. After fusing the interactive item features and interactive behavior features, the fused feature sequence is input into the transformer decoding layer. The transformer decoding layer extracts the deep enhancement feature sequence corresponding to the fused feature sequence through the network architecture shown in FIG8 . The deep enhancement feature sequence is then input into the output layer. The output layer predicts a probability distribution representing the degree of interaction corresponding to each candidate item identification. The item identification with the highest degree of interaction is then selected as the next item identification to be pushed to the user, and the item information is pushed to the user's terminal. In other words, the sparse hybrid expert network is used to better distinguish the modalities and scenario behavior information belonging to different item identifications, thereby better fitting the user's sequential behavior characteristics in multiple modalities and scenarios, thereby achieving more accurate information push services.

[0139] In one embodiment, the interactive item identification sequence includes an interactive video identification sequence, the interactive video identification sequence includes each interactive video identification of the user, and each preset candidate item identification includes each preset candidate video identification. The information push method further includes the steps of:

[0140] Obtain an enhanced feature sequence corresponding to the interactive video identification sequence, and predict the degree of interaction corresponding to each preset candidate video identification based on the enhanced feature sequence corresponding to the interactive video identification sequence; screen each preset candidate video identification based on the degree of interaction corresponding to each preset candidate video identification to obtain a target video identification; and push video information of the target video identification.

[0141] The interactive video identifier uniquely identifies an interactive video, which is a video in which a user interacts. The preset candidate video identifier is a pre-set candidate video identifier. The target video identifier is the candidate video identifier with the highest degree of interaction.

[0142] Specifically, the server performs a video information push service. When pushing video information, the server obtains the user's interactive video identifier sequence. This sequence is then input into an information push model, resulting in the interaction levels corresponding to each of the preset candidate video identifiers. The candidate video identifier with the highest interaction level is then selected to obtain the target video identifier. Finally, the server pushes the video information of the target video identifier to the user's terminal. Upon receiving the video information pushed by the server, the user's terminal can display the video information, such as playing the video content, displaying the video title, displaying the video introduction, and displaying the video type.

[0143] In the above embodiment, by obtaining an interactive video identification sequence, then predicting a target video identification based on the interactive video identification sequence, and pushing the video information of the target video identification, the accuracy of video information push is improved, thereby increasing the possibility of users interacting with the pushed video, thereby improving the conversion rate of the pushed video and saving push resources.

[0144] In a specific embodiment, as shown in FIG10 , a schematic diagram of the principle of an information push method is provided. Specifically, a server obtains a user's interactive item identification sequence, searches for corresponding interactive item features and interactive behavior features based on each interactive item identification in the interactive item identification sequence, obtains an interactive item feature sequence and an interactive behavior feature sequence, and then fuses the interactive item feature sequence and the interactive behavior feature sequence to obtain a fused feature sequence. The server further extracts an autocorrelation feature sequence from the fused feature sequence and identifies at least two feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information. The server then uses all identified feature enhancement information to enhance the autocorrelation feature sequence to obtain an enhanced feature sequence. In this case, the enhanced feature sequence is used as a representation of the user. The enhanced feature sequence can then be used for information push. Specifically, by enhancing the autocorrelation feature sequence using multiple different feature enhancement information, the accuracy of the enhanced feature sequence is improved. The enhanced feature sequence is then used for information push, thereby improving the accuracy of the information push.

[0145] In a specific embodiment, as shown in FIG11 , a flow chart of an information push method is provided, which is executed by a computer device, which can be a server or a terminal. In this embodiment, the computer device is used as an example for description, and specifically includes the following steps:

[0146] S1102: Obtain an interactive item identification sequence for the user to be pushed, input the interactive item identification sequence into the information push model, and use a feature extraction network to search for interactive item features corresponding to each interactive item identification in each pre-extracted interactive item feature based on each interactive item identification in the interactive item identification sequence to obtain an interactive item feature sequence.

[0147] S1104, obtaining at least two pieces of interaction behavior information for each interactive item identifier from at least two interaction scenarios to which each interactive item identifier belongs through a feature extraction network, extracting embedded representations of at least two pieces of interaction behavior information for each interactive item identifier to obtain at least two behavior representations, and fusing the at least two behavior representations to obtain an interaction behavior feature sequence.

[0148] S1106 , fusing the interactive item features in the interactive item feature sequence with the corresponding interactive behavior features in the interactive behavior feature sequence through a feature extraction network to obtain a fused feature sequence.

[0149] S1108, linearly transform the fused feature sequence through the autocorrelation feature extraction network to obtain the target query sequence, the target key sequence and the target value sequence, calculate the correlation between the target query sequence and the target key sequence to obtain the correlation sequence, and transform the target value sequence based on the correlation sequence to obtain the autocorrelation feature sequence.

[0150] S1110, performing information selection calculation on the autocorrelation feature sequence through the feature enhancement network to obtain the selection degree corresponding to each preset scene feature enhancement information, screening each preset scene feature enhancement information based on the selection degree corresponding to each preset scene feature enhancement information to obtain at least two scene feature enhancement information corresponding to the autocorrelation feature sequence, and performing feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain at least two current enhanced feature sequences.

[0151] S1112, obtaining the information quantity of at least two scene feature enhancement information through the feature enhancement network, updating the selection degree corresponding to each preset scene feature enhancement information based on the information quantity, obtaining the update degree corresponding to each preset scene feature enhancement information, normalizing the update degree corresponding to each preset scene feature enhancement information, and obtaining the target selection degree corresponding to each preset scene feature enhancement information.

[0152] S1114: Determine, by a feature enhancement network, target selection degrees corresponding to at least two pieces of scene feature enhancement information from the target selection degrees corresponding to the preset scene feature enhancement information, and fuse the at least two current enhanced feature sequences according to the target selection degrees corresponding to the at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence;

[0153] S1116, linearly transforming the enhanced feature sequence according to each preset candidate item identifier through the information push network to obtain the linear transformation features corresponding to each preset candidate item identifier, and performing interaction degree mapping on the linear transformation features corresponding to each preset candidate item identifier to obtain the interaction degree corresponding to each preset candidate item identifier.

[0154] S1118 , based on the interaction levels corresponding to the preset candidate item identifiers, the preset candidate item identifiers are screened to obtain a target item identifier, and the item information of the target item identifier is pushed to the terminal of the user to be pushed.

[0155] In the above embodiment, by using the information push model to push the item information identified by the target item to the terminal of the user to be pushed, the accuracy of the pushed item information can be improved on the basis of improving the efficiency of the item information push, and the possibility of the user interacting with the pushed item information can be increased, thereby improving the conversion rate of the pushed item information, reducing the push of item information with low interaction possibility, and thus saving push resources.

[0156] In a specific embodiment, the information push method can be applied to a product information push platform. Specifically, the server of the product information push platform obtains a product push request sent by a user's terminal, searches the database for the user's historical interaction product sequence based on the product push request, and then inputs the historical interaction product sequence into the deployed information push model. The information push model predicts the interaction level corresponding to each candidate product identifier, and then selects the product identifier with the maximum interaction level to obtain the target product identifier. Finally, the server of the product information push platform pushes the product information of the target product identifier to the user's terminal. The user's terminal receives the product information and displays it on the product information push platform. The user can then interact with the displayed product information through the product information push platform, avoiding pushing product information with low user interaction levels to the user, thereby improving the accuracy of product information push. The information push method can also be applied to a live broadcast information push platform, using the user's interactive live broadcast identifier sequence to predict the interaction level of each candidate live broadcast identifier, and then selecting the live broadcast identifier with the maximum interaction level to push the corresponding live broadcast information, thereby improving the accuracy of live broadcast information push. This information push method can also be applied to a news information push platform to use the user's interactive news identifier sequence to predict the degree of interaction of each candidate news identifier, and then select the news identifier with the maximum interaction degree to push the corresponding news information, thereby improving the accuracy of news information push.

[0157] In a specific embodiment, the information push method can be applied to a music playback platform. Specifically, users can browse and play various types of music pushed through the music playback platform. When the server of the music playback platform pushes music to the user's terminal, it can obtain a music push request and the user's interactive music sequence, then call the deployed information push model, input the interactive music sequence into the information push model for push prediction, and obtain the interaction level corresponding to each candidate music identifier output. The server can then select the candidate music identifiers to be pushed in descending order of interaction level, and push the music corresponding to each selected candidate music identifier to the music playback platform in the user's terminal. At this point, the user can play the pushed music through the music playback platform. That is, by pushing music with a high possibility of user interaction, the possibility of users playing the pushed music can be increased, thereby improving the conversion rate of pushed music and reducing the push of music with a low possibility of user play, thereby saving push resources.

[0158] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0159] Based on the same inventive concept, the embodiments of the present application also provide an information push device for implementing the aforementioned information push method. The implementation solution provided by this device is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations of one or more information push device embodiments provided below can be referred to the limitations of the information push method above and will not be repeated here.

[0160] In one embodiment, as shown in FIG12 , an information push device 1200 is provided, comprising: an identification sequence acquisition module 1202 , a feature sequence acquisition module 1204 , a feature extraction module 1206 , and a push module 1208 , wherein:

[0161] An identification sequence acquisition module 1202 is used to acquire an interactive item identification sequence, where the interactive item identification sequence includes each interactive item identification 1204 of the user;

[0162] Feature sequence acquisition module 1204 is configured to acquire an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identifier sequence. The interactive item feature sequence is obtained by extracting features from the item information corresponding to each interactive item identifier. The interactive behavior feature sequence is obtained by extracting features from the interactive behavior information corresponding to each interactive item identifier. The interactive behavior information is acquired from the interactive scene to which the interactive item identifier belongs.

[0163] The feature extraction module 1206 is used to fuse the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, and extract the autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence;

[0164] Push module 1208 is used to identify at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and perform feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence; the enhanced feature sequence is used to predict the interaction degree corresponding to each preset candidate item identifier, and push item information based on the interaction degree corresponding to each preset candidate item identifier.

[0165] In one embodiment, the feature sequence acquisition module 1204 is further configured to search for interactive item features corresponding to each interactive item identifier in each pre-extracted interactive item feature to obtain an interactive item feature sequence; obtain at least two interactive behavior information corresponding to each interactive item identifier from at least two interactive scenes to which each interactive item identifier belongs; extract embedded representations of at least two interactive behavior information for each interactive item identifier to obtain at least two behavior representations, fuse the at least two behavior representations to obtain an interactive behavior feature of each interactive item identifier, and obtain an interactive behavior feature sequence based on the interactive behavior feature of each interactive item identifier.

[0166] In one embodiment, the feature extraction module 1206 is further configured to fuse the interactive item features in the interactive item feature sequence with the corresponding interactive behavior features in the interactive behavior feature sequence to obtain a fused feature sequence.

[0167] In one embodiment, the feature extraction module 1206 is further used to perform a linear transformation on the fused feature sequence to obtain a target query sequence, a target key sequence, and a target value sequence; calculate the correlation between the target query sequence and the target key sequence to obtain a correlation sequence; and transform the target value sequence based on the correlation sequence to obtain an autocorrelation feature sequence.

[0168] In one embodiment, the push module 1208 is also used to perform information selection calculations based on the autocorrelation feature sequence to obtain the selection degree corresponding to each preset scene feature enhancement information; based on the selection degree corresponding to each preset scene feature enhancement information, each preset scene feature enhancement information is screened to obtain at least two scene feature enhancement information corresponding to the autocorrelation feature sequence; based on the at least two scene feature enhancement information, the autocorrelation feature sequence is feature enhanced to obtain at least two current enhanced feature sequences; according to the selection degree corresponding to the at least two scene feature enhancement information, the at least two current enhanced feature sequences are fused to obtain an enhanced feature sequence.

[0169] In one embodiment, the push module 1208 is also used to obtain the information quantity of at least two scene feature enhancement information, update the selection degree corresponding to each preset scene feature enhancement information based on the information quantity, and obtain the update degree corresponding to each preset scene feature enhancement information; normalize the update degree corresponding to each preset scene feature enhancement information to obtain the target selection degree corresponding to each preset scene feature enhancement information; determine the target selection degree corresponding to at least two scene feature enhancement information from the target selection degree corresponding to each preset scene feature enhancement information; and fuse at least two current enhancement feature sequences according to the target selection degree corresponding to at least two scene feature enhancement information to obtain an enhanced feature sequence.

[0170] In one embodiment, the push module 1208 is also used to determine each first selection degree and each second selection degree from the selection degrees corresponding to each preset scene feature enhancement information based on the amount of information, each first selection degree is greater than the second selection degree, and the number of each first selection degree is the same as the amount of information; each first selection degree remains unchanged, and each second selection degree is updated to the preset target value to obtain the update degree corresponding to each preset scene feature enhancement information.

[0171] In one embodiment, the push module 1208 is further configured to weight the corresponding current enhanced feature sequence according to the target selection degree for each scene feature enhancement information to obtain each weighted feature sequence; and fuse each weighted feature sequence to obtain an enhanced feature sequence.

[0172] In one embodiment, the information push device 1200 further includes:

[0173] The deep feature extraction module is used to use the enhanced feature sequence as the fused feature sequence, and return the autocorrelation information of the fused features in the extracted fused feature sequence. The steps of obtaining the autocorrelation feature sequence are iteratively executed until the deep enhancement completion condition is met, and the deep enhanced feature sequence is obtained; the deep enhanced feature sequence is used to predict the depth interaction degree corresponding to each preset candidate item identifier, and push item information based on the depth interaction degree corresponding to each preset candidate item identifier.

[0174] In one embodiment, the information push device 1200 further includes:

[0175] The item information push module is used to perform a linear transformation on the enhanced feature sequence based on each preset candidate item identifier to obtain the linear transformation features corresponding to each preset candidate item identifier; perform interaction degree mapping on the linear transformation features corresponding to each preset candidate item identifier to obtain the interaction degree corresponding to each preset candidate item identifier; based on the interaction degree corresponding to each preset candidate item identifier, screen each preset candidate item identifier to obtain a target item identifier; and push the item information of the target item identifier.

[0176] In one embodiment, the information push device 1200 further includes:

[0177] The model push module is used to input the interactive item identification sequence into the information push model, which includes a feature extraction network, an autocorrelation feature extraction network, a feature enhancement network and an information push network; the interactive item feature sequence and the interactive behavior feature sequence corresponding to the interactive item identification sequence are obtained through the feature extraction network, and the interactive item feature sequence and the interactive behavior feature sequence are fused to obtain a fused feature sequence; the autocorrelation information of the fused features in the fused feature sequence is extracted through the autocorrelation feature extraction network to obtain an autocorrelation feature sequence; the feature enhancement network is used to identify at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and the autocorrelation feature sequence is feature enhanced based on the at least two scene feature enhancement information to obtain an enhanced feature sequence; the information push network is used to predict the degree of interaction corresponding to each preset candidate item identification based on the enhanced feature sequence, and item information is pushed based on the degree of interaction corresponding to each preset candidate item identification.

[0178] In one embodiment, the information push device 1200 further includes:

[0179] The model training module is used to obtain a historical interactive item identification sequence, and determine a training interactive item identification sequence and an item identification training label based on the historical interactive item identification sequence; input the training interactive item identification sequence into the initial information push model to obtain the training interaction degree corresponding to each training candidate item identification; calculate the loss based on the training interaction degree and the item identification training label corresponding to each training candidate item identification to obtain loss information; train the initial information push model based on the loss information, and obtain the information push model when the training completion conditions are met.

[0180] In one embodiment, the interactive item identification sequence includes an interactive video identification sequence, the interactive video identification sequence includes each interactive video identification of the user, each preset candidate item identification includes each preset candidate video identification, and the information push device 1200 further includes:

[0181] The video push module is used to obtain the enhanced feature sequence corresponding to the interactive video identification sequence, and predict the interaction degree corresponding to each preset candidate video identification based on the enhanced feature sequence corresponding to the interactive video identification sequence; based on the interaction degree corresponding to each preset candidate video identification, each preset candidate video identification is screened to obtain the target video identification; and the video information of the target video identification is pushed.

[0182] Each module in the above-mentioned information push device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0183] In one embodiment, a computer device is provided, which may be a server. Its internal structure may be as shown in FIG13 . The computer device includes a processor, memory, an input / output (I / O) interface, and a communication interface. The processor, memory, and I / O interface are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and computer program stored in the non-volatile storage medium. The database of the computer device stores data such as a sequence of interactive item identifiers, item information associated with each interactive item identifier, interactive behavior information associated with each interactive item identifier, a sequence of interactive behavior characteristics, and the interactive scenarios to which the interactive item identifiers belong. The I / O interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements an information push method.

[0184] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be shown in Figure 14. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via wired or wireless communication, and the wireless communication may be implemented via Wi-Fi, a mobile cellular network, NFC (near field communication), or other technologies. When the computer program is executed by the processor, it implements an information push method. The display unit of the computer device is used to form a visually visible image, and can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse, etc.

[0185] Those skilled in the art will understand that the structure shown in Figure 13 or Figure 14 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0186] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0187] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0188] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0189] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0190] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0191] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0192] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. An information push method, characterized in that: The method comprises: Obtaining an interactive item identification sequence, wherein the interactive item identification sequence includes identifications of each interactive item of the user; Obtaining an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identifier sequence, wherein the interactive item feature sequence is obtained by extracting features from item information corresponding to each interactive item identifier, and the interactive behavior feature sequence is obtained by extracting features from interactive behavior information corresponding to each interactive item identifier, wherein the interactive behavior information is obtained from the interactive scene to which the interactive item identifier belongs; Fusing the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, and extracting autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence; Identifying at least two pieces of scene feature enhancement information corresponding to the autocorrelation feature sequence from each piece of preset scene feature enhancement information, and performing feature enhancement on the autocorrelation feature sequence based on the at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence; The enhanced feature sequence is used to predict the interaction degree corresponding to each preset candidate item identifier, and push item information based on the interaction degree corresponding to each preset candidate item identifier.

2. The method according to claim 1, characterized in that The acquiring of the interactive item feature sequence and the interactive behavior feature sequence based on the interactive item identification sequence includes: searching for the interactive item features corresponding to the respective interactive item identifiers in the pre-extracted interactive item features to obtain the interactive item feature sequence; Acquiring at least two pieces of interaction behavior information corresponding to each interactive item identifier from at least two interaction scenarios to which each interactive item identifier belongs; For each interactive item identifier, at least two embedded representations of interactive behavior information are extracted to obtain at least two behavior representations, and the at least two behavior representations are fused to obtain an interactive behavior feature of each interactive item identifier, and the interactive behavior feature sequence is obtained based on the interactive behavior feature of each interactive item identifier.

3. The method according to any one of claims 1 to 2, characterized in that The step of fusing the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence includes: The interactive item features in the interactive item feature sequence are fused with the corresponding interactive behavior features in the interactive behavior feature sequence to obtain the fused feature sequence.

4. The method according to any one of claims 1 to 3, characterized in that Extracting the autocorrelation information of the fused features in the fused feature sequence to obtain the autocorrelation feature sequence includes: Performing linear transformation on the fused feature sequence to obtain a target query sequence, a target key sequence and a target value sequence; The correlation between the target query sequence and the target key sequence is calculated to obtain a correlation sequence, and the target value sequence is transformed based on the correlation sequence to obtain the autocorrelation feature sequence.

5. The method according to any one of claims 1 to 4, characterized in that The identifying at least two pieces of scene feature enhancement information corresponding to the autocorrelation feature sequence from each piece of preset scene feature enhancement information, and performing feature enhancement on the autocorrelation feature sequence based on the at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence, includes: Performing information selection calculation based on the autocorrelation feature sequence to obtain the selection degree corresponding to each preset scene feature enhancement information; Based on the selection degree corresponding to each of the preset scene feature enhancement information, the preset scene feature enhancement information is screened to obtain at least two pieces of scene feature enhancement information corresponding to the autocorrelation feature sequence; Performing feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain at least two current enhanced feature sequences; The at least two current enhanced feature sequences are fused according to the selection degrees corresponding to the at least two pieces of scene feature enhancement information to obtain the enhanced feature sequence.

6. The method according to any one of claims 1 to 5, characterized in that The fusing the at least two current enhanced feature sequences according to the respective selection degrees corresponding to the at least two scene feature enhancement information to obtain the enhanced feature sequence includes: Acquiring information quantities of the at least two scene feature enhancement information, and updating the selection degrees corresponding to the respective preset scene feature enhancement information based on the information quantities to obtain update degrees corresponding to the respective preset scene feature enhancement information; Normalizing the update degrees corresponding to the respective preset scene feature enhancement information to obtain the target selection degrees corresponding to the respective preset scene feature enhancement information; Determining the target selection degrees corresponding to the at least two scene feature enhancement information respectively from the target selection degrees corresponding to the respective preset scene feature enhancement information; The at least two current enhanced feature sequences are fused according to target selection degrees respectively corresponding to at least two pieces of scene feature enhancement information to obtain the enhanced feature sequence.

7. The method according to any one of claims 1 to 6, characterized in that The updating of the selection degrees corresponding to the respective preset scene feature enhancement information based on the information quantity to obtain the updating degrees corresponding to the respective preset scene feature enhancement information includes: Determining first selection degrees and second selection degrees from the selection degrees corresponding to the respective preset scene feature enhancement information based on the amount of information, wherein each first selection degree is greater than each second selection degree, and the number of each first selection degree is the same as the amount of information; The first selection degrees are kept unchanged, and the second selection degrees are updated to preset target values, so as to obtain the update degrees corresponding to the preset scene feature enhancement information.

8. The method according to any one of claims 1 to 7, characterized in that The fusing the at least two current enhanced feature sequences according to the target selection degrees respectively corresponding to the at least two scene feature enhancement information to obtain the enhanced feature sequence includes: For each scene feature enhancement information, the corresponding current enhancement feature sequence is weighted according to the target selection degree to obtain each weighted feature sequence; The weighted feature sequences are fused to obtain the enhanced feature sequence.

9. The method according to any one of claims 1 to 8, characterized in that After the step of performing feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence, the method further includes: The enhanced feature sequence is used as the fused feature sequence, and the steps of extracting the autocorrelation information of the fused features in the fused feature sequence and obtaining the autocorrelation feature sequence are iteratively performed until a depth enhancement completion condition is met, thereby obtaining a depth enhanced feature sequence; The depth enhancement feature sequence is used to predict the depth interaction degree corresponding to each preset candidate item identifier, and push item information based on the depth interaction degree corresponding to each preset candidate item identifier.

10. The method according to any one of claims 1 to 9, characterized in that After identifying at least two pieces of scene feature enhancement information corresponding to the autocorrelation feature sequence from the respective preset scene feature enhancement information, and performing feature enhancement on the autocorrelation feature sequence based on the at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence, the method further includes: Performing a linear transformation on the enhanced feature sequence based on each of the preset candidate item identifiers to obtain linear transformation features corresponding to each of the preset candidate item identifiers; Performing interaction degree mapping on the linear transformation features corresponding to the respective preset candidate item identifiers to obtain the interaction degrees corresponding to the respective preset candidate item identifiers; Based on the interaction levels corresponding to the preset candidate item identifiers, the preset candidate item identifiers are screened to obtain a target item identifier; The item information of the target item identifier is pushed.

11. The method according to any one of claims 1 to 10, characterized in that The method further comprises: Inputting the interactive item identification sequence into an information push model, wherein the information push model includes a feature extraction network, an autocorrelation feature extraction network, a feature enhancement network, and an information push network; Acquire an interactive item feature sequence and an interactive behavior feature sequence corresponding to the interactive item identification sequence through the feature extraction network, and fuse the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence; Extracting the autocorrelation information of the fused features in the fused feature sequence through the autocorrelation feature extraction network to obtain an autocorrelation feature sequence; Identifying, by the feature enhancement network, at least two pieces of scene feature enhancement information corresponding to the autocorrelation feature sequence from each piece of preset scene feature enhancement information, and performing feature enhancement on the autocorrelation feature sequence based on the at least two pieces of scene feature enhancement information to obtain an enhanced feature sequence; The information push network predicts the interaction levels corresponding to the respective preset candidate item identifiers based on the enhanced feature sequence, and pushes item information based on the interaction levels corresponding to the respective preset candidate item identifiers.

12. The method according to any one of claims 1 to 11, characterized in that The training of the information push model includes the following steps: Acquire a historical interactive item identification sequence, and determine a training interactive item identification sequence and an item identification training label based on the historical interactive item identification sequence; Inputting the training interaction item identification sequence into the initial information push model to obtain the training interaction degree corresponding to each training candidate item identification; Calculating loss based on the training interaction degree corresponding to each of the candidate training item identifiers and the item identifier training label to obtain loss information; The initial information push model is trained based on the loss information, and when a training completion condition is met, the information push model is obtained.

13. The method according to any one of claims 1 to 12, characterized in that The interactive item identification sequence includes an interactive video identification sequence, the interactive video identification sequence includes each interactive video identification of the user, and the each preset candidate item identification includes each preset candidate video identification. The method further includes: Obtaining an enhanced feature sequence corresponding to the interactive video identification sequence, and predicting the interaction degree corresponding to each of the preset candidate video identifications based on the enhanced feature sequence corresponding to the interactive video identification sequence; Based on the interaction levels corresponding to the preset candidate video identifiers, the preset candidate video identifiers are screened to obtain a target video identifier; The video information of the target video identifier is pushed.

14. An information push device, characterized in that: The device comprises: an identification sequence acquisition module, configured to acquire an interactive item identification sequence, wherein the interactive item identification sequence includes identifications of each interactive item of the user; a feature sequence acquisition module, configured to acquire an interactive item feature sequence and an interactive behavior feature sequence based on the interactive item identification sequence, wherein the interactive item feature sequence is obtained by extracting features from item information corresponding to each interactive item identification, and the interactive behavior feature sequence is obtained by extracting features from interactive behavior information corresponding to each interactive item identification, wherein the interactive behavior information is obtained from the interactive scene to which the interactive item identification belongs; a feature extraction module, configured to fuse the interactive item feature sequence with the interactive behavior feature sequence to obtain a fused feature sequence, and extract autocorrelation information of the fused features in the fused feature sequence to obtain an autocorrelation feature sequence; The push module is used to identify at least two scene feature enhancement information corresponding to the autocorrelation feature sequence from each preset scene feature enhancement information, and perform feature enhancement on the autocorrelation feature sequence based on the at least two scene feature enhancement information to obtain an enhanced feature sequence; the enhanced feature sequence is used to predict the interaction degree corresponding to each preset candidate item identification, and push item information based on the interaction degree corresponding to each preset candidate item identification.

15. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 13 are implemented.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 13 are implemented.

17. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 13 are implemented.

Images