Content recommendation method and related devices

Abstract

This application discloses a content recommendation method and related devices. Related embodiments can be applied to various scenarios such as cloud technology, artificial intelligence, and intelligent transportation. It is possible to obtain the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information. Deep feature extraction is performed on the object attribute information and the historical interaction content sequence to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period. Based on the deep content feature information and the interaction time information, interest analysis is performed on the target object to obtain the interest characteristics of the target object at different time periods. According to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information, target recommended content is selected from the candidate recommended content for recommendation. This application can improve the accuracy of content recommendation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular, to a content recommendation method and related devices. Background Art

[0002] With the rapid development of Internet technology, the content on the Internet has increased explosively, and it has become increasingly important to screen out the content that users are interested in from the vast amount of content and recommend it to users.

[0003] In the current related technologies, generally, personalized content recommendations are made for users based on the historical interaction content of the users. The historical interaction content of the users is the content that the users have interacted with, which can represent the interests of the users to a certain extent. However, the current content recommendation methods of this kind do not consider the temporal characteristics of the historical interaction content, resulting in a low accuracy rate of content recommendation. Summary of the Invention

[0004] Embodiments of this application provide a content recommendation method and related devices. The related devices may include a content recommendation device, an electronic device, a computer-readable storage medium, and a computer program product, which can improve the accuracy of content recommendation.

[0005] Embodiments of this application provide a content recommendation method, including:

[0006] Obtaining the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence;

[0007] Performing deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period;

[0008] Based on the deep content feature information and the interaction time information of the historical interaction content sequence, performing interest analysis on the target object to obtain the interest characteristics of the target object in different time periods;

[0009] Selecting target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information.

[0010] Correspondingly, embodiments of this application provide a content recommendation device, including:

[0011] An obtaining unit, configured to obtain the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence;

[0012] An extraction unit for deeply extracting features from the object attribute information and the historical interaction content sequence within the historical time period, to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period;

[0013] An analysis unit for performing interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence, to obtain the interest characteristics of the target object in different time periods;

[0014] A selection unit for selecting target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information.

[0015] Optionally, in some embodiments of the present application, the historical interaction content sequence within at least one historical time period includes the historical interaction content sequences within the first historical time period and the second historical time period; the second historical time period is earlier than the first historical time period;

[0016] The analysis unit may include a short-term interest analysis subunit and a long-term interest analysis subunit, as follows:

[0017] The short-term interest analysis subunit is used to perform short-term interest analysis on the target object according to the deep content feature information of the candidate recommended content, the deep content feature information corresponding to the historical interaction content sequence within the first historical time period, and the interaction time information corresponding to the historical interaction content sequence within the first historical time period, to obtain the short-term interest characteristics of the target object;

[0018] The long-term interest analysis subunit is used to perform long-term interest analysis on the target object according to the deep content feature information corresponding to the historical interaction content sequence within the second historical time period and the interaction time information corresponding to the historical interaction content sequence within the second historical time period, to obtain the long-term interest characteristics of the target object.

[0019] Optionally, in some embodiments of the present application, the selection unit may include a first fusion subunit, a shallow feature extraction subunit, a second fusion subunit, and a selection subunit, as follows:

[0020] The first fusion subunit is used to fuse the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information to obtain deep feature information;

[0021] A shallow feature extraction subunit, configured to perform shallow feature extraction on the object attribute information and the candidate recommended content, so as to obtain shallow attribute feature information of the object attribute information and shallow content feature information of the candidate recommended content;

[0022] A second fusion subunit, configured to fuse the shallow attribute feature information and the shallow content feature information of the candidate recommended content to obtain shallow feature information;

[0023] A selection subunit, configured to select target recommended content from the candidate recommended content for recommendation based on the deep feature information and the shallow feature information.

[0024] Optionally, in some embodiments of the present application, the selection subunit may specifically be configured to fuse the deep feature information and the shallow feature information to obtain target feature information; predict a recommendation index corresponding to the candidate recommended content according to the target feature information; and select target recommended content from the candidate recommended content for recommendation according to the recommendation index.

[0025] Optionally, in some embodiments of the present application, the deep feature information includes deep sub-features of at least one dimension;

[0026] The step of "fusing the deep feature information and the shallow feature information to obtain target feature information" may include:

[0027] Performing cross-operation on the deep sub-features of each dimension to obtain second-order cross-feature information corresponding to the deep feature information;

[0028] Performing fully-connected processing on the deep feature information to obtain fully-connected feature information corresponding to the deep feature information;

[0029] Fusing the second-order cross-feature information, the fully-connected feature information, the deep feature information, and the shallow feature information to obtain target feature information.

[0030] Optionally, in some embodiments of the present application, the short-term interest analysis subunit may specifically be configured to fuse the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fusion feature information; perform a normalization operation on the fusion feature information and the interaction time information corresponding to the historical interaction content sequence within the first historical time period to obtain weight information corresponding to the historical interaction content sequence within the first historical time period; and fuse each historical interaction content of the historical interaction content sequence within the first historical time period according to the weight information to obtain the short-term interest feature of the target object.

[0031] Optionally, in some embodiments of the present application, the step of "fusing the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain fused feature information" may include:

[0032] Performing a summation operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain a summation operation result;

[0033] Performing a multiplication operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain a multiplication operation result;

[0034] Fusing the summation operation result, the multiplication operation result, the deep content feature information of the candidate recommended content, and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain fused feature information.

[0035] Optionally, in some embodiments of the present application, the historical interaction content sequence in the second historical time period includes historical interaction content sequences in at least one dimension in the second historical time period;

[0036] The long-term interest analysis subunit may specifically be configured to perform attention processing on the deep content feature information corresponding to the historical interaction content sequence in each dimension in the second historical time period to obtain initial weight information corresponding to the historical interaction content sequence in the dimension; based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period, process the historical interaction content sequence in the dimension to obtain sequence feature information of the historical interaction content sequence in the dimension; and fuse the sequence feature information of the historical interaction content sequences in each dimension in the second historical time period to obtain the long-term interest feature of the target object.

[0037] Optionally, in some embodiments of the present application, the step of "processing the historical interaction content sequence in the dimension based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period to obtain sequence feature information of the historical interaction content sequence in the dimension" may include:

[0038] Fusing the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period to obtain target weight information corresponding to the historical interaction content sequence in the dimension;

[0039] Process the historical interaction content sequence under the dimension based on the target weight information to obtain the sequence feature information of the historical interaction content sequence under the dimension.

[0040] An electronic device provided by an embodiment of the present application includes a processor and a memory. The memory stores multiple instructions, and the processor loads the instructions to execute the steps in the content recommendation method provided by the embodiment of the present application.

[0041] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in the content recommendation method provided by the embodiment of the present application are implemented.

[0042] In addition, an embodiment of the present application further provides a computer program product, including a computer program or instruction. When the computer program or instruction is executed by a processor, the steps in the content recommendation method provided by the embodiment of the present application are implemented.

[0043] An embodiment of the present application provides a content recommendation method and related devices, which can obtain the object attribute information of a target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence; perform deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period; perform interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence to obtain the interest characteristics of the target object at different time periods; select target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information. The present application can perform content recommendation to a target object based on the interest characteristics of the target object at different time periods, improving the accuracy of content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained without creative efforts based on these drawings.

[0045] Figure 1a It is a schematic diagram of the scenario of the content recommendation method provided by the embodiment of the present application;

[0046] Figure 1bIt is a flowchart of the content recommendation method provided by the embodiment of the present application;

[0047] Figure 1c It is a page schematic diagram of the content recommendation method provided by the embodiment of the present application;

[0048] Figure 1d It is a model structure diagram of the content recommendation method provided by the embodiment of the present application;

[0049] Figure 1e It is another model structure diagram of the content recommendation method provided by the embodiment of the present application;

[0050] Figure 1f It is another model structure diagram of the content recommendation method provided by the embodiment of the present application;

[0051] Figure 2 It is another flowchart of the content recommendation method provided by the embodiment of the present application;

[0052] Figure 3 It is a structural schematic diagram of the content recommendation device provided by the embodiment of the present application;

[0053] Figure 4 It is a structural schematic diagram of the electronic device provided by the embodiment of the present application. Detailed implementation manners

[0054] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0055] The embodiment of the present application provides a content recommendation method and related devices. The related devices may include a content recommendation device, an electronic device, a computer-readable storage medium, and a computer program product. The content recommendation device may be specifically integrated in the electronic device, and the electronic device may be a device such as a terminal or a server.

[0056] It can be understood that the content recommendation method in this embodiment can be executed on the terminal, can also be executed on the server, or can be jointly executed by the terminal and the server. The above examples should not be construed as limitations to the present application.

[0057] As Figure 1a shown, taking the joint execution of the content recommendation method by the terminal and the server as an example. The content recommendation system provided by the embodiment of the present application includes a terminal 10, a server 11, etc.; the terminal 10 and the server 11 are connected through a network, for example, through a wired or wireless network connection, etc., where the content recommendation device may be integrated in the server.

[0058] Among them, the server 11 can be used to: obtain the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence; perform deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period; perform interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence to obtain the interest characteristics of the target object in different time periods; select target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information. Among them, the server 11 can be a single server, or a server cluster or cloud server composed of multiple servers. The content recommendation method or device disclosed in this application, where multiple servers can form a blockchain, and the server is a node on the blockchain.

[0059] Among them, the terminal 10 can be used to: receive the target recommended content sent by the server 11 and display the target recommended content on the corresponding recommended content page. Among them, the terminal 10 can include a mobile phone, a smart TV, a tablet computer, a laptop computer, or a personal computer (PC, Personal Computer), etc. A client can also be set on the terminal 10, and the client can be an application client or a browser client, etc.

[0060] The step of the above server 11 obtaining the recommended content can also be executed by the terminal 10.

[0061] The content recommendation method provided by the embodiments of this application involves computer vision technology and natural language processing in the field of artificial intelligence.

[0062] Among them, artificial intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable machines to have the functions of perception, reasoning, and decision-making. Artificial intelligence technology is an interdisciplinary subject that involves a wide range of fields, including both hardware-level and software-level technologies. Among them, artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, intelligent transportation, and other major directions.

[0063] Among them, computer vision technology (CV): Computer vision is a science that studies how to enable machines to "see". Further, it refers to using cameras and computers to replace human eyes for target recognition, measurement, and other machine vision, and further performing graphic processing to make the computer-processed images more suitable for human eyes to observe or be transmitted to instruments for detection. As a scientific discipline, computer vision studies related theories and technologies and attempts to build artificial intelligence systems that can obtain information from images or multi-dimensional data. Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, autonomous driving, intelligent transportation, and other technologies, as well as common biometric recognition technologies such as face recognition and fingerprint recognition.

[0064] Among them, natural language processing (NLP) is an important direction in the field of computer science and artificial intelligence. It studies various theories and methods that can achieve effective communication between humans and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, the research in this field will involve natural language, that is, the language people use in daily life, so it has a close connection with the research of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph, and other technologies.

[0065] The following will be described in detail respectively. It should be noted that the description order of the following embodiments does not limit the preferred order of the embodiments.

[0066] This embodiment will be described from the perspective of a content recommendation device, which can be specifically integrated in an electronic device, such as a server or a terminal.

[0067] It can be understood that in the specific implementation of this application, user information is involved, such as data related to the user's historical interaction content, etc. When the above embodiments of this application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0068] The content recommendation method of the embodiments of this application can be applied to various scenarios that require content recommendation, such as video recommendation, text recommendation, etc. This embodiment can be applied to various scenarios such as cloud technology, artificial intelligence, intelligent transportation, and assisted driving.

[0069] Such as Figure 1b As shown, the specific process of this content recommendation method can be as follows:

[0070] 101. Obtain the object attribute information of the target object, the historical interaction content sequence in at least one historical time period, and the interaction time information of the historical interaction content sequence.

[0071] Among them, the target object is the object for which content is to be recommended. Through the content recommendation method provided by this application, content can be recommended to the target object. The object attribute information can specifically include object characteristics such as the age and gender of the target object, and this embodiment does not limit this.

[0072] Among them, the historical interaction content sequence can include at least one historical interaction content. For example, it can include multiple historical interaction contents arranged in chronological order of interaction. Among them, each historical interaction content can specifically be the content corresponding to the interaction behavior of the target object at a certain time within the historical time period. Here, the interaction behavior can specifically refer to behaviors such as browsing and clicking; the content interacted with can include information in various modalities such as video, audio, text, and image, and this embodiment does not limit this. Specifically, for each historical interaction content, the historical interaction content can include content-related information in at least one dimension, or rather, each historical interaction content includes content-related information in multiple fields. For example, it can include content-related information in dimensions such as the content itself, the publisher of the content, the content category, the content title, and the identification number of the content.

[0073] Among them, the interaction time information of the historical interaction content sequence can include the interaction time information specifically corresponding to each historical interaction content in the historical interaction content sequence. Generally speaking, the historical interaction content closer to the current time is more helpful for predicting the next interaction behavior of the target object.

[0074] As shown Figure 1c in the figure, it may be a subscription message page displayed by the content recommendation method based on the present application, which may include videos and graphic messages recommended to the target object. The present application can be applied to the fine ranking side in the recommendation process and modeled with the click-through rate as the target, specifically modeled with the recommendation accuracy as the target.

[0075] 102. Deep feature extraction is performed on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period.

[0076] Among them, deep feature extraction can be performed on the object attribute information and the historical interaction content sequence through a neural network model. The neural network model may include a Visual Geometry Group Network (VGGNet), a Residual Network (ResNet), a Dense Convolutional Network (DenseNet), and so on. However, it should be understood that the neural network model in this embodiment is not limited to the several types listed above.

[0077] Specifically, compared with shallow feature extraction, the scale of the feature information extracted by deep feature extraction is relatively large. The larger the scale of the feature information, the richer the information it contains.

[0078] 103. Based on the deep content feature information and the interaction time information of the historical interaction content sequence, interest analysis is performed on the target object to obtain the interest characteristics of the target object at different time periods.

[0079] Specifically, in this embodiment, the historical interaction content of the target object in different historical time periods can be separately analyzed to obtain the interest characteristics at different time periods. The historical interaction content in different historical time periods can reflect different interest characteristics of the target object. For example, the historical interaction content closer to the current time has a more significant indication for predicting the next interaction content of the target object. The historical interaction content closer to the current time can be called short-term historical interaction content; while the historical interaction content farther from the current time may reflect the long-term interest of the target object, has no strong hint for predicting the next interaction content of the target object, but reflects the potential and stable interest of the target object. The historical interaction content farther from the current time can be called long-term historical interaction content.

[0080] Optionally, in this embodiment, the historical interaction content sequence within at least one historical time period includes the historical interaction content sequences within the first historical time period and the second historical time period; the second historical time period is earlier than the first historical time period;

[0081] The step of "performing interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence to obtain the interest characteristics of the target object at different time periods" may include:

[0082] Performing short-term interest analysis on the target object according to the deep content feature information of the candidate recommended content, the deep content feature information corresponding to the historical interaction content sequence within the first historical time period, and the interaction time information corresponding to the historical interaction content sequence within the first historical time period to obtain the short-term interest characteristics of the target object;

[0083] Performing long-term interest analysis on the target object according to the deep content feature information corresponding to the historical interaction content sequence within the second historical time period and the interaction time information corresponding to the historical interaction content sequence within the second historical time period to obtain the long-term interest characteristics of the target object.

[0084] Specifically, the time length of the first historical time period may be less than the time length of the second historical time period.

[0085] For example, the first historical time period may be the historical time period corresponding to within 3 days from the current time, and the second historical time period may be the time period corresponding to the 3rd to 15th days before the current time. This embodiment may construct the historical interaction content sequence within the first historical time period based on the historical interaction content generated by the target object within 3 days; construct the historical interaction content sequence within the second historical time period based on the historical interaction content generated by the target object within the 3rd to 15th days before the current time.

[0086] Among them, the deep content feature information of the candidate recommended content may be obtained by performing deep feature extraction on the candidate recommended content. Specifically, the candidate recommended content may be any content in the preset content database.

[0087] In a specific embodiment, the object attribute information may include attribute information of m fields (domains). A field can be specifically understood as a feature group. For example, age is a field, and gender is another field. The candidate recommended content and each historical interaction content have content information of K fields.

[0088] Among them, the candidate recommended content can be recorded as the target item (a piece of news), the historical interaction content can be recorded as the item in the historical interaction content sequence, and the historical interaction content sequence can be regarded as the sequence of items clicked by the target object in a certain scenario within a certain historical time period. For each item in the sequence, 5 types of features can be specifically mined, namely: item ID (identity, identification information), author ID, category of the item, tag of the item, and title of the item.

[0089] Suppose the historical interaction content sequence in the first historical time period contains n historical interaction contents, that is, the length of the short-term sequence is n; the historical interaction content sequence in the second historical time period contains p historical interaction contents, that is, the length of the long-term sequence is p; the scale size of the feature information extracted by the deep feature extraction is D, then the deep attribute feature information of the object attribute information can be recorded as The deep content feature information of the candidate recommended content can be recorded as The deep content feature information corresponding to the short-term sequence can be recorded as The deep content feature information corresponding to the long-term sequence can be recorded as In addition, the interaction time information corresponding to the historical interaction content sequence in the first historical time period can be recorded as Specifically, it is the time when each item in the short-term sequence occurs; the interaction time information corresponding to the historical interaction content sequence in the second historical time period can be recorded as Specifically, it is the time when each item in the long-term sequence occurs. Among them, represents the real number field.

[0090] Optionally, in this embodiment, the step of "performing short-term interest analysis on the target object according to the deep content feature information of the candidate recommended content, the deep content feature information corresponding to the historical interaction content sequence in the first historical time period, and the interaction time information corresponding to the historical interaction content sequence in the first historical time period, to obtain the short-term interest feature of the target object" may include:

[0091] Fusing the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain fused feature information;

[0092] Performing a normalization operation on the fused feature information and the interaction time information corresponding to the historical interaction content sequence in the first historical time period to obtain the weight information corresponding to the historical interaction content sequence in the first historical time period;

[0093] Fusing each piece of historical interaction content in the historical interaction content sequence within the first historical time period according to the weight information to obtain the short-term interest feature of the target object.

[0094] Among them, there are various fusion methods for the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period, and this embodiment does not limit this. For example, this fusion method can be concatenation, weighted operation, etc.

[0095] Among them, normalization operations can be performed through the softmax function, and the softmax function can convert the output values of multi-classification into a probability distribution with a range of [0, 1] and a sum of 1.

[0096] Among them, the weight information can specifically include the weights corresponding to each piece of historical interaction content in the historical interaction content sequence within the first historical time period. Based on this weight information, weighted operations can be performed on the historical interaction content in the historical interaction content sequence within the first historical time period, so as to obtain the short-term interest feature of the target object.

[0097] Optionally, in this embodiment, the step of "fusing the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain the fusion feature information" may include:

[0098] Performing a summation operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain the result of the summation operation;

[0099] Performing a multiplication operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain the result of the multiplication operation;

[0100] Fusing the result of the summation operation, the result of the multiplication operation, the deep content feature information of the candidate recommended content, and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain the fusion feature information.

[0101] Among them, there are various ways to fuse the result of the summation operation, the result of the multiplication operation, the deep content feature information of the candidate recommended content, and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period, and this embodiment does not limit this. For example, this fusion method can be splicing processing, and the splicing order is not limited, or this fusion method can also be weighted operation, etc.

[0102] Specifically, the deep content feature information of the candidate recommended content can be denoted as The deep content feature information corresponding to the historical interaction content sequence within the first historical time period, i.e., the short-term sequence feature, can be denoted as

[0103] In a specific embodiment, the short-term interest analysis of the target object can be performed through a short-term interest analysis model. Among them, there are various types of short-term interest analysis models, and this embodiment does not limit this. For example, the short-term interest analysis model can be a neural network structure of DIN Attention (deep interest network, Attention, attention mechanism), and the model structure diagram is as Figure 1d shown. The short-term interest analysis model can complete sequence feature modeling by establishing an attention relationship between the candidate recommended content and the short-term historical behavior of the target object.

[0104] Among them, before fusing the deep content feature information X I of the candidate recommended content and the deep content feature information X s corresponding to the historical interaction content sequence within the first historical time period, X I can be processed repeatedly first to make X I and X s have the same dimension. Specifically, in this embodiment, the Tile function can be used. The function of the Tile function is to repeat a certain array. For example, Tile(A, n), the function is to repeat the array A n times to form a new array. Through the Tile function, X I can be repeated n times along the row direction to obtain the deep content feature information of the candidate recommended content after repeated processing: has the same shape as X s .

[0105] Then, since the short-term sequence behavior has a more significant indication for the next action of the target object, the candidate recommended content can be associated with the items in the short-term sequence. Among them, specifically, (denoted as ) and X s (denoted as R) can be used with two operations of "summation" and "product", as shown in equations (1) and (2):

[0106]

[0107]

[0108] Among them, represents the result of the summation operation, represents the result of the product operation. Then, through the connection (concatenate) layer, After fusing with R (specifically, it can be splicing processing), fused feature information is obtained Then for F s After dimensionality reduction through several fully connected layers (also known as multi-layer perceptron, MLP, Multilayer Perceptron), Since the weighted coefficient of each item in the short-term sequence is not only affected by the feature correlation degree (obtained through attention), but also affected by the time interval between the occurrence time of the item in this sequence and the current time, so the time series feature, that is, the interaction time information T corresponding to the short-term sequence s and S s After normalization operation through the logistic regression layer, the weighted scores of each item in the short-term sequence are obtained That is, the weight information W corresponding to the historical interaction content sequence in the first historical time period in the above embodiment s The calculation process is specifically shown in formula (3):

[0109] W s = Softmax(S s + T s ) (3)

[0110] Among them, the Softmax function can convert the output values of multi-classification into a probability distribution in the range of [0, 1]. After obtaining the weight information W s After that, W can be used s to perform weighted operations on each row of R to obtain the weighted operation result and use the reduce_sum function to perform summation processing to obtain the final output O s , O s That is, the short-term interest feature of the target object. Among them, the function reduce_sum() can be used to sum the elements on the specified dimension in the vector and can reduce the dimension after summation.

[0111] Optionally, in this embodiment, the historical interaction content sequence in the second historical time period includes the historical interaction content sequence under at least one dimension in the second historical time period;

[0112] The step of "performing long-term interest analysis on the target object according to the deep content feature information corresponding to the historical interaction content sequence in the second historical time period and the interaction time information corresponding to the historical interaction content sequence in the second historical time period, and obtaining the long-term interest feature of the target object" may include:

[0113] For the historical interaction content sequences in each dimension during the second historical time period, perform attention processing on the deep content feature information corresponding to the historical interaction content sequences in the dimension to obtain the initial weight information corresponding to the historical interaction content sequences in the dimension;

[0114] Based on the initial weight information and the interaction time information corresponding to the historical interaction content sequences during the second historical time period, process the historical interaction content sequences in the dimension to obtain the sequence feature information of the historical interaction content sequences in the dimension;

[0115] Fuse the sequence feature information of the historical interaction content sequences in each dimension during the second historical time period to obtain the long-term interest feature of the target object.

[0116] Specifically, each historical interaction content may include content information in multiple dimensions. Therefore, the long-term sequence can be divided into long-term sequences in multiple dimensions, and each long-term sequence in a dimension may include each historical interaction content in the corresponding dimension during the second historical time period.

[0117] There are various ways to fuse the sequence feature information of the historical interaction content sequences in each dimension during the second historical time period, and this embodiment does not limit this. For example, the fusion method can be splicing processing or weighted operation, etc.

[0118] Specifically, the deep content feature information corresponding to the historical interaction content sequences during the second historical time period can be denoted as The interaction time information corresponding to the long-term sequence can be denoted as Among them, the input X l Each item of can have the characteristics of K fields. Therefore, these K fields can be separated by the split function to form i = 1,..., K, That is, the deep content feature information corresponding to the historical interaction content sequences in each dimension during the second historical time period, and then perform attention processing on each respectively, so that specific attention can be performed on features of different types (i.e., different dimensions). Among them, the split function can slice the string by specifying the delimiter.

[0119] Optionally, in this embodiment, the step "Based on the initial weight information and the interaction time information corresponding to the historical interaction content sequences during the second historical time period, process the historical interaction content sequences in the dimension to obtain the sequence feature information of the historical interaction content sequences in the dimension" may include:

[0120] Fuse the initial weight information with the interaction time information corresponding to the historical interaction content sequence in the second historical time period to obtain the target weight information corresponding to the historical interaction content sequence in this dimension;

[0121] Based on the target weight information, process the historical interaction content sequence in this dimension to obtain the sequence feature information of the historical interaction content sequence in this dimension.

[0122] Among them, since the weighting coefficient of each item in the long-term sequence is affected not only by the feature correlation degree but also by the time interval between the occurrence time of the item in this sequence and the current time, therefore, the corresponding target weight information can be obtained by fusing the interaction time information. There are various ways to fuse the initial weight information with the interaction time information corresponding to the historical interaction content sequence in the second historical time period, and this embodiment does not limit this. For example, this fusion method can be weighted operation or splicing processing, etc.

[0123] In a specific embodiment, a long-term interest analysis model can be used to perform long-term interest analysis on the target object. Among them, there are various types of long-term interest analysis models, and this embodiment does not limit this. For example, this long-term interest analysis model can be a network structure of Transformer (transformer), and the model result diagram is as Figure 1e shown. The network structure of Transformer is mainly composed of an attention mechanism. It belongs to the structure of multi-head self-attention and can act on long-term sequence features.

[0124] Among them, the specific structure of Attention can be as Figure 1e shown by the "Multi-Head Attention Mechanism (AttentionHeader)" on the right. Take each as the input, and input Use three projection matrices and to project onto the three elements of attention respectively to obtain query key and value above. Their shapes and sizes are all p×D. Then, multiply by the transposed matrix of using the matmul function to obtain the initial weight matrix That is, the initial weight information corresponding to the historical interaction content sequence in the i-th dimension in the second historical time period. Its calculation process is shown in formula (4):

[0125]

[0126] Attention is a weighting mechanism widely used in both the fields of CV (Computer Vision) and NLP (Natural Language Processing). It consists of key, query, and value. Among them, key (key vector) refers to the reference feature, usually the item feature; while query (query vector) and value (value vector) are both obtained by linearly transforming the input sequence features. The initial weight information can be obtained through the inner product of key and query, and the initial weight information is applied to the value, thereby obtaining the weighted feature output. If key, query, and value are all the same input feature, this attention is called "self-attention".

[0127] Among them, the interaction time information can be preprocessed first, and the time series features are subjected to "diff" calculation, which can be understood as l the elements in T 2 subtracting each other pairwise (a total of p elements), and then transformed into a p×p shape through the reshape function to obtain

[0128] as the "time difference" information between long-term sequence items. Among them, diff can compare the similarities and differences of text files line by line. The reshape function is a function that transforms a specified matrix into a matrix with a specific dimension, and it can be used to adjust the number of rows, columns, and dimensions of the matrix.

[0128] Then, can be summed with the initial weight information , and then the sum result is normalized to obtain attention scores (attention scores) that is, the target weight information corresponding to the historical interaction content sequence in the i-th dimension during the second historical time period. Its specific calculation process is shown in formula (5):

[0129]

[0130] Among them, the Softmax function can convert the output values of multi-classification into a probability distribution in the range of [0, 1]. After obtaining the target weight information, and can be multiplied through the Matmul function to obtain the weighted sequence feature information that is, the sequence feature information corresponding to the historical interaction content sequence in the i-th dimension during the second historical time period. Its specific calculation process is shown in formula (6):

[0131]

[0132] Among them, Matmul can be used to return the matrix product of two arrays. After obtaining each dimension, they can be concatenated through a connection layer, and an output consistent with the shape of the input X l can be obtained. Finally, is subjected to a "feed-forward" operation (feed-forward consists of two fully connected layers), and the final output shape is still p×D*K. Then, through the reduce_sum function, a summation operation is performed on the first dimension, and the final output i.e., the long-term interest feature of the target object can be obtained. Among them, the function reduce_sum() can be used to perform a summation operation on the elements in the specified dimension of the vector and can reduce the dimension after summation.

[0133] 104. Select target recommended content from the candidate recommended content for recommendation according to the deep content feature information, the interest feature, and the deep attribute feature information of the candidate recommended content.

[0134] Optionally, in this embodiment, the step of "selecting target recommended content from the candidate recommended content for recommendation according to the deep content feature information, the interest feature, and the deep attribute feature information of the candidate recommended content" may include:

[0135] Fuse the deep content feature information, the interest feature, and the deep attribute feature information of the candidate recommended content to obtain deep feature information;

[0136] Perform shallow feature extraction on the object attribute information and the candidate recommended content to obtain the shallow attribute feature information of the object attribute information and the shallow content feature information of the candidate recommended content;

[0137] Fuse the shallow attribute feature information and the shallow content feature information of the candidate recommended content to obtain shallow feature information;

[0138] Based on the deep feature information and the shallow feature information, select target recommended content from the candidate recommended content for recommendation.

[0139] Among them, there are various ways to fuse the deep content feature information, the interest feature, and the deep attribute feature information of the candidate recommended content, and this embodiment does not limit this. For example, this fusion method can be weighted operation or splicing processing, etc. Among them, this interest feature may include long-term interest feature and short-term interest feature.

[0140] Among them, there are also various ways to fuse the shallow attribute feature information and the shallow content feature information of the candidate recommended content, and this embodiment does not limit this. For example, the fusion method can be weighted operation or splicing processing, etc.

[0141] In a specific embodiment, after the short-term sequence passes through DIN Attention, the short-term interest feature with a dimension of D*K can be output As shown in formula (7), it can be mathematically expressed as:

[0142] O s = DIN(X I , X s , T s ) (7)

[0143] After the long-term sequence passes through Transformer, the output long-term interest feature can be as shown in formula (8):

[0144] O l = Transformer(X l , T l ) (8)

[0145] Optionally, in this embodiment, the step of "selecting a target recommended content from the candidate recommended content for recommendation based on the deep feature information and the shallow feature information" may include:

[0146] Fusing the deep feature information and the shallow feature information to obtain target feature information;

[0147] Predicting the recommendation index corresponding to the candidate recommended content according to the target feature information;

[0148] Selecting a target recommended content from the candidate recommended content for recommendation according to the recommendation index.

[0149] Among them, the recommendation index corresponding to the candidate recommended content can be predicted by a fully connected deep neural network (DNN, Deep Neual Networks), or can be predicted by a support vector machine (SVM, Support Vector Machine), and this embodiment does not limit this.

[0150] Among them, in some embodiments, candidate recommended content with a recommendation index greater than a preset value can be selected as target recommended content and recommended to the target object. In other embodiments, based on the recommendation index, each candidate recommended content can be sorted, for example, sorted from large to small, to obtain the sorted candidate recommended content, and then the first n candidate recommended content in the sorted candidate recommended content can be selected as target recommended content and recommended to the target object.

[0151] Among them, fusion refers to feature fusion. Fusing deep feature information and shallow feature information can improve the representation ability of features. Deep feature information contains more detailed information, but it may contain more noise and lower semanticity; while shallow feature information has stronger semantic information, but it loses more detailed information. By fusing deep feature information and shallow feature information of different scales, the accuracy of content recommendation can be improved.

[0152] In a specific embodiment, the deep feature information can be denoted as quad embedding, and the shallow feature information can be denoted as wide embedding. "Wide embedding" refers to an embedding with a size of 1, representing a shallow embedding expression; while "quad embedding" refers to an embedding with a size of D, representing a deep embedding expression. Among them, embedding can refer to projecting the input discrete features or continuous features into continuous vectors.

[0153] Specifically, the shallow attribute feature information of the object attribute information can be denoted as The shallow content feature information of the candidate recommended content can be denoted as represents the real number field, m represents the m-dimensional attribute information included in the object attribute information, and K represents the K-dimensional content information included in the candidate recommended content.

[0154] Optionally, in this embodiment, the deep feature information includes at least one dimension of deep sub-features;

[0155] The step of "fusing the deep feature information and the shallow feature information to obtain target feature information" may include:

[0156] Performing a cross operation on the deep sub-features of each dimension to obtain the second-order cross feature information corresponding to the deep feature information;

[0157] Performing a fully connected process on the deep feature information to obtain the fully connected feature information corresponding to the deep feature information;

[0158] Fuse the second-order cross feature information, the fully-connected feature information, the deep feature information, and the shallow feature information to obtain target feature information.

[0159] Among them, the cross operation can specifically be the Hadamard product operation, and this embodiment does not limit this.

[0160] Among them, there are various ways to fuse the second-order cross feature information, the fully-connected feature information, the deep feature information, and the shallow feature information, and this embodiment does not limit this. For example, the fusion method can be a weighted operation or a splicing process, etc.

[0161] In a specific embodiment, a content recommendation model can be used to recommend content to a target object, and the overall framework of the content recommendation model can be as Figure 1f shown and is specifically described as follows:

[0162] Among them, the deep feature information can be denoted as It can be obtained by splicing the deep content feature information X I , the short-term interest feature O s , the long-term interest feature O l , and the deep attribute feature information X u through a connection layer, and its calculation method can be as shown in formula (9):

[0163] X A = Concat(X u , X I , O s , O l ) (9)

[0164] Among them, the Concat function represents a splicing process. Since different sequence features are different dimensional representations of object features, this embodiment does not weight the features output by DIN / Transformer (i.e., the short-term interest feature and the long-term interest feature), but treats them as features of multiple fields for splicing.

[0165] Among them, the shallow feature information can be denoted as Z A , which is obtained by splicing the shallow attribute feature information Z u and the shallow content feature information Z I of the candidate recommended content through a connection layer, and its calculation method can be as shown in formula (10):

[0166] Z A = Concat(Z u , Z I ) (10)

[0167] Obtain the deep feature information After that, cross operations can be performed on the deep sub-features in each dimension of the deep feature information. Specifically, can be regarded as composed of the features of H fields, and the feature of the i-th field is denoted as Then, perform Hadamard product pairwise on to obtain second-order cross feature information The calculation process is shown in formula (11):

[0168]

[0169] Among them, the Hadamard product is a type of operation on matrices. It multiplies the elements in the matrices bit by bit and outputs a vector of the same dimension. Therefore, when using the Hadamard product, it is required that the dimensions of the input vectors are consistent.

[0170] In addition, the deep feature information X A can also be processed through a DeepFM Framework. The Deep layer in DeepFM can be composed of l fully connected layers (also known as multi-layer perceptrons, MLP, MultilayerPerceptron). Denote the parameters of the i-th fully connected layer as and The fully connected layer can project a vector to another dimension through a projection matrix, and generally a non-linear function such as Relu, Sigmoid, etc. will be connected behind.

[0171] Among them, the sigmoid function, that is, the S-shaped growth curve, can be used as an activation function in a neural network or in logistic regression processing to map variables to the numerical range from zero to one. The Relu function, that is, the Rectified Linear Unit (ReLU), is a commonly used activation function in artificial neural networks.

[0172] Specifically, the Deep layer can be as shown in formula (12):

[0173]

[0174] Among them, M A represents the fully connected feature information corresponding to the deep feature information.

[0175] Then, the second-order cross feature information I A the fully connected feature information M A the deep feature information X A and the shallow feature information Z AFusion, such as splicing processing, can obtain the final feature output C, which is the target feature information in the above embodiments. Its calculation process is shown in formula (13):

[0176] C = Concat(X A , Z A , M A , I A ) (13)

[0177] Finally, after passing C through a fully connected layer FC (parameters are W C , b C ), project it onto a 1-dimensional scalar value and perform sigmoid conversion to obtain the prediction score Y in the range of [0, 1], as shown in formula (14):

[0178] Y = Sigmoid(FC(C, W C , b C )) (14)

[0179] Among them, the prediction score Y is also the recommendation index corresponding to the candidate recommended content.

[0180] Specifically, the training samples used by this content recommendation model can come from the real log data of the subscription account scenario. The scenarios corresponding to the historical interaction content sequences within the historical time period can include the picture - text scenario and the video scenario, and long - short - term sequences are divided for them. By expanding the sequence features of the object, the behavioral interests of the object can be characterized more accurately.

[0181] This application can divide the historical interaction content sequence of the target object into two parts: "short - term sequence" and "long - term sequence". Considering the role of time sequence, establish attention between the short - term sequence and the candidate recommended content; at the same time, for the long - term sequence, use multi - head self - attention, that is, the transformer network to model its internal correlation to characterize the long - term and stable interests of the target object. Finally, fuse the sequence features of the two parts to obtain rich historical behavior information of the target object, improving the offline AUC (a model evaluation index in the field of machine learning) and the online ctr (click - through rate) index.

[0182] In a specific scenario, such as conducting experiments on the fine - ranking of the subscription account video scenario. In this scenario, there are approximately 15 million log data per hour. Considering the sparse features, 10 - hour data of a certain day can be used as the training set, and then the data of the next hour is used as the validation set. The comparison models adopted in this embodiment are listed as follows, and the feature size of all models is 64:

[0183] 1. The DeepFM model divides long-term and short-term sequences and directly sums the sequence features through reduce sum;

[0184] 2. The pure DIN model (DIN-1) does not divide long-term and short-term sequences and has no time series features;

[0185] 3. The pure DIN model (DIN-2) divides long-term and short-term sequences but has no time series features;

[0186] 4. The DIN-Transformer model (DIN-T) provided by the content recommendation method of the present application divides long-term and short-term sequences and uses time series features.

[0187] The experimental results of the above four models are shown in Table 1:

[0188] Table 1

[0189] Model Auc DeepFM 0.765 DIN-1 0.770 DIN-2 0.774 DIN-T 0.781

[0190] As can be seen from Table 1, compared with DeepFM that directly performs reduce sum on sequence features, the DIN series models that use attention to perform weighted sum on sequences are all better than DeepFM. In addition, the auc of DIN-2 is higher than that of DIN-1, indicating that after dividing long-term and short-term sequences, the characterization of user interests is more accurate, and it also shows that users are different in short-term and long-term interests and should be distinguished. Finally, the DIN-T algorithm proposed in the present application achieves the optimal effect, fully demonstrating the effectiveness of the Transformer structure in modeling long-term interests and also indicating the importance of time series features; combined with the ability of DIN to model short-term interests, DIN-T can more accurately predict user click behavior.

[0191] In addition, DIN-T and the DeepFM model were deployed online for A / B Test (testing), and after continuously observing for 7 days, it was found that the ctr (exposure click-through rate) corresponding to the DIN-T model increased by an average of 4%.

[0192] As can be seen from the above, in this embodiment, the object attribute information of the target object, the historical interaction content sequence in at least one historical time period, and the interaction time information of the historical interaction content sequence can be obtained; deep feature extraction is performed on the object attribute information and the historical interaction content sequence in the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence in the historical time period; based on the deep content feature information and the interaction time information of the historical interaction content sequence, interest analysis is performed on the target object to obtain the interest characteristics of the target object in different time periods; according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information, target recommended content is selected from the candidate recommended content for recommendation. This application can perform content recommendation to the target object based on the interest characteristics of the target object in different time periods, improving the accuracy of content recommendation.

[0193] According to the method described in the previous embodiment, the following will take the specific integration of the content recommendation device in the server as an example for further detailed description.

[0194] An embodiment of the present application provides a content recommendation method, as Figure 2 shown, the specific process of this content recommendation method can be as follows:

[0195] 201. The server obtains the object attribute information of the target object, the historical interaction content sequence in at least one historical time period, and the interaction time information of the historical interaction content sequence.

[0196] Among them, the historical interaction content sequence may include at least one historical interaction content. For example, it may include a plurality of historical interaction contents arranged in the order of interaction time. Among them, each historical interaction content may specifically be the content corresponding to the interaction behavior of the target object at a certain time in the historical time period, and the interacted content may include information in various modalities such as video, audio, text, and image. This embodiment does not limit this. Specifically, for each historical interaction content, the historical interaction content may include at least one dimension of content-related information, or in other words, each historical interaction content includes content-related information in multiple fields. For example, it may include content-related information in dimensions such as the content itself, the publisher of the content, the content category, the content title, and the identification number of the content.

[0197] Among them, the interaction time information of the historical interaction content sequence may include the specific interaction time information corresponding to each historical interaction content in the historical interaction content sequence. Generally speaking, the historical interaction content closer to the current time is more helpful for predicting the next interaction behavior of the target object.

[0198] 202. The server performs deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period, to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period.

[0199] Specifically, compared with shallow feature extraction, the scale of the feature information extracted by deep feature extraction is relatively large. The larger the scale of the feature information, the richer the information it contains.

[0200] 203. The server performs interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence, to obtain the interest characteristics of the target object in different time periods.

[0201] Specifically, in this embodiment, the historical interaction content of the target object in different historical time periods can be separately analyzed to obtain the interest characteristics in different time periods. The historical interaction content in different historical time periods can reflect different interest characteristics of the target object. For example, the historical interaction content closer to the current time has a more significant indicative meaning for predicting the next interaction content of the target object, and the historical interaction content closer to the current time can be called short-term historical interaction content; while the historical interaction content farther from the current time may reflect the long-term interest of the target object, and has no strong hint for predicting the next interaction content of the target object, but reflects the potential and stable interest of the target object. The historical interaction content farther from the current time can be called long-term historical interaction content.

[0202] Optionally, in this embodiment, the historical interaction content sequence within at least one historical time period includes the historical interaction content sequences within the first historical time period and the second historical time period; the second historical time period is earlier than the first historical time period.

[0203] The step of "performing interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence, to obtain the interest characteristics of the target object in different time periods" may include:

[0204] Performing short-term interest analysis on the target object according to the deep content feature information of the candidate recommended content, the deep content feature information corresponding to the historical interaction content sequence within the first historical time period, and the interaction time information corresponding to the historical interaction content sequence within the first historical time period, to obtain the short-term interest characteristics of the target object;

[0205] Performing long-term interest analysis on the target object based on the deep content feature information corresponding to the historical interaction content sequence within the second historical time period and the interaction time information corresponding to the historical interaction content sequence within the second historical time period to obtain the long-term interest characteristics of the target object.

[0206] Specifically, the time length of the first historical time period can be less than that of the second historical time period.

[0207] Optionally, in this embodiment, the step of "performing short-term interest analysis on the target object according to the deep content feature information of the candidate recommended content, the deep content feature information corresponding to the historical interaction content sequence within the first historical time period, and the interaction time information corresponding to the historical interaction content sequence within the first historical time period to obtain the short-term interest characteristics of the target object" may include:

[0208] Fusing the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information;

[0209] Performing a normalization operation on the fused feature information and the interaction time information corresponding to the historical interaction content sequence within the first historical time period to obtain the weight information corresponding to the historical interaction content sequence within the first historical time period;

[0210] Fusing each historical interaction content of the historical interaction content sequence within the first historical time period according to the weight information to obtain the short-term interest characteristics of the target object.

[0211] Specifically, the weight information may include the weights corresponding to each historical interaction content of the historical interaction content sequence within the first historical time period. Based on this weight information, the historical interaction content of the historical interaction content sequence within the first historical time period can be weighted to obtain the short-term interest characteristics of the target object.

[0212] Optionally, in this embodiment, the step of "fusing the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information" may include:

[0213] Performing a summation operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain the result of the summation operation;

[0214] Perform a product operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain a product operation result;

[0215] Fuse the sum operation result, the product operation result, the deep content feature information of the candidate recommended content, and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information.

[0216] Optionally, in this embodiment, the historical interaction content sequence within the second historical time period includes the historical interaction content sequences in at least one dimension within the second historical time period;

[0217] The step of "performing long-term interest analysis on the target object according to the deep content feature information corresponding to the historical interaction content sequence within the second historical time period and the interaction time information corresponding to the historical interaction content sequence within the second historical time period to obtain the long-term interest feature of the target object" may include:

[0218] For the historical interaction content sequences in each dimension within the second historical time period, perform attention processing on the deep content feature information corresponding to the historical interaction content sequence in the dimension to obtain the initial weight information corresponding to the historical interaction content sequence in the dimension;

[0219] Based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence within the second historical time period, process the historical interaction content sequence in the dimension to obtain the sequence feature information of the historical interaction content sequence in the dimension;

[0220] Fuse the sequence feature information of the historical interaction content sequences in each dimension within the second historical time period to obtain the long-term interest feature of the target object.

[0221] Specifically, each historical interaction content may include content information in multiple dimensions. Therefore, the long-term sequence can be divided into long-term sequences in multiple dimensions, and each long-term sequence in a dimension may include each historical interaction content in the corresponding dimension within the second historical time period.

[0222] Among them, there are various ways to fuse the sequence feature information of the historical interaction content sequences in each dimension within the second historical time period, and this embodiment does not limit this. For example, this fusion method can be splicing processing or weighted operation, etc.

[0223] Optionally, in this embodiment, the step of "processing the historical interaction content sequence in the dimension based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period to obtain the sequence feature information of the historical interaction content sequence in the dimension" may include:

[0224] Fuse the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period to obtain the target weight information corresponding to the historical interaction content sequence in the dimension;

[0225] Based on the target weight information, process the historical interaction content sequence in the dimension to obtain the sequence feature information of the historical interaction content sequence in the dimension.

[0226] Among them, since the weighting coefficient of each item in the long-term sequence is affected not only by the feature correlation degree but also by the time interval between the occurrence time of the item in the sequence and the current time, therefore, the corresponding target weight information can be obtained by fusing the interaction time information. There are various ways to fuse the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period, and this embodiment does not limit this. For example, the fusion method can be weighted operation or splicing processing, etc.

[0227] 204. The server fuses the deep content feature information, the interest feature, and the deep attribute feature information of the candidate recommended content to obtain deep feature information.

[0228] Among them, there are various ways to fuse the deep content feature information, the interest feature, and the deep attribute feature information of the candidate recommended content, and this embodiment does not limit this. For example, the fusion method can be weighted operation or splicing processing, etc. Among them, the interest feature may include long-term interest feature and short-term interest feature.

[0229] 205. The server performs shallow feature extraction on the object attribute information and the candidate recommended content to obtain the shallow attribute feature information of the object attribute information and the shallow content feature information of the candidate recommended content.

[0230] 206. The server fuses the shallow attribute feature information and the shallow content feature information of the candidate recommended content to obtain shallow feature information.

[0231] Among them, there are also various ways to fuse the shallow attribute feature information and the shallow content feature information of the candidate recommended content, and this embodiment does not limit this. For example, the fusion method can be weighted operation or splicing processing, etc.

[0232] 207. The server selects target recommended content from the candidate recommended content for recommendation based on the deep feature information and the shallow feature information.

[0233] Optionally, in this embodiment, the step of "selecting target recommended content from the candidate recommended content for recommendation based on the deep feature information and the shallow feature information" may include:

[0234] Fuse the deep feature information and the shallow feature information to obtain target feature information;

[0235] Predict the recommendation index corresponding to the candidate recommended content according to the target feature information;

[0236] Select target recommended content from the candidate recommended content for recommendation according to the recommendation index.

[0237] Among them, fusion refers to feature fusion. Fusing deep feature information and shallow feature information can improve the representation ability of features. Deep feature information contains more detailed information, but it may contain more noise and lower semanticity; while shallow feature information has stronger semantic information, but it loses more detailed information. By fusing deep feature information and shallow feature information at different scales, the accuracy of content recommendation can be improved.

[0238] Among them, in some embodiments, candidate recommended content with a recommendation index greater than a preset value can be selected as the target recommended content and recommended to the target object. In other embodiments, based on the recommendation index, each candidate recommended content can be sorted, such as sorting from large to small, to obtain the sorted candidate recommended content, and then the first n candidate recommended content in the sorted candidate recommended content can be selected as the target recommended content and recommended to the target object.

[0239] As can be seen from the above, in this embodiment, the object attribute information of the target object, the historical interaction content sequence in at least one historical time period, and the interaction time information of the historical interaction content sequence can be obtained through the server; deep feature extraction is performed on the object attribute information and the historical interaction content sequence in the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence in the historical time period; based on the deep content feature information and the interaction time information of the historical interaction content sequence, interest analysis is performed on the target object to obtain the interest characteristics of the target object in different time periods; the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information are fused to obtain deep feature information; shallow feature extraction is performed on the object attribute information and the candidate recommended content to obtain the shallow attribute feature information of the object attribute information and the shallow content feature information of the candidate recommended content; the shallow attribute feature information and the shallow content feature information of the candidate recommended content are fused to obtain shallow feature information; based on the deep feature information and the shallow feature information, the target recommended content is selected from the candidate recommended content for recommendation. The present application can perform content recommendation to the target object based on the interest characteristics of the target object in different time periods, improving the accuracy of content recommendation.

[0240] To better implement the above method, an embodiment of the present application further provides a content recommendation device, as Figure 3 shown. The content recommendation device may include an acquisition unit 301, an extraction unit 302, an analysis unit 303, and a selection unit 304, as follows:

[0241] (1) Acquisition unit 301;

[0242] The acquisition unit is configured to acquire the object attribute information of the target object, the historical interaction content sequence in at least one historical time period, and the interaction time information of the historical interaction content sequence.

[0243] (2) Extraction unit 302;

[0244] The extraction unit is configured to perform deep feature extraction on the object attribute information and the historical interaction content sequence in the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence in the historical time period.

[0245] (3) Analysis unit 303;

[0246] An analysis unit for performing interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence, so as to obtain the interest characteristics of the target object in different time periods.

[0247] Optionally, in some embodiments of the present application, the historical interaction content sequence within at least one historical time period includes the historical interaction content sequences within the first historical time period and the second historical time period; the second historical time period is earlier than the first historical time period;

[0248] The analysis unit may include a short-term interest analysis subunit and a long-term interest analysis subunit, as follows:

[0249] The short-term interest analysis subunit is configured to perform short-term interest analysis on the target object according to the deep content feature information of the candidate recommended content, the deep content feature information corresponding to the historical interaction content sequence within the first historical time period, and the interaction time information corresponding to the historical interaction content sequence within the first historical time period, so as to obtain the short-term interest characteristics of the target object;

[0250] The long-term interest analysis subunit is configured to perform long-term interest analysis on the target object according to the deep content feature information corresponding to the historical interaction content sequence within the second historical time period and the interaction time information corresponding to the historical interaction content sequence within the second historical time period, so as to obtain the long-term interest characteristics of the target object.

[0251] Optionally, in some embodiments of the present application, the short-term interest analysis subunit may specifically be configured to fuse the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information; perform a normalization operation on the fused feature information and the interaction time information corresponding to the historical interaction content sequence within the first historical time period to obtain the weight information corresponding to the historical interaction content sequence within the first historical time period; and fuse each historical interaction content of the historical interaction content sequence within the first historical time period according to the weight information to obtain the short-term interest characteristics of the target object.

[0252] Optionally, in some embodiments of the present application, the step of "fusing the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information" may include:

[0253] Performing a summation operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain a summation operation result;

[0254] Perform a product operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain a product operation result;

[0255] Fuse the summation operation result, the product operation result, the deep content feature information of the candidate recommended content, and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information.

[0256] Optionally, in some embodiments of the present application, the historical interaction content sequence within the second historical time period includes historical interaction content sequences in at least one dimension within the second historical time period;

[0257] The long-term interest analysis subunit may specifically be used to perform attention processing on the deep content feature information corresponding to the historical interaction content sequence in each dimension within the second historical time period to obtain initial weight information corresponding to the historical interaction content sequence in the dimension; based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence within the second historical time period, process the historical interaction content sequence in the dimension to obtain sequence feature information of the historical interaction content sequence in the dimension; fuse the sequence feature information of the historical interaction content sequences in each dimension within the second historical time period to obtain the long-term interest feature of the target object.

[0258] Optionally, in some embodiments of the present application, the step of "processing the historical interaction content sequence in the dimension based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence within the second historical time period to obtain sequence feature information of the historical interaction content sequence in the dimension" may include:

[0259] Fuse the initial weight information and the interaction time information corresponding to the historical interaction content sequence within the second historical time period to obtain target weight information corresponding to the historical interaction content sequence in the dimension;

[0260] Process the historical interaction content sequence in the dimension based on the target weight information to obtain sequence feature information of the historical interaction content sequence in the dimension.

[0261] (4) Selection unit 304;

[0262] The selection unit is used to select target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest feature, and the deep attribute feature information.

[0263] Optionally, in some embodiments of the present application, the selection unit may include a first fusion subunit, a shallow feature extraction subunit, a second fusion subunit, and a selection subunit, as follows:

[0264] The first fusion subunit is configured to fuse the deep content feature information of the candidate recommended content, the interest feature, and the deep attribute feature information to obtain deep feature information;

[0265] The shallow feature extraction subunit is configured to perform shallow feature extraction on the object attribute information and the candidate recommended content to obtain the shallow attribute feature information of the object attribute information and the shallow content feature information of the candidate recommended content;

[0266] The second fusion subunit is configured to fuse the shallow attribute feature information and the shallow content feature information of the candidate recommended content to obtain shallow feature information;

[0267] The selection subunit is configured to select target recommended content from the candidate recommended content for recommendation based on the deep feature information and the shallow feature information.

[0268] Optionally, in some embodiments of the present application, the selection subunit may specifically be configured to fuse the deep feature information and the shallow feature information to obtain target feature information; predict a recommendation index corresponding to the candidate recommended content according to the target feature information; and select target recommended content from the candidate recommended content for recommendation according to the recommendation index.

[0269] Optionally, in some embodiments of the present application, the deep feature information includes deep sub-features of at least one dimension;

[0270] The step of "fusing the deep feature information and the shallow feature information to obtain target feature information" may include:

[0271] Performing a cross operation on the deep sub-features of each dimension to obtain second-order cross feature information corresponding to the deep feature information;

[0272] Performing a fully connected process on the deep feature information to obtain fully connected feature information corresponding to the deep feature information;

[0273] Fusing the second-order cross feature information, the fully connected feature information, the deep feature information, and the shallow feature information to obtain target feature information.

[0274] As can be seen from the above, in this embodiment, the acquisition unit 301 can acquire the object attribute information of the target object, the historical interaction content sequence in at least one historical time period, and the interaction time information of the historical interaction content sequence; the extraction unit 302 can perform deep feature extraction on the object attribute information and the historical interaction content sequence in the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence in the historical time period; the analysis unit 303 can perform interest analysis on the target object based on the deep content feature information and the interaction time information of the historical interaction content sequence to obtain the interest features of the target object in different time periods; the selection unit 304 can select target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest features, and the deep attribute feature information. This application can perform content recommendation to the target object based on the interest features of the target object in different time periods, improving the accuracy of content recommendation.

[0275] An embodiment of this application also provides an electronic device, such as Figure 4 shown, which shows a schematic structural diagram of the electronic device involved in the embodiment of this application. The electronic device can be a terminal or a server, etc. Specifically:

[0276] The electronic device can include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art can understand that Figure 4 the structural diagram of the electronic device shown in

[0277] does not constitute a limitation on the electronic device. It can include more or fewer components than shown, or combine certain components, or have different component arrangements. Among them:

[0278] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the electronic device. In addition, the memory 402 can include high-speed random access memory and can also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices. Correspondingly, the memory 402 can also include a memory controller to provide the processor 401 with access to the memory 402.

[0279] The electronic device further includes a power supply 403 for powering each component. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to implement functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 403 can also include any components such as one or more DC or AC power supplies, a recharge system, a power failure detection circuit, a power converter or inverter, and a power status indicator.

[0280] The electronic device may further include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.

[0281] Although not shown, the electronic device may further include a display unit, etc., which will not be elaborated here. Specifically, in this embodiment, the processor 401 in the electronic device will load the executable files corresponding to the processes of one or more application programs into the memory 402 according to the following instructions, and the processor 401 will run the application programs stored in the memory 402 to implement various functions as follows:

[0282] Obtain the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence; perform deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period; based on the deep content feature information and the interaction time information of the historical interaction content sequence, perform interest analysis on the target object to obtain the interest characteristics of the target object at different time periods; select target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information.

[0283] For the specific implementation of each of the above operations, reference may be made to the previous embodiments, which will not be elaborated herein.

[0284] As can be seen from the above, in this embodiment, the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence can be obtained; deep feature extraction is performed on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period; based on the deep content feature information and the interaction time information of the historical interaction content sequence, interest analysis is performed on the target object to obtain the interest characteristics of the target object at different time periods; according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information, target recommended content is selected from the candidate recommended content for recommendation. This application can perform content recommendation to the target object based on the interest characteristics of the target object at different time periods, improving the accuracy of content recommendation.

[0285] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions or by controlling related hardware through instructions. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0286] Therefore, an embodiment of this application provides a computer-readable storage medium, which stores multiple instructions that can be loaded by a processor to execute the steps in any content recommendation method provided by the embodiment of this application. For example, the instructions can execute the following steps:

[0287] Obtain the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence; perform deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period; based on the deep content feature information and the interaction time information of the historical interaction content sequence, perform interest analysis on the target object to obtain the interest characteristics of the target object in different time periods; select target recommended content from the candidate recommended content for recommendation according to the deep content feature information of the candidate recommended content, the interest characteristics, and the deep attribute feature information.

[0288] For the specific implementation of each of the above operations, reference may be made to the previous embodiments and will not be elaborated here.

[0289] Among them, the computer-readable storage medium may include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, etc.

[0290] Since the instructions stored in the computer-readable storage medium can execute the steps in any of the content recommendation methods provided in the embodiments of the present application, the beneficial effects that can be achieved by any of the content recommendation methods provided in the embodiments of the present application can be realized. For details, reference may be made to the previous embodiments and will not be elaborated here.

[0291] According to one aspect of the present application, there is provided a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in various alternative implementation manners of the above content recommendation aspect.

[0292] The above has introduced in detail a content recommendation method and related devices provided in the embodiments of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A content recommendation method, characterized in that, Including: Obtain the object attribute information of the target object, the historical interaction content sequence within at least one historical time period, and the interaction time information of the historical interaction content sequence. The historical interaction content sequence within at least one historical time period includes the historical interaction content sequence within the first historical time period and the historical interaction content sequence within the second historical time period; the second historical time period is earlier than the first historical time period. The historical interaction content sequence includes at least one historical interaction content, and each historical interaction content includes content information in multiple dimensions. The historical interaction content sequence within the second historical time period includes the historical interaction content sequence in multiple dimensions within the second historical time period, and the historical interaction content sequence in each dimension within the second historical time period includes each historical interaction content in the corresponding dimension within the second historical time period; Perform deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period to obtain the deep attribute feature information of the object attribute information and the deep content feature information of the historical interaction content sequence within the historical time period; Fuse the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information; Perform a normalization operation on the fused feature information and the interaction time information corresponding to the historical interaction content sequence within the first historical time period to obtain the weight information corresponding to the historical interaction content sequence within the first historical time period. The interaction time information includes the interaction time of each historical interaction content in the historical interaction content sequence within the first historical time period, and the weight information includes the weight corresponding to each historical interaction content in the historical interaction content sequence within the first historical time period; According to the weight information, fuse each historical interaction content in the historical interaction content sequence within the first historical time period to obtain the short-term interest feature of the target object; For the historical interaction content sequence in each dimension within the second historical time period, perform attention processing on the deep content feature information corresponding to the historical interaction content sequence in the dimension to obtain the initial weight information corresponding to the historical interaction content sequence in the dimension; Based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence within the second historical time period, process the historical interaction content sequence in the dimension to obtain the sequence feature information of the historical interaction content sequence in the dimension; Fuse the sequence feature information of the historical interaction content sequence in each dimension within the second historical time period to obtain the long-term interest feature of the target object; Fuse the deep content feature information of the candidate recommended content, the short-term interest feature, the long-term interest feature, and the deep attribute feature information to obtain deep feature information; Perform shallow feature extraction on the object attribute information and the candidate recommended content to obtain the shallow attribute feature information of the object attribute information and the shallow content feature information of the candidate recommended content; Fuse the shallow attribute feature information and the shallow content feature information of the candidate recommended content to obtain shallow feature information; Fuse the deep feature information and the shallow feature information to obtain target feature information; Predict the recommendation index corresponding to the candidate recommended content according to the target feature information; Select target recommended content from the candidate recommended content for recommendation according to the recommendation index.

2. The method according to claim 1, characterized in that, The deep feature information includes deep sub-features in at least one dimension; The fusing the deep feature information and the shallow feature information to obtain target feature information includes: Perform cross-operation on the deep sub-features in each dimension to obtain second-order cross-feature information corresponding to the deep feature information; Perform fully-connected processing on the deep feature information to obtain fully-connected feature information corresponding to the deep feature information; Fuse the second-order cross-feature information, the fully-connected feature information, the deep feature information and the shallow feature information to obtain target feature information.

3. The method according to claim 1, characterized in that, The fusing the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain fusion feature information includes: Perform a summation operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain a summation operation result; Perform a product operation on the deep content feature information of the candidate recommended content and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain a product operation result; Fuse the summation operation result, the product operation result, the deep content feature information of the candidate recommended content, and the deep content feature information corresponding to the historical interaction content sequence in the first historical time period to obtain fusion feature information.

4. The method according to claim 1, characterized in that, The processing the historical interaction content sequence in the dimension based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period to obtain the sequence feature information of the historical interaction content sequence in the dimension includes: Fuse the initial weight information and the interaction time information corresponding to the historical interaction content sequence in the second historical time period to obtain target weight information corresponding to the historical interaction content sequence in the dimension; Process the historical interaction content sequence in the dimension based on the target weight information to obtain the sequence feature information of the historical interaction content sequence in the dimension.

5. A content recommendation device, characterized in that, Include: An acquisition unit for acquiring object attribute information of a target object, a historical interaction content sequence within at least one historical time period, and interaction time information of the historical interaction content sequence, where the historical interaction content sequence within at least one historical time period includes historical interaction content sequences within a first historical time period and a second historical time period; the second historical time period is earlier than the first historical time period, the historical interaction content sequence includes at least one historical interaction content, each historical interaction content includes content information in multiple dimensions, the historical interaction content sequence within the second historical time period includes historical interaction content sequences in multiple dimensions within the second historical time period, and the historical interaction content sequence in each dimension within the second historical time period includes each historical interaction content in the corresponding dimension within the second historical time period; An extraction unit for performing deep feature extraction on the object attribute information and the historical interaction content sequence within the historical time period to obtain deep attribute feature information of the object attribute information and deep content feature information of the historical interaction content sequence within the historical time period; An analysis unit for fusing deep content feature information of candidate recommended content and deep content feature information corresponding to the historical interaction content sequence within the first historical time period to obtain fused feature information; performing a normalization operation on the fused feature information and interaction time information corresponding to the historical interaction content sequence within the first historical time period to obtain weight information corresponding to the historical interaction content sequence within the first historical time period, where the interaction time information includes the interaction time of each historical interaction content in the historical interaction content sequence within the first historical time period, and the weight information includes the weights corresponding to each historical interaction content in the historical interaction content sequence within the first historical time period; According to the weight information, fusing each historical interaction content in the historical interaction content sequence within the first historical time period to obtain short-term interest features of the target object; for the historical interaction content sequences in each dimension within the second historical time period, performing attention processing on the deep content feature information corresponding to the historical interaction content sequence in the dimension to obtain initial weight information corresponding to the historical interaction content sequence in the dimension; based on the initial weight information and the interaction time information corresponding to the historical interaction content sequence within the second historical time period, processing the historical interaction content sequence in the dimension to obtain sequence feature information of the historical interaction content sequence in the dimension; Fusing the sequence feature information of the historical interaction content sequences in each dimension within the second historical time period to obtain long-term interest features of the target object; A selection unit for fusing deep content feature information of candidate recommended content, the short-term interest features, the long-term interest features, and the deep attribute feature information to obtain deep feature information; Perform shallow feature extraction on the object attribute information and the candidate recommended content to obtain the shallow attribute feature information of the object attribute information and the shallow content feature information of the candidate recommended content; Fuse the shallow attribute feature information and the shallow content feature information of the candidate recommended content to obtain shallow feature information; Fuse the deep feature information and the shallow feature information to obtain target feature information; Predict the recommendation index corresponding to the candidate recommended content according to the target feature information; select the target recommended content from the candidate recommended content for recommendation according to the recommendation index.

6. An electronic device, characterized in that, Comprising a memory and a processor; the memory stores an application program, and the processor is configured to run the application program in the memory to execute the operations in the content recommendation method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores multiple instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the content recommendation method according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instruction is executed by a processor, the steps in the content recommendation method according to any one of claims 1 to 4 are implemented.

Images