A content recall method, device, equipment and storage medium
By integrating users' short-term and long-term interest characteristics into the recommendation system and utilizing content viewing records and user profile information, the problem of inaccurate content recall in existing technologies is solved, achieving more efficient content recommendation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TENCENT TECHNOLOGY (SHENZHEN) CO LTD
- Filing Date
- 2021-06-04
- Publication Date
- 2026-05-22
AI Technical Summary
In existing recommendation systems, the accuracy of content retrieval based on users' browsing history is low, and it cannot fully represent users' interests and preferences, resulting in inaccurate recommendations.
By extracting target content sequences from content viewing records, obtaining target attribute information, and combining it with user profile information, feature extraction models such as RNN, LSTM, and Transformer are used to integrate users' short-term and long-term interest features to determine target interest features, thereby recalling content.
It improves the accuracy of content recall and enhances the precision of content recommendation, thus better meeting users' interest needs.
Smart Images
Figure CN115439770B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to a content retrieval method, apparatus, device and storage medium. Background Technology
[0002] With the proliferation of various applications today, to meet the diverse needs of users, they typically offer a vast amount of content (such as videos, images, and articles) for users to choose from. Because applications offer so much content, users need to spend a significant amount of time searching and browsing for content that interests them. To save users' time, some applications use recommendation systems to personalize content recommendations. Specifically, recommendation systems determine content that a user might like based on their browsing history, and then retrieve similar content from the content library. However, users' interests change over time, and their browsing history cannot fully represent their interests; therefore, content retrieval based solely on browsing history has relatively low accuracy. Summary of the Invention
[0003] This application provides a content retrieval method, apparatus, device, and storage medium to improve the accuracy of content retrieval.
[0004] On the one hand, embodiments of this application provide a content recall method, the method comprising:
[0005] Extract the target content sequence from the content viewing history of the target object, and obtain the target attribute information corresponding to each target content in the target content sequence;
[0006] Feature extraction is performed on the target content sequence to obtain the first interest feature of the target object;
[0007] Feature extraction is performed on the target profile information of the target object to obtain the second interest feature of the target object;
[0008] Based on the first interest feature, the second interest feature, and the obtained target attribute information, the target interest feature of the target object is determined.
[0009] Based on the target interest characteristics, at least one target recall content is determined from each content to be recalled.
[0010] On one hand, embodiments of this application provide a content recall device, which includes:
[0011] The acquisition module is used to extract a target content sequence from the content viewing records of the target object, and to obtain the target attribute information corresponding to each target content in the target content sequence;
[0012] The first feature extraction module is used to extract features from the target content sequence to obtain the first interest feature of the target object;
[0013] The second feature extraction module is used to extract features from the target profile information of the target object to obtain the second interest features of the target object;
[0014] The fusion module is used to determine the target interest features of the target object based on the first interest feature, the second interest feature, and the obtained target attribute information;
[0015] The content recall module is used to determine at least one target recall content from each content to be recalled based on the target interest characteristics.
[0016] Optionally, the second feature extraction module is specifically used for:
[0017] Feature extraction is performed on each portrait tag in the target portrait information to obtain the tag features corresponding to each portrait tag;
[0018] Based on the weights of each portrait label and the label features corresponding to each portrait label, the second interest features of the target object are obtained.
[0019] Optionally, the fusion module is specifically used for:
[0020] The first interest feature and the second interest feature are fused to obtain the target fusion feature of the target object;
[0021] Based on the target fusion features and the target attribute information, the third interest feature of the target object is obtained;
[0022] The first interest feature, the second interest feature, and the third interest feature are fused to obtain the target interest feature of the target object.
[0023] Optionally, the fusion module is specifically used for:
[0024] Feature extraction is performed on each of the target attribute information to obtain the candidate attribute features corresponding to each target attribute information;
[0025] Based on the target fusion feature, at least one target attribute feature is obtained from each candidate attribute feature by comparing the similarity between the target fusion feature and the obtained candidate attribute features.
[0026] Based on the at least one target attribute feature and the weight corresponding to each of the at least one target attribute feature, the third interest feature of the target object is obtained.
[0027] Optionally, the fusion module is specifically used for:
[0028] The first interest feature, the second interest feature, the third interest feature, and the contextual features corresponding to the target object are fused to obtain the target interest feature of the target object.
[0029] Optionally, the content retrieval module is specifically used for:
[0030] Each of the contents to be recalled is subjected to feature extraction to obtain the content features corresponding to each of the contents to be recalled.
[0031] Based on the target interest features, and the similarity between them and the obtained content features, at least one target recall content is determined from the various content to be recalled.
[0032] Optionally, it also includes a content recommendation module;
[0033] The content recommendation module is specifically used for:
[0034] After determining at least one target recall content from each content to be recalled based on the target interest features, the at least one target recall content is filtered according to a preset rule to obtain at least one candidate recommendation content.
[0035] Based on the target interest features, the similarity between the target and at least one candidate recommended content is used to rank the at least one candidate recommended content to obtain a recommendation ranking result.
[0036] Based on the recommendation ranking results, at least one candidate recommendation content is recommended to the target object.
[0037] Optionally, the first interest feature is used to characterize the short-term interest features of the target object, and the second interest feature is used to characterize the long-term interest features of the target object.
[0038] On one hand, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described content recall method.
[0039] On one hand, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the above-described content recall method.
[0040] In this embodiment, feature extraction is performed on the target content sequence in the content viewing record to obtain the first interest feature of the target object. Feature extraction is performed on the target profile information of the target object to obtain the second interest feature of the target object. Then, the first interest feature, the second interest feature, and the target attribute information corresponding to each target content in the target content sequence are combined to determine the target interest feature of the target object. This makes the obtained target interest feature of the target object more complete and accurate in representing the target object's interests and hobbies. Therefore, when recalling content based on the target object's target interest feature, the accuracy of content recall is effectively improved, thereby improving the accuracy of content recommendation. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 A schematic diagram of a system architecture provided for an embodiment of this application;
[0043] Figure 2 A flowchart illustrating a content recall method provided in an embodiment of this application;
[0044] Figure 3 This application provides a schematic diagram of the structure of a layer in a Transformer.
[0045] Figure 4 A flowchart illustrating a method for extracting a first interest feature provided in an embodiment of this application;
[0046] Figure 5 A flowchart illustrating a method for extracting a second interest feature provided in an embodiment of this application;
[0047] Figure 6 A flowchart illustrating a method for extracting a second interest feature provided in an embodiment of this application;
[0048] Figure 7 A flowchart illustrating a method for extracting a second interest feature provided in an embodiment of this application;
[0049] Figure 8 A flowchart illustrating a method for obtaining target interest features provided in an embodiment of this application;
[0050] Figure 9 A flowchart illustrating a method for obtaining target interest features provided in an embodiment of this application;
[0051] Figure 10 A flowchart illustrating a method for obtaining target interest features provided in an embodiment of this application;
[0052] Figure 11 A flowchart illustrating a method for obtaining target interest features provided in an embodiment of this application;
[0053] Figure 12 A schematic diagram of a recommendation interface provided in an embodiment of this application;
[0054] Figure 13 A schematic diagram of a network structure for a Transformer-based model provided in an embodiment of this application;
[0055] Figure 14 This is a schematic diagram of the structure of a content recall device provided in an embodiment of this application;
[0056] Figure 15 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0057] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0058] For ease of understanding, the terms used in the embodiments of this invention are explained below.
[0059] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.
[0060] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.
[0061] An item (ITEM) refers to a piece of content such as an article, a video, or an image.
[0062] Attributes (ATTR): These refer to the attributes of a piece of content, such as an article, a video, or an image. They usually include multiple attributes, such as tags and categories.
[0063] Context: Contextual information, that is, the context in which the user is currently located.
[0064] Collaborative filtering (CF) uses the preferences of a group of like-minded people with shared experiences to recommend information that users may be interested in. Individuals respond to information to a certain extent through a collaborative mechanism (such as rating) and record it to achieve the purpose of filtering, thereby helping others to filter information. The responses are not limited to those that are particularly interested; recording information that is particularly uninteresting is also quite important.
[0065] UserCF: Mines user similarity information and recommends items liked by similar users.
[0066] ItemCF: Calculates item similarity by mining item co-occurrence information, and then uses the item similarity.
[0067] Item2Vec: Assigns a dense vector to each item, which, in contrast to the one-hot representation, preserves the semantic dimension information between items.
[0068] RNN (Recurrent Neural Network) is an artificial neural network in which nodes are connected in a directed loop. The internal state of this network can exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can utilize their internal memory to process input sequences of arbitrary temporal order.
[0069] LSTM (Long Short-Term Memory) is a type of temporal recurrent neural network suitable for processing and predicting important events with relatively long intervals and delays in time series.
[0070] GRU (Gated Recurrent Unit) controls the propagation of historical information by introducing a Reset Gate and an Update Gate.
[0071] Transformer: A network structure composed of Multi-Head Attention and Feed Forward Network, which can process time series data in parallel.
[0072] Cosine similarity: Also known as cosine similarity, it is used to evaluate the similarity between two vectors by calculating the cosine of the angle between them.
[0073] The design concept of the embodiments of this application will be introduced below.
[0074] Currently, when recommending content to users in a personalized way, the system often determines the content a user might like based on their browsing history, then retrieves similar content from the content library, and finally makes recommendations based on the retrieved content. However, users' interests change over time, and their browsing history cannot fully represent their interests. Therefore, the accuracy of content retrieval based solely on browsing history is relatively low.
[0075] Analysis revealed that while older browsing history doesn't necessarily reflect a user's current interests, recent viewing history can indicate their short-term interests. User attributes such as age, gender, and city are inherent and relatively stable, thus representing long-term interests. Furthermore, the attributes of recently viewed content also contain rich information that can be used to characterize content-related interests. Combining these dimensions—long-term, short-term, and content-related interests—leads to a better understanding of user preferences and improves the accuracy of content retrieval and recommendation.
[0076] In view of this, embodiments of this application provide a content recall method. In this method, a target content sequence is extracted from the content viewing history of a target object, and target attribute information corresponding to each target content in the target content sequence is obtained. Then, feature extraction is performed on the target content sequence to obtain the first interest feature of the target object, and feature extraction is performed on the target profile information of the target object to obtain the second interest feature of the target object. Based on the first interest feature, the second interest feature, and the obtained target attribute information, the target interest feature of the target object is determined. Finally, based on the target interest feature, at least one target recall content is determined from the various content to be recalled.
[0077] In this embodiment, feature extraction is performed on the target content sequence in the content viewing record to obtain the first interest feature of the target object. Feature extraction is then performed on the target profile information of the target object to obtain the second interest feature of the target object. Finally, the first interest feature, the second interest feature, and the target attribute information corresponding to each target content in the target content sequence are combined to determine the target interest feature of the target object. Since the target interest feature of the target object integrates the target object's interest features across multiple dimensions, it more completely and accurately represents the target object's interests and hobbies. Therefore, when recalling content based on the target object's target interest feature, the accuracy of content recall can be effectively improved, thereby improving the accuracy of content recommendation.
[0078] refer to Figure 1 This is a system architecture diagram applicable to the embodiments of this application. The system architecture includes at least a terminal device 101 and a server 102.
[0079] The terminal device 101 pre-installs a target application with content recommendation functionality. This target application can be a client application, a web application, a mini-program application, etc. The terminal device 101 may include one or more processors 1011, a memory 1012, an I / O interface 1013 for interacting with the server 102, and a display panel 1014, etc. The terminal device 101 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, in-vehicle device, etc., but is not limited to these.
[0080] Server 102 is the backend server for the target application. Server 102 may include one or more processors 1021, memory 1022, and I / O interfaces 1023 for interacting with terminal device 101. Furthermore, server 102 may be configured with a database 1024. Server 102 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. Terminal device 101 and server 102 can be connected directly or indirectly via wired or wireless communication, which is not limited herein.
[0081] The content recall method in this application embodiment can be executed by terminal device 101 or server 102.
[0082] In the first scenario, the content recall method in this application embodiment can be executed by the terminal device 101.
[0083] Terminal device 101 extracts a target content sequence from the target object's content viewing history and obtains the target attribute information corresponding to each target content in the target content sequence. Then, it performs feature extraction on the target content sequence to obtain the target object's first interest feature, and performs feature extraction on the target object's target profile information to obtain the target object's second interest feature. Based on the first interest feature, the second interest feature, and the obtained target attribute information, it determines the target object's target interest feature. Then, based on the target interest feature, it determines at least one target recall content from the various content to be recalled. Further, terminal device 101 filters the at least one target recall content to obtain at least one candidate recommendation content. Then, it ranks the at least one candidate recommendation content and displays it on the target application's recommendation interface according to the ranking result.
[0084] In the second scenario, the content recall method in this application embodiment can be executed by server 102.
[0085] Server 102 extracts target content sequences from the target object's content viewing history and obtains the target attribute information corresponding to each target content in the target content sequence. Then, it performs feature extraction on the target content sequence to obtain the target object's first interest feature and extracts features from the target object's target profile information to obtain the target object's second interest feature. Based on the first interest feature, the second interest feature, and the obtained target attribute information, it determines the target object's target interest feature. Next, based on the target interest feature, it determines at least one target recall content from the various content to be recalled. Further, server 102 filters the at least one target recall content to obtain at least one candidate recommendation content. Then, it ranks the at least one candidate recommendation content to obtain a recommendation ranking result. Server 102 sends the at least one candidate recommendation content and the recommendation ranking result to terminal device 101. Terminal device 101 displays the at least one candidate recommendation content in the target application's recommendation interface according to the recommendation ranking result.
[0086] It should be noted that the content retrieval method in this embodiment can also be executed interactively by the terminal device 101 and the server 102, which will not be elaborated here.
[0087] based on Figure 1 The system architecture diagram shown in this application illustrates the flow of a content retrieval method. Figure 2 As shown, the process of this method is executed by a computer device, which can be... Figure 1 The terminal device 101 or server 102 shown includes the following steps:
[0088] Step S201: Extract the target content sequence from the content viewing record of the target object, and obtain the target attribute information corresponding to each target content in the target content sequence.
[0089] Specifically, the target object can be a user account, device ID, operating system ID, etc. The content in this embodiment can be articles, images, videos, etc. Correspondingly, the content viewing record includes at least one of the following: article viewing record, image viewing record, video viewing record, etc. Users can generate content viewing records through clicks, double-clicks, long presses, etc. The target content sequence is a sequence of content viewed by the target object within a target time period, where the target time period can be the most recent preset duration or other time periods. The target attribute information of the target content includes the category of the target content, the attribute tags of the target content, etc. The server stores the attribute information of each piece of content and updates the attribute information of each piece of content periodically or in real time. The attribute information of the content can be obtained through manual labeling or through speech recognition or image recognition.
[0090] For example, setting the current time as 10:00, user account M clicked on football match video A, football match highlights video B, news report C about football star 1, and news report D about football star 2 between 9:00 and 10:00. The target content sequence extracted from user account M's viewing history between 9:00 and 10:00 is: football match video A, football match highlights video B, news report C about football star 1, and news report D about football star 2. Then, the target attribute information for each of these target contents is obtained. Specifically, the target attribute information for football match video A is: category {video}, attribute tag {sports, football}; the target attribute information for football match highlights video B is: category {video}, attribute tag {sports, football}; the target attribute information for news report C about football star 1 is: category {article}, attribute tag {sports, football, football star 1}; and the target attribute information for news report D about football star 2 is: category {article}, attribute tag {sports, football, football star 2}.
[0091] Step S202: Extract features from the target content sequence to obtain the first interest feature of the target object.
[0092] Specifically, features can be extracted from the target content sequence using RNN, LSTM, GRU, Transformer, embedding, and other methods to obtain the first interest feature of the target object. Since the first interest feature is obtained by extracting features from the target content sequence over a period of time, it can be used to characterize the short-term interest features of the target object over a given time period.
[0093] Taking the Transformer model as an example, the Transformer model is an efficient parallel computing model that can process time series. It consists of an Encoder and a Decoder. In this embodiment, the Encoder is used to extract features from the target content sequence to obtain the first interest feature of the target object.
[0094] Specifically, the Encoder consists of 6 identical layers, each with the following structure: Figure 3 As shown, each layer consists of two sub-layers: a multi-head self-attention mechanism and a fully connected feed-forward network. Each sub-layer incorporates residual connections and normalization. Therefore, the output of the sub-layer can be expressed as the following formula (1):
[0095] …….(1)
[0096] in, This represents the output of sub_layer. Indicates normalization, This represents the input to sub_layer. This represents the input for the residual connection.
[0097] The structure of these two sub-layers will be introduced below. The two sub-layers are sub-layer_1 and sub-layer_2.
[0098] sub_layer_1:
[0099] sub_layer_1 uses Multi-head self-attention as its main structure, where the calculation process of attention is as follows (2):
[0100] …….(2)
[0101] Where Q represents the query vector, Indicates the weight coefficient key. This represents the value to be merged.
[0102] Multi-head attention projects Q, K, and V through h different linear transformations and then concatenates the different attention results.
[0103] sub_layer_2:
[0104] sub_layer_2 uses feed-forward networks (FFN) as its main structure. The role of this FFN is spatial transformation. The FFN contains two linear transformation layers, and the activation function between the two linear transformation layers is ReLU.
[0105] In this embodiment of the application, the target content sequence is input into the Encoder of the trained Transformer model, and the Encoder extracts features from the target content sequence to obtain the first interest feature of the target object.
[0106] For example, such as Figure 4 As shown, the target content sequence of user account M, namely: football match video A, football match highlights video B, news report of football star 1 C, and news report of football star 2 D, is input into the Encoder of the trained Transformer model. The Encoder outputs the first interest feature of user account M.
[0107] Since users' information consumption patterns remain relatively consistent over a period of time, modeling short-term user interests based on the sequence of target content clicked within that timeframe is the most effective approach, allowing the model to capture the user's current point of interest. Furthermore, because the Transformer model can be computed in parallel and considers all content within the target content sequence, using it for feature extraction from the target content sequence provides an even better model of the user's short-term interests.
[0108] Step S203: Extract features from the target profile information of the target object to obtain the second interest features of the target object.
[0109] Specifically, the target profile information includes profile tags such as the target object's age, gender, and city. By extracting features from each profile tag, the target object's secondary interest features are obtained. Since each profile tag in the target profile information is an inherent attribute of the target object and does not easily change, the secondary interest features obtained by feature extraction from the target profile information can be used to represent the target object's long-term interest characteristics.
[0110] In practice, methods such as GRU model, Transformer model, and embedding can be used to extract features from the target profile information of the target object and obtain the second interest features of the target object.
[0111] Step S204: Based on the first interest feature, the second interest feature, and the obtained target attribute information, determine the target interest feature of the target object.
[0112] Specifically, the target attribute information corresponding to each target content in the target content sequence can be used to represent the user's preference for content attributes. By combining the short-term interest features represented by the first interest feature, the long-term interest features represented by the second interest feature, and the user's preference for content attributes, target interest features that can completely represent the user's interests can be obtained.
[0113] Step S205: Based on the target interest characteristics, determine at least one target recall content from each content to be recalled.
[0114] Specifically, the content to be recalled includes videos, images, and articles. Each piece of content to be recalled is pre-stored in a content library, which is updated periodically or in real time.
[0115] In this embodiment, feature extraction is performed on the target content sequence in the content viewing record to obtain the first interest feature of the target object. Feature extraction is then performed on the target profile information of the target object to obtain the second interest feature of the target object. Finally, the first interest feature, the second interest feature, and the target attribute information corresponding to each target content in the target content sequence are combined to determine the target interest feature of the target object. Since the target interest feature of the target object integrates the target object's interest features across multiple dimensions, it more completely and accurately represents the target object's interests and hobbies. Therefore, when recalling content based on the target object's target interest feature, the accuracy of content recall can be effectively improved, thereby improving the accuracy of content recommendation.
[0116] Optionally, in step S203 above, when extracting features from the target image information of the target object to obtain the second interest feature of the target object, the embodiments of this application provide the following implementation methods:
[0117] Implementation Method 1, such as Figure 5 As shown, feature extraction is performed on each portrait label in the target portrait information to obtain the label features corresponding to each portrait label. Based on the weight of each portrait label and the label features corresponding to each portrait label, the second interest features of the target object are obtained.
[0118] Specifically, embedding is performed on each profile tag to obtain the tag features corresponding to each profile tag. Since the importance of each profile tag in the target profile information varies, in order to make the obtained secondary interest features more accurately represent the user's long-term interests, it is necessary to perform weighted fusion of the tag features corresponding to each profile tag based on their importance to obtain the secondary interest features of the target object.
[0119] In practice, the attention mechanism network is pre-trained to acquire weights for each image label. After training, the attention mechanism network is used to perform weighted fusion of the label features corresponding to each image label based on their weights, thereby obtaining the second interest feature of the target object. This attention mechanism network can be a multi-head self-attention mechanism network in the Transformer model, or a standalone self-attention network.
[0120] For example, such as Figure 6 As shown, the target profile information for user account M is defined by the following tags: 25 years old, male, Shanghai, programmer. Each tag is input into the encoder of the trained Transformer model, and the encoder outputs the second interest feature of user account M.
[0121] In this embodiment, the important procedures of each portrait tag are distinguished by the weight of each portrait tag. Then, based on the weight of each portrait tag, the tag features corresponding to each portrait tag are weighted and fused so that the second interest features of the target object more accurately represent the user's long-term interests.
[0122] Implementation Method 2, such as Figure 7 As shown, features are extracted from each image tag in the target image information to obtain the tag features corresponding to each image tag. The tag features corresponding to each image tag are added together to obtain the second interest feature of the target object.
[0123] Specifically, embedding is performed on each profile tag to obtain the tag features corresponding to each profile tag. Then, the tag features corresponding to each profile tag are directly added together to obtain the second interest feature used to represent the user's long-term interests.
[0124] In this application, the method of directly adding the label features corresponding to each portrait label to obtain the second interest feature of the target object has low requirements for the complexity of the fusion network and fast processing speed.
[0125] Optionally, in step S204 above, when determining the target interest features of the target object based on the first interest feature, the second interest feature, and the obtained target attribute information, this application embodiment provides at least the following implementation methods:
[0126] Implementation method one, such as Figure 8 As shown, the first interest feature and the second interest feature are fused to obtain the target fusion feature of the target object. Then, based on the target fusion feature and the information of each target attribute, the third interest feature of the target object is obtained. Finally, the first interest feature, the second interest feature, and the third interest feature are fused to obtain the target interest feature of the target object.
[0127] Specifically, the first interest feature and the second interest feature can be directly added together to obtain the target fusion feature of the target object. Alternatively, an attention mechanism network can be used to perform a weighted summation of the first interest feature and the second interest feature to obtain the target fusion feature of the target object.
[0128] Optionally, feature extraction is performed on each target attribute information to obtain candidate attribute features corresponding to each target attribute information. Then, based on the similarity between the target fusion feature and each of the obtained candidate attribute features, at least one target attribute feature is obtained from each candidate attribute feature. Finally, based on at least one target attribute feature and its corresponding weight, a third interest feature of the target object is obtained.
[0129] Specifically, embedding is performed on each target attribute information to obtain candidate attribute features corresponding to each target attribute information. Based on the target fusion feature, nearest neighbor search is performed on the candidate attribute features corresponding to each target attribute information to determine the similarity between the target fusion feature and each obtained candidate attribute feature, and at least one target attribute feature is obtained from each candidate attribute feature based on the similarity.
[0130] Specifically, the candidate attribute features with the highest similarity ranking (N) can be selected as the target attribute features, where N is a positive integer. Alternatively, candidate attribute features with similarity greater than a preset threshold can be selected as the target attribute features. The similarity can be Cosine similarity, Euclidean distance, etc. The calculation of Cosine similarity is as follows (3):
[0131] ……(3)
[0132] in, Cosine similarity between feature vector u and feature vector v Let be the eigenvalue of the i-th dimension in the eigenvector u. Let be the eigenvalue of the i-th dimension in the eigenvector v.
[0133] An attention mechanism network is employed to fuse at least one target attribute feature based on its respective weights, thereby obtaining a third interest feature for the target object. The weights for each target attribute feature are obtained by training the attention mechanism network. This third interest feature represents the user's content attribute interest. Alternatively, average-pooling or max-pooling methods can be used to fuse the first, second, and third interest features to obtain the target interest feature for the target object.
[0134] For example, such as Figure 9 As shown, user account M recently watched two videos and read two articles. The target attributes for these videos and articles are: video, article, sports, and football. Embedding is performed on each target attribute to obtain candidate attribute feature 1, candidate attribute feature 2, candidate attribute feature 3, and candidate attribute feature 4. The first and second interest features of user account M are fused to obtain the target fused feature. Nearest neighbor search is performed on each candidate attribute feature using the target fused feature to obtain the two most similar target attribute features, which are candidate attribute feature 3 and candidate attribute feature 4. Candidate attribute feature 3 and candidate attribute feature 4 are then input into an attention mechanism network to obtain the third interest feature. Finally, the first, second, and third interest features are fused to obtain the target interest feature of user account M.
[0135] In this embodiment, based on the user's long-term and short-term interests, the most relevant target attribute features are selected from the target attribute features corresponding to each target content. Then, the target attribute features are weighted and fused to make the obtained third interest features more accurately represent the user's preference for content attributes. At the same time, the third interest features are aligned with the distribution of the user's long-term and short-term interests, balancing the user's long-term and short-term interests, strengthening the characterization of the collaborative relationship between the various contents viewed by the user, and improving the recommendation effect for users with sparse behavior.
[0136] Implementation Method 2, such as Figure 10 As shown, feature extraction is performed on each target attribute information separately to obtain the target attribute features corresponding to each target attribute information. The obtained target attribute features are then fused to obtain the third interest feature of the target object. The first interest feature, second interest feature, and third interest feature are then fused to obtain the target interest feature of the target object.
[0137] Specifically, each target attribute is embedded to obtain its corresponding target attribute feature. An attention mechanism network is then used to fuse these features based on their respective weights, resulting in the third interest feature of the target object. The weights for each target attribute feature are obtained by training the attention mechanism network. Alternatively, the target attribute features can be directly added together to obtain the third interest feature of the target object. This third interest feature is used to represent the user's content attribute interest.
[0138] Alternatively, avg-pooling or max-pooling can be used to fuse the first interest feature, the second interest feature, and the third interest feature to obtain the target interest feature of the target object.
[0139] For example, such as Figure 11 As shown, user account M recently watched two videos and read two articles. The target attributes for the videos and articles are: video, article, sports, and football. Embedding is performed on each target attribute to obtain target attribute feature 1, target attribute feature 2, target attribute feature 3, and target attribute feature 4. These four features are then input into an attention mechanism network to obtain a third interest feature. Finally, the first, second, and third interest features are fused to obtain the target interest features of user account M.
[0140] In this embodiment, the target interest features of the target object are obtained by combining the user's long-term interests, short-term interests, and content attribute interests from multiple dimensions. This allows the obtained target interest features to more comprehensively and accurately represent the user's interests and hobbies. Therefore, when recalling content based on target interest features, the accuracy of content recall can be improved, thereby improving the accuracy of content recommendation.
[0141] Optionally, in addition to users' long-term interests, short-term interests, and content attribute interests, contextual information such as the time, device, and network information of users viewing content also has certain reference value for content retrieval and recommendation. Therefore, in this embodiment, the first interest feature, the second interest feature, the third interest feature, and the contextual features corresponding to the target object are fused to obtain the target interest feature of the target object.
[0142] Specifically, the contextual environment information corresponding to the target object is pre-acquired, including environmental parameters such as time information, device information, and network information. Embedding is performed on each environmental parameter to obtain its corresponding environmental features. Then, the obtained environmental features are fused using either direct addition or a weighted summation method employing an attention mechanism network to obtain the contextual environment features corresponding to the target object. Finally, either average pooling or max pooling is used to fuse the first interest feature, second interest feature, third interest feature, and the contextual environment features corresponding to the target object to obtain the target interest feature.
[0143] By combining multiple dimensions of user's long-term interests, short-term interests, content attribute interests, and contextual information, the target interest features of the target object can be obtained. This allows the obtained target interest features to more comprehensively and accurately represent the user's interests and hobbies. Therefore, when recalling content based on target interest features, the accuracy of content recall can be improved, thereby improving the accuracy of content recommendation.
[0144] It should be noted that when obtaining the target interest characteristics of the target object, it is not limited to using the first interest characteristics, second interest characteristics, third interest characteristics and the contextual environment characteristics corresponding to the target object. It may also include the user's specific content viewing format (such as playing videos, sharing videos, liking videos, commenting on videos, etc.), user feedback information (such as positive and negative evaluations of videos, etc.), and other user behavioral information. This application does not make specific limitations in this regard.
[0145] Optionally, in step S205 above, during content recall, features are extracted from each piece of content to be recalled to obtain content features corresponding to each piece of content to be recalled. Then, based on the similarity between the target interest features and each of the obtained content features, at least one target content to be recalled is determined from each piece of content to be recalled.
[0146] Specifically, embedding is performed on each piece of content to be recalled to obtain the content features corresponding to each piece of content. Similarity can be cosine similarity, Euclidean distance, etc. The pieces of content to be recalled can be sorted in descending order of similarity, and then the top L pieces of content to be recalled are selected as the target content to be recalled, where L is a preset positive integer. Alternatively, a candidate threshold can be preset, and the pieces of content to be recalled with similarity greater than the candidate threshold can be selected as the target content to be recalled.
[0147] In this embodiment, based on the target interest features, the similarity between the target interest features and the obtained content features is used to determine at least one target recall content that is most similar to the target interest features from each content to be recalled, so that the target recall content is close to the user's interests and the accuracy of content recall is improved.
[0148] Optionally, after identifying at least one target content from the various content to be recalled, the at least one target content is filtered according to preset rules to obtain at least one candidate recommended content. Then, based on the similarity between the target's interest features and the at least one candidate recommended content, the at least one candidate recommended content is ranked to obtain a recommendation ranking result. Finally, according to the recommendation ranking result, the at least one candidate recommended content is recommended to the target object.
[0149] Specifically, the preset rules can be content relevance, content timeliness, content geographic location, content diversity, etc. By filtering at least one target content, at least one candidate recommendation can be obtained, reducing the scale of subsequent recommendation ranking. The server ranks the at least one candidate recommendation according to the similarity from highest to lowest, obtaining the recommendation ranking result. Then, the recommendation ranking result and at least one candidate recommendation are sent to the terminal device, and the terminal device displays at least one candidate recommendation in the application's recommendation interface according to the recommendation ranking result.
[0150] For example, the server uses the content recall method in this embodiment to determine five target recall contents: Video 1, Video 2, Video 3, Article 4, and Article 5. Then, it filters these five target recall contents based on content relevance. Since the content relevance of Video 1 and Video 2 is greater than a preset threshold, and the content relevance of Article 4 and Article 5 is also greater than a preset threshold, to avoid recommending duplicate content to users and causing user aversion, Video 2 and Article 5 are removed, leaving Video 1, Video 3, and Article 4 as candidate recommendation contents. Then, based on the similarity between the target interest features and the above three candidate recommendation contents, the three candidate recommendation contents are ranked, resulting in a recommendation ranking of Video 3, Article 4, and Video 1. The server sends Video 1, Video 3, Article 4, and the recommendation ranking result to the terminal device. The terminal device displays Video 1, Video 3, and Article 4 in the application's recommendation interface according to the recommendation ranking result, specifically as follows: Figure 12 As shown, the recommendation interface displays Video 3, Article 4, and Video 1 from top to bottom.
[0151] It should be noted that, in addition to ranking at least one candidate recommendation based on the similarity between the target interest features and at least one candidate recommendation, other methods can be used to rank each candidate recommendation after obtaining the ranking result. For example, a click-through rate prediction model can be trained to predict the probability of a user clicking on each candidate recommendation, and then the recommendations can be ranked in descending order of probability.
[0152] In this embodiment, by combining multiple dimensions of user characteristics, such as long-term interests, short-term interests, and content attribute interests, the target interest features of the target object are obtained. This allows the obtained target interest features to more comprehensively and accurately represent the user's interests and hobbies, thus improving the accuracy of content retrieval when recalling content based on target interest features. Furthermore, by filtering, sorting, and recommending the recalled content, the accuracy of content recommendation is also improved, thereby enhancing the user experience.
[0153] To better explain the embodiments of this application, the following describes a content retrieval method provided by the embodiments of this application, in conjunction with the model structure. This method is executed by the server. First, the network structure of the Transformer-based model in the embodiments of this application is introduced, such as... Figure 13 As shown, it includes the content sequence module (ItemSeq), the object feature module (UserProf), the content attribute module (Attr2Item), the online service module (Online Serving), the offline training module (OfflineTraining), the shared embedding module (Shared Embeds), and various embedding lookup layers (Various Embeds LookupLayer).
[0154] The content sequence module (ItemSeq) includes two Transformer layers, the object feature module (UserProf) includes one Transformer layer, and the content attribute module (Attr2Item) includes one attention mechanism layer (TopN-attention layer). The shared embedding module provides various feature embeddings for various embedding lookup layers, specifically including attribute embeddings (Attr Embeds), content embeddings (Item Embeds), action embeddings (Action Embeds), and context embeddings (Context Embeds).
[0155] First, the Transformer-based model is trained offline. The offline training module calculates the sampling loss between the predicted sample interest features and the reference interest features (Targrt Embed). Training ends when the sampling loss meets a preset condition.
[0156] The specific process of content retrieval using a pre-trained Transformer-based model is as follows:
[0157] Extract the sequence of target content clicked by the user in the recent period from the content viewing history of the target user account (Item Seq Inputs), obtain the target attribute information (Item Attr Inputs) corresponding to each target content in the target content sequence, and obtain the object feature information (User Prof Inputs) of the target user account.
[0158] The target content sequence (Item Seq Inputs) is input into the content sequence module (ItemSeq) through various embedding lookup layers. The content sequence module extracts features from the object's feature information through sequence embedding (Seq Embeds Pooler), Transformer layer, and sequence units (Seq Units) to obtain the first interest feature. ).
[0159] The object feature information (User Prof Inputs) is input into the object feature module (UserProf) through various embedding lookup layers. The object feature module extracts features from the target content sequence through profile embeddings (Prof Embeds), a Transformer layer, and profile units (Prof Units) to obtain the second interest feature. ).
[0160] The first and second interest features are aggregated to obtain the long and short interest features. These long and short interest features are then input into the content attribute module (Attr2Item), along with the candidate attribute features corresponding to each target attribute (Item Attr Inputs). The content attribute module uses the long and short interest features as its activation signal to perform nearest neighbor search. From the candidate attribute features corresponding to each target attribute, it selects the N target attribute features most similar to the long and short interest features. Then, an attention mechanism is used to weight-sum pooling these N target attribute features to obtain the third interest feature. ).
[0161] An average pooling approach is used to fuse the first, second, and third interest features to obtain the target interest feature, which is then input into the online serving module. The online serving module calculates the cosine similarity between the target interest feature and each piece of content to be recalled in the content library. Based on the cosine similarity, the Top K target content pieces are selected from the remaining content.
[0162] Furthermore, according to preset rules, the TopK target recall content is filtered to obtain at least one candidate recommendation content. A click-through rate prediction model is trained to estimate the probability of a user clicking on each candidate recommendation content, and then the recommendations are sorted in descending order of probability to obtain the recommendation ranking result. Based on the recommendation ranking result, at least one candidate recommendation content is recommended to the target user account.
[0163] In this embodiment, feature extraction is performed on the target content sequence in the content viewing record to obtain the first interest feature of the target object. Feature extraction is then performed on the target profile information of the target object to obtain the second interest feature of the target object. Finally, the first interest feature, the second interest feature, and the target attribute information corresponding to each target content in the target content sequence are combined to determine the target interest feature of the target object. Since the target interest feature of the target object integrates the target object's interest features across multiple dimensions, it more completely and accurately represents the target object's interests and hobbies. Therefore, when recalling content based on the target object's target interest feature, the accuracy of content recall can be effectively improved, thereby improving the accuracy of content recommendation.
[0164] Based on the same technical concept, embodiments of this application provide a content recall device, such as... Figure 14 As shown, the device 1400 includes:
[0165] The acquisition module 1401 is used to extract a target content sequence from the content viewing record of the target object, and to acquire the target attribute information corresponding to each target content in the target content sequence;
[0166] The first feature extraction module 1402 is used to extract features from the target content sequence to obtain the first interest feature of the target object;
[0167] The second feature extraction module 1403 is used to extract features from the target profile information of the target object to obtain the second interest feature of the target object.
[0168] The fusion module 1404 is used to determine the target interest features of the target object based on the first interest feature, the second interest feature, and the obtained target attribute information;
[0169] The content recall module 1405 is used to determine at least one target recall content from each content to be recalled based on the target interest characteristics.
[0170] Optionally, the second feature extraction module 1403 is specifically used for:
[0171] Feature extraction is performed on each portrait tag in the target portrait information to obtain the tag features corresponding to each portrait tag;
[0172] Based on the weights of each portrait label and the label features corresponding to each portrait label, the second interest features of the target object are obtained.
[0173] Optionally, the fusion module 1404 is specifically used for:
[0174] The first interest feature and the second interest feature are fused to obtain the target fusion feature of the target object;
[0175] Based on the target fusion features and the target attribute information, the third interest feature of the target object is obtained;
[0176] The first interest feature, the second interest feature, and the third interest feature are fused to obtain the target interest feature of the target object.
[0177] Optionally, the fusion module 1404 is specifically used for:
[0178] Feature extraction is performed on each of the target attribute information to obtain the candidate attribute features corresponding to each target attribute information;
[0179] Based on the target fusion feature, at least one target attribute feature is obtained from each candidate attribute feature by comparing the similarity between the target fusion feature and the obtained candidate attribute features.
[0180] Based on the at least one target attribute feature and the weight corresponding to each of the at least one target attribute feature, the third interest feature of the target object is obtained.
[0181] Optionally, the fusion module 1404 is specifically used for:
[0182] The first interest feature, the second interest feature, the third interest feature, and the contextual features corresponding to the target object are fused to obtain the target interest feature of the target object.
[0183] Optionally, the content retrieval module 1405 is specifically used for:
[0184] Each of the contents to be recalled is subjected to feature extraction to obtain the content features corresponding to each of the contents to be recalled.
[0185] Based on the target interest features, and the similarity between them and the obtained content features, at least one target recall content is determined from the various content to be recalled.
[0186] Optionally, it also includes a content recommendation module 1406;
[0187] The content recommendation module 1406 is specifically used for:
[0188] Based on the target interest characteristics, after determining at least one target recall content from each content to be recalled, the at least one target recall content is filtered according to preset rules to obtain at least one candidate recommendation content;
[0189] Based on the target interest features, the similarity between the target and at least one candidate recommended content is used to rank the at least one candidate recommended content to obtain a recommendation ranking result.
[0190] Based on the recommendation ranking results, at least one candidate recommendation content is recommended to the target object.
[0191] Optionally, the first interest feature is used to characterize the short-term interest features of the target object, and the second interest feature is used to characterize the long-term interest features of the target object.
[0192] In this embodiment, feature extraction is performed on the target content sequence in the content viewing record to obtain the first interest feature of the target object. Feature extraction is then performed on the target profile information of the target object to obtain the second interest feature of the target object. Finally, the first interest feature, the second interest feature, and the target attribute information corresponding to each target content in the target content sequence are combined to determine the target interest feature of the target object. Since the target interest feature of the target object integrates the target object's interest features across multiple dimensions, it more completely and accurately represents the target object's interests and hobbies. Therefore, when recalling content based on the target object's target interest feature, the accuracy of content recall can be effectively improved, thereby improving the accuracy of content recommendation.
[0193] Based on the same technical concept, embodiments of this application provide a computer device, which may be a terminal or a server, such as... Figure 15As shown, it includes at least one processor 1501 and a memory 1502 connected to at least one processor. In this embodiment, the specific connection medium between the processor 1501 and the memory 1502 is not limited. Figure 15 Taking the connection between processor 1501 and memory 1502 via a bus as an example, the bus can be divided into address bus, data bus, control bus, etc.
[0194] In this embodiment of the application, the memory 1502 stores instructions that can be executed by at least one processor 1501. By executing the instructions stored in the memory 1502, at least one processor 1501 can perform the steps included in the above-described content recall method.
[0195] The processor 1501 is the control center of the computer device. It can connect to various parts of the computer device using various interfaces and lines. It performs content retrieval and recommendation by running or executing instructions stored in the memory 1502 and calling data stored in the memory 1502. Optionally, the processor 1501 may include one or more processing units. The processor 1501 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 1501. In some embodiments, the processor 1501 and the memory 1502 can be implemented on the same chip; in some embodiments, they can also be implemented on separate chips.
[0196] Processor 1501 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.
[0197] Memory 1502, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 1502 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 1502 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In the embodiments of this application, memory 1502 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.
[0198] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of the above-described content recall method.
[0199] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0200] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0201] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0202] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0203] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0204] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A content recall method, characterized in that, include: Extract the target content sequence from the content viewing history of the target object, and obtain the target attribute information corresponding to each target content in the target content sequence; Feature extraction is performed on the target content sequence to obtain the first interest feature of the target object; Feature extraction is performed on the target profile information of the target object to obtain the second interest feature of the target object; The first interest feature and the second interest feature are fused to obtain the target fusion feature of the target object; Furthermore, based on the target fusion features and the obtained target attribute information, the third interest feature of the target object is obtained; The first interest feature, the second interest feature, and the third interest feature are fused to obtain the target interest feature of the target object. Based on the target interest characteristics, at least one target recall content is determined from each content to be recalled.
2. The method as described in claim 1, characterized in that, The step of extracting features from the target profile information of the target object to obtain the second interest feature of the target object includes: Feature extraction is performed on each portrait tag in the target portrait information to obtain the tag features corresponding to each portrait tag; Based on the weights of each portrait label and the label features corresponding to each portrait label, the second interest features of the target object are obtained.
3. The method as described in claim 1, characterized in that, The process of obtaining the third interest feature of the target object based on the target fusion features and the obtained target attribute information includes: Feature extraction is performed on each of the target attribute information to obtain the candidate attribute features corresponding to each target attribute information; Based on the target fusion feature, at least one target attribute feature is obtained from each candidate attribute feature by comparing the similarity between the target fusion feature and the obtained candidate attribute features. Based on the at least one target attribute feature and the weight corresponding to each of the at least one target attribute feature, the third interest feature of the target object is obtained.
4. The method as described in claim 1, characterized in that, The step of fusing the first interest feature, the second interest feature, and the third interest feature to obtain the target interest feature of the target object includes: The first interest feature, the second interest feature, the third interest feature, and the contextual features corresponding to the target object are fused to obtain the target interest feature of the target object.
5. The method as described in claim 1, characterized in that, The step of determining at least one target recall content from each list of content to be recalled based on the target interest features includes: Each of the contents to be recalled is subjected to feature extraction to obtain the content features corresponding to each of the contents to be recalled. Based on the target interest features, and the similarity between them and the obtained content features, at least one target recall content is determined from the various content to be recalled.
6. The method according to any one of claims 1 to 5, characterized in that, After determining at least one target content to be recalled from the various content to be recalled based on the target interest features, the method further includes: According to preset rules, the at least one target recall content is filtered to obtain at least one candidate recommendation content; Based on the target interest features, the similarity between the target and at least one candidate recommended content is used to rank the at least one candidate recommended content to obtain a recommendation ranking result. Based on the recommendation ranking results, at least one candidate recommendation content is recommended to the target object.
7. The method as described in claim 6, characterized in that, The first interest feature is used to characterize the short-term interest features of the target object, and the second interest feature is used to characterize the long-term interest features of the target object.
8. A content recall device, characterized in that, include: The acquisition module is used to extract a target content sequence from the content viewing records of the target object, and to obtain the target attribute information corresponding to each target content in the target content sequence; The first feature extraction module is used to extract features from the target content sequence to obtain the first interest feature of the target object; The second feature extraction module is used to extract features from the target profile information of the target object to obtain the second interest features of the target object; The fusion module is used to fuse the first interest feature and the second interest feature to obtain the target fusion feature of the target object; Furthermore, based on the target fusion features and the obtained target attribute information, a third interest feature of the target object is obtained; the first interest feature, the second interest feature, and the third interest feature are fused to obtain the target interest feature of the target object. The content recall module is used to determine at least one target recall content from each content to be recalled based on the target interest characteristics.
9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, It stores a computer program executable by a computer device, which, when run on the computer device, causes the computer device to perform the steps of any one of claims 1 to 7.