Content recommendation method and device, equipment, medium and program product

Abstract

The invention discloses a content recommendation method and device, equipment, a medium and a program product, and relates to the field of artificial intelligence. The method comprises the steps that sequence features are obtained, candidate content features are obtained, the sequence features are sequence features of behavior sequences of single-end interaction behaviors of a user account in a first device and cross-end interaction behaviors of the user account in at least two second devices, and the candidate content features are embedding features of candidate content; sequence modeling is executed on the basis of the sequence features and the candidate content features, interest vectors are obtained, and the interest vectors are related to the candidate content features and the sequence features; mapping the candidate content features and the interest vectors to obtain interest degree scores of candidate contents corresponding to the candidate content features; and recommending the target content from the candidate content based on the interestingness score of the candidate content.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and particularly to a content recommendation method, apparatus, device, medium and program product. Background Art

[0002] In a recommendation system, content (item) that a user account may be interested in is exposed to the user account, and the user account can perform a series of actions, such as click, play, swipe away, like, comment, share, reward, repeat play, exit. These actions express the degree of preference of the user account for the content. The key point of sequence modeling is to utilize the sequence of these actions to obtain an effective interest vector to represent the user's interest.

[0003] In the related art, a recommendation system model is used to perform sequence modeling. Based on the sequence of each historical action of the user account, the similarity with the candidate content is calculated, and the similarity is used as the weight of the historical action. The weighted historical actions are aggregated to obtain an interest vector of the user account for the candidate content.

[0004] However, for a recommendation system model, at the feature input level, it often involves multiple sequences, and these sequences may be cross-terminal. For example, the interaction material sequence and search content sequence of the user account on the current device, and the interaction material sequence and search content sequence of the user account on at least two other devices. When using the method of the related art to perform sequence modeling on each sequence, it is easy to have the problem of noise interference caused by cross-terminal differences, resulting in the interest vector being unable to accurately represent the user's interest. Summary of the Invention

[0005] The present application provides a content recommendation method, apparatus, device, medium and program product. The technical solution is as follows:

[0006] On the one hand, a content recommendation method is provided, and the method includes:

[0007] Obtain sequence features, and obtain candidate content features, where the sequence features are the sequence features of the behavior sequence of the single-terminal interaction behavior of the user account on the first device and the cross-terminal interaction behavior on at least two second devices, and the candidate content features are the embedding features of the candidate content;

[0008] Perform sequence modeling based on the sequence features and the candidate content features to obtain an interest vector, where the interest vector is related to the sequence features and the candidate content features;

[0009] Map the candidate content features and the interest vector to obtain the interest degree score of the candidate content corresponding to the candidate content features;

[0010] Recommend target content from the candidate content based on the interest score of the candidate content.

[0011] On the other hand, a content recommendation device is provided, and the device includes:

[0012] An acquisition module, configured to acquire sequence features and acquire candidate content features, where the sequence features are sequence features of a behavior sequence of a single-end interaction behavior of a user account in a first device and a cross-end interaction behavior in at least two second devices, and the candidate content features are embedding features of candidate content;

[0013] A modeling module, configured to perform sequence modeling based on the sequence features and the candidate content features to obtain an interest vector, where the interest vector is related to the sequence features and the candidate content features;

[0014] A mapping module, configured to map the candidate content features and the interest vector to obtain the interest score of the candidate content corresponding to the candidate content features;

[0015] A recommendation module, configured to recommend target content from the candidate content based on the interest score of the candidate content.

[0016] On the other hand, a computer device is provided, and the computer device includes: a processor and a memory, where the memory stores a computer program, and the computer program is loaded and executed by the processor to implement the content recommendation method as described above.

[0017] On the other hand, a computer-readable storage medium is provided, and the computer-readable storage medium stores a computer program, and the computer program is loaded and executed by a processor to implement the content recommendation method as described above.

[0018] On the other hand, a computer program product is provided, and the computer program product includes computer instructions, where the computer instructions are stored in a computer-readable storage medium, and a processor obtains the computer instructions from the computer-readable storage medium, so that the processor loads and executes to implement the content recommendation method as described above.

[0019] The beneficial effects brought by the technical solutions provided in the embodiments of the present application at least include:

[0020] A new content recommendation method is provided. The sequence features of the single - end interaction behavior of the user account in the first device and the cross - end interaction behavior in at least two second devices are introduced, and sequence modeling is performed based on the sequence features and the candidate content features to obtain an interest vector. It can fully capture and fuse the interaction habits and preferences corresponding to the single - end interaction behavior and cross - end interaction behavior of the user account, suppress the problem of noise interference caused by cross - end and cross - domain differences, reduce the effect loss caused by noise interference, enable the interest vector to accurately represent the user's interest, and thus improve the accuracy of content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following - described drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0022] Figure 1 It is a block diagram of the structure of a computer system provided by an exemplary embodiment of the present application;

[0023] Figure 2 It is a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0024] Figure 3 It is a schematic diagram of the architecture of a content recommendation model provided by an exemplary embodiment of the present application;

[0025] Figure 4 It is a schematic diagram of the architecture of a content recommendation model provided by an exemplary embodiment of the present application;

[0026] Figure 5 It is a schematic diagram of the architecture of a content recommendation model provided by an exemplary embodiment of the present application;

[0027] Figure 6 It is a flowchart of a content recommendation method provided by an exemplary embodiment of the present application;

[0028] Figure 7 It is an overall schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0029] Figure 8 It is a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0030] Figure 9 It is a schematic diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0031] Figure 10It is a flowchart of a content recommendation method provided by an exemplary embodiment of the present application;

[0032] Figure 11 It is a schematic application diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0033] Figure 12 It is a schematic application diagram of a content recommendation method provided by an exemplary embodiment of the present application;

[0034] Figure 13 It is a schematic diagram of metrics of a content recommendation method provided by an exemplary embodiment of the present application;

[0035] Figure 14 It is a schematic diagram of metrics of a content recommendation method provided by an exemplary embodiment of the present application;

[0036] Figure 15 It is a block diagram of a content recommendation device provided by an exemplary embodiment of the present application;

[0037] Figure 16 It is a block diagram of the structure of a computer device provided by an exemplary embodiment of the present application. Detailed implementation manners

[0038] To make the objectives, technical solutions, and advantages of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the accompanying drawings.

[0039] Here, the exemplary embodiments will be described in detail, and the examples are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0040] The terms used in the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "the", and "said" used in the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0041] It should be understood that although the terms first, second, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, the first parameter may also be referred to as the second parameter, and similarly, the second parameter may also be referred to as the first parameter. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to a determination".

[0042] It should be noted that before and during the process of collecting relevant data of the user (for example, a series of behaviors, behavior sequences) in this application, a prompt interface, a pop-up window or voice prompt information can be displayed. The prompt interface, the pop-up window or the voice prompt information is used to prompt the user that their relevant data is being collected currently, so that this application only starts to execute the relevant steps of obtaining the user's relevant data after obtaining the confirmation operation issued by the user for the prompt interface or the pop-up window. Otherwise, when the confirmation operation issued by the user for the prompt interface or the pop-up window is not obtained, the relevant steps of obtaining the user's relevant data are ended, that is, the relevant data of the user is not obtained. In other words, all user data collected in this application is collected with the consent and authorization of the user, and the collection, use and processing of the relevant user data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0043] First, briefly introduce the nouns involved in the embodiments of this application:

[0044] Cross-Attention: It is an operation based on the attention mechanism, used to establish a connection between two different input behavior sequences. It allows each element in one behavior sequence to be weighted according to the elements in the other behavior sequence, so as to achieve cross-sequence information interaction. Cross-Attention is often used when it is necessary to fuse features from different sources. By effective cross-sequence information fusion, the expression ability of the model can be enhanced, especially suitable for scenarios dealing with heterogeneous data or multi-source information. The attention mechanism involves the following three matrices: query (q) matrix, key (k) matrix, and value (v) matrix.

[0045] Behavior sequence: Also known as a sequence, it is usually composed of a series of interaction behaviors of a user account over a period of time. These behaviors can be at least one of the following: click, browse, purchase, comment. The interaction behaviors in the embodiments of this application include: single-terminal interaction behaviors and cross-terminal interaction behaviors. A single-terminal interaction behavior is an interaction behavior of a user account in a single first terminal. For example, the user account logs in to a video client in the first terminal and plays a video. A cross-terminal interaction behavior is an interaction behavior of a user account in at least two second terminals. For example, the user account logs in to a video client in one second terminal, plays a video, logs in to the same video client in another second terminal, and continues to play the video based on the previous play progress.

[0046] In some embodiments, based on different application scenarios, the specific content of the interaction behavior is different. Taking the single-terminal interaction behavior of a user account in a first device as an example, in the video playback scenario, the single-terminal interaction behavior is a playback behavior, including at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit. In the e-commerce consumption scenario, the single-terminal interaction behavior is a purchase behavior, including at least one of the following: click, swipe away, favorite, unfavorite, add to cart, delete, purchase, cancel purchase, comment, share, send a gift, receive a gift, exit.

[0047] In some embodiments, at least two second devices may include the first device or may be other devices other than the first device. The first device and at least two second devices may also be respectively referred to as: user devices or terminals. The device types include at least one of the following: mobile phone, tablet computer, in-vehicle terminal (car computer), wearable device, personal computer, unmanned reservation terminal, smart home appliance, intelligent voice interaction device, unmanned vending terminal. For the first device and at least two second devices, and for at least two second devices, the device types may be the same, partially the same, or different. When the device types of the first device and at least two second devices are completely the same, it can be understood as: cross-terminal. When the types of the first device and at least two second devices are partially the same or different, it can be understood as: cross-terminal, cross-domain.

[0048] Sequence modeling: It is a way to perform modeling and analysis on the behavior sequence of the interaction behavior of a user account to capture the interaction habits and preferences of the user account. The sequence modeling in the embodiments of this application includes: single-end sequence modeling and cross-end sequence modeling. Single-end sequence modeling is also called: the first sequence modeling, which is used to perform modeling and analysis on the behavior sequence of the single-end interaction behavior of a user account in a first device to capture the interaction habits and preferences of the user account in the first device. Cross-end sequence modeling is also called: the second sequence modeling, cross-end behavior sequence modeling, which is used to perform unified modeling and analysis on the behavior sequence of the cross-end interaction behavior of a user account in at least two second devices to capture the interaction habits and preferences of the user account between at least two second devices. By integrating the cross-end behaviors of the user account in at least two second devices into a continuous behavior sequence, the understanding of the user's interests and intentions is improved, and the effect of improving the recommendation system or personalized service is achieved. In practical applications, the main challenges faced by cross-end sequence modeling at least include: the diversity and cross-end differences of interaction behaviors, cross-domain differences, and how to process and fuse behavior sequences from different sources. The second sequence modeling in this embodiment also involves the first sequence features, that is, it will fuse the relevant information of the first sequence features to guide the second sequence modeling of the second sequence features through the first sequence features and suppress the problem of noise interference.

[0049] Content recommendation model: It is a neural network model provided in the embodiments of this application, which is used to perform sequence modeling and determine the interest degree score, so as to achieve content recommendation. Specifically, the content recommendation model is used to perform sequence modeling based on candidate content features and sequence features to obtain an interest vector; map the candidate content features and the interest vector to obtain the interest degree score of the candidate content corresponding to the candidate content features; and recommend target content from the candidate content based on the interest degree score of the candidate content.

[0050] Figure 1 It is a structural block diagram of a computer system provided by an exemplary embodiment of this application. The computer system 100 can be implemented as the system architecture of the content recommendation method. The computer system 100 includes: a terminal 120 and a server 140.

[0051] The terminal 120 may be an electronic device such as a mobile phone, a tablet computer, an in-vehicle terminal (car computer), a wearable device, a personal computer (PC), an unattended reservation terminal, a smart home appliance, a smart voice interaction device, an unattended vending terminal, etc. A client for running the target application may be installed in the terminal 120. The target application may be an application for performing content recommendation, an application for performing and / or training a content recommendation model, an application providing a content recommendation function, or other applications providing a content recommendation model training function, which is not limited herein. In addition, the form of the target application is not limited, including but not limited to an application (App) installed in the terminal 120, a mini-program, etc., and may also be in the form of a web page. In some embodiments, the terminal 120 may also store a content recommendation model.

[0052] The server 140 may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The server 140 may be the background server of the above target application, and is used to provide background services for the client of the target application. In some alternative embodiments, the server 140 may also be implemented as a node in a blockchain system. In some embodiments, the server 140 may also store a content recommendation model.

[0053] Communication may be performed between the terminal 120 and the server 140 via a network, such as a wired or wireless network.

[0054] In the content recommendation method provided by the embodiments of the present application, the execution subject of each step may be a computer device, and the computer device refers to an electronic device having data calculation, processing, and storage capabilities. Taking Figure 1 the solution implementation environment shown as an example, the content recommendation method may be executed by the terminal 120. For example, the client of the target application installed and running in the terminal 120 executes the content recommendation method, or the content recommendation method may be executed by the server 140, or executed by the interaction and cooperation of the terminal 120 and the server 140, which is not limited herein.

[0055] Those skilled in the art may be aware that the number of terminals 120 may be more or less. For example, there may be only one terminal 120, or there may be dozens or hundreds of terminals 120, or more. The embodiments of the present application do not limit the number and device type of the terminals 120.

[0056] The recommendation system model includes the Deep Interest Network (DIN). The Deep Interest Network is a neural network model for sequence modeling in the recommendation system, mainly used to capture the short-term interests of user accounts. Through the attention mechanism, the Deep Interest Network selects behaviors related to the current candidate content / item from the behavior sequence of the user account, dynamically focuses on important behaviors, and improves the recommendation effect. The Deep Interest Network uses an adaptive local activation network to dynamically adjust the weights of the behaviors of the user account according to the current candidate item, achieving the purpose of filtering and weighting.

[0057] Specifically, expose the content (item) that the user account may be interested in to the user account, and the user account can perform a series of behaviors, such as click, play, swipe away, like, comment, share, reward, repeat play, exit. Based on the sequence of each historical behavior of the user account, the Deep Interest Network calculates the similarity between the historical behavior and the candidate content, and uses the similarity as the weight of the historical behavior. Aggregate the weighted historical behaviors to obtain the interest representation of the user account for the candidate content, also known as the interest vector or interest feature.

[0058] However, for the recommendation system model, at the feature input level, it often involves multiple sequences, and these sequences may be cross-terminal. For example, the interaction material sequence and search content sequence of the user account on the current device, and the interaction material sequence and search content sequence of the user account on other devices. When using the methods of related technologies to perform sequence modeling on each sequence, problems such as cross-terminal differences and noise interference caused by cross-domains are likely to occur, resulting in the inability of the interest vector to accurately represent the user's interests.

[0059] The embodiments of the present application provide a content recommendation method. By introducing multiple single-terminal and cross-terminal sequences to perform sequence modeling, an interest vector is obtained, the loss of effect caused by the problem of noise interference is suppressed, the interest vector accurately represents the user's interests, and the accuracy of content recommendation is improved.

[0060] Figure 2 It is a schematic diagram of the content recommendation method provided by an exemplary embodiment of the present application. This method is executed by a computer device, taking the computer device as the server 142 as an example for illustration. The steps are briefly described as follows:

[0061] Step 1: Obtain sequence features and obtain candidate content features. The sequence features are the sequence features of the behavior sequence of the single-terminal interaction behavior of the user account on the first device and the cross-terminal interaction behavior on at least two second devices. The candidate content features are the embedding features of the candidate content; refer to Figure 2, taking the first device as the terminal 132 and at least two second devices as the terminal 132 and the terminal 134 as an example, the sequence feature is the sequence feature of the single - end interaction behavior of the user account in the terminal 132 and the cross - end interaction behavior in the terminal 132 and the terminal 134, and the candidate content feature is the embedding feature of the recalled candidate content; for example, in the video - playing scenario, the single - end interaction behavior is the playing behavior of the user account in the terminal 132, and the cross - end interaction behavior is the playing behavior of the user account in the terminal 132 and the playing behavior in the terminal 134. Optionally, in the terminal 134, the playing behavior can be continued at the progress of the playing behavior in the terminal 132. The playing behavior includes at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit.

[0062] Step 2: Perform sequence modeling based on the sequence feature and the candidate content feature to obtain an interest vector; the interest vector is related to the sequence feature and the candidate content feature. Refer to Figure 2 , and the interest vector is used to represent the interest of the user account in the candidate content corresponding to the candidate content feature.

[0063] Step 3: Map the candidate content feature and the interest vector to obtain the interest degree score of the candidate content corresponding to the candidate content feature; the higher the interest degree score, the greater the possibility that the user account is interested in the candidate content.

[0064] Step 4: Recommend target content from the candidate content based on the interest degree score of the candidate content.

[0065] For the detailed embodiment content, please refer to the following text.

[0066] · Model Architecture

[0067] Figure 3 is a schematic diagram of the architecture of the content recommendation model provided by an exemplary embodiment of the present application.

[0068] Refer to Figure 3 , the content recommendation model 200 for performing the content recommendation method includes: a sequence modeling network 10 and a mapping network 20 cascaded in sequence; wherein, the sequence modeling network 10 is used to perform sequence modeling based on the candidate content feature and the sequence feature to obtain an interest vector; the mapping network 20 is used to map the candidate content feature and the interest vector to obtain the interest degree score of the candidate content corresponding to the candidate content feature.

[0069] Continue to refer to Figure 3, the sequence modeling network 10 includes: a first sequence modeling network 11 and a second sequence modeling network 12 connected in parallel. Among them, the first sequence modeling network 11 is used to perform first sequence modeling on the first sequence feature based on the candidate content feature to obtain a first interest vector; the second sequence modeling network 12 is used to perform second sequence modeling on the second sequence feature based on the candidate content feature and the first sequence feature to obtain a second interest vector.

[0070] In some embodiments, the first sequence feature is also referred to as: the main sequence feature. The first sequence modeling network 11 is also referred to as at least one of the following names: single-ended sequence modeling network, main sequence modeling network, main sequence target attention modeling network. The second sequence feature is also referred to as: the auxiliary sequence feature. The second sequence modeling network 12 is also referred to as at least one of the following names: cross-ended sequence modeling network, auxiliary sequence modeling network, auxiliary sequence double-layer double-query matrix cross attention modeling network.

[0071] Figure 4 It is a schematic diagram of the architecture of a content recommendation model provided by an exemplary embodiment of the present application.

[0072] Reference Figure 4 , the second sequence modeling network 12 includes: a second numerical signal embedding network (Engagement Network) 121, a dot product cross attention network 122, and a second target attention network 123 connected in series in sequence. Among them, the second numerical signal embedding network 121 is used to perform signal strength weighting on the second sequence feature to obtain a second content feature; the dot product cross attention network 122 is used to perform dot product cross attention processing on the first sequence feature and the second content feature to obtain a cross attention feature; the second target attention network 123 is used to perform target attention processing on the candidate content feature and the cross attention feature to obtain a second interest vector.

[0073] Figure 5 It is a schematic diagram of the architecture of a content recommendation model provided by an exemplary embodiment of the present application.

[0074] Reference Figure 5, the first sequence modeling network 11 includes: a first numerical signal embedding network (Engagement Network) 111 and a first target attention network 112 cascaded in sequence. Among them, the first numerical signal embedding network 111 is used to perform signal strength weighting on the first sequence feature to obtain the first content feature; the first target attention network 112 is used to perform target attention processing on the candidate content feature and the first content feature to obtain the first interest vector.

[0075] In a possible implementation manner, the mapping network 20 is implemented based on a multilayer perceptron (MLP), or based on a more complex recommendation system model capable of mapping to obtain an interest degree score, or a multi-objective and multi-scenario neural network model. The multilayer perceptron is used to capture the complex mapping relationship from input to output, so as to realize the prediction of the interest degree score. The structures of the first numerical signal embedding network 111 and the second numerical signal embedding network 121 may be the same, and at least include a fully connected layer (FC Layer) in their structures. The structures of the first target attention network 112 and the second target attention network 123 may be the same, and both are implemented based on the target attention mechanism. The dot-product cross-attention network 122 is implemented based on the dot-product cross-attention mechanism.

[0076] · Inference stage

[0077] Figure 6 is a flowchart of a content recommendation method provided by an exemplary embodiment of the present application. This method is executed by a computer device, and the computer device stores a trained content recommendation model. Specifically, this method can be executed by this trained content recommendation model. Optionally, the computer device may be Figure 1 the terminal 120 and / or the server 140 shown. This method includes at least some of the steps: step 220, step 240, step 260, and step 280:

[0078] Step 220, obtain the sequence feature, and obtain the candidate content feature. The sequence feature is the sequence feature of the single-end interaction behavior of the user account in the first device and the cross-end interaction behavior in at least two second devices, and the candidate content feature is the embedding feature of the candidate content.

[0079] The sequence feature is the sequence feature of the behavior sequence of the interaction behavior of the user account. In this embodiment, the interaction behavior of the user account includes the single-end interaction behavior in the first device and the cross-end interaction behavior in at least two second devices. The single-end interaction behavior is the interaction behavior of the user account in a single first terminal. For example, the user account logs in to the video client in the first terminal and plays a video. The cross-end interaction behavior is the interaction behavior of the user account in at least two second terminals. For example, the user account logs in to the video client in one second terminal and plays a video, and logs in to the same video client in another second terminal and continues to play the video based on the previous play progress.

[0080] In some embodiments, based on different application scenarios, the specific content of the interaction behavior is different. Taking the single-end interaction behavior of the user account in the first device as an example, in the video playback scenario, the single-end interaction behavior is the playback behavior, including at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit. In the e-commerce consumption scenario, the single-end interaction behavior is the purchase behavior, including at least one of the following: click, swipe away, favorite, unfavorite, add to cart, delete, purchase, cancel purchase, comment, share, send a gift, receive a gift, exit.

[0081] In some embodiments, at least two second devices may include the first device or may be other devices other than the first device. The first device and at least two second devices may also be respectively referred to as: user devices or terminals, and the device types include at least one of the following: mobile phone, tablet computer, in-vehicle terminal (carputer), wearable device, personal computer, unmanned reservation terminal, smart home appliance, smart voice interaction device, unmanned vending terminal. For the first device and at least two second devices, and for at least two second devices, the device types may be the same, partially the same, or different. When the device types of the first device and at least two second devices are completely the same, it can be understood as: cross-end. When the types of the first device and at least two second devices are partially the same or different, it can be understood as: cross-end, cross-domain.

[0082] In one possible implementation, the single-end interaction behavior of the user account in the first device and the cross-end interaction behavior in at least two second devices are combined into one behavior sequence. In another possible implementation, the single-end interaction behavior of the user account in the first device is combined into one behavior sequence, also referred to as the first sequence or the main sequence. The cross-end interaction behavior of the user account in at least two second devices is combined into one behavior sequence, also referred to as the second sequence or the auxiliary sequence.

[0083] It should also be noted that considering that a user usually logs in with a user account, the user account involved in the above-mentioned first device and the user accounts involved in at least two second devices are the same user account. In some other possible cases, the user account involved in the above-mentioned first device refers to all user accounts logged in on the first device, and the user accounts involved in at least two second devices are one of these all user accounts logged in on the first device. For example, the user accounts involved in the first device are user account 1 and user account 2, and the user accounts involved in at least two second devices are user account 1.

[0084] The candidate content feature is the embedding feature of the candidate content (item). Among them, the candidate content is pre-determined, which can be common candidate content or recalled candidate content. The candidate content can be one or multiple. In the case where the candidate content is multiple, the method of this embodiment is respectively executed for each candidate content to obtain the interest degree score for each candidate content.

[0085] In some embodiments, based on different application scenarios, the specific content of the candidate content is different. For example, in the video playback scenario, the candidate content includes at least one of the following: TV series, variety shows, animations, movies, talk shows, documentaries, short videos. In the e-commerce consumption scenario, the candidate content includes at least one of the following: daily necessities, clothing, furniture supplies, shoes and socks, cosmetics.

[0086] In some embodiments, the feature dimensions of the sequence feature and the candidate content feature are the same, and the feature dimensions can be specifically determined according to actual technical needs. For example, in one example, the feature dimension is 128 dimensions.

[0087] Step 240, perform sequence modeling based on the sequence feature and the candidate content feature to obtain an interest vector, and the interest vector is related to the sequence feature and the candidate content feature.

[0088] Sequence modeling is a way to perform modeling and analysis on the behavior sequence of the interaction behavior of the user account to capture the interaction habits and preferences of the user account. In this embodiment, the sequence modeling is performed based on the sequence feature and the candidate content feature, and can fully capture and fuse the interaction habits and preferences corresponding to the single-end interaction behavior and cross-end interaction behavior of the user account.

[0089] The interest vector is an interest representation of the user account for the candidate content corresponding to the candidate content feature, and is used to characterize the interest of the user account in the candidate content corresponding to the candidate content feature. The interest vector is related to the sequence feature and the candidate content feature. That is, by fully capturing and fusing the interaction habits and preferences corresponding to the single-end interaction behavior and cross-end interaction behavior of the user account, the effective representation of the interest vector is realized.

[0090] Exemplarily, the computer device performs sequence modeling based on sequence features and candidate content features through the sequence modeling network of the content recommendation model to obtain an interest vector.

[0091] Step 260: Map the candidate content features and the interest vector to obtain the interest degree score of the candidate content corresponding to the candidate content features.

[0092] Mapping is used to capture the relationship between the candidate content features and the interest vector, and the interest degree score is predicted. The interest degree score is predicted based on the candidate content features and the interest vector, and is used to characterize the interest degree of the user account in the candidate content corresponding to the candidate content features. When the interest degree score is higher, the greater the possibility that the user account is interested in the candidate content, and the greater the possibility of performing an interaction behavior on the candidate content.

[0093] Exemplarily, the computer device maps the candidate content features and the interest vector through the mapping network of the content recommendation model to obtain the interest degree score of the candidate content corresponding to the candidate content features.

[0094] In some embodiments, in addition to the sequence features, other features are also involved. The other features are used to characterize the needs and preferences of the user account. The other features may be at least one of the following: the portrait features of the user account, the statistical features of the behavioral content of the user account. Among them, the portrait features are used to characterize the personal needs and preferences of the user account. The statistical features of the behavioral content are used to characterize the needs and preferences of the user account for the content.

[0095] Optionally, obtain other features of the user account, map the other features, the candidate content features and the interest vector, and obtain the interest degree score of the candidate content corresponding to the candidate content features. In this embodiment, by obtaining other features of the user account, the needs and preferences of the user account can be further characterized from different dimensions and levels, which is beneficial to improving the accuracy of the interest degree score and thus improving the accuracy of content recommendation.

[0096] Step 280: Recommend target content from the candidate content based on the interest degree score of the candidate content.

[0097] The target content is the candidate content recommended to the user account. The target content may be one or more.

[0098] In some embodiments, the computer device determines candidate content with an interest score greater than the score threshold as target content, and recommends the target content to the user account. Alternatively, the interest scores of the candidate content are sorted in descending order, and the top N candidate content are determined as target content, and the target content is recommended to the user account. N is greater than or equal to 1. Optionally, the method of recommending the target content to the user account includes at least one of the following: displaying the target content through a pop-up window, prompting the target content through a text message, and rendering the target content in the current interface.

[0099] In summary, for the content recommendation method provided in the embodiments of the present application, the computer device obtains a sequence feature and a candidate content feature. The sequence feature is the sequence feature of the behavior sequence of the single-end interaction behavior of the user account in the first device and the cross-end interaction behavior in at least two second devices. The candidate content feature is the embedding feature of the candidate content; sequence modeling is performed based on the sequence feature and the candidate content feature to obtain an interest vector, and the interest vector is related to the sequence feature and the candidate content feature; the candidate content feature and the interest vector are mapped to obtain the interest score of the candidate content corresponding to the candidate content feature; based on the interest score of the candidate content, target content is recommended from the candidate content. Accordingly, a new content recommendation method is provided. By introducing the sequence feature of the behavior sequence of the single-end interaction behavior of the user account in the first device and the cross-end interaction behavior in at least two second devices, and performing sequence modeling based on the sequence feature and the candidate content feature to obtain an interest vector, it can fully capture and integrate the interaction habits and preferences corresponding to the single-end interaction behavior and cross-end interaction behavior of the user account, suppress the problem of noise interference caused by cross-end and cross-domain differences, reduce the effect loss caused by noise interference, enable the interest vector to accurately represent the user's interest, and thus improve the accuracy of content recommendation.

[0100] In some embodiments, the single-end interaction behaviors of the user account in the first device are combined into a behavior sequence, which is also called the first sequence or the main sequence. The cross-end interaction behaviors of the user account in at least two second devices are combined into a behavior sequence, which is also called the second sequence or the auxiliary sequence.

[0101] The sequence feature includes: a first sequence feature and a second sequence feature. The first sequence feature can also be called: the main sequence feature, the single-end sequence feature. The second sequence feature can also be called: the auxiliary sequence feature, the cross-end sequence feature. The first sequence feature is the sequence feature of the behavior sequence of the single-end interaction behavior of the user account in the first device, and the second sequence feature is the sequence feature of the behavior sequence of the cross-end interaction behavior of the user account in at least two second devices. The interest vector includes: a first interest vector and a second interest vector. The first interest vector is related to the first sequence feature and the candidate content feature, and the second interest vector is related to the first sequence feature, the second sequence feature, and the candidate content feature.

[0102] Specifically, step 260 is implemented as step 320 and step 340 (not shown in the figure). Among them, step 320 and step 340 can be executed successively or synchronously, and there is no limitation on this.

[0103] Step 320, based on the candidate content features, performs first sequence modeling on the first sequence features to obtain a first interest vector; and step 340, based on the candidate content features and the first sequence features, performs second sequence modeling on the second sequence features to obtain a second interest vector; where the first sequence features are the sequence features of the behavior sequence of the single-end interaction behavior of the user account in the first device, and the second sequence features are the sequence features of the behavior sequence of the cross-end interaction behavior of the user account in at least two second devices.

[0104] The first sequence modeling is used to perform modeling and analysis on the behavior sequence of the single-end interaction behavior of the user account in the first device. The first interest vector is the interest vector obtained through the first sequence modeling. The first interest vector is related to the first sequence features and the candidate content features.

[0105] The second sequence modeling is used to perform modeling and analysis on the behavior sequence of the cross-end interaction behavior of the user account in at least two second devices. The second sequence modeling in this embodiment also involves the first sequence features, that is, it will fuse the relevant information of the first sequence features to guide the second sequence modeling of the second sequence features through the first sequence features, suppress the problem of noise interference, and improve the accuracy of the second interest vector. The second interest vector is the interest vector obtained through the second sequence modeling. The second interest vector is related to the first sequence features, the second sequence features, and the candidate content features.

[0106] Exemplarily, the computer device performs first sequence modeling on the first sequence features based on the candidate content features through the first sequence modeling network of the content recommendation model to obtain a first interest vector. And the computer device performs second sequence modeling on the second sequence features based on the candidate content features and the first sequence features through the second sequence modeling network of the content recommendation model to obtain a second interest vector.

[0107] This embodiment distinguishes the first sequence features and the second sequence features, and provides a specific way to perform sequence modeling based on the sequence features and the candidate content features. Among them, for the first sequence modeling, it can fully capture the interaction habits and preferences corresponding to the single-end interaction behavior of the user account and improve the accuracy of the first interest vector. For the second sequence modeling, it can guide the second sequence modeling of the second sequence features through the first sequence features, suppress the problem of noise interference, and improve the accuracy of the second interest vector, so as to accurately represent the user's interest and facilitate improving the accuracy of content recommendation.

[0108] For the second sequence modeling:

[0109] Specifically, step 340 is implemented as step 342, step 344, and step 346 (not shown in the figure):

[0110] Step 342, perform signal strength weighting on the second sequence features to obtain second content features, where the second content features are the embedding features of the behavior content corresponding to the cross-terminal interaction behaviors in the second sequence features;

[0111] Step 344, perform dot product cross-attention processing based on the first sequence features and the second content features to obtain cross-attention features;

[0112] Step 346, perform target attention processing based on the candidate content features and the cross-attention features to obtain a second interest vector.

[0113] The second content features are obtained by performing signal strength weighting on the second sequence features. The second content features are the embedding features (item embeddings) of the behavior content (item id) corresponding to the cross-terminal interaction behaviors in the second sequence features. Among them, the behavior content refers to the specific content of the cross-terminal interaction behavior. For example, in the video playback scenario, the cross-terminal interaction behavior is the playback behavior, and the behavior content includes at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit. Performing signal strength weighting on the second sequence features is implemented through the second numerical signal embedding network of the content recommendation model.

[0114] Dot product cross-attention processing can use the first sequence features to filter the interest relevance of the second content features, so as to fuse the relevant information of the first sequence features, and guide the second sequence modeling of the second sequence features through the first sequence features, thereby suppressing noise interference. The cross-attention features are obtained by performing dot product cross-attention processing on the first sequence features and the second content features. Dot product cross-attention processing is implemented through the dot product cross-attention network of the content recommendation model.

[0115] Target attention processing can use the candidate content features to model the cross-attention features and extract a second interest vector. The second interest vector is obtained by performing target attention processing on the candidate content features and the cross-attention features. Target attention processing is implemented through the second target attention network of the content recommendation model.

[0116] Exemplarily, the computer device uses the second sequence modeling network of the content recommendation model and the second numerical signal embedding network to perform signal strength weighting on the second sequence features to obtain second content features; uses the dot-product cross-attention network to perform dot-product cross-attention processing based on the first sequence features and the second content features to obtain cross-attention features; uses the second target attention network to perform target attention processing based on the candidate content features and the cross-attention features to obtain a second interest vector.

[0117] This embodiment provides a specific method for second sequence modeling. The second sequence modeling is a double-layer double-query matrix cross-attention modeling. The first sequence features and candidate content features are introduced during the second sequence modeling, and the first sequence features are fully utilized to perform interest relevance filtering on the second sequence features, which can improve the accuracy of the second interest vector.

[0118] In some embodiments, step 344 is specifically implemented as step 3441 and step 3442 (not shown in the figure):

[0119] Step 3441, perform a pooling operation on the first sequence features to obtain pooled features;

[0120] Step 3442, perform dot-product cross-attention processing based on the pooled features and the second content features to obtain cross-attention features.

[0121] The pooling operation is used to reduce the feature size and computational amount. The pooling operation can be at least one of the following types: max pooling and mean pooling. In this embodiment, the pooling operation adopted is mean pooling.

[0122] Exemplarily, the computer device performs a pooling operation on the first sequence features through the content recommendation model to obtain pooled features; uses the dot-product cross-attention network to perform dot-product cross-attention processing based on the pooled features and the second content features to obtain cross-attention features.

[0123] This embodiment provides a method for introducing the first sequence features into the second sequence modeling. Among them, the pooled features obtained by performing a pooling operation on the first sequence features are used as the input of the dot-product cross-attention processing, which can realize the full utilization of the first sequence features.

[0124] In some embodiments, the first Q matrix (query1) corresponding to the dot-product cross-attention processing is obtained based on the pooled features, and the first K matrix and the first V matrix are obtained based on the second content features. Exemplarily, the first Q matrix is the pooled features, and the first K matrix and the first V matrix are the second content features. As an example, taking the video playback scenario as an example, the first Q matrix is expressed as:

[0125] q1 = MeanPooling(mainplaylist)

[0126] Wherein, q1 is the first Q matrix, MeanPooling is mean pooling, and main playlist is the first sequence feature, specifically the behavior content (item id) indicating single-end interaction behavior, which refers to the play behavior in the video playback scenario. Step 3442 is specifically implemented as steps 3442-1, 3442-2, 3442-3, 3442-4, 3442-5, and 3442-6 (not shown in the figure). Among them, steps 3442-1 and 3442-2 can be executed sequentially or synchronously, and there is no restriction on this.

[0127] Step 3442-1: Multiply the first Q matrix by the Q weight matrix to obtain a Q product result;

[0128] Step 3442-2: Multiply the first K matrix by the K weight matrix to obtain a K product result;

[0129] Step 3442-3: Dot-multiply the Q product result and the K product result to obtain a dot product result;

[0130] Step 3442-4: Normalize the dot product result to obtain a normalized result;

[0131] Step 3442-5: Multiply the first V matrix by the V weight matrix to obtain a V product result;

[0132] Step 3442-6: Dot-multiply the V product result and the normalized result to obtain a cross-attention feature.

[0133] The Q weight matrix, K weight matrix, and V weight matrix are learnable, specifically obtained through the training stage of the content recommendation model. The normalization is implemented through the softmax function. The cross-attention feature is obtained by filtering the interest correlation of the second content feature using the first Q matrix.

[0134] Exemplarily, the computer device multiplies the first Q matrix by the Q weight matrix through the dot-product cross-attention network of the content recommendation model to obtain a Q product result; multiplies the first K matrix by the K weight matrix to obtain a K product result; dot-multiplies the Q product result and the K product result to obtain a dot product result; normalizes the dot product result to obtain a normalized result; multiplies the first V matrix by the V weight matrix to obtain a V product result; dot-multiplies the V product result and the normalized result to obtain a cross-attention feature.

[0135] As an example, the cross-attention feature is expressed as:

[0136]

[0137] Among them, output represents the cross-attention feature, softmax is normalization, and W q is the Q weight matrix, and W K is the K weight matrix, and W V is the V weight matrix, q1 is the first Q matrix, K is the first K matrix, V is the first V matrix, and ⊙ represents dot product.

[0138] This embodiment provides a specific method for dot-product cross-attention processing in second sequence modeling, which can perform interest correlation filtering on second sequence features using the first Q matrix determined based on first sequence features, realizing the ability of cross-end interaction behavior in second sequence features to perform interest correlation filtering based on single-end interaction behavior representation in first sequence features, thereby suppressing noise interference and facilitating further obtaining a second interest vector through target attention processing and improving the accuracy of the second interest vector.

[0139] In some embodiments, the second Q matrix (query2) corresponding to target attention processing is obtained based on candidate content features, and the second K matrix and second V matrix are obtained based on cross-attention features. As an example, the second Q matrix is the candidate content feature, and the second K matrix and second V matrix are the cross-attention features. Step 346 is specifically implemented as steps 3461 and 3462 (not shown in the figure):

[0140] Step 3461, obtain an attention weight matrix based on the second Q matrix and the second K matrix;

[0141] Step 3462, perform weighted summation on the second V matrix based on the attention weight matrix to obtain a second interest vector.

[0142] Target attention processing is used to capture the relationship between each element in candidate content features and cross-attention features. For each element in the cross-attention features, target attention processing calculates the interest correlation with candidate content features, and each element respectively obtains a weight value related to the candidate content features. Based on the weight values, weighted summation is performed on each element to obtain the second interest vector.

[0143] Exemplarily, the computer device obtains an attention weight matrix based on the second Q matrix and the second K matrix through a second target attention network; performs weighted summation on the second V matrix based on the attention weight matrix to obtain a second interest vector.

[0144] This embodiment provides a specific method for target attention processing in second sequence modeling, which can obtain a second interest vector and improve the accuracy of the second interest vector.

[0145] In some embodiments, step 342 is specifically implemented as steps 3421, 3422, 3423, and 3424 (not shown in the figure):

[0146] Step 3421: Determine a second numerical signal corresponding to the second sequence feature, where the second numerical signal is used to characterize the intensity of the behavior content corresponding to the cross-terminal interaction behavior in the second sequence feature;

[0147] Step 3422: Sequentially perform non-linear transformation and stacking on the second numerical signal to obtain a second signal feature;

[0148] Step 3423: Map the second signal feature to obtain a second signal mapping feature;

[0149] Step 3424: Perform dot product on the second sequence feature and the second signal mapping feature to obtain a second content feature.

[0150] The second numerical signal is a numerical signal corresponding to the second sequence feature. The second numerical signal is used to characterize the intensity of the behavior content (item id) corresponding to the cross-terminal interaction behavior in the second sequence feature. For example, in the field of video playback, the numerical signal includes at least one of the following: playback duration, playback times, number of viewers, average playback duration per viewer, number of member views, 10% / 30% / 50% / 90% playback completion numbers. Mapping the second signal feature can be implemented through a fully connected layer.

[0151] Exemplarily, the computer device sequentially performs non-linear transformation and stacking (stack) on the second numerical signal through the second numerical signal embedding network to obtain a second signal feature; maps the second signal feature through the fully connected layer in the second numerical signal embedding network to obtain a second signal mapping feature; performs dot product on the second sequence feature and the second signal mapping feature to obtain a second content feature.

[0152] This embodiment provides a method for weighting the signal intensity of the behavior content itself in the second sequence feature. Since the second numerical signal involved in the second sequence feature has important value for the behavior content itself and can characterize the consumption intensity of the behavior content itself, performing signal intensity weighting at the shallowest layer of the second sequence modeling can highlight the behavior content with stronger intensity, which is beneficial to capturing user interests and improving the accuracy of the second interest vector.

[0153] For the first sequence modeling:

[0154] Specifically, step 320 is implemented as steps 322 and 324 (not shown in the figure):

[0155] Step 322: Perform signal strength weighting on the first sequence feature to obtain a first content feature, where the first content feature is the embedded feature of the behavior content corresponding to the single - end interaction behavior in the first sequence feature;

[0156] Step 324: Perform target attention processing based on the candidate content feature and the first content feature to obtain a first interest vector.

[0157] The first content feature is obtained by performing signal strength weighting on the first sequence feature. The first content feature is the embedded feature (item embedding) of the behavior content (item id) corresponding to the single - end interaction behavior in the first sequence feature. Here, the behavior content refers to the specific content of the single - end interaction behavior. For example, in the video playback scenario, the single - end interaction behavior is the playback behavior, and the behavior content includes at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit. Performing signal strength weighting on the first sequence feature is implemented through the first numerical signal embedding network of the content recommendation model.

[0158] Target attention processing can model the first content feature using the candidate content feature and extract the first interest vector. The first interest vector is obtained by performing target attention processing on the candidate content feature and the first content feature. Target attention processing is implemented through the first target attention network of the content recommendation model.

[0159] Exemplarily, the computer device, through the first sequence modeling network of the content recommendation model and the first numerical signal embedding network, performs signal strength weighting on the first sequence feature to obtain the first content feature; through the first target attention network, performs target attention processing based on the candidate content feature and the first content feature to obtain the first interest vector.

[0160] This embodiment provides a specific method for the first sequence modeling, which can make full use of the first sequence feature and the candidate content feature to improve the accuracy of the first interest vector.

[0161] In some embodiments, the Q matrix corresponding to the target attention processing is obtained based on the candidate content feature, and the K matrix and the V matrix are obtained based on the first content feature. Exemplarily, the Q matrix is the candidate content feature, and the K matrix and the V matrix are the first content feature. Step 324 is specifically implemented as Step 3241 and Step 3242 (not shown in the figure):

[0162] Step 3241: Based on the Q matrix and the K matrix, obtain an attention weight matrix;

[0163] Step 3242: Perform weighted summation on the V matrix based on the attention weight matrix to obtain the first interest vector.

[0164] The target attention processing is used to capture the relationships between the candidate content features and the respective elements in the first content features. For each element in the first content features, the target attention processing calculates the interest correlation with the candidate content features, and each element separately obtains a weight value related to the candidate content features. By performing weighted summation on each element based on the weight values, the first interest vector can be obtained.

[0165] Exemplarily, the computer device obtains an attention weight matrix based on the Q matrix and the K matrix through the first target attention network; performs weighted summation on the V matrix based on the attention weight matrix to obtain the first interest vector.

[0166] This embodiment provides a specific manner of target attention processing in the first sequence modeling, which can obtain the first interest vector and improve the accuracy of the first interest vector.

[0167] In some embodiments, step 322 is specifically implemented as step 3221, step 3222, step 3223, and step 3224 (not shown in the figure):

[0168] Step 3221, determine the first numerical signal corresponding to the first sequence feature, where the first numerical signal is used to characterize the intensity of the behavior content corresponding to the single-end interaction behavior in the first sequence feature;

[0169] Step 3222, perform non-linear transformation and stacking on the first numerical signal in sequence to obtain the first signal feature;

[0170] Step 3223, map the first signal feature to obtain the first signal mapping feature;

[0171] Step 3224, perform dot product on the first sequence feature and the first signal mapping feature to obtain the first content feature.

[0172] The first numerical signal is the numerical signal corresponding to the first sequence feature. The first numerical signal is used to characterize the intensity of the behavior content (item id) corresponding to the single-end interaction behavior in the first sequence feature. For example, in the field of video playback, the numerical signal includes at least one of the following: playback duration, playback times, number of viewers, average playback duration per viewer, number of member views, 10% / 30% / 50% / 90% playback completion number. Mapping the second signal feature can be achieved through a fully connected layer.

[0173] Exemplarily, the computer device performs non-linear transformation and stacking on the first numerical signal in sequence through the first numerical signal embedding network to obtain the first signal feature; maps the first signal feature through the fully connected layer in the first numerical signal embedding network to obtain the first signal mapping feature; performs dot product on the first sequence feature and the first signal mapping feature to obtain the first content feature.

[0174] This embodiment provides a method for weighting the signal strength of the behavior content itself in the first sequence feature. Since the first numerical signal involved in the first sequence feature has important value for the behavior content itself and can represent the consumption intensity of the behavior content itself, performing signal strength weighting at the shallowest layer of the first sequence modeling can highlight the behavior content with stronger intensity, which is beneficial to capturing the user's interest and improving the accuracy of the first interest vector.

[0175] As an example, the content recommendation method provided by the embodiments of the present application will be described as a whole with reference to the schematic diagram.

[0176] Figure 7 It is a schematic diagram of the overall content recommendation method provided by an exemplary embodiment of the present application.

[0177] The computer device stores a content recommendation model for executing the content recommendation method. The content recommendation model includes: a first sequence modeling network 11 and a second sequence modeling network 12 connected in parallel. The outputs of the first sequence modeling network 11 and the second sequence modeling network 12 are also respectively connected to a Multilayer Perceptron (MLP) 22. In some embodiments, the first sequence feature is also referred to as: the main sequence feature. The second sequence feature is also referred to as: the auxiliary sequence feature. The first sequence modeling network 11 is also referred to as the main sequence Target Attention modeling network, and the second sequence modeling network 12 is also referred to as the auxiliary sequence double-layer double-query (query) matrix Cross Attention modeling network.

[0178] The first sequence modeling network 11 includes: a first numerical signal embedding network (EngagementNetwork) 111 and a first Target Attention network 112 cascaded in sequence. The second sequence modeling network 12 includes: a second numerical signal embedding network (Engagement Network) 121, a dot product CrossAttention network 122, and a second Target Attention network 123 cascaded in sequence.

[0179] Taking the first sequence feature as an example, with reference to Figure 7 , the first sequence feature in the first sequence modeling network 11 is represented by squares, and each square represents a vector of the behavior content of a single-ended interaction behavior in the first sequence feature. For example, if the length of the first sequence feature (20, 128) is 20 and the dimension is 128, then there are 20 corresponding squares, and each square represents a vector (1, 128) of the behavior content of a single-ended interaction behavior.

[0180] The structures of the first numerical signal embedding network 111 and the second numerical signal embedding network 121 are the same, and both can refer to Figure 8 The numerical signal embedding network shown. Among them, the input of the numerical signal embedding network is a numerical signal and sequence features. Taking the video playback scenario as an example, the numerical signals are playback duration, playback times, and so on. The output of the numerical signal embedding network is content features. Since the numerical signals involved in the sequence features have important value for the behavior content itself, and this numerical signal is used to characterize the intensity of the behavior content, therefore, at the shallowest layer of sequence modeling, the numerical signal is used to perform signal intensity weighting on the behavior content itself in the sequence features.

[0181] The dot-product cross-attention network 122 can refer to Figure 9 Shown. The input of the dot-product cross-attention network is the first Q matrix (query1) and content features, and the output is cross-attention features. In this embodiment, the first sequence feature and the second sequence feature are distinguished, and through the dot-product cross-attention network, the first sequence feature is used to perform interest correlation filtering on the second sequence feature. The calculation method is:

[0182] Use mean pooling (Mean Pooling) to calculate the pooled feature of the first sequence feature as the query of the dot-product cross-attention network, denoted as q1, which is expressed as:

[0183] q1 = MeanPooling(main playlist)

[0184] Use q1 to perform interest correlation filtering on the second sequence feature, which is expressed as:

[0185]

[0186] The structures of the first target attention network 112 and the second target attention network 123 are the same, and both are target attention networks. The candidate content feature (target item) is used to perform modeling on the sequence features to extract the interest vector. Inputting the interest vector into the multi-layer perceptron 22 can obtain the interest score. Specifically:

[0187] Obtain the first sequence feature, the second sequence feature, the candidate content feature, and other features (other features).

[0188] For the first sequence modeling network 11: Through the first numerical signal embedding network 111, perform signal intensity weighting on the first sequence feature to obtain the first content feature; through the first target attention network 112, perform target attention processing on the candidate content feature and the first content feature to obtain the first interest vector, where the Q matrix (query) used by the first target attention network 112 is determined based on the candidate content feature.

[0189] For the second sequence modeling network 12: Through the second numerical signal embedding network 121, signal strength weighting is performed on the second sequence features to obtain second content features; through the dot product cross-attention network 122, dot product cross-attention processing is performed on the pooled features and the second content features to obtain cross-attention features, where the first Q matrix (query1) used by the cross-attention network 122 is determined based on the pooled features, and the pooled features are obtained by performing average pooling operation on the first sequence features; through the second target attention network 123, target attention processing is performed on the candidate content features and the cross-attention features to obtain a second interest vector, where the second Q matrix (query2) used by the second target attention network 123 is determined based on the candidate content features.

[0190] For the multi-layer perceptron 22: Input the first interest vector, the second interest vector, the candidate content features, and other features into the multi-layer perceptron 22, and output the interest degree score of the candidate content corresponding to the candidate content features. The computer device can continue to recommend target content according to the interest degree score of the candidate content.

[0191] · Training stage

[0192] Figure 10 It is a flowchart of a content recommendation method provided by an exemplary embodiment of the present application. This method is executed by a computer device, and the computer device stores a content recommendation model to be trained. Specifically, this method can be executed by the content recommendation model to be trained. Optionally, the computer device can be Figure 1 the terminal 120 and / or the server 140 shown in the figure. It should also be noted that the computer device in the training stage of the content recommendation model and the computer device in the inference stage can be the same or different, which is not limited here. This method includes at least some of the steps of step 420, step 440, step 460, and step 480:

[0193] Step 420, obtain sample sequence features, and obtain sample content features. The sample sequence features are the sequence features of the behavior sequence of the single-end interaction behavior of the sample user account in the first sample device and the cross-end interaction behavior in at least two second sample devices, and the sample content features are the embedding features of the sample content, and the sample content features correspond to the sample interest degree score.

[0194] The sample sequence feature is the sequence feature of the behavior sequence of the interaction behavior of the sample user account. In this embodiment, the interaction behavior of the sample user account includes single-end interaction behavior in the first sample device and cross-end interaction behavior in at least two second sample devices. The single-end interaction behavior is the interaction behavior of the sample user account in a single first sample terminal. For example, the sample user account logs in to the video client and plays a video in the first sample terminal. The cross-end interaction behavior is the interaction behavior of the sample user account in at least two second sample terminals. For example, the sample user account logs in to the video client in one second sample terminal, plays a video, logs in to the same video client in another second sample terminal, and continues to play the video based on the previous play progress.

[0195] In some embodiments, based on different application scenarios, the specific content of the interaction behavior is different. Taking the single-end interaction behavior of the sample user account in the first sample device as an example, in the video playback scenario, the single-end interaction behavior is the playback behavior, including at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit. In the e-commerce consumption scenario, the single-end interaction behavior is the purchase behavior, including at least one of the following: click, swipe away, favorite, unfavorite, add to cart, delete, purchase, cancel purchase, comment, share, give a gift, receive a gift, exit.

[0196] In some embodiments, the at least two second sample devices may include the first sample device or other sample devices other than the first sample device. The first sample device and the at least two second sample devices may also be respectively referred to as: sample user devices or sample terminals, and the device types include at least one of the following: mobile phone, tablet computer, vehicle-mounted terminal (carputer), wearable device, personal computer, unmanned reservation terminal, smart home appliance, smart voice interaction device, unmanned vending terminal. For the first sample device and the at least two second sample devices, and for the at least two second sample devices, the device types may be the same, partially the same, or different. When the device types of the first sample device and the at least two second sample devices are completely the same, it can be understood as: cross-end. When the types of the first sample device and the at least two second sample devices are partially the same or different, it can be understood as: cross-end, cross-domain.

[0197] In one possible implementation, the single-end interaction behavior of the sample user account in the first sample device and the cross-end interaction behavior in the at least two second sample devices are combined into one behavior sequence. In another possible implementation, the single-end interaction behavior of the sample user account in the first sample device is combined into one behavior sequence, also referred to as the first sequence or the main sequence. The cross-end interaction behavior of the sample user account in the at least two second sample devices is combined into one behavior sequence, also referred to as the second sequence or the auxiliary sequence.

[0198] It should also be noted that considering that a sample user usually logs in with a sample user account, the sample user account involved in the above-mentioned first sample device and the sample user accounts involved in at least two second sample devices are the same sample user account. In some other possible cases, the sample user account involved in the above-mentioned first sample device refers to all the sample user accounts logged in on the first sample device, and the sample user accounts involved in at least two second sample devices are one of all these sample user accounts logged in on the first sample device. For example, the sample user accounts involved in the first sample device are sample user account 1 and sample user account 2, and the sample user accounts involved in at least two second sample devices are sample user account 1.

[0199] The sample content feature is the embedding feature of the sample content. Among them, the sample content is pre-determined, which can be common sample content or recalled sample content. The sample content can be one or multiple. In the case where the sample content is multiple, the method of this embodiment is respectively executed for each sample content to obtain the predicted interest score for each sample content. In the training stage, the sample content feature corresponds to the sample interest score, and the sample interest score is the labeled information.

[0200] In some embodiments, based on different application scenarios, the specific content of the sample content is different. For example, in the video playback scenario, the sample content includes at least one of the following: TV series, variety shows, animations, movies, talk shows, documentaries, short videos. In the e-commerce consumption scenario, the sample content includes at least one of the following: daily necessities, clothing, furniture, shoes and socks, cosmetics.

[0201] Step 440, perform sequence modeling based on the sample sequence feature and the sample content feature to obtain a sample interest vector, and the sample interest vector is related to the sample sequence feature and the sample content feature.

[0202] The sample sequence modeling is to perform modeling and analysis on the behavior sequence of the interaction behavior of the sample user account to capture the interaction habits and preferences of the sample user account. In this embodiment, the sample sequence modeling is performed based on the sample sequence feature and the sample content feature, and can fully capture and integrate the interaction habits and preferences corresponding to the single-end interaction behavior and cross-end interaction behavior of the sample user account.

[0203] The sample interest vector is an interest representation of the sample user account for the sample content corresponding to the sample content feature, and is used to characterize the interest of the sample user account in the sample content corresponding to the sample content feature. The sample interest vector is related to the sample sequence feature and the sample content feature. That is, by fully capturing and integrating the interaction habits and preferences corresponding to the single-end interaction behavior and cross-end interaction behavior of the sample user account, an effective representation of the sample interest vector is achieved.

[0204] Exemplarily, the computer device performs sequence modeling based on the sample sequence feature and the sample content feature through the sequence modeling network of the content recommendation model to obtain the sample interest vector.

[0205] Step 460: Map the sample content feature and the sample interest vector to obtain the predicted interest degree score of the sample content corresponding to the sample content feature.

[0206] Mapping is used to capture the relationship between the sample content feature and the sample interest vector, and the predicted interest degree score is obtained through prediction. The predicted interest degree score is predicted based on the sample content feature and the sample interest vector, and is used to characterize the interest degree of the sample user account in the sample content corresponding to the sample content feature. When the predicted interest degree score is closer to the sample interest degree score, it indicates that the sample interest vector is more accurate and effective. When the model accuracy is high enough, the predicted interest degree score should be very close to or even the same as the sample interest degree score.

[0207] Exemplarily, the computer device maps the sample content feature and the sample interest vector through the mapping network of the content recommendation model to obtain the sample interest degree score of the sample content corresponding to the sample content feature.

[0208] In some embodiments, in addition to the sample sequence feature, other sample features (otherfeatures) are also involved. Other sample features are used to characterize the needs and preferences of the sample user account. Other sample features may be at least one of the following: the portrait feature of the sample user account, the behavioral content statistical feature of the sample user account. Among them, the portrait feature is used to characterize the personal needs and preferences of the sample user account. The behavioral content statistical feature is used to characterize the needs and preferences of the sample user account for content.

[0209] Optionally, obtain other sample features of the sample user account, map the other sample features, the sample content feature and the sample interest vector, and obtain the predicted interest degree score of the sample content corresponding to the sample content feature. In this embodiment, by obtaining other sample features of the sample user account, the needs and preferences of the sample user account can be further characterized from different dimensions and levels, which is beneficial to improving the accuracy of the predicted interest degree score, thereby improving the model accuracy of the content recommendation model.

[0210] Step 480: Optimize the model parameters of the content recommendation model with the objective of reducing the difference between the predicted interest score and the sample interest score.

[0211] Determine a loss function based on the predicted interest score and the sample interest score. Optimize the model parameters of the content recommendation model with the objective of reducing the difference between the predicted interest score and the sample interest score. When the optimization termination condition is reached, obtain the trained content recommendation model. Optionally, the optimization termination condition includes at least one of the following: the loss function reaches the minimum, the set number of iterations is reached, and the preset accuracy is reached.

[0212] This embodiment provides a specific training method for the content recommendation model, so that a trained content recommendation model can be obtained, which can effectively capture the interaction habits and preferences of the sample user account, enabling the trained content recommendation model to effectively represent the interest vector and improve the model accuracy of the content recommendation model.

[0213] In some embodiments, the sample sequence features include: a first sample sequence feature and a second sample sequence feature, and the sample interest vectors include: a first sample interest vector and a second sample interest vector.

[0214] Specifically, step 460 is implemented as step 520 and step 540 (not shown in the figure). Optionally, step 520 and step 540 can be executed sequentially or synchronously, and no limitation is imposed thereon.

[0215] Step 520: Perform first sequence modeling on the first sample sequence feature based on the sample content feature to obtain a first sample interest vector; and step 540: Perform second sequence modeling on the second sample sequence feature based on the sample content feature and the first sample sequence feature to obtain a second sample interest vector; wherein the first sample sequence feature is the sequence feature of the single-end interaction behavior sequence of the sample user account in the first sample device, and the second sample sequence feature is the sequence feature of the cross-end interaction behavior sequence of the sample user account in at least two second sample devices.

[0216] The first sequence modeling is used to perform modeling and analysis on the single-end interaction behavior sequence of the sample user account in the first sample device. The first sample interest vector is the sample interest vector obtained through the first sequence modeling. The first sample interest vector is related to the first sample sequence feature and the sample content feature.

[0217] The second sequence modeling is used to perform modeling and analysis on the behavior sequence of the cross - end interaction behavior of the sample user account in at least two second sample devices. The second sequence modeling in this embodiment also involves the first sample sequence feature, that is, it will fuse the relevant information of the first sample sequence feature to guide the second sequence modeling of the second sample sequence feature through the first sample sequence feature, suppress the problem of noise interference, and improve the accuracy of the second sample interest vector. The second sample interest vector is the sample interest vector obtained through the second sequence modeling. The second sample interest vector is related to the first sample sequence feature, the second sample sequence feature, and the sample content feature.

[0218] Exemplarily, the computer device performs the first sequence modeling on the first sample sequence feature based on the sample content feature through the first sequence modeling network of the content recommendation model to obtain the first sample interest vector. And, the computer device performs the second sequence modeling on the second sample sequence feature based on the sample content feature and the first sample sequence feature through the second sequence modeling network of the content recommendation model to obtain the second sample interest vector.

[0219] This embodiment distinguishes the first sample sequence feature and the second sample sequence feature, and provides a specific way to perform sequence modeling based on the sample sequence feature and the sample content feature. Among them, for the first sequence modeling, it can fully capture the interaction habits and preferences corresponding to the single - end interaction behavior of the sample user account, and improve the accuracy of the first sample interest vector. For the second sequence modeling, it can guide the second sequence modeling of the second sample sequence feature through the first sample sequence feature, suppress the problem of noise interference, and improve the accuracy of the second sample interest vector, thereby accurately representing the user interest and being conducive to improving the model accuracy of the content recommendation model.

[0220] For the second sequence modeling:

[0221] Specifically, step 540 is implemented as step 542, step 544, and step 546 (not shown in the figure):

[0222] Step 542, perform signal strength weighting on the second sample sequence feature to obtain the second sample content feature, where the second sample content feature is the embedded feature of the behavior content corresponding to the cross - end interaction behavior in the second sample sequence feature;

[0223] Step 544, perform dot - product cross - attention processing based on the first sample sequence feature and the second sample content feature to obtain the sample cross - attention feature;

[0224] Step 546, perform target attention processing based on the sample content feature and the cross - attention feature to obtain the second sample interest vector.

[0225] The second sample content feature is obtained by performing signal strength weighting on the second sample sequence feature. The second sample content feature is the embedding feature (item embedding) of the behavior content (item id) corresponding to the cross-terminal interaction behavior in the second sample sequence feature. Among them, the behavior content refers to the specific content of the cross-terminal interaction behavior. For example, in the video playback scenario, the cross-terminal interaction behavior is the playback behavior, and the behavior content includes at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit. Performing signal strength weighting on the second sample sequence feature is implemented through the second numerical signal embedding network of the content recommendation model.

[0226] Dot product cross-attention processing can use the first sample sequence feature to perform interest correlation filtering on the second sample content feature, so as to fuse the relevant information of the first sample sequence feature, and guide the second sequence modeling of the second sample sequence feature through the first sample sequence feature, thereby suppressing noise interference. The sample cross-attention feature is obtained by performing dot product cross-attention processing on the first sample sequence feature and the second sample content feature. Dot product cross-attention processing is implemented through the dot product cross-attention network of the content recommendation model.

[0227] Target attention processing can use the sample content feature to model the sample cross-attention feature and extract the second sample interest vector. The second sample interest vector is obtained by performing target attention processing on the sample content feature and the sample cross-attention feature. Target attention processing is implemented through the second target attention network of the content recommendation model.

[0228] Exemplarily, the computer device, through the second sequence modeling network of the content recommendation model and the second numerical signal embedding network, performs signal strength weighting on the second sample sequence feature to obtain the second sample content feature; through the dot product cross-attention network, performs dot product cross-attention processing based on the first sample sequence feature and the second sample content feature to obtain the sample cross-attention feature; through the second target attention network, performs target attention processing based on the sample content feature and the cross-attention feature to obtain the second sample interest vector.

[0229] This embodiment provides a specific manner of second sequence modeling. The second sequence modeling is double-layer double-query matrix cross-attention modeling. When performing the second sequence modeling, the first sample sequence feature and the sample content feature are introduced, and the first sample sequence feature is fully utilized to perform interest correlation filtering on the second sample sequence feature, which can improve the accuracy of the second sample interest vector and the model accuracy of the content recommendation model.

[0230] In some embodiments, step 544 is specifically implemented as step 5441 and step 5442 (not shown in the figure):

[0231] Step 5441: Perform a pooling operation on the first sample sequence feature to obtain a sample pooled feature;

[0232] Step 5442: Perform dot product cross-attention processing based on the sample pooled feature and the second sample content feature to obtain a sample cross-attention feature.

[0233] The pooling operation is used to reduce the feature size and the amount of computation. The pooling operation can be at least one of the following types: Max Pooling and Mean Pooling. In this embodiment, the pooling operation adopted is Mean Pooling, which is consistent with the inference stage.

[0234] Exemplarily, the computer device performs a pooling operation on the first sample sequence feature through a content recommendation model to obtain a sample pooled feature; through a dot product cross-attention network, dot product cross-attention processing is performed based on the sample pooled feature and the second sample content feature to obtain a sample cross-attention feature.

[0235] This embodiment provides a way to introduce the first sample sequence feature into the second sequence modeling. Among them, the sample pooled feature obtained by performing a pooling operation on the first sample sequence feature is used as the input of the dot product cross-attention processing, which can realize the full utilization of the first sample sequence feature.

[0236] In some embodiments, the first sample Q matrix corresponding to the dot product cross-attention processing is obtained based on the sample pooled feature, and the first sample K matrix and the first sample V matrix are obtained based on the second sample content feature. Step 5442 is specifically implemented as steps 5442-1, 5442-2, 5442-3, 5442-4, 5442-5, 5442-6 (not shown in the figure). Among them, steps 5442-1 and 5442-2 can be executed sequentially or synchronously, and there is no limitation on this.

[0237] Step 5442-1: Multiply the first sample Q matrix by the sample Q weight matrix to obtain a sample Q product result;

[0238] Step 5442-2: Multiply the first sample K matrix by the sample K weight matrix to obtain a sample K product result;

[0239] Step 5442-3: Dot multiply the sample Q product result and the sample K product result to obtain a sample dot product result;

[0240] Step 5442-4: Normalize the sample dot product result to obtain a sample normalization result;

[0241] Step 5442-5: Multiply the first sample V matrix by the sample V weight matrix to obtain the sample V product result.

[0242] Step 5442-6: Dot-multiply the sample V product result by the sample normalization result to obtain the sample cross-attention feature.

[0243] The sample Q weight matrix, the sample K weight matrix, and the sample V weight matrix are learnable. Normalization is achieved through the softmax function. The sample cross-attention feature is obtained by filtering the interest correlation of the second sample content feature using the first sample Q matrix.

[0244] Exemplarily, the computer device multiplies the first sample Q matrix by the sample Q weight matrix through the dot-product cross-attention network of the content recommendation model to obtain the sample Q product result; multiplies the first sample K matrix by the sample K weight matrix to obtain the sample K product result; dot-multiplies the sample Q product result by the sample K product result to obtain the sample dot-product result; normalizes the sample dot-product result to obtain the sample normalization result; multiplies the first sample V matrix by the sample V weight matrix to obtain the sample V product result; dot-multiplies the sample V product result by the sample normalization result to obtain the sample cross-attention feature.

[0245] This embodiment provides a specific method for dot-product cross-attention processing in the second sequence modeling, which can perform interest correlation filtering on the second sample sequence feature using the first sample Q matrix determined based on the first sample sequence feature, realizing the ability of interest correlation filtering of the cross-end interaction behavior in the second sample sequence feature based on the single-end interaction behavior representation in the first sample sequence feature, thereby suppressing noise interference, which is beneficial to further obtaining the second sample interest vector through target attention processing and improving the accuracy of the second sample interest vector.

[0246] In some embodiments, the second sample Q matrix corresponding to the target attention processing is obtained based on the sample content feature, and the second sample K matrix and the second sample V matrix are obtained based on the sample cross-attention feature. Step 546 is specifically implemented as Step 5461 and Step 5462 (not shown in the figure):

[0247] Step 5461: Based on the second sample Q matrix and the second sample K matrix, obtain the sample attention weight matrix.

[0248] Step 5462: Perform weighted summation on the second sample V matrix based on the sample attention weight matrix to obtain the second sample interest vector.

[0249] The target attention processing is used to capture the relationship between each element in the sample content feature and the sample cross-attention feature. For each element in the sample cross-attention feature, the target attention processing calculates the interest correlation with the sample content feature, and each element respectively obtains a weight value related to the sample content feature. By performing weighted summation on each element based on the weight value, the second sample interest vector can be obtained.

[0250] Exemplarily, the computer device obtains a sample attention weight matrix based on the second sample Q matrix and the second sample K matrix through the second target attention network; performs weighted summation on the second sample V matrix based on the sample attention weight matrix to obtain the second sample interest vector.

[0251] This embodiment provides a specific method for target attention processing in the second sequence modeling, which can obtain the second sample interest vector and improve the accuracy of the second sample interest vector.

[0252] In some embodiments, step 542 is specifically implemented as step 5421, step 5422, step 5423, and step 5424 (not shown in the figure):

[0253] Step 5421, determine the second numerical signal corresponding to the second sample sequence feature, where the second numerical signal is used to characterize the intensity of the behavior content corresponding to the cross-terminal interaction behavior in the second sample sequence feature;

[0254] Step 5422, perform non-linear transformation and stacking on the second numerical signal in sequence to obtain the second signal feature;

[0255] Step 5423, map the second signal feature to obtain the second signal mapping feature;

[0256] Step 5424, perform dot product on the second sample sequence feature and the second signal mapping feature to obtain the second sample content feature.

[0257] The second numerical signal is the numerical signal corresponding to the second sample sequence feature. The second numerical signal is used to characterize the intensity of the behavior content (item id) corresponding to the cross-terminal interaction behavior in the second sample sequence feature. For example, in the field of video playback, the numerical signal includes at least one of the following: playback duration, playback times, number of viewers, average playback duration per person, number of member views, 10% / 30% / 50% / 90% playback completion number. Mapping the second signal feature can be implemented through a fully connected layer.

[0258] Exemplarily, the computer device embeds the second numerical signal into the network, performs non-linear transformation and stacking on the second numerical signal in sequence to obtain the second signal feature; maps the second signal feature through the fully-connected layer in the second numerical signal embedding network to obtain the second signal mapping feature; multiplies the second sample sequence feature and the second signal mapping feature to obtain the second sample content feature.

[0259] This embodiment provides a method for signal strength weighting of the behavior content itself in the second sample sequence feature. Since the second numerical signal involved in the second sample sequence feature has important value for the behavior content itself and can represent the consumption intensity of the behavior content itself, signal strength weighting is performed at the shallowest layer of the second sequence modeling, which can highlight the behavior content with stronger intensity, is beneficial to capturing user interests, and improves the accuracy of the second sample interest vector.

[0260] For the first sequence modeling:

[0261] Specifically, step 520 is implemented as step 522 and step 524 (not shown in the figure):

[0262] In step 522, signal strength weighting is performed on the first sample sequence feature to obtain the first sample content feature, and the first sample content feature is the embedding feature of the behavior content corresponding to the single-ended interaction behavior in the first sample sequence feature;

[0263] In step 524, target attention processing is performed based on the sample content feature and the first sample content feature to obtain the first sample interest vector.

[0264] The first sample content feature is obtained by performing signal strength weighting on the first sample sequence feature. The first sample content feature is the embedding feature (item embedding) of the behavior content (item id) corresponding to the single-ended interaction behavior in the first sample sequence feature. Among them, the behavior content refers to the specific content of the single-ended interaction behavior. For example, in the video playback scenario, the single-ended interaction behavior is the playback behavior, and the behavior content includes at least one of the following: click, play, swipe away, like, comment, share, reward, repeat play, exit. Performing signal strength weighting on the first sample sequence feature is implemented through the first numerical signal embedding network of the content recommendation model.

[0265] Target attention processing can model the first sample content feature using the sample content feature and extract the first sample interest vector. The first sample interest vector is obtained by performing target attention processing on the sample content feature and the first sample content feature. Target attention processing is implemented through the first target attention network of the content recommendation model.

[0266] Exemplarily, the computer device performs signal strength weighting on the first sample sequence feature through the first sequence modeling network of the content recommendation model and the first numerical signal embedding network to obtain the first sample content feature; and performs target attention processing on the sample content feature and the first sample content feature through the first target attention network to obtain the first sample interest vector.

[0267] This embodiment provides a specific manner of the first sequence modeling, which can make full use of the first sample sequence feature and the sample content feature to improve the accuracy of the first sample interest vector.

[0268] In some embodiments, the sample Q matrix corresponding to the target attention processing is obtained based on the sample content feature, and the sample K matrix and the sample V matrix are obtained based on the first sample content feature. Step 524 is specifically implemented as Step 5241 and Step 5242 (not shown in the figure):

[0269] Step 5241, obtaining a sample attention weight matrix based on the sample Q matrix and the sample K matrix;

[0270] Step 5242, performing weighted summation on the sample V matrix based on the sample attention weight matrix to obtain the first sample interest vector.

[0271] The target attention processing is used to capture the relationship between each element in the sample content feature and the first sample content feature. For each element in the first sample content feature, the interest correlation with the sample content feature is calculated, and each element respectively obtains a weight value related to the sample content feature. Based on the weight value, weighted summation is performed on each element to obtain the first sample interest vector.

[0272] Exemplarily, the computer device obtains a sample attention weight matrix based on the sample Q matrix and the sample K matrix through the first target attention network; and performs weighted summation on the sample V matrix based on the sample attention weight matrix to obtain the first sample interest vector.

[0273] This embodiment provides a specific manner of the target attention processing in the first sequence modeling, which can obtain the first sample interest vector and improve the accuracy of the first sample interest vector.

[0274] In some embodiments, Step 522 is specifically implemented as Step 5221, Step 5222, Step 5223, and Step 5224 (not shown in the figure):

[0275] Step 5221, determining a first numerical signal corresponding to the first sample sequence feature, where the first numerical signal is used to characterize the intensity of the behavior content corresponding to the single-ended interaction behavior in the first sample sequence feature;

[0276] Step 5222: Perform non-linear transformation and stacking on the first numerical signal in sequence to obtain the first signal feature;

[0277] Step 5223: Map the first signal feature to obtain the first signal mapping feature;

[0278] Step 5224: Perform dot product on the first sample sequence feature and the first signal mapping feature to obtain the first sample content feature.

[0279] The first numerical signal is the numerical signal corresponding to the first sample sequence feature. The first numerical signal is used to characterize the intensity of the behavior content (item id) corresponding to the single-ended interaction behavior in the first sample sequence feature. For example, in the field of video playback, the numerical signal includes at least one of the following: playback duration, number of plays, number of viewers, average playback duration per person, number of member views, 10% / 30% / 50% / 90% playback completion number. Mapping the second signal feature can be achieved through a fully connected layer.

[0280] Exemplarily, the computer device passes the first numerical signal through the first numerical signal embedding network, performs non-linear transformation and stacking (stack) on the first numerical signal in sequence to obtain the first signal feature; maps the first signal feature through the fully connected layer in the first numerical signal embedding network to obtain the first signal mapping feature; performs dot product on the first sample sequence feature and the first signal mapping feature to obtain the first sample content feature.

[0281] This embodiment provides a method for weighting the signal intensity of the behavior content itself in the first sample sequence feature. Since the first numerical signal involved in the first sample sequence feature has important value for the behavior content itself and can characterize the consumption intensity of the behavior content itself, performing signal intensity weighting at the shallowest layer of the first sequence modeling can highlight the behavior content with stronger intensity, which is beneficial to capturing user interests and improving the accuracy of the first sample interest vector.

[0282] The following is a general description of the content recommendation method provided in the embodiments of the present application with reference to the schematic diagram.

[0283] · Application scenario

[0284] (1) The application scenario is a video playback scenario and can be applied to the selected page of the video client. Among them, the video client can be at least one of the following: mobile terminal, PC terminal, TV terminal. When the user account logs in to the video client, refer to Figure 11As shown, it automatically jumps to the featured page of the video client. Generally, the featured page includes: video content tags such as "TV dramas, movies, variety shows, animations, children's programs, documentaries", video promotional images, and video content that "you are following". The recall and ranking models related to the recommendation system are important components of the recommendation algorithm for the video featured page. By using the content recommendation model and content recommendation method of this embodiment, target content can be more accurately recommended to user accounts. Refer to Figure 12 As shown, the target content is displayed in the recommendation area 132 of the featured page. The target content can be talk shows, movies, TV dramas, documentaries, etc.

[0285] (2) The application scenario is an e-commerce consumption scenario and can be applied to the home page of the e-commerce client. Among them, the e-commerce client can be at least one of the following: mobile terminal, PC terminal, TV terminal. When the user account logs in to the e-commerce client, it automatically jumps to the home page of the e-commerce client. Generally, the home page includes: function controls, e-commerce content tags such as "new arrivals, outfits, food", e-commerce promotional images, and recommended target content. The target content can be clothing, cosmetics, daily necessities, etc. By using the content recommendation model and content recommendation method of this embodiment, target content can be more accurately recommended to user accounts.

[0286] (3) The application scenario is a social network scenario and can be applied to the home page, applet page or social dynamic page of the social network client. Among them, the social network client can be at least one of the following: mobile terminal, PC terminal, TV terminal. When the user account logs in to the social network client, it automatically jumps to the home page of the social network client, and the applet page or social dynamic page can be displayed after interaction operations. Recommended target content can be displayed on the home page, applet page or social dynamic page. The target content can be advertisements, applets, friend accounts, social dynamics, promotional posters, etc. By using the content recommendation model and content recommendation method of this embodiment, target content can be more accurately recommended to user accounts.

[0287] · Beneficial effects

[0288] Taking the video playback scenario as an example, the content recommendation method has achieved significant benefits in the core indicators of the experiment on the featured page of the video client. The detailed content is as follows:

[0289] (1) Experiment background and objectives: The basic fine-ranking model is the traditional Deep Interest Network (DIN), or a fine-ranking model that performs sequence modeling on auxiliary sequence features based on Target Attention. The experimental model is the content recommendation model proposed in this embodiment.

[0290] (2) Experiment configuration: The experiment is carried out in the fine-ranking layer of the strong purpose area recommendation traffic on the video PC side.

[0291] (3)Expected project benefits: Improve core indicators such as the play duration in the strong target area.

[0292]

[0293] (4)Experimental analysis:

[0294] Indicators in the strong target area:

[0295] Per capita positive video vv on the selected page in the strong target area: +0.78% (significant)

[0296] Per capita positive video play duration on the selected page in the strong target area: +0.79% (significant)

[0297] Per capita vv of materials on the selected page in the strong target area: -0.12% (not significant, p-value = 0.78)

[0298] Per capita play duration of materials on the selected page in the strong target area: +0.06% (not significant, p-value = 0.91)

[0299] Indicators by module:

[0300] Focus map:

[0301] Per capita positive video vv of the focus map on the main player on the selected page: +0.26% (not significant, p-value)

[0302] Per capita positive video play duration (seconds) of the focus map on the main player on the selected page: +0.56% (not significant, p-value = 0.41)

[0303] Chasing drama indicators:

[0304] Per capita attributed positive video play vv of the heavyweight dramas being chased on the selected page: +0.98% (significant)

[0305] Per capita attributed positive video play duration of the heavyweight dramas being chased on the selected page: +0.65% (not significant, p-value = 0.16)

[0306] Heavyweight new and popular indicators:

[0307] Per capita attributed positive video play vv of the heavyweight new and popular on the selected page: +1.40% (significant)

[0308] Per capita attributed positive video play duration of the heavyweight new and popular on the selected page: +2.11% (significant)

[0309] Indicators on the selected page:

[0310] Per capita play duration (seconds) of the main player on the selected page: +0.78% (significant)

[0311] Per capita vv of the main player on the selected page: +0.95% (significant)

[0312] Main Player_Play UTR_Selected Page: +0.21% (not significant, p-value = 0.09)

[0313] Global Metrics:

[0314] Main Player_Average Play Duration per Person (seconds)_Global: +0.39% (not significant, p-value = 0.14)

[0315] Main Player_Average VV per Person_Global: +0.47% (not significant, p-value = 0.24)

[0316] Module-Specific Metrics:

[0317] Focus Map:

[0318] Focus Map_Main Player_Average Positive-Film VV per Person_Selected Page: +0.26% (not significant, p-value = 0.68)

[0319] Focus Map_Main Player_Average Positive-Film Play Duration per Person (seconds)_Selected Page: +0.56% (not significant, p-value

[0320] = 0.41)

[0321] TV Drama Following Metrics:

[0322] Average Attributed Positive-Film Play VV per Person_Selected Page_Blockbuster_In Progress: +0.98% (significant, p-value = 0.08)

[0323] Average Attributed Positive-Film Play Duration per Person_Selected Page_Blockbuster_In Progress: +0.65% (not significant, p-value = 0.16)

[0324] Blockbuster New and Popular Metrics:

[0325] Average Attributed Positive-Film Play VV per Person_Selected Page_Blockbuster_New and Popular: +1.40% (significant)

[0326] Average Attributed Positive-Film Play Duration per Person_Selected Page_Blockbuster_New and Popular: +2.11% (significant)

[0327] Selected Page Metrics:

[0328] Main Player_Average Play Duration per Person (seconds)_Selected Page: +0.78% (significant)

[0329] Main Player_Average VV per Person_Selected Page: +0.95% (significant, p-value = 0.06)

[0330] Main Player_Play UTR_Selected Page: +0.21% (not significant, p-value = 0.09)

[0331] Global Metrics:

[0332] Main Player_Average Play Duration per User (seconds)_Global: +0.39% (not significant, p-value = 0.14)

[0333] Main Player_Average VV per User_Global: +0.47% (not significant, p-value = 0.24)

[0334] Commercialization and Payment Metrics:

[0335] Total Average Revenue per Member (guid): -3.92% (not significant, p-value = 0.80)

[0336] Member Payment Conversion Rate (guid): -0.47% (not significant, p-value = 0.65)

[0337] Total Average Advertising Revenue per User_Global: -1.76% (not significant, p-value = 0.87)

[0338] For the specific data and schematic diagrams of each of the above metrics, reference can also be made to Figure 13 、 Figure 14 as shown. In Figure 13 and Figure 14 , the first column is the metric name, the second column is the data of the control group, the third column is the data of the experimental group, the fourth column is the difference ratio of the data, and the fifth and sixth columns are the bar chart and line chart corresponding to the data. In summary, it can be seen that the content recommendation model and content recommendation method provided in this embodiment can effectively represent user interests, improve the accuracy of the interest vector, and thus improve the accuracy of content recommendation.

[0339] Figure 15 is a block diagram of a content recommendation device provided by an exemplary embodiment of the present application. The content recommendation device 800 includes at least some of the following modules: an acquisition module 820, a modeling module 840, a mapping module 860, and a recommendation module 880.

[0340] The acquisition module 820 is configured to acquire sequence features and candidate content features. The sequence features are the sequence features of the single-end interaction behavior of the user account on the first device and the cross-end interaction behavior on at least two second devices, and the candidate content features are the embedding features of the candidate content.

[0341] The modeling module 840 is configured to perform sequence modeling based on the sequence features and the candidate content features to obtain an interest vector, where the interest vector is related to the sequence features and the candidate content features.

[0342] The mapping module 860 is configured to map the candidate content features and the interest vector to obtain the interest degree score of the candidate content corresponding to the candidate content features.

[0343] A recommendation module 880 for recommending target content from the candidate content based on the interest score of the candidate content.

[0344] In some embodiments, the sequence features include: a first sequence feature and a second sequence feature, and the interest vectors include: a first interest vector and a second interest vector; the modeling module 840 is configured to:

[0345] Perform a first sequence modeling on the first sequence feature based on the candidate content feature to obtain the first interest vector; and perform a second sequence modeling on the second sequence feature based on the candidate content feature and the first sequence feature to obtain the second interest vector;

[0346] Wherein, the first sequence feature is the sequence feature of the behavior sequence of the single - end interaction behavior of the user account in the first device, and the second sequence feature is the sequence feature of the behavior sequence of the cross - end interaction behavior of the user account in the at least two second devices.

[0347] In some embodiments, the modeling module 840 is configured to:

[0348] Perform signal strength weighting on the second sequence feature to obtain a second content feature, where the second content feature is the embedding feature of the behavior content corresponding to the cross - end interaction behavior in the second sequence feature;

[0349] Perform dot - product cross - attention processing based on the first sequence feature and the second content feature to obtain a cross - attention feature;

[0350] Perform target attention processing based on the candidate content feature and the cross - attention feature to obtain the second interest vector.

[0351] In some embodiments, the modeling module 840 is configured to:

[0352] Perform a pooling operation on the first sequence feature to obtain a pooling feature;

[0353] Perform the dot - product cross - attention processing based on the pooling feature and the second content feature to obtain the cross - attention feature.

[0354] In some embodiments, the first Q matrix corresponding to the dot - product cross - attention processing is obtained based on the pooling feature, and the first K matrix and the first V matrix are obtained based on the second content feature; the modeling module 840 is configured to:

[0355] Multiply the first Q matrix by a Q weight matrix to obtain a Q product result;

[0356] Multiply the first K matrix by the K weight matrix to obtain a K product result;

[0357] Dot multiply the Q product result and the K product result to obtain a dot product result;

[0358] Normalize the dot product result to obtain a normalized result;

[0359] Multiply the first V matrix by the V weight matrix to obtain a V product result;

[0360] Dot multiply the V product result and the normalized result to obtain the cross-attention feature.

[0361] In some embodiments, the second Q matrix corresponding to the target attention processing is obtained based on the candidate content feature, and the second K matrix and the second V matrix are obtained based on the cross-attention feature; the modeling module 840 is configured to:

[0362] Based on the second Q matrix and the second K matrix, obtain an attention weight matrix;

[0363] Perform weighted summation on the second V matrix based on the attention weight matrix to obtain the second interest vector.

[0364] In some embodiments, the modeling module 840 is configured to:

[0365] Determine a second numerical signal corresponding to the second sequence feature, where the second numerical signal is used to characterize the intensity of the behavior content corresponding to the cross-end interaction behavior in the second sequence feature;

[0366] Perform non-linear transformation and stacking on the second numerical signal in sequence to obtain a second signal feature;

[0367] Map the second signal feature to obtain a second signal mapping feature;

[0368] Dot multiply the second sequence feature and the second signal mapping feature to obtain the second content feature.

[0369] In some embodiments, the modeling module 840 is configured to:

[0370] Perform signal strength weighting on the first sequence feature to obtain a first content feature, where the first content feature is an embedded feature of the behavior content corresponding to the single-end interaction behavior in the first sequence feature;

[0371] Perform target attention processing based on the candidate content feature and the first content feature to obtain the first interest vector.

[0372] In some embodiments, the Q matrix corresponding to the target attention processing is obtained based on the candidate content features, and the K matrix and the V matrix are obtained based on the first content features; the modeling module 840 is configured to:

[0373] Based on the Q matrix and the K matrix, obtain an attention weight matrix;

[0374] Perform weighted summation on the V matrix based on the attention weight matrix to obtain the first interest vector.

[0375] In some embodiments, the modeling module 840 is configured to:

[0376] Determine a first numerical signal corresponding to the first sequence feature, where the first numerical signal is used to characterize the intensity of the behavior content corresponding to the single-end interaction behavior in the first sequence feature;

[0377] Perform non-linear transformation and stacking on the first numerical signal in sequence to obtain a first signal feature;

[0378] Map the first signal feature to obtain a first signal mapping feature;

[0379] Perform dot product on the first sequence feature and the first signal mapping feature to obtain the first content feature.

[0380] In some embodiments, the content recommendation model for performing the content recommendation method includes: a sequence modeling network and a mapping network connected in series in sequence;

[0381] Wherein, the sequence modeling network is configured to perform sequence modeling based on the sequence feature and the candidate content feature to obtain an interest vector; the mapping network is configured to map the candidate content feature and the interest vector to obtain the interest degree score of the candidate content corresponding to the candidate content feature.

[0382] In some embodiments, the sequence modeling network includes: a first sequence modeling network and a second sequence modeling network connected in parallel;

[0383] Wherein, the first sequence modeling network is configured to perform first sequence modeling on the first sequence feature based on the candidate content feature to obtain the first interest vector; the second sequence modeling network is configured to perform second sequence modeling on the second sequence feature based on the candidate content feature and the first sequence feature to obtain the second interest vector.

[0384] In some embodiments, the second sequence modeling network includes: a second numerical signal embedding network, a dot product cross-attention network, and a second target attention network connected in series in sequence;

[0385] Among them, the second numerical signal embedding network is used to perform signal strength weighting on the second sequence feature to obtain a second content feature; the dot product cross-attention network is used to perform dot product cross-attention processing on the first sequence feature and the second content feature to obtain a cross-attention feature; the second target attention network is used to perform target attention processing on the candidate content feature and the cross-attention feature to obtain the second interest vector.

[0386] In some embodiments, the first sequence modeling network includes: a first numerical signal embedding network and a first target attention network cascaded in sequence;

[0387] Among them, the first numerical signal embedding network is used to perform signal strength weighting on the first sequence feature to obtain a first content feature; the first target attention network is used to perform target attention processing on the candidate content feature and the first content feature to obtain the first interest vector.

[0388] In some embodiments, the acquisition module 820 is used for:

[0389] Acquire a sample sequence feature and a sample content feature, where the sample sequence feature is the sequence feature of the single-end interaction behavior of the sample user account in the first sample device and the cross-end interaction behavior in at least two second sample devices, and the sample content feature is the embedding feature of the sample content, and the sample content feature corresponds to a sample interest score;

[0390] In some embodiments, the modeling module 840 is used for:

[0391] Perform sequence modeling based on the sample sequence feature and the sample content feature to obtain a sample interest vector, where the sample interest vector is related to the sample sequence feature and the sample content feature;

[0392] In some embodiments, the mapping module 860 is used for:

[0393] Map the sample content feature and the sample interest vector to obtain the predicted interest score of the sample content corresponding to the sample content feature;

[0394] In some embodiments, it further includes a training module; the training module is used for:

[0395] Taking reducing the difference between the predicted interest score and the sample interest score as the training objective, optimize the model parameters of the content recommendation model.

[0396] It should be noted that the specific limitations in the above-described embodiments of the one or more content recommendation apparatuses 800 can be referred to the limitations on the content recommendation method in the foregoing text, and will not be elaborated herein. Each module of the above apparatus can be implemented in whole or in part by software, hardware, and their combination. Each module can be embedded in the processor of the computer device in the form of hardware or be independent of it, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

[0397] An embodiment of the present application also provides a computer device, which includes: a processor and a memory, and a computer program is stored in the memory; the processor is configured to execute the computer program in the memory to implement the content recommendation method provided by each of the above method embodiments.

[0398] Exemplarily, Figure 16 is a structural block diagram of a computer device provided by an exemplary embodiment of the present application. Optionally, taking the computer device as a server as an example, the computer device is the server 1000.

[0399] Generally, the server 1000 includes: a processor 1001 and a memory 1002.

[0400] The processor 1001 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 1001 can be implemented in at least one of the hardware forms of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 1001 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the wake state, also known as the Central Processing Unit (CPU); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 1001 may be integrated with a Graphics Processing Unit (GPU), and the GPU is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 1001 may further include an Artificial Intelligence (AI) processor, and the AI processor is used to process computational operations related to machine learning.

[0401] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory, as well as non-volatile memory, such as one or more disk storage devices and flash storage devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1002 is used to store at least one instruction, and the at least one instruction is used to be executed by the processor 1001 to implement the content recommendation method provided by the above method embodiments.

[0402] In some embodiments, the server 1000 may further optionally include: an input interface 1003 and an output interface 1004. The processor 1001, the memory 1002, the input interface 1003, and the output interface 1004 may be connected through a bus or signal lines. Each peripheral device may be connected to the input interface 1003 and the output interface 1004 through a bus, signal lines, or a circuit board. The input interface 1003 and the output interface 1004 may be used to connect at least one peripheral device related to input / output (I / O) to the processor 1001 and the memory 1002. In some embodiments, the processor 1001, the memory 1002, the input interface 1003, and the output interface 1004 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1001, the memory 1002, the input interface 1003, and the output interface 1004 may be implemented on a separate chip or circuit board, and the embodiments of the present application do not limit this.

[0403] Those skilled in the art can understand that Figure 16 the structure shown in

[0404] does not constitute a limitation on the computer device, and may include more or fewer components than shown in the figure, or combine certain components, or adopt different component arrangements.

[0405] The embodiments of the present application further provide a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and the computer program is loaded and executed by a processor to implement the content recommendation method provided by the above method embodiments.

[0406] The embodiments of the present application also provide a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the processor of the computer device loads and executes to implement the content recommendation method provided in each of the above method embodiments.

[0407] The serial numbers of the embodiments of the present application above are only for description and do not represent the superiority or inferiority of the embodiments.

[0408] Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware or can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium, and the computer-readable storage medium mentioned above can be a read-only memory, a disk, an optical disc, etc.

[0409] Those skilled in the art should be able to realize that in one or more of the above examples, the functions described in the embodiments of the present application can be implemented by hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. The computer-readable medium includes a computer storage medium and a communication medium, where the communication medium includes any medium that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0410] The above are only optional embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A content recommendation method, characterized in that: The method comprises: Acquire sequence features, and acquire candidate content features, wherein the sequence features are sequence features of a behavior sequence of a single-end interaction behavior of a user account in a first device and a cross-end interaction behavior in at least two second devices, and the candidate content features are embedded features of the candidate content; Perform sequence modeling based on the sequence feature and the candidate content feature to obtain an interest vector, where the interest vector is related to the sequence feature and the candidate content feature; Mapping the candidate content feature and the interest vector to obtain an interest score of the candidate content corresponding to the candidate content feature; Based on the interest scores of the candidate contents, target contents are recommended from the candidate contents.

2. The method according to claim 1, characterized in that The sequence features include: a first sequence feature and a second sequence feature, and the interest vector includes: a first interest vector and a second interest vector; The performing sequence modeling based on the sequence feature and the candidate content feature to obtain an interest vector includes: Based on the candidate content features, performing first sequence modeling on the first sequence features to obtain the first interest vector; and based on the candidate content features and the first sequence features, performing second sequence modeling on the second sequence features to obtain the second interest vector; Among them, the first sequence feature is a sequence feature of a behavior sequence of a single-end interactive behavior of the user account in the first device, and the second sequence feature is a sequence feature of a behavior sequence of a cross-end interactive behavior of the user account in the at least two second devices.

3. The method according to claim 2, characterized in that The performing second sequence modeling on the second sequence feature based on the candidate content feature and the first sequence feature to obtain the second interest vector includes: Performing signal strength weighting on the second sequence feature to obtain a second content feature, where the second content feature is an embedded feature of the behavior content corresponding to the cross-end interaction behavior in the second sequence feature; Performing a dot-wise cross-attention process based on the first sequence feature and the second content feature to obtain a cross-attention feature; Target attention processing is performed based on the candidate content features and the cross-attention features to obtain the second interest vector.

4. The method according to claim 3, characterized in that The performing point-wise cross-attention processing based on the first sequence feature and the second content feature to obtain a cross-attention feature includes: Performing a pooling operation on the first sequence feature to obtain a pooled feature; The dot-wise cross-attention processing is performed based on the pooled feature and the second content feature to obtain the cross-attention feature.

5. The method according to claim 4, characterized in that The first Q matrix corresponding to the point-by-point cross attention processing is obtained based on the pooling feature, and the first K matrix and the first V matrix are obtained based on the second content feature; The performing the dot-wise cross-attention processing based on the pooling feature and the second content feature to obtain the cross-attention feature includes: Multiplying the first Q matrix by a Q weight matrix to obtain a Q product result; and multiplying the first K matrix by a K weight matrix to obtain a K product result; dot-multiplying the Q product result and the K product result to obtain a dot-multiplication result; and normalizing the dot-multiplication result to obtain a normalized result; Multiply the first V matrix by the V weight matrix to obtain a V product result; Dot product the V product result and the normalized result to obtain the cross-attention feature.

6. The method according to claim 3, characterized in that The second Q matrix corresponding to the target attention processing is obtained based on the candidate content features, and the second K matrix and the second V matrix are obtained based on the cross attention features; The performing target attention processing based on the candidate content feature and the cross-attention feature to obtain the second interest vector includes: Based on the second Q matrix and the second K matrix, obtaining an attention weight matrix; A weighted sum is performed on the second V matrix based on the attention weight matrix to obtain the second interest vector.

7. The method according to claim 3, characterized in that The performing signal strength weighting on the second sequence feature to obtain a second content feature includes: determining a second numerical signal corresponding to the second sequence feature, where the second numerical signal is used to characterize the intensity of the behavior content corresponding to the cross-end interaction behavior in the second sequence feature; Perform nonlinear transformation and stacking on the second numerical signal in sequence to obtain a second signal feature; Mapping the second signal feature to obtain a second signal mapping feature; The second content feature is obtained by performing a dot product of the second sequence feature and the second signal mapping feature.

8. The method according to claim 2, characterized in that: The performing first sequence modeling on the first sequence feature based on the candidate content feature to obtain the first interest vector includes: Performing signal strength weighting on the first sequence feature to obtain a first content feature, where the first content feature is an embedded feature of a behavior content corresponding to a single-end interactive behavior in the first sequence feature; Target attention processing is performed based on the candidate content features and the first content features to obtain the first interest vector.

9. The method according to claim 8, characterized in that The Q matrix corresponding to the target attention processing is obtained based on the candidate content features, and the K matrix and the V matrix are obtained based on the first content features; The performing target attention processing based on the candidate content feature and the first content feature to obtain the first interest vector includes: Based on the Q matrix and the K matrix, an attention weight matrix is ​​obtained; A weighted sum is performed on the V matrix based on the attention weight matrix to obtain the first interest vector.

10. The method according to claim 8, characterized in that The performing signal strength weighting on the first sequence feature to obtain a first content feature includes: determining a first numerical signal corresponding to the first sequence feature, where the first numerical signal is used to characterize the intensity of the behavior content corresponding to the single-end interaction behavior in the first sequence feature; Perform nonlinear transformation and stacking on the first numerical signals in sequence to obtain a first signal feature; Mapping the first signal feature to obtain a first signal mapping feature; The first content feature is obtained by performing a dot product of the first sequence feature and the first signal mapping feature.

11. The method according to any one of claims 1 to 10, characterized in that: The content recommendation model for executing the content recommendation method comprises: a sequence modeling network and a mapping network cascaded in sequence; The sequence modeling network is used to perform sequence modeling based on the sequence features and the candidate content features to obtain an interest vector; the mapping network is used to map the candidate content features and the interest vector to obtain an interest score of the candidate content corresponding to the candidate content features.

12. The method according to claim 11, characterized in that The sequence modeling network includes: a first sequence modeling network and a second sequence modeling network connected in parallel; The first sequence modeling network is used to perform first sequence modeling on the first sequence features based on the candidate content features to obtain the first interest vector; the second sequence modeling network is used to perform second sequence modeling on the second sequence features based on the candidate content features and the first sequence features to obtain the second interest vector.

13. The method according to claim 12, characterized in that The second sequence modeling network includes: a second numerical signal embedding network, a point-wise cross attention network, and a second target attention network, which are cascaded in sequence; Among them, the second numerical signal embedding network is used to perform signal strength weighting on the second sequence feature to obtain the second content feature; the point multiplication cross attention network is used to perform point multiplication cross attention processing on the first sequence feature and the second content feature to obtain the cross attention feature; the second target attention network is used to perform target attention processing on the candidate content feature and the cross attention feature to obtain the second interest vector.

14. The method according to claim 12, characterized in that The first sequence modeling network includes: a first numerical signal embedding network and a first target attention network cascaded in sequence; Among them, the first numerical signal embedding network is used to perform signal strength weighting on the first sequence feature to obtain a first content feature; the first target attention network is used to perform target attention processing on the candidate content feature and the first content feature to obtain the first interest vector.

15. The method according to any one of claims 12 to 14, characterized in that: The method further comprises: Acquire a sample sequence feature, and acquire a sample content feature, wherein the sample sequence feature is a sequence feature of a behavior sequence of a single-end interaction behavior of a sample user account in a first sample device and a cross-end interaction behavior in at least two second sample devices, and the sample content feature is an embedded feature of the sample content, and the sample content feature corresponds to a sample interest score; Perform sequence modeling based on the sample sequence feature and the sample content feature to obtain a sample interest vector, where the sample interest vector is related to the sample sequence feature and the sample content feature; Mapping the sample content feature and the sample interest vector to obtain a predicted interest score of the sample content corresponding to the sample content feature; The model parameters of the content recommendation model are optimized by taking reducing the difference between the predicted interest score and the sample interest score as a training goal.

16. A content recommendation device, characterized in that: The device comprises: an acquisition module, configured to acquire sequence features and candidate content features, wherein the sequence features are sequence features of a behavior sequence of a single-end interaction behavior of a user account in a first device and a cross-end interaction behavior in at least two second devices, and the candidate content features are embedded features of the candidate content; A modeling module, configured to perform sequence modeling based on the sequence feature and the candidate content feature to obtain an interest vector, wherein the interest vector is related to the sequence feature and the candidate content feature; A mapping module, used to map the candidate content features and the interest vector to obtain an interest score of the candidate content corresponding to the candidate content features; A recommendation module is used to recommend target content from the candidate content based on the interest scores of the candidate content.

17. A computer device, characterized in that: The computer device comprises: a processor and a memory, wherein the memory stores a computer program, and the computer program is loaded and executed by the processor to implement the content recommendation method according to any one of claims 1 to 15.

18. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is loaded and executed by a processor to implement the content recommendation method according to any one of claims 1 to 15.

19. A computer program product, characterized in that The computer program product includes computer instructions, which are stored in a computer-readable storage medium. The processor obtains the computer instructions from the computer-readable storage medium, so that the processor loads and executes them to implement the content recommendation method as described in any one of claims 1 to 15.

Images