Content recommendation method, apparatus, server, and storage medium

Abstract

The application discloses a content recommendation method and device, a server and a storage medium, and relates to the field of artificial intelligence. The method comprises the following steps: acquiring a cross-domain content interaction sequence, wherein the cross-domain content interaction sequence comprises historical interaction behaviors corresponding to different domain contents, the domain contents comprise first domain contents belonging to a first domain and second domain contents belonging to at least one second domain; performing feature extraction on the cross-domain content interaction sequence based on the correlation between the historical interaction behaviors, to obtain a cross-domain content interaction feature sequence; performing cross-domain interest extraction on the cross-domain content interaction feature sequence, to obtain a first domain interest vector corresponding to the first domain and a second domain interest vector corresponding to the second domain; and determining target recommendation contents from candidate recommendation contents based on the first domain interest vector and the second domain interest vector, wherein the candidate recommendation contents belong to the first domain. The scheme improves the accuracy and diversity of the content recommendation result.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a content recommendation method, apparatus, server, and storage medium. Background Technology

[0002] With the development of mobile internet technology and the explosion of information, recommendation systems have emerged. Recommendation systems use big data information to predict users' preferences for specific types of information resources and filter information accordingly.

[0003] In related technologies, interest recommendation algorithms are mostly limited to extracting user behavior data in specific domains. When the behavior data is limited, the user cold start problem is likely to occur, making it impossible to accurately model user interests. Correspondingly, as the behavior data increases, modeling user interests solely based on specific domains can lead to a narrowing of user interests, causing them to fall into information cocoons, which in turn affects the accuracy and diversity of recommendation results. Summary of the Invention

[0004] This application provides a content recommendation method, apparatus, server, and storage medium, which can improve the accuracy and diversity of content recommendation results. The technical solution is as follows:

[0005] On the one hand, embodiments of this application provide a content recommendation method, the method comprising:

[0006] Obtain a cross-domain content interaction sequence, wherein the cross-domain content interaction sequence contains historical interaction behaviors corresponding to content in different domains, and the historical interaction behaviors are ordered according to the execution order, wherein the domain content includes first domain content belonging to a first domain and second domain content belonging to at least one second domain;

[0007] Based on the correlation between the historical interaction behaviors, feature extraction is performed on the cross-domain content interaction sequence to obtain a cross-domain content interaction feature sequence, which includes the cross-domain content interaction features corresponding to the historical interaction behaviors.

[0008] Cross-domain interest extraction is performed on the cross-domain content interaction feature sequence to obtain the first domain interest vector corresponding to the first domain and the second domain interest vector corresponding to the second domain.

[0009] Based on the first domain interest vector and the second domain interest vector, the target recommended content is determined from the candidate recommended content, wherein the candidate recommended content belongs to the first domain.

[0010] On the other hand, embodiments of this application provide a content recommendation device, the device comprising:

[0011] The acquisition module is used to acquire cross-domain content interaction sequences, which include historical interaction behaviors corresponding to different domain content, and the historical interaction behaviors are ordered according to the execution order. The domain content includes first domain content belonging to a first domain and second domain content belonging to at least one second domain.

[0012] The feature extraction module is used to extract features from the cross-domain content interaction sequence based on the correlation between the historical interaction behaviors, and obtain a cross-domain content interaction feature sequence, wherein the cross-domain content interaction feature sequence contains the cross-domain content interaction features corresponding to the historical interaction behaviors.

[0013] The interest extraction module is used to extract cross-domain interests from the cross-domain content interaction feature sequence to obtain a first domain interest vector corresponding to the first domain and a second domain interest vector corresponding to the second domain.

[0014] The determination module is used to determine the target recommended content from the candidate recommended content based on the first domain interest vector and the second domain interest vector, wherein the candidate recommended content belongs to the first domain.

[0015] On the other hand, embodiments of this application provide a server, a server processor and a memory, wherein the memory stores at least one program, which is loaded and executed by the processor to implement the content recommendation method as described above.

[0016] On the other hand, embodiments of this application provide a computer-readable storage medium storing at least one program that is loaded and executed by a processor to implement the content recommendation method as described above.

[0017] On the other hand, embodiments of this application provide a computer program product including computer instructions stored in a computer-readable storage medium. A server's processor reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the server to perform the content recommendation method described above.

[0018] In this embodiment, the server determines cross-domain content interaction features representing user preferences based on cross-domain content interaction sequences, thereby determining the user's interest vectors in different domains, and combining multi-domain interest vectors to determine target recommended content that matches the user's interests. This application enriches behavioral data and improves the diversity of user interests by introducing historical interaction behaviors corresponding to content in different domains, realizing the transfer and application of user interests between different domains. Furthermore, when behavioral data in the domain corresponding to the content recommendation task is limited, this embodiment can effectively alleviate the user cold start problem and overcome the limitations of deep learning for content recommendation caused by limited user data in a single domain. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic diagram of an implementation environment provided by an exemplary embodiment of this application is shown;

[0021] Figure 2 A flowchart illustrating a content recommendation method provided in an exemplary embodiment of this application is shown;

[0022] Figure 3 A schematic diagram illustrating historical interaction behavior provided by an exemplary embodiment of this application is shown;

[0023] Figure 4 This illustration shows a schematic diagram of a cross-domain content interaction sequence provided in an exemplary embodiment of this application;

[0024] Figure 5 A flowchart of a feature extraction method provided in an exemplary embodiment of this application is shown;

[0025] Figure 6 A schematic diagram of a word vector embedding network provided in an exemplary embodiment of this application is shown;

[0026] Figure 7 A schematic diagram of a feature extraction network provided in an exemplary embodiment of this application is shown;

[0027] Figure 8 A schematic diagram of an attention mechanism provided in an exemplary embodiment of this application is shown;

[0028] Figure 9 A flowchart illustrating a pre-training process provided in an exemplary embodiment of this application is shown;

[0029] Figure 10 A schematic diagram of a pre-training process provided in an exemplary embodiment of this application is shown;

[0030] Figure 11 This illustration shows a schematic diagram of portrait feature extraction provided by an exemplary embodiment of this application;

[0031] Figure 12 A schematic diagram of an attention mechanism provided by another exemplary embodiment of this application is shown;

[0032] Figure 13 A schematic diagram illustrating a hard search provided in an exemplary embodiment of this application is shown;

[0033] Figure 14 A schematic diagram illustrating a soft search provided by an exemplary embodiment of this application is shown;

[0034] Figure 15 A schematic diagram illustrating a content recommendation process provided in an exemplary embodiment of this application is shown;

[0035] Figure 16 A flowchart illustrating task-specific training provided in an exemplary embodiment of this application is shown;

[0036] Figure 17 This illustration shows a schematic diagram of a specific task training provided by an exemplary embodiment of this application;

[0037] Figure 18 A schematic diagram of a method for obtaining training loss for a specific task provided in an exemplary embodiment of this application is shown;

[0038] Figure 19 This invention provides a structural block diagram of a content recommendation apparatus according to an exemplary embodiment of the present application.

[0039] Figure 20 A schematic diagram of the structure of a server provided in an exemplary embodiment of this application is shown. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0041] For ease of understanding, the terms used in the embodiments of this application will be explained below.

[0042] A behavior sequence refers to a collection of behavioral objects (items) that the user interacts with the system. Each behavioral object includes its own ID and its side information. In this embodiment, for an advertising scenario, the attributes of the behavioral object can be advertiser ID, industry ID, category ID, etc.

[0043] Approximate Nearest Neighbors (ANN) retrieval leverages the characteristic that data clusters together as the volume of data increases. It uses data clustering analysis to classify or encode data in a database. During the retrieval process, for target data, its data category is predicted based on its characteristics, and then some or all of the data in that category are returned as the retrieval results. Approximate nearest neighbor retrieval is not limited to returning the most likely data; rather, it improves retrieval efficiency by retrieving approximate data within data clusters, sacrificing acceptable precision. Depending on the retrieval method, ANN retrieval can be implemented using methods such as hashing and Product Quantization (PQ).

[0044] Recommendation algorithms are now widely used in various fields. Compared with traditional recommendation algorithms, which can only represent user interests through a single user vector, multi-interest recommendation algorithms can represent the diversity of user interests by modeling multiple user interests, thereby improving the diversity of recommended content.

[0045] In related technologies, multi-interest recommendation algorithms enrich the diversity of user interests by acquiring their interests at different time stages. However, the data sources are often limited to a single domain, that is, the domain corresponding to the recommended content. In specific application scenarios, this leads to the user cold start problem. When behavioral data in the domain corresponding to the recommended content is scarce, multi-time-period interest recommendation algorithms cannot effectively improve interest diversity, thus failing to accurately model user interests. Furthermore, being limited to information from a single domain can trap users in information cocoons, narrowing their interests. This application acquires users' historical interaction behavior across domains, incorporating user interests from different domains during content recommendation. This enriches user data, improves the diversity of user interest modeling, and ensures good recommendation results even when user interest data in the domain corresponding to the recommended content is sparse.

[0046] Please refer to Figure 1This illustration shows a schematic diagram of an implementation environment provided by an exemplary embodiment of this application. This embodiment uses the application of a content recommendation model to advertising recommendation as an example. This implementation environment includes a computer device 110 and a server 120. The computer device 110 and the server 120 communicate via a communication network. Optionally, the communication network can be a wired network or a wireless network, and the communication network can be at least one of a local area network (LAN), a metropolitan area network (MAN), and a wide area network (WAN).

[0047] Computer device 110 is an electronic device with content recommendation requirements. The recommended content can be within an application; for example, in an e-commerce application, the recommended content can be product information within the e-commerce application. This embodiment of the application does not limit this.

[0048] Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. In this embodiment, server 120 is a content recommendation server for applications in computer device 110, used to determine recommended content for the applications.

[0049] like Figure 1 As shown, computer device 110 reports the user's historical interaction behavior with content in different domains to server 120. Server 120 collects the historical interaction behavior in server 110 and then constructs a cross-domain content interaction sequence 121. Server 120 is equipped with a pre-trained content recommendation model 122 and recommended content in specific domains. With the cross-domain content interaction sequence 121 as input, server 120 determines the interest vectors of different domains through content recommendation model 122. Based on the interest vectors, server 120 retrieves the target recommended content 123 from the recommended content and further feeds back the target recommended content 123 to server 110.

[0050] For ease of explanation, the following examples illustrate how content recommendation and the training of the content recommendation model are both performed by the server.

[0051] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user data. These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their data is being collected. This ensures that the application only begins the steps for collecting user data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without user confirmation), the steps for collecting user data end, meaning no user data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of related user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0052] Please refer to Figure 2 This document illustrates a flowchart of a content recommendation method provided in an exemplary embodiment of this application. This embodiment uses the method applied to a server as an example, and the method includes the following steps:

[0053] Step 201: Obtain cross-domain content interaction sequence. The cross-domain content interaction sequence contains historical interaction behaviors corresponding to different domain content, and the historical interaction behaviors are ordered according to the execution order. The domain content includes first domain content belonging to the first domain and second domain content belonging to at least one second domain.

[0054] When the server recommends content for a specific area, such as Figure 3 As shown, since the number of historical interaction behaviors of users in a single domain is limited and it is difficult to reflect the diversity of user interests, and based on the fact that users have different behaviors in different domains, and that these behaviors all indicate user interests and preferences, in this embodiment of the application, the server combines historical interaction behavior data from multiple different domains to obtain cross-domain content interaction sequences, thereby enriching user interest data. For example... Figure 4 As shown, the cross-domain content interaction sequence 401 includes first-domain content 402 belonging to the first domain. The first domain is the domain corresponding to the content recommendation task pair, and the historical interaction behaviors corresponding to the first-domain content can include user clicks, browsing, etc., within the first domain. Correspondingly, the second domain refers to other domains besides the one corresponding to the content recommendation task, and the historical interaction behaviors corresponding to the second-domain content 403 can include user clicks, browsing, shopping, and other interaction behaviors performed within the second domain.

[0055] In one possible implementation, when determining the server's input, the cross-domain content interaction sequence can be represented as:

[0056]

[0057] in, Let M represent the content interaction sequence in the domain M, and have b i The i-th behavior object is denoted by T, which is the behavior length, i.e., the number of historical interaction behaviors of the user in the domain M.

[0058] To illustrate, when the server recommends ads using a recommendation algorithm, the ad domain is the first domain. Correspondingly, the second domain belongs to other interaction domains besides the ad domain. For example, the second domain may include the shopping domain and the content domain. Combining the user's click behavior, browsing behavior and other interaction behaviors in the ad domain, shopping domain and content domain, the server obtains a cross-domain content interaction sequence containing the above-mentioned historical interaction information, and the historical interaction behaviors in the sequence are ordered according to the execution order.

[0059] Step 202: Based on the correlation between historical interaction behaviors, feature extraction is performed on the cross-domain content interaction sequence to obtain the cross-domain content interaction feature sequence, which contains the cross-domain content interaction features corresponding to the historical interaction behaviors.

[0060] In one possible implementation, a user's historical interaction behavior and the correlation between these behaviors indicate, to some extent, the user's interests and preferences. The server can then obtain the user's interest information based on the cross-domain content interaction sequence. The server extracts features from the cross-domain content interaction sequence using word embedding networks and feature extraction networks, determining the interest features indicated by each historical interaction behavior and the correlations between these features. In other words, through feature extraction, the server can determine the concrete behavioral features corresponding to abstract historical interactions, thereby obtaining a cross-domain content interaction feature sequence corresponding to the cross-domain content interaction sequence, enabling the server to further determine the user's interests and preferences.

[0061] Step 203: Extract cross-domain interests from the cross-domain content interaction feature sequence to obtain the first domain interest vector corresponding to the first domain and the second domain interest vector corresponding to the second domain.

[0062] In cross-domain content interaction feature sequences, the content interaction features corresponding to content from different domains are still arranged according to the execution order of their corresponding historical interaction behaviors; that is, the behavioral features from different domains are arranged in an overlapping manner. Based on the strong correlation between behavioral features within the same domain, combining the behavioral features indicated by historical interaction behaviors within the same domain to perform user interest analysis can ensure the accuracy of user interests corresponding to that domain.

[0063] In one possible implementation, the server determines a first domain interest vector based on content interaction features corresponding to the first domain in a cross-domain content interaction feature sequence. The first domain interest vector indicates the user's interest preferences in the first domain. Correspondingly, the server determines a second domain interest vector based on content interaction features of the second domain, and the second domain interest vector indicates the user's interest preferences in the second domain.

[0064] Step 204: Based on the first domain interest vector and the second domain interest vector, determine the target recommended content from the candidate recommended content. The candidate recommended content belongs to the first domain.

[0065] The first domain interest vector and the second domain interest vector indicate the user's interest preferences in the first domain and the second domain, respectively, and thus indicate the user's probability of preferring the recommended content. Therefore, the server can determine the target recommended content that matches the user's interest preferences based on the first domain interest vector and the second domain interest vector.

[0066] Because, based on the determined first interest vector, the server in this embodiment combines a second domain interest vector to determine the target recommended content, even when historical interaction behavior in the first domain is limited, i.e., when the accuracy of the first domain interest vector is limited, the server can still ensure that the target recommended content has a good effect, which can alleviate the user cold start problem. Similarly, the server determines the target recommended content based on multi-domain interest vectors, which improves the diversity of user interests and avoids the information cocoon problem caused by being confined to a single domain, thus narrowing user interests.

[0067] In summary, in this embodiment, the server determines cross-domain content interaction features representing user preferences based on cross-domain content interaction sequences, thereby determining the user's interest vectors in different domains, and comprehensively determining target recommended content that matches the user's interests by integrating multi-domain interest vectors. This application enriches behavioral data and improves the diversity of user interests by introducing historical interaction behaviors corresponding to content in different domains, realizing the transfer and application of user interests between different domains. Furthermore, when behavioral data in the corresponding domain of the content recommendation task is limited, this embodiment can effectively alleviate the user cold start problem and overcome the limitations of deep learning for content recommendation caused by the limited user data in a single domain.

[0068] In the above embodiments, feature extraction and cross-domain interest extraction are performed by an interest extraction model. Please refer to [link / reference]. Figure 5 Regarding the way the interest extraction module processes cross-domain content interaction sequences, it may include the following steps:

[0069] Step 501: The server performs word embedding processing on the cross-domain content interaction sequence through a word embedding network to obtain a word embedding vector sequence, which contains word embedding vectors corresponding to each historical interaction behavior.

[0070] In one possible implementation, such as Figure 6 As shown, the server takes the cross-domain content interaction sequence 610 as input and performs word embedding processing on the content identifier or attribute identifier 611 in it through the word embedding network 620, unifying the historical interaction behaviors in multiple domains into a low-dimensional space, and obtaining a word embedding vector sequence 630 composed of word embedding vectors 631. The word embedding vectors 631 represent the features of the corresponding historical interaction behaviors.

[0071] Step 502: The server extracts features from the word embedding vector sequence through a feature extraction network to obtain a cross-domain content interaction feature sequence. The feature extraction network is a neural network used to model sequence relationships.

[0072] Optionally, the feature extraction network can be a Transformer network, or any one or a combination of deep learning networks such as RNN (Recurrent Neural Network), CNN (Convolutional Neural Network), LSTM (Long Short-Term Memory Network), and GRU (Gated Recurrent Unit). This application does not limit this; the embodiments of this application use a Transformer network as an example for illustration.

[0073] In one possible implementation, such as Figure 7 As shown, the server takes the word embedding vector sequence 710 as input and extracts features from it through the feature extraction network 720. Based on the correlation between historical interaction behaviors indicated by the word embedding vectors 711, the feature extraction network 720 learns the cross-relationship between each word embedding vector 711 and determines its weight, and then outputs a weighted cross-domain content interaction feature sequence 730. This weight represents the reference value of the historical interaction behaviors corresponding to the word embedding vectors in content recommendation.

[0074] During feature extraction, the server combines historical interaction behaviors from multiple domains to comprehensively determine the features that users are more interested in, so as to better, more comprehensively, and more accurately identify user interests and preferences and improve the accuracy of recommended content.

[0075] Step 503: The server generates a first domain interest vector based on the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence through an attention mechanism.

[0076] In one possible implementation, such as Figure 8 As shown, the server performs weighted fusion on the cross-domain content interaction features 811 of the first domain in the cross-domain content interaction feature sequence 810 to obtain the first domain interest vector 821. This weighted fusion is implemented using an attention mechanism. The attention mechanism focuses on the content interaction features of a single domain and determines the weight of each content interaction feature in the content recommendation process. Further, based on these weights, it fuses to obtain a content interaction feature in the single domain that can accurately indicate the user's interests and preferences, thus obtaining the first domain interest vector. The server comprehensively considers historical interaction behaviors across multiple domains to determine behavioral correlation and weights. During the interest extraction process, the server only focuses on a single domain, ensuring the accuracy of the interest vector.

[0077] Step 504: The server generates a second-domain interest vector based on the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence through an attention mechanism.

[0078] In one possible implementation, such as Figure 8 As shown, the server performs weighted fusion on the cross-domain content interaction features 812 of the second domain in the cross-domain content interaction feature sequence 810 to obtain the second domain interest vector 822.

[0079] In summary, in this embodiment of the application, the server models cross-domain content interaction sequences through word embedding networks and feature extraction networks. While improving the diversity of user interests based on multi-domain user data, it also enhances the ability to express user interests by utilizing cross-domain content interaction feature sequences.

[0080] In the above embodiments, the server uses a cross-domain content interaction sequence as initial input to determine the target recommendation content. This cross-domain interaction sequence contains historical interaction behaviors from different domains, and these historical interactions exhibit significant heterogeneity and differences. To improve the processing performance of the word embedding network and feature extraction network when performing word embedding or feature extraction on heterogeneous historical interaction behaviors, this embodiment of the application improves the convergence speed of the interest extraction model by pre-training the word embedding network and feature extraction network. Specifically, the pre-training method for the word embedding network and feature extraction network in this embodiment of the application may include the following steps:

[0081] 1. Based on the cross-domain content interaction sequence of the samples, the word embedding network is pre-trained using the Word2vec algorithm.

[0082] Because cross-domain content interaction sequences contain historical interaction behaviors from multiple different domains, these historical interactions exhibit significant heterogeneity and differences across different domains. This application's embodiments pre-train a word embedding network, specifically using the Word2vec algorithm to predict word embedding vectors belonging to different domains, thereby unifying word embedding vectors from multiple heterogeneous vector spaces into a unified multi-domain space.

[0083] Here, the cross-domain content interaction sequence of the samples refers to the input sequence used for pre-training, and the Word2vec algorithm is the algorithm used to train the word embedding network. Optionally, the Word2vec algorithm can be skip-gram or CBOW. The following explanation uses the skip-gram algorithm on the server as an example.

[0084] In one possible implementation, such as Figure 6 As shown, the server obtains a word embedding vector sequence 630 corresponding to the cross-domain content interaction sequence 610 through the word embedding network 620. Based on the target word embedding vector 632 in the word embedding vector sequence, the server predicts the two word embedding vectors 633 to the left and right of the target word embedding vector with a window size W=2, that is, it calculates the probability of each of the four word embedding vectors 633. Through the above prediction task, the server pre-trains the word embedding network to perform clustering-like processing on similar domain content, thereby improving the overall convergence speed.

[0085] 2. With the word embedding network pre-trained, the word embedding network and feature extraction network are pre-trained based on the cross-domain content interaction sequence of the samples.

[0086] To further mitigate the impact of cross-domain content interaction sequences on the convergence speed and recommendation effectiveness in the content recommendation process, the server further pre-trains the word embedding network and feature extraction network through a language masking task.

[0087] In one possible implementation, such as Figure 9 As shown, this step may include the following sub-steps:

[0088] Step 901: Mask the sample cross-domain content interaction sequence to obtain a sample mask sequence. Masking is used to cover the sample interaction behavior at the mask position in the sample cross-domain content interaction sequence.

[0089] Optionally, the cross-domain content interaction sequence of the sample includes the content identifier of the domain content and the attribute identifier of the content attribute of the domain content.

[0090] In one possible implementation, such as Figure 10 As shown, when masking the sample cross-domain content interaction sequence 1001, the server selects a certain proportion of content identifiers or attribute identifiers from all sample cross-domain content interaction sequences 1001 to participate in the gradient update during the masking process. For example, the server can determine 20% of the content identifiers or attribute identifiers in the cross-domain content interaction sequence as mask positions 1002. Specifically, when masking the content identifiers or attribute identifiers participating in the gradient update, the server first determines a certain proportion of the first mask position, the second mask position, and the third mask position. For example, 80% of the mask positions are determined as the first mask position, 10% as the second mask position, and 10% as the third mask position. Corresponding to the content identifiers or attribute identifiers within different mask positions, the server's masking methods can include the following:

[0091] (1) The server replaces the content identifier and attribute identifier at the first mask position with the mask identifier.

[0092] The server completes the language masking task by predicting the mask identifier at the first mask position, so that the word embedding network and the feature extraction network can improve the convergence speed while ensuring the effectiveness of the cross-domain content interaction feature sequence when extracting features from heterogeneous cross-domain content interaction sequences.

[0093] (2) The server replaces the first content identifier at the second mask position with the second content identifier, or replaces the first attribute identifier at the second mask position with the second attribute identifier.

[0094] The server replaces the first content identifier or the first attribute identifier at the second mask location with the second content identifier or the second attribute identifier.

[0095] (3) Maintain the content identifier and attribute identifier at the third mask position.

[0096] The server keeps the content identifier or attribute identifier at the third mask position unchanged so that it can be used as the prediction target for the pre-training task, thereby improving the convergence speed.

[0097] It should be noted that, compared to related technologies that replace all content identifiers or attribute identifiers at all mask positions with mask identifiers, in this embodiment, the server determines a certain proportion of first, second, and third mask positions within the mask positions and applies different processing to these three mask positions. This forces the server to focus on all mask positions during training, improving convergence speed and avoiding coupling issues between the word embedding network and the feature extraction network when the content of the pre-training task and the online / test service task differs. Furthermore, this prediction task can train the word embedding network to improve its fine-grained alignment performance in the semantic space when obtaining word embedding vectors from different domains.

[0098] Step 902: The word embedding network obtained through pre-training performs word embedding processing on the sample mask sequence to obtain the sample word embedding vector sequence.

[0099] The word embedding network is pre-trained using the Word2vec algorithm, and the sample mask sequence is a cross-domain content interaction sequence that has been masked.

[0100] Step 903: Extract features from the sample word embedding vector sequence using a feature extraction network to obtain the sample cross-domain content interaction feature sequence.

[0101] This step is the same as step 402, and will not be repeated here.

[0102] Step 904: Predict the mask position based on the cross-domain content interaction feature sequence of the sample to obtain the predicted interaction behavior.

[0103] like Figure 10 As shown, in the sample cross-domain content interaction feature sequence 1004 obtained by the feature extraction network 1003, the first mask position is the mask identifier, the second mask position is the identifier feature corresponding to the second content identifier or the second attribute identifier, and the third mask position is the feature identifier corresponding to the content identifier or attribute identifier at that position. Since the content identifier or attribute identifier corresponding to the third mask position remains unchanged, the third mask position can be used as a pre-training target to participate in gradient update.

[0104] In one possible implementation, such as Figure 10As shown, based on the language masking task, the server predicts all mask positions 1002 and determines the predicted interaction behavior 1005. Given a word set size M corresponding to each mask position, the feature extraction network outputs M features. The server takes these M features as input and fully connects them through a fully connected layer of dimension M. Furthermore, it outputs the probability distribution of the identifier corresponding to each mask position using the Softmax function. Based on this probability distribution, the server determines the predicted interaction behavior, i.e., the index with the highest probability in the probability distribution, using the Argmax function.

[0105] Step 905: Using sample interaction behavior as supervision for predicting interaction behavior, train the word embedding network and the feature extraction network.

[0106] Here, sample interaction behavior is identified by vector or attribute identifiers in one-hot form. In this embodiment, the server calculates the loss value of the prediction process using a cross-entropy function, that is, using sample interaction behavior as supervision for predicting interaction behavior. The process of calculating the loss value using the cross-entropy function can be expressed by the following formula:

[0107]

[0108] Where M is the size of the word set corresponding to the mask position, y ic The indicator determines whether the sample interaction behavior i and the predicted interaction behavior c match. When the predicted interaction behavior is the same as the sample interaction behavior, i.e., the prediction result is accurate, y... ic Select 1 otherwise select 0, p ic The predicted value indicates the probability that the sample interaction behavior matches the predicted interaction behavior.

[0109] Optionally, the sample interaction behavior includes the content identifier of the domain content and the attribute identifier of the content attributes of the domain content.

[0110] It should be noted that the parameters of the word embedding network and feature extraction module obtained during the training phase can be applied to the specific task training phase in the following embodiments as their initialization parameters.

[0111] In determining user preferences, similar to historical interaction behavior, users' static attributes indicate their interests to some extent. Therefore, when the server generates first-domain and second-domain interest vectors through the attention mechanism, it can further introduce profile features that represent users' static attributes. This allows for the evaluation of the credibility of the obtained interest vectors based on the profile features, thereby enriching the dimensions of user data and improving the accuracy of interest extraction results.

[0112] In one possible implementation, such as Figure 11As shown, the server takes user static attribute 1110 as input, where user static attribute 1110 includes attribute information 1111. The word embedding network 1120 obtains a word embedding vector sequence corresponding to the attribute information 1111. The word embedding vector sequence indicates the user's static features. Then, the word embedding vectors are concatenated by the concatenation layer 1130, and the profile features 1150 are determined by the fully connected layer 1140. The profile features 1150 indicate the user's static features and, to a certain extent, indicate the possibility of the user's interests and preferences.

[0113] Optionally, in the server, user static attributes can be in the form of a static attribute sequence P. u Input in the form of P u The data may include, but is not limited to, static attributes such as the user's age, gender, and region.

[0114] Furthermore, having obtained the profile features, the server generates a first-domain interest vector based on the profile features and the first-domain cross-domain content interaction features in the cross-domain content interaction feature sequence through an attention mechanism.

[0115] In one possible implementation, such as Figure 12 As shown, through the attention mechanism 1220, the server performs weighted fusion of the cross-domain content interaction features 1211 of the first domain in the cross-domain content interaction feature sequence 1210 to obtain the first fused feature 1231. Further, based on the profile features representing user static features, the server performs weighted fusion of the first fused feature 1231 and the profile feature 1233 through the attention mechanism 1220 to obtain the first domain interest vector 1241. Guided by the profile feature 1233, the server evaluates the credibility of the first fused feature 1231 in the content recommendation process through the attention mechanism 1220, thereby obtaining the weighted first domain interest vector 1241.

[0116] Correspondingly, the server generates a second-domain interest vector based on the profile features and the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence through an attention mechanism.

[0117] In one possible implementation, such as Figure 12 As shown, through the attention mechanism 1220, the server performs weighted fusion on the cross-domain content interaction features 1212 of the second domain in the cross-domain content interaction feature sequence 1210 to obtain the second fused feature 1232. Further, the server performs weighted fusion on the second fused feature 1232 and the profile feature 1233 through the attention mechanism 1220 to obtain the second domain interest vector 1242.

[0118] Based on cross-domain content interaction sequences, and through the word embedding network, feature extraction network, and attention mechanism described in the above embodiments, the server can obtain a first-domain interest vector and a second-domain interest vector that represent the user's interest preferences in different domains. During the online / test service phase, the server can determine recommended content that matches the user's interest preferences based on the first and second interest vectors. Optionally, the server can implement recall, coarse ranking, and fine ranking stages in the recommendation system through the embodiments of this application. The server can determine the recommended content in any of the following ways:

[0119] 1. Hard search.

[0120] In one possible implementation, such as Figure 13 As shown, the server determines recommended content for domain A from the candidate recommended content based on the domain A interest vector. Correspondingly, the server determines recommended content for domain B from the candidate recommended content based on the domain B interest vector, and determines recommended content for domain C based on the domain C interest vector. Domains A, B, and C include the first domain. Given the recommended content for domains A, B, and C, the server selects the k highest-scoring recommended content from these multiple domains as the target recommended content.

[0121] Optionally, based on the first domain interest vector, the server can use an ANN retrieval to determine the top k highest-scoring candidate recommendations as the first recommended content. Similarly, the server can use an ANN retrieval to determine the top k highest-scoring candidate recommendations based on the second domain interest vector as the second recommended content. Further, based on the fusion queue composed of the first and second recommended content, the server determines the top k highest-scoring content as the target recommended content by sorting within the fusion queue. In this process, the server performs content retrieval based on the interest vector corresponding to each domain. By determining and fusing the first and second recommended content, the target recommended content fully expresses the user's interest preferences in both the first and second domains, ensuring that the content recommendation results reflect the diversity of user interests.

[0122] To illustrate, in the case where the first domain is the advertising domain and the second domain is the content domain, the server determines the interest vector for the first domain as "hotel" and the interest vector for the second domain as "travel guide". Then, based on the interest vector for the first domain as "hotel", the server determines the top k recommended content with the highest scores in the candidate recommended content corresponding to the advertising domain. Correspondingly, based on the interest vector for the second domain, the server determines the second recommended content as travel package ads and hotel ads.

[0123] 2. Soft search.

[0124] In one possible implementation, such as Figure 14 As shown, the server uses an attention mechanism to weightedly fuse the interest vectors from domain A, domain B, and domain C to obtain a fused interest vector. Further, based on this fused interest vector, the server identifies the k highest-scoring recommendations from the candidate recommendations as the target recommendations.

[0125] Through an attention mechanism, the server first determines the weights of the first-domain interest vector and the second-domain interest vector, that is, the credibility of the interest vector in representing the user's interest preferences. Then, it performs a weighted fusion calculation on the various interest vectors to determine the fused interest vector. The fused interest vector can more accurately represent the user's interest preferences in the first domain.

[0126] Optionally, based on the fused interest vectors, the server can use an ANN to retrieve the k highest-scoring content from the candidate recommendations as the target recommendations.

[0127] Please refer to Figure 15 In an illustrative example, taking domain A as the game domain, domain B as the video domain, and domain C as the content domain, with domain A being the first domain, the content recommendation process in this application is explained.

[0128] First, the server obtains user usage information uploaded by the computer device, namely historical interaction behavior. This historical interaction behavior can include user behavior in the gaming domain, viewing behavior in the video domain (e.g., watching basketball games), and browsing behavior in the content domain (e.g., browsing automotive-related public accounts). These behaviors constitute a cross-domain content interaction sequence.

[0129] Furthermore, when cross-domain content interaction sequences are used as input, the content recommendation model pre-trains word embedding networks and feature extraction networks to unify multi-domain historical interaction behaviors in heterogeneous spaces into at least a few unified vector spaces based on their similarity, thereby reducing the impact of the heterogeneity of historical interaction behaviors on model convergence and recommendation performance.

[0130] The pre-trained word embedding network performs word embedding processing on the domain content corresponding to historical interaction behaviors to obtain content interaction features. These content interaction features are used to indicate behavioral features. The domain content includes content identifiers or attribute identifiers. For example, in the gaming domain, the word embedding network performs word embedding processing based on the game identifier (content identifier) ​​and attribute identifiers such as game duration, resulting in the word embedding vector "game". Similarly, in the video domain, it performs word embedding processing based on the video identifier and attribute identifiers such as number of views, resulting in the word embedding vector "basketball". For browsing behaviors in the content domain, it performs word embedding processing on the car identifier and attribute identifiers such as cost-effectiveness, resulting in the word embedding vector "car". Furthermore, the word embedding network obtains word embedding vectors corresponding to behavioral features, and the word embedding vectors from each domain constitute a word embedding vector sequence.

[0131] Given the available information, to enhance the model's expressive power and thus improve its recommendation performance, the server establishes connections between various content interaction features through a feature extraction network and assigns weights to them. Based on the weighted cross-domain content interaction feature sequences, the server uses an attention mechanism to determine interest vector A in the game domain, interest vector B in the video domain, and interest vector C in the content domain, respectively.

[0132] Furthermore, the server uses an ANN search to identify recommended content (A, B, and C) from the candidate recommendations that match interest vectors A, B, and C, respectively. The recommended content for each domain is the k highest-scoring recommendations from the candidate recommendations corresponding to their respective domain's interest vector score. By identifying recommendations that align with user interests in different domains, the server effectively represents the diversity of user interests. Subsequently, the server merges and sorts the recommendations from each domain, and uses an ANN search to determine the top k highest-scoring candidate recommendations as the target recommendations.

[0133] To improve the accuracy of recommended content, that is, to increase the alignment between the target recommended content and user interests and preferences, the server can set specific training tasks to train the prediction process in the embodiments of this application. Please refer to... Figure 16 This document illustrates a flowchart of a training task provided by an exemplary embodiment of this application. This embodiment uses the method applied to a server as an example for illustration. The method includes the following steps:

[0134] Step 1601: Obtain the cross-domain content interaction sequence of the samples. The cross-domain content interaction sequence of the samples contains sample interaction behaviors corresponding to different sample contents, and the sample interaction behaviors are ordered according to the execution order. The sample contents include first-domain sample contents belonging to the first domain and second-domain sample contents belonging to the second domain.

[0135] Here, the cross-domain content interaction sequence of the samples refers to the content interaction sequence used as input for this specific training task. This step is the same as step 201, and will not be repeated here.

[0136] Step 1602: Preprocess the cross-domain content interaction sequence of the sample. The preprocessing is used to cover the last sample content belonging to the first domain in the cross-domain content interaction sequence of the sample.

[0137] In this embodiment, the server uses the prediction task of the next first domain content as its training task. Then, based on the sample cross-domain content interaction sequence ordered by execution sequence, the server uses the last sample content of the first domain in the sample cross-domain content interaction sequence as the target recommended content, and uses the sample cross-domain content interaction sequence other than the target recommended content as input for content recommendation, i.e., target recommended content prediction. Therefore, the server needs to mask the target recommended content through preprocessing.

[0138] Step 1603: Input the preprocessed sample cross-domain content interaction sequence into the interest extraction model to obtain the sample first domain interest vector and sample second domain interest vector output by the interest extraction model.

[0139] This step is the same as steps 202 and 203, and will not be repeated here.

[0140] Step 1604: Based on the first domain interest vector and the second domain interest vector of the sample, determine the predicted recommendation content from the sample candidate recommendation content, wherein the sample candidate recommendation content belongs to the first domain.

[0141] In one possible implementation, such as Figure 17 As shown, during the training process for a specific task, the server fuses the first domain interest vector 1721 and the second domain interest vector 1722 of the sample through a vector fusion network 1710 to determine a fused interest vector 1730, and then determines the predicted recommendation content based on the fused interest vector 1730. Optionally, the vector fusion network can be a pooling layer.

[0142] Step 1605: Train the interest extraction model by using the masked sample content as supervision for predicting recommended content.

[0143] In one possible implementation, the server determines the Sampled Softmax Loss through in-batch negative sampling. For example... Figure 18As shown, during the in-batch negative sampling process, the computer determines the other N sample candidate recommendation contents 1821 in the same batch of data 1810 as negative samples of this sample. Therefore, a sample content sequence 1830 includes one sample content 1822 as a positive sample and N sample candidate recommendation contents 1821 as negative samples, that is, the sample content sequence 1830 contains 1+N recommendation contents.

[0144] Furthermore, the server performs a dot product between the predicted recommended content and the sample content sequence, and calculates the sigmoid function to obtain 1+N estimated scores p. i The predicted score range is 0 ≤ p. i <1. Using the estimated score as the predicted value, the server calculates the loss value using the binary cross-entropy function. The calculation process can be expressed by the following formula:

[0145]

[0146] Among them, y i Let y be the identifier for sample i. When sample i is a positive sample, y i =1, and when sample i is a negative sample, y i =0, predicted score p i It represents the probability that sample i obtained from the prediction task is a positive sample.

[0147] Please refer to Figure 19 The diagram illustrates a structural block diagram of a content recommendation apparatus provided in an exemplary embodiment of this application, the apparatus comprising:

[0148] The acquisition module 1901 is used to acquire a cross-domain content interaction sequence, wherein the cross-domain content interaction sequence includes historical interaction behaviors corresponding to different domain content, and the historical interaction behaviors are sorted according to the execution order, wherein the domain content includes first domain content belonging to a first domain and second domain content belonging to at least one second domain.

[0149] Feature extraction module 1902 is used to extract features from the cross-domain content interaction sequence based on the correlation between the historical interaction behaviors to obtain a cross-domain content interaction feature sequence, wherein the cross-domain content interaction feature sequence contains the cross-domain content interaction features corresponding to the historical interaction behaviors.

[0150] The interest extraction module 1903 is used to perform cross-domain interest extraction on the cross-domain content interaction feature sequence to obtain the first domain interest vector corresponding to the first domain and the second domain interest vector corresponding to the second domain.

[0151] The determination module 1904 is used to determine the target recommended content from the candidate recommended content based on the first domain interest vector and the second domain interest vector, wherein the candidate recommended content belongs to the first domain.

[0152] Optionally, the feature extraction module 1902 is further configured to:

[0153] The cross-domain content interaction sequence is processed by word embedding network to obtain word embedding vector sequence, which contains word embedding vectors corresponding to each of the historical interaction behaviors.

[0154] The word embedding vector sequence is subjected to feature extraction by a feature extraction network to obtain the cross-domain content interaction feature sequence. The feature extraction network is a neural network used to model sequence relationships.

[0155] Optionally, the feature extraction module 1902 is further configured to:

[0156] Based on the cross-domain content interaction sequences of the samples, the word embedding network is pre-trained using the word2vec algorithm;

[0157] After the word embedding network has completed pre-training, the word embedding network and the feature extraction network are pre-trained based on the cross-domain content interaction sequence of the samples.

[0158] Optionally, when the word embedding network and the feature extraction network are pre-trained based on the cross-domain content interaction sequence of the samples, the feature extraction module 1902 is further configured to:

[0159] The sample cross-domain content interaction sequence is masked to obtain a sample mask sequence. The masking process is used to cover the sample interaction behavior at the mask position in the sample cross-domain content interaction sequence.

[0160] The word embedding network obtained through pre-training is used to perform word embedding processing on the sample mask sequence to obtain a sample word embedding vector sequence.

[0161] The feature extraction network is used to extract features from the sample word embedding vector sequence to obtain the sample cross-domain content interaction feature sequence.

[0162] Based on the cross-domain content interaction feature sequence of the sample, the mask position is predicted to obtain the predicted interaction behavior;

[0163] The word embedding network and the feature extraction network are trained using the sample interaction behavior as supervision for the predicted interaction behavior.

[0164] Optionally, when the sample interaction behavior includes the content identifier of the domain content and the attribute identifier of the content attribute of the domain content, the feature extraction module 1902 is further configured to:

[0165] Replace the content identifier and the attribute identifier at the first mask position with the mask identifier;

[0166] Replace the content identifier or attribute identifier at the second mask position with the mask identifier;

[0167] Maintain the content identifier and the attribute identifier at the third mask position.

[0168] Optionally, the interest extraction module 1903 is further configured to:

[0169] An interest vector for the first domain is generated based on the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence through an attention mechanism.

[0170] By using an attention mechanism, an interest vector for the second domain is generated based on the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence.

[0171] Optionally, when generating a first-domain interest vector based on the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence through an attention mechanism, the interest extraction module 1903 is further configured to:

[0172] By using an attention mechanism, the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence are weighted and fused to obtain the first domain interest vector;

[0173] The step of generating a second-domain interest vector based on the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence through an attention mechanism includes:

[0174] By using an attention mechanism, the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence are weighted and fused to obtain the interest vector of the second domain.

[0175] Optionally, the interest extraction module 1903 is further configured to:

[0176] By using an attention mechanism, an interest vector for the first domain is generated based on the profile features and the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence.

[0177] By using an attention mechanism, an interest vector for the second domain is generated based on the profile features and the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence.

[0178] Optionally, when generating a first-domain interest vector based on profile features and the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence through an attention mechanism, the interest extraction module 1903 is further configured to:

[0179] By using an attention mechanism, the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence are weighted and fused to obtain the first fused feature;

[0180] The first fusion feature and the portrait feature are weighted and fused using an attention mechanism to obtain the first domain interest vector;

[0181] When generating a second-domain interest vector based on the portrait features and the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence through an attention mechanism, the interest extraction module 1903 is further configured to:

[0182] By using an attention mechanism, the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence are weighted and fused to obtain the second fused feature;

[0183] The second fusion feature and the portrait feature are weighted and fused using an attention mechanism to obtain the second domain interest vector.

[0184] Optionally, the determining module 1904 is further configured to:

[0185] Based on the first domain interest vector, a first recommended content is determined from the candidate recommended content; based on the second domain interest vector, a second recommended content is determined from the candidate recommended content; and based on the first recommended content and the second recommended content, the target recommended content is determined.

[0186] or,

[0187] The first domain interest vector and the second domain interest vector are weighted and fused to obtain a fused interest vector; based on the fused interest vector, the target recommended content is determined from the candidate recommended content.

[0188] Optionally, when feature extraction and cross-domain interest extraction are performed by an interest extraction model, the acquisition module 1901 is further configured to:

[0189] Obtain a cross-domain content interaction sequence of samples, wherein the cross-domain content interaction sequence of samples contains sample interaction behaviors corresponding to different sample contents, and the sample interaction behaviors are ordered in the order of execution, wherein the sample contents include first-domain sample contents belonging to the first domain and second-domain sample contents belonging to the second domain.

[0190] The device further includes a preprocessing module, which is used to preprocess the sample cross-domain content interaction sequence, and the preprocessing is used to cover the last sample content belonging to the first domain in the sample cross-domain content interaction sequence.

[0191] The interest extraction module 1903 is also used for:

[0192] The preprocessed cross-domain content interaction sequence of the sample is input into the interest extraction model to obtain the first domain interest vector and the second domain interest vector of the sample output by the interest extraction model.

[0193] The determining module 1904 is further configured to:

[0194] Based on the first domain interest vector and the second domain interest vector of the sample, the predicted recommendation content is determined from the sample candidate recommendation content, wherein the sample candidate recommendation content belongs to the first domain;

[0195] The device further includes a training module for training the interest extraction model under supervision, using the masked sample content as the supervision of the predicted recommendation content.

[0196] In summary, in this embodiment, the server uses the acquisition module to obtain the domain content interaction sequence, and the feature extraction module determines the cross-domain content interaction features representing user preferences based on the cross-domain content interaction sequence. Then, the interest extraction module determines the user's interest vectors in different domains. Based on the comprehensive multi-domain interest vectors, the determination module obtains target recommended content that matches the user's interests. This application enriches behavioral data and improves the diversity of user interests by introducing historical interaction behaviors corresponding to different domains, realizing the transfer and application of user interests across different domains. Furthermore, when behavioral data in the corresponding domain of the content recommendation task is limited, this embodiment can effectively alleviate the user cold start problem and overcome the limitations of deep learning for content recommendation caused by limited user data in a single domain.

[0197] Please refer to Figure 20This illustration shows a schematic diagram of the structure of a server provided in an exemplary embodiment of this application. Specifically, the server 2000 includes a Central Processing Unit (CPU) 2001, a system memory 2004 including a random access memory 2002 and a read-only memory 2003, and a system bus 2005 connecting the system memory 2004 and the CPU 2001. The server 2000 may also include a basic input / output system (I / O system) 2006 that facilitates the transfer of information between various devices within the computer, and a mass storage device 2007 for storing the operating system 2013, application programs 2014, and other program modules 2015.

[0198] In some embodiments, the basic input / output system 2006 may include a display 2008 for displaying information and an input device 2009 for user input of information, such as a mouse or keyboard. Both the display 2008 and the input device 2009 are connected to the central processing unit 2001 via an input / output controller 2010 connected to the system bus 2005. The basic input / output system 2006 may also include the input / output controller 2010 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 2010 also provides output to a display screen, printer, or other types of output devices.

[0199] The mass storage device 2007 is connected to the central processing unit 2001 via a mass storage controller (not shown) connected to the system bus 2005. The mass storage device 2007 and its associated computer-readable media provide non-volatile storage for the server 2000. That is, the mass storage device 2007 may include computer-readable media (not shown) such as a hard disk or drive.

[0200] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include random access memory (RAM), read-only memory (ROM), flash memory or other solid-state storage technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage media are not limited to the above-mentioned types. The system memory 2004 and mass storage device 2007 described above can be collectively referred to as memory.

[0201] The memory stores one or more programs, which are configured to be executed by one or more central processing units 2001. The one or more programs contain instructions for implementing the methods described above, and the central processing unit 2001 executes the one or more programs to implement the methods provided in the various method embodiments described above.

[0202] According to various embodiments of this application, the server 2000 can also be connected to a remote computer on a network, such as the Internet. That is, the server 2000 can be connected to the network 2012 via the network interface unit 2011 connected to the system bus 2005, or the network interface unit 2011 can be used to connect to other types of networks or remote computer systems (not shown).

[0203] This application also provides a computer-readable storage medium storing at least one program that is executed by a processor to implement the content recommendation method as described in the above embodiments.

[0204] This application also provides a computer program product including computer instructions stored in a computer-readable storage medium. A server's processor reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the server to perform the content recommendation method provided in the above embodiments.

[0205] Those skilled in the art will recognize that the functions described in the embodiments of this application in one or more of the above examples can be implemented using 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 code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of a computer program from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0206] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A content recommendation method, characterized in that, The method includes: Based on cross-domain content interaction sequences of samples, the word embedding network is pre-trained using the word2vec algorithm; After the word embedding network has completed pre-training, the sample cross-domain content interaction sequence is masked to obtain a sample mask sequence. The masking process is used to cover sample interaction behaviors at masked positions in the sample cross-domain content interaction sequence. The masked positions include a first masked position, a second masked position, and a third masked position. The sample interaction behaviors include content identifiers of domain content and attribute identifiers of the content attributes of the domain content. Specifically, the masking process includes replacing the content identifier and the attribute identifier at the first masked position with a mask identifier; replacing the content identifier or the attribute identifier at the second masked position with a mask identifier; and maintaining the content identifier and the attribute identifier at the third masked position. The word embedding network, obtained through pre-training, performs word embedding processing on the sample mask sequence to obtain a sample word embedding vector sequence. A feature extraction network then extracts features from the sample word embedding vector sequence to obtain a sample cross-domain content interaction feature sequence. Based on the sample cross-domain content interaction feature sequence, the mask position is predicted to obtain a predicted interaction behavior. The word embedding network and the feature extraction network are trained using the sample interaction behavior as supervision for the predicted interaction behavior. Obtain a cross-domain content interaction sequence, wherein the cross-domain content interaction sequence contains historical interaction behaviors corresponding to content in different domains, and the historical interaction behaviors are ordered according to the execution order, wherein the domain content includes first domain content belonging to a first domain and second domain content belonging to at least one second domain; The cross-domain content interaction sequence is processed by word embedding network to obtain word embedding vector sequence, which contains word embedding vectors corresponding to each of the historical interaction behaviors; the word embedding vector sequence is processed by feature extraction network to obtain cross-domain content interaction feature sequence, which is a neural network used to model sequence relationships, and the cross-domain content interaction feature sequence contains cross-domain content interaction features corresponding to the historical interaction behaviors. By using an attention mechanism, the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence are weighted and fused to obtain a first domain interest vector; by using an attention mechanism, the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence are weighted and fused to obtain a second domain interest vector. Based on the first domain interest vector and the second domain interest vector, the target recommended content is determined from the candidate recommended content, wherein the candidate recommended content belongs to the first domain.

2. The method according to claim 1, characterized in that, The method further includes: By using an attention mechanism, the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence are weighted and fused to obtain a first fused feature; by using an attention mechanism, the first fused feature and the profile features are weighted and fused to obtain the first domain interest vector. By using an attention mechanism, the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence are weighted and fused to obtain a second fused feature; by using an attention mechanism, the second fused feature and the profile features are weighted and fused to obtain a second domain interest vector.

3. The method according to claim 1, characterized in that, The step of determining target recommended content from candidate recommended content based on the first domain interest vector and the second domain interest vector, wherein the candidate recommended content belongs to the first domain, includes: Based on the first domain interest vector, a first recommended content is determined from the candidate recommended content; based on the second domain interest vector, a second recommended content is determined from the candidate recommended content; and based on the first recommended content and the second recommended content, the target recommended content is determined. or, The first domain interest vector and the second domain interest vector are weighted and fused to obtain a fused interest vector; based on the fused interest vector, the target recommended content is determined from the candidate recommended content.

4. The method according to claim 1, characterized in that, The feature extraction and cross-domain interest extraction are performed by the interest extraction model; The method further includes: Obtain a cross-domain content interaction sequence of samples, wherein the cross-domain content interaction sequence of samples contains sample interaction behaviors corresponding to different sample contents, and the sample interaction behaviors are ordered in the order of execution, wherein the sample contents include first-domain sample contents belonging to the first domain and second-domain sample contents belonging to the second domain. The sample cross-domain content interaction sequence is preprocessed, and the preprocessing is used to cover the last sample content belonging to the first domain in the sample cross-domain content interaction sequence. The preprocessed cross-domain content interaction sequence of the sample is input into the interest extraction model to obtain the first domain interest vector and the second domain interest vector of the sample output by the interest extraction model. Based on the first domain interest vector and the second domain interest vector of the sample, the predicted recommendation content is determined from the sample candidate recommendation content, wherein the sample candidate recommendation content belongs to the first domain; The interest extraction model is trained by using the masked sample content as supervision for the predicted recommendation content.

5. A content recommendation device, characterized in that, The device includes: The feature extraction module is used to pre-train the word embedding network based on the cross-domain content interaction sequence of the samples using the word2vec algorithm; The feature extraction module is further configured to perform masking processing on the sample cross-domain content interaction sequence after the word embedding network has completed pre-training, to obtain a sample mask sequence. The masking processing is used to cover sample interaction behaviors at mask positions in the sample cross-domain content interaction sequence. The mask positions include a first mask position, a second mask position, and a third mask position. The sample interaction behaviors include content identifiers of domain content and attribute identifiers of the content attributes of the domain content. The masking processing includes replacing the content identifier and the attribute identifier at the first mask position with a mask identifier; replacing the content identifier or the attribute identifier at the second mask position with a mask identifier; and maintaining the content identifier and the attribute identifier at the third mask position. The feature extraction module is further configured to: perform word embedding processing on the sample mask sequence using the pre-trained word embedding network to obtain a sample word embedding vector sequence; extract features from the sample word embedding vector sequence using the feature extraction network to obtain a sample cross-domain content interaction feature sequence; predict the mask position based on the sample cross-domain content interaction feature sequence to obtain a predicted interaction behavior; and train the word embedding network and the feature extraction network using the sample interaction behavior as supervision for the predicted interaction behavior. The acquisition module is used to acquire cross-domain content interaction sequences, which include historical interaction behaviors corresponding to different domain content, and the historical interaction behaviors are ordered according to the execution order. The domain content includes first domain content belonging to a first domain and second domain content belonging to at least one second domain. The feature extraction module is further configured to perform word embedding processing on the cross-domain content interaction sequence through a word embedding network to obtain a word embedding vector sequence, wherein the word embedding vector sequence contains word embedding vectors corresponding to each of the historical interaction behaviors; and to perform feature extraction on the word embedding vector sequence through a feature extraction network to obtain a cross-domain content interaction feature sequence, wherein the feature extraction network is a neural network used to model sequence relationships, and the cross-domain content interaction feature sequence contains cross-domain content interaction features corresponding to the historical interaction behaviors; The interest extraction module is used to perform weighted fusion of the cross-domain content interaction features of the first domain in the cross-domain content interaction feature sequence through an attention mechanism to obtain a first domain interest vector; and to perform weighted fusion of the cross-domain content interaction features of the second domain in the cross-domain content interaction feature sequence through an attention mechanism to obtain a second domain interest vector. The determination module is used to determine the target recommended content from the candidate recommended content based on the first domain interest vector and the second domain interest vector, wherein the candidate recommended content belongs to the first domain.

6. A server, characterized in that, The server includes a processor and a memory, the memory storing at least one program that is loaded and executed by the processor to implement the content recommendation method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one program, which is loaded and executed by a processor to implement the content recommendation method as described in any one of claims 1 to 4.

8. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium; the server's processor reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the server to perform the content recommendation method as described in any one of claims 1 to 4.

Images