A content recommendation method, related device and equipment based on a deep learning model

Abstract

The present application discloses a content recommendation method and related device based on a deep learning model, which can be applied to various scenarios such as cloud technology, artificial intelligence, intelligent transportation, and assisted driving. The present application includes: obtaining a comprehensive feature vector and an activity feature vector; based on the comprehensive feature vector, obtaining N expert vectors through N expert networks; based on the activity feature vector, obtaining an activity weight vector through an activity gating network; multiplying the elements of the expert vectors by the activity weight vector to obtain weighted expert vectors; based on the weighted expert vectors, obtaining task scores of candidate content through a prediction network; if it is determined according to the K task scores that the candidate content belongs to the content to be pushed, pushing the candidate content to the recommendation object. The present application enables the deep learning model to better learn the data distribution differences of objects with different activity levels, thereby improving the recommendation accuracy and efficiency of the recommendation system for objects with different activity levels.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular, to a content recommendation method, related device, and equipment based on a deep learning model. Background Art

[0002] With the development of Internet technology and the increasing scale of network data, the needs of users have become more and more diverse and personalized. Recommendation systems have become a common solution for the public when facing the problem of screening massive Internet information. By reasonably controlling the content presented to users through content recommendation, the content value can be better realized.

[0003] Currently, a general deep learning model is usually obtained by uniformly modeling all recommendation objects. In the content recommendation stage, the recommendation object features and candidate content features are used as the input of the deep learning model. The deep learning model outputs the score of each candidate content, and relevant candidate content is pushed to the recommendation object based on the score.

[0004] However, the inventor found that there are at least the following problems in the existing solutions: The sample size of highly active objects is often much larger than that of low-active objects, so that the richness of sample features of highly active objects is also significantly higher than that of low-active objects. Therefore, the deep learning model is easily dominated by the sample features of highly active objects, resulting in the recommendation system not being able to serve low-active objects well. Summary of the Invention

[0005] The embodiments of this application provide a content recommendation method, related device, and equipment based on a deep learning model. This application enables the deep learning model to better learn the data distribution differences of objects with different activity levels, thereby improving the recommendation accuracy and efficiency of the recommendation system for objects with different activity levels.

[0006] In view of this, on the one hand, this application provides a content recommendation method based on a deep learning model. The deep learning model includes an activity gating network and K task networks. Each task network includes N expert networks and a prediction network. K≥1, N≥1, and both K and N are integers. The content recommendation method includes:

[0007] In response to a content recommendation request for a recommendation object, obtain a comprehensive feature vector and an activity feature vector. The comprehensive feature vector includes the feature vector of the recommendation object and the feature vector of the candidate content, and the activity feature vector characterizes the activity of the recommendation object in the historical period;

[0008] For each task network, based on the comprehensive feature vector, obtain N expert vectors through N expert networks;

[0009] Based on the activity feature vector, an activity weight vector is obtained through an activity gating network;

[0010] For each task network, at least one expert vector among the N expert vectors is element-wise multiplied by the activity weight vector to obtain at least one weighted expert vector;

[0011] For each task network, based on at least one weighted expert vector, a task score of the candidate content is obtained through a prediction network;

[0012] If it is determined according to the K task scores that the candidate content belongs to the content to be pushed, the candidate content is pushed to the recommended object, where the K task scores are the output results of the K task networks.

[0013] On the other hand, the present application provides a content recommendation device based on a deep learning model. The deep learning model includes an activity gating network and K task networks. Each task network includes N expert networks and a prediction network. K≥1, N≥1, and both K and N are integers. The content recommendation device includes:

[0014] An acquisition module, configured to respond to a content recommendation request for a recommended object, and acquire a comprehensive feature vector and an activity feature vector, where the comprehensive feature vector includes the feature vector of the recommended object and the feature vector of the candidate content, and the activity feature vector characterizes the activity of the recommended object in a historical period;

[0015] The acquisition module is further configured to, for each task network, acquire N expert vectors through the N expert networks based on the comprehensive feature vector;

[0016] The acquisition module is further configured to, based on the activity feature vector, obtain an activity weight vector through the activity gating network;

[0017] The acquisition module is further configured to, for each task network, element-wise multiply at least one expert vector among the N expert vectors by the activity weight vector to obtain at least one weighted expert vector;

[0018] The acquisition module is further configured to, for each task network, obtain a task score of the candidate content through the prediction network based on at least one weighted expert vector;

[0019] A recommendation module, configured to push the candidate content to the recommended object if it is determined according to the K task scores that the candidate content belongs to the content to be pushed, where the K task scores are the output results of the K task networks.

[0020] In a possible design, in another implementation manner of the other aspect of the embodiments of the present application,

[0021] An acquisition module, specifically configured to respond to a content recommendation request for a recommended object, and acquire object data of the recommended object, where the object data includes object activity data and object association data, and the object association data includes one or more of object basic data, object consumption sequence data, and object description data;

[0022] Acquire candidate content from the candidate content set corresponding to the recommended object, where the candidate content set includes Q candidate contents, Q≥1, and Q is an integer;

[0023] Acquire content data of the candidate content, where the content data includes one or more of content basic data and content statistical data;

[0024] Perform feature extraction on the object association data and the content data to obtain a comprehensive feature vector;

[0025] Perform feature extraction on the object activity data to obtain an activity feature vector.

[0026] In a possible design, in another implementation manner of another aspect of the embodiments of the present application, the deep learning model further includes a feature weighting network;

[0027] The acquisition module is further configured to, after acquiring the comprehensive feature vector and the activity feature vector in response to a content recommendation request for a recommended object, acquire a target comprehensive feature vector based on the comprehensive feature vector through the feature weighting network;

[0028] The acquisition module is specifically configured to, for each task network, acquire N expert vectors through N expert networks based on the target comprehensive feature vector.

[0029] In a possible design, in another implementation manner of another aspect of the embodiments of the present application,

[0030] The acquisition module is specifically configured to perform pooling processing on the feature vectors of each feature domain in the comprehensive feature vector to obtain a compressed feature vector, where the comprehensive feature vector includes feature vectors of T feature domains, the compressed feature vector includes T elements, T>1, and T is an integer;

[0031] Based on the compressed feature vector, acquire a feature weight vector through an activation network, where the feature weight vector includes T elements;

[0032] Multiply the feature vectors of the T feature domains element by element using the feature weight vector to obtain a target comprehensive feature vector.

[0033] In a possible design, in another implementation manner of another aspect of the embodiments of the present application, the feature weighting network includes X gating networks and X multi-layer perceptrons, X≥1, and X is an integer;

[0034] An acquisition module, specifically configured to acquire screened feature vectors corresponding to the screened data, where the screened data is data screened from object-related data and content data;

[0035] Based on the screened feature vectors, obtain X gated feature vectors through X gated networks, where the input of each gated network is the screened feature vector;

[0036] Based on the comprehensive feature vector and the X gated feature vectors, obtain the target comprehensive feature vector through X multi-layer perceptrons, where the input of each multi-layer perceptron is different.

[0037] In a possible design, in another implementation manner on the other hand of the embodiment of the present application,

[0038] An acquisition module, specifically configured to, for each task network, multiply each weight value in the activity weight vector by the corresponding expert vector among the N expert vectors element by element to obtain N weighted expert vectors, where the activity weight vector includes N weight values, or the activity weight vector and the expert vector have the same dimension.

[0039] In a possible design, in another implementation manner on the other hand of the embodiment of the present application,

[0040] An acquisition module, specifically configured to obtain an activity weight vector based on the N expert vectors and the activity feature vector through an activity gated network, where the activity gated network is a network using an attention mechanism, and the activity weight vector includes N weight values;

[0041] An acquisition module, specifically configured to, for each task network, multiply each weight value in the activity weight vector by the corresponding expert vectors among the N expert vectors element by element to obtain N weighted expert vectors.

[0042] In a possible design, in another implementation manner on the other hand of the embodiment of the present application, the content recommendation device based on the deep learning model further includes a determination module;

[0043] The determination module is configured to, after obtaining the activity weight vector based on the activity feature vector through the activity gated network, for each task network, determine the largest P weight values from the activity weight vector, where the activity weight vector includes N weight values, each weight value corresponds to an expert network, N > P ≥ 1, and P is an integer;

[0044] For each task network, determine the corresponding P expert networks from the N expert networks according to the P weight values;

[0045] An acquisition module, specifically configured to, for each task network, multiply each of the P weight values with the corresponding expert vectors among the P expert vectors element by element to obtain P weighted expert vectors, where the P expert vectors are the output results of P expert networks.

[0046] In a possible design, in another implementation manner of another aspect of the embodiments of the present application, K>1, and the N expert networks include a shared expert network and exclusive expert networks, and each task network further includes a task gating network;

[0047] An acquisition module, specifically configured to, for each task network, obtain a task feature vector through the task gating network based on the comprehensive feature vector;

[0048] For each task network, multiply the task feature vector with at least one weighted expert vector element by element to obtain a target feature vector;

[0049] For each task network, obtain the task score of the candidate content through the prediction network based on the target feature vector.

[0050] In a possible design, in another implementation manner of another aspect of the embodiments of the present application, K>1, and the N expert networks are N shared expert networks, and each task network further includes a task gating network;

[0051] An acquisition module, specifically configured to, for each task network, obtain a task feature vector through the task gating network based on the comprehensive feature vector;

[0052] For each task network, multiply the task feature vector with at least one weighted expert vector element by element to obtain a target feature vector;

[0053] For each task network, obtain the task score of the candidate content through the prediction network based on the target feature vector.

[0054] In a possible design, in another implementation manner of another aspect of the embodiments of the present application, K>1, and the N expert networks are N shared expert networks;

[0055] An acquisition module, specifically configured to, for each task network, perform matrix transformation on at least one weighted expert vector through a matrix set to obtain a target feature vector, where the matrix set includes a transformation matrix or an identity matrix;

[0056] For each task network, obtain the task score of the candidate content through the prediction network based on the target feature vector.

[0057] In a possible design, in another implementation manner of another aspect of the embodiments of the present application,

[0058] The determination module is further configured to, after obtaining the task score of the candidate content through the prediction network based on at least one weighted expert vector, determine the sorting score of the candidate content according to the K task scores;

[0059] The acquisition module is further configured to acquire the sorting scores of each candidate content among other candidate contents, where the other candidate contents are from the candidate content set of the recommended object;

[0060] The determination module is further configured to determine the score sorting result of the candidate content according to the sorting score of the candidate content and the sorting scores of each candidate content among other candidate contents;

[0061] The determination module is further configured to determine that the candidate content belongs to the content to be pushed if the score sorting result of the candidate content is within the recommended sorting range.

[0062] In a possible design, in another implementation manner of another aspect of the embodiments of the present application,

[0063] The determination module is specifically configured to, if K = 1, use the K task scores as the sorting score of the candidate content;

[0064] If K > 1, the sorting score of the candidate content is calculated according to the K task scores and the task weights of each task score among the K task scores.

[0065] In a possible design, in another implementation manner of another aspect of the embodiments of the present application, the K task scores include a duration score, an interaction score, a completion rate score, and a fast - slide rate score;

[0066] The determination module is specifically configured to weight the duration score with a duration weight to obtain a first score;

[0067] Weight the interaction score with an interaction weight to obtain a second score;

[0068] Weight the completion rate score with a completion rate weight to obtain a third score;

[0069] Weight the fast - slide rate score with a fast - slide rate weight to obtain a fourth score;

[0070] The sorting score of the candidate content is calculated according to the first score, the second score, the third score, and the fourth score, where the sorting score of the candidate content is positively correlated with the duration score, the interaction score, and the completion rate score, and the sorting score of the candidate content is negatively correlated with the fast - slide rate score.

[0071] In a possible design, in another implementation manner of another aspect of the embodiments of the present application, the content recommendation device based on a deep - learning model further includes a processing module and a training module;

[0072] An acquisition module, which is further configured to acquire a set of data to be trained, where each piece of data to be trained includes object data of a sample object, content data of a sample content, and K task labels;

[0073] A processing module, configured to perform feature extraction on the object data of the sample object and the content data of the sample content for each piece of data to be trained, so as to obtain a sample feature vector of the data to be trained, where the sample feature vector includes a comprehensive sample feature vector and an activity sample feature vector;

[0074] The acquisition module is further configured to obtain K prediction scores for each piece of data to be trained through a deep learning model based on the sample feature vector of each piece of data to be trained;

[0075] A determination module, which is further configured to determine a total loss result according to the K prediction scores of each piece of data to be trained and the corresponding K task labels;

[0076] A training module, configured to update the model parameters of the deep learning model according to the total loss result.

[0077] In a possible design, in another implementation manner of another aspect of this application,

[0078] The determination module is specifically configured to, for each of the K tasks, calculate a task loss result by using a first loss function according to the prediction score of each piece of data to be trained and the corresponding task label;

[0079] If K = 1, then use the task loss result as the total loss result;

[0080] If K > 1, then sum the task loss results of each task to obtain the total loss result;

[0081] Or,

[0082] The determination module is specifically configured to, for each of the K tasks, calculate a task loss result by using a second loss function according to the sample weight, the corresponding prediction score, and the corresponding task label of each piece of data to be trained, where the sample weight is related to the historical behavior data generated by the sample object for the sample content;

[0083] If K = 1, then use the task loss result as the total loss result;

[0084] If K > 1, then sum the task loss results of each task to obtain the total loss result.

[0085] Another aspect of this application provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the methods in the above aspects are implemented.

[0086] Another aspect of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the methods of the above aspects are implemented.

[0087] Another aspect of the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, the methods of the above aspects are implemented.

[0088] As can be seen from the above technical solutions, the embodiments of the present application have the following advantages:

[0089] In the embodiments of the present application, a content recommendation method based on a deep learning model is provided. First, in response to a content recommendation request for a recommended object, a comprehensive feature vector and an activity feature vector are obtained. On the one hand, based on the comprehensive feature vector, N expert vectors can be obtained through N expert networks. On the other hand, based on the activity feature vector, an activity weight vector can be obtained through an activity gating network. Then, the activity weight vector is used to perform element-wise multiplication on the expert vectors among the N expert vectors to obtain weighted expert vectors. Then, based on at least one weighted expert vector, a task score of the candidate content is obtained through a prediction network. In this way, it is determined whether to push the candidate content to the recommended object according to the task scores respectively output by each task network. Through the above method, the historical behavior data of the object is used to construct its corresponding activity feature vector as the input of the activity gating network, and then the activity gating network is used to weight each expert network, so that the deep learning model can better learn the data distribution differences of objects with different activities. Therefore, the recommendation accuracy and efficiency of the recommendation system for objects with different activities are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 It is a schematic diagram of an interface of a content recommendation scenario in the embodiments of the present application;

[0091] Figure 2 It is a schematic diagram of an implementation environment of the content recommendation method in the embodiments of the present application;

[0092] Figure 3 It is a schematic flowchart of the content recommendation method based on a deep learning model in the embodiments of the present application;

[0093] Figure 4 It is a schematic diagram of a structure based on a single-task single-expert network in the embodiments of the present application;

[0094] Figure 5 It is a schematic diagram of a structure based on a multi-task single-expert network in the embodiments of the present application;

[0095] Figure 6It is a schematic structural diagram of a multi-task and multi-expert network in an embodiment of the present application;

[0096] Figure 7 It is a schematic structural diagram of introducing a feature weighting network in an embodiment of the present application;

[0097] Figure 8 It is a schematic diagram of obtaining a target comprehensive feature vector based on a feature weighting network in an embodiment of the present application;

[0098] Figure 9 It is another schematic diagram of obtaining a target comprehensive feature vector based on a feature weighting network in an embodiment of the present application

[0099] Figure 10 It is a schematic diagram of realizing feature weighting based on an activity gating network in an embodiment of the present application;

[0100] Figure 11 It is another schematic diagram of realizing feature weighting based on an activity gating network in an embodiment of the present application;

[0101] Figure 12 It is another schematic diagram of realizing feature weighting based on an activity gating network in an embodiment of the present application;

[0102] Figure 13 It is a schematic structural diagram of a deep learning model in an embodiment of the present application;

[0103] Figure 14 It is another schematic structural diagram of a deep learning model in an embodiment of the present application;

[0104] Figure 15 It is another schematic structural diagram of a deep learning model in an embodiment of the present application;

[0105] Figure 16 It is a schematic flowchart of applying a deep learning model in an embodiment of the present application;

[0106] Figure 17 It is a schematic structural diagram of a multi-task deep learning model in an embodiment of the present application;

[0107] Figure 18 It is a schematic flowchart of training a deep learning model in an embodiment of the present application;

[0108] Figure 19 It is a schematic diagram of a content recommendation device based on a deep learning model in an embodiment of the present application;

[0109] Figure 20 It is a schematic structural diagram of a computer device in an embodiment of the present application. Detailed implementation manners

[0110] Embodiments of the present application provide a content recommendation method, related apparatus, and device based on a deep learning model. The present application enables the deep learning model to better learn the data distribution differences of objects with different activity levels, thereby improving the recommendation accuracy and efficiency of the recommendation system for objects with different activity levels.

[0111] The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "corresponding to" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily limit to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0112] The rapid growth of the Internet scale and coverage has brought about the problem of information overload. The presentation of excessive information makes users pay too high a price to search for useful information for themselves, thereby reducing the efficiency of information use. As an important means of information filtering, personalized recommendation is a very promising method to solve the problem of information overload currently. Personalized recommendation recommends content (such as pictures and texts, short videos, goods, etc.) that the object is interested in according to the interest characteristics of the object through machine learning (ML) algorithms based on artificial intelligence (AI). Among them, ML is a multi-disciplinary intersection discipline, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve its own performance. ML is the core of AI and the fundamental way to make computers intelligent, and its applications cover all fields of AI. ML and deep learning generally include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, and inductive learning.

[0113] The content recommendation scenario will be introduced below in conjunction with the drawings. For ease of understanding, please refer to Figure 1 , Figure 1 which is a schematic diagram of an interface for the content recommendation scenario in the embodiments of the present application. As shown in Figure 1As shown in Figure (A), a bottom navigation bar indicated by A1 is provided on the message page of an application (APP). Among them, a content recommendation control indicated by A2 is also displayed on the bottom navigation bar. After clicking on the content recommendation control, the user enters the main page as shown in Figure (B) of Figure 1, and the default immersive video mode can be experienced through the main page. This video mode includes short videos recommended to the user by a personalized recommendation algorithm. Each time a video is exposed, the user can perform operations such as swiping, liking, commenting, coin-voting, and forwarding according to their own interests.

[0114] In a recommendation system, the data distributions of objects with different levels of activity vary greatly, mainly reflected in sample differences, feature differences, and behavior differences. Compared with low-active objects, high-active objects have a much larger sample size, a higher degree of feature richness, and a larger number of historical behaviors. Therefore, the model is easily dominated by the data of high-active objects, resulting in the recommendation system being unable to serve low-active objects well. However, low-active objects are often the core group that determines the ecosystem of a recommendation system. If the recommendation system cannot serve the low-active object group well, it will have a negative impact on aspects such as user growth, user stickiness, user retention rate, and the product user ecosystem.

[0115] Based on this, the present application proposes a content ranking method weighted by object activity, which is used to solve the problem that the model is dominated by the data of highly active users in the ranking scenario of the recommendation system, resulting in the inability to take into account the behavior differences of objects with different levels of activity, and further improve the consumption indicators of object groups with different levels of activity in the recommendation system. The main purpose of the present application is to construct a recommendation model in the fine-ranking stage that can take into account objects with different levels of activity, so as to fully explore the data distribution differences of objects with different levels of activity, display more accurate content to the user, and improve the accuracy and efficiency of the recommendation system. The method provided by the present application can be applied to Figure 2 the implementation environment shown in the figure. This implementation environment includes a terminal 110 and a server 120, and the terminal 110 and the server 120 are connected through a communication network 130. Among them, the terminal 110 and the server 120 can communicate through the communication network 130. Among them, the communication network 130 uses standard communication technologies and / or protocols, usually the Internet, but can also be any network, including but not limited to any combination of Bluetooth, local area network (LAN), metropolitan area network (MAN), wide area network (WAN), mobile, private network, or virtual private network). In some embodiments, customized or dedicated data communication technologies can be used to replace or supplement the above data communication technologies.

[0116] The terminal 110 involved in this application includes, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc. Among them, the client is deployed on the terminal 110. The client can run on the terminal 110 in the form of a browser, or can also run on the terminal 110 in the form of an independent APP, etc. The server 120 involved in this application can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (CDN), and big data and AI platforms. It should be noted that this application takes the configuration and deployment of the deep learning model 140 on the server 120 as an example for illustration. In some embodiments, the configuration of the deep learning model 140 can also be deployed on the terminal 110. In some embodiments, part of the configuration of the deep learning model 140 is deployed on the terminal 110, and part of the configuration is deployed on the server 120.

[0117] Combined with the above implementation environment, the terminal 110 sends a content recommendation request to the server 120. Among them, the content recommendation request can be used to request at least one of recommended videos, news, and commodities. The terminal 110 sends a content recommendation request to the server 120 through the communication network 130, and instructs the server 120 to sort and recommend candidate content based on the content recommendation request and the object data of the target account logged in to the terminal 110. The server 120 feeds back the recommended content to the terminal 110 through the communication network 130 according to the content sorting result.

[0118] In view of the fact that this application involves some terms related to professional fields, for the convenience of understanding, the following explanations will be given.

[0119] (1) Object activity: In the recommendation system, the frequencies of different behaviors of different objects are different, and the activity of an object can be defined by certain behavior frequencies. For example, based on the historical view count or historical like count of the object, etc. Usually, according to the statistical value of a certain type of behavior of the object, the object is defined as a group of high-activity objects, medium-activity objects, low-activity objects, etc.

[0120] (2) Multi-object sorting model: In the recommendation system, the multi-object sorting model is a type of sorting model used to solve the problem of having multiple objective functions and needing to make multiple objectives reach the overall optimum.

[0121] (3) Expert Network: It refers to a small network located between the input layer and the output layer, which provides a better expression for the input layer and, at the same time, provides more appropriate inputs for the output layer. Generally, expert networks are divided into shared expert networks and specific expert networks. The shared expert network is input into the output layer networks of all targets, while the specific expert network is only input into the output layer network of a specific target.

[0122] Combined with the above introduction, the content recommendation method based on the deep learning model in this application will be introduced below. Please refer to Figure 3 , in the embodiments of this application, the content recommendation method based on the deep learning model can be completed independently by the server, or independently by the terminal, or jointly completed by the terminal and the server. Among them, the deep learning model includes an activity gating network and K task networks. Each task network includes N expert networks and a prediction network, where K≥1, N≥1, and both K and N are integers. The content recommendation method includes:

[0123] 210. In response to a content recommendation request for a recommended object, obtain a comprehensive feature vector and an activity feature vector. Among them, the comprehensive feature vector includes the feature vector of the recommended object and the feature vector of the candidate content, and the activity feature vector characterizes the activity of the recommended object in the historical period;

[0124] In one or more embodiments, the recommended object triggers a content recommendation request through the terminal, and the content recommendation request carries the object identifier of the recommended object. Based on this, according to the object identifier carried in the content recommendation request, the data related to the recommended object and the data related to the candidate content can be extracted, and then the comprehensive feature vector and the activity feature vector can be constructed according to the data related to the recommended object and the data related to the candidate content. Among them, the data related to the recommended object needs to include the activity of the recommended object in the historical period (for example, the number of historical views, the number of historical likes, etc.).

[0125] 220. For each task network, based on the comprehensive feature vector, obtain N expert vectors through N expert networks;

[0126] In one or more embodiments, the deep learning model can be a single-object ranking model (i.e., K = 1) or a multi-object ranking model (i.e., K > 1). Among them, the N expert networks in the single-object ranking model do not need to distinguish between exclusive expert networks and shared expert networks. The N expert networks in the multi-object ranking model include exclusive expert networks and / or shared expert networks. Taking one task network as an example, the comprehensive feature vector is respectively used as the input of the N expert networks, and an expert vector is output through each expert network. Taking two task networks as an example, assuming that each task network includes 5 shared expert networks, then these 5 shared expert networks belong to the N expert networks of one task network and also belong to the N expert networks of another task network. At this time, N should be an integer greater than or equal to 5.

[0127] It should be noted that the expert vector and the comprehensive feature vector can have the same dimension. For example, both are 128-dimensional. The expert network can adopt one or more layers of multilayer perceptron (MLP), Factor Machine (FM), or Deep Interest Network (DIN), etc., which is not limited here.

[0128] 230. Based on the activity feature vector, obtain the activity weight vector through the activity gating network;

[0129] In one or more embodiments, the activity feature vector is used as the input of the activity gating network, and an activity weight vector is output through the activity gating network. Exemplarily, the activity weight vector and the expert vector can have the same dimension. For example, both are 128-dimensional. Exemplarily, the dimension of the activity weight vector can be set to N or P, where P is an integer less than N.

[0130] It should be noted that the execution order between step 230 and step 220 is not limited.

[0131] 240. For each task network, perform element-wise multiplication on at least one of the N expert vectors using the activity weight vector to obtain at least one weighted expert vector;

[0132] In one or more embodiments, taking a task network as an example, after obtaining the activity weight vector, at least one of the N expert vectors can be multiplied element-wise to obtain the corresponding at least one weighted expert vector. Exemplarily, if there is no need to filter the N expert vectors, the activity weight vector is directly used to multiply the N expert vectors element-wise to obtain N weighted expert vectors. Exemplarily, if it is necessary to filter the N expert vectors, first P expert vectors are selected from the N expert vectors, and then the activity weight vector is used to multiply the P expert vectors element-wise to obtain P weighted expert vectors.

[0133] 250. For each task network, based on at least one weighted expert vector, obtain the task score of the candidate content through the prediction network;

[0134] In one or more embodiments, taking a task network as an example, the N weighted expert vectors (or, P weighted expert vectors) can be directly concatenated, and then the concatenated feature vector is used as the input of the prediction network, and a task score is output through the prediction network. The value range of the task score is between 0 and 1, and this task score is used to characterize the degree of interest of the recommended object in the candidate content in a task (such as a duration task or an interaction task, etc.). Generally, the more interested the recommended object is in the candidate content, the higher the task score.

[0135] It should be noted that the prediction network can adopt one or more layers of MLP. That is, after passing through the prediction network, the task score is normalized by the sigmoid function. The prediction network in this application can also be understood as the "Tower network layer".

[0136] 260. If it is determined that the candidate content belongs to the content to be pushed according to the K task scores, the candidate content is pushed to the recommended object, where the K task scores are the output results of the K task networks.

[0137] In one or more embodiments, a task network outputs a task score. Based on this, the K task networks can output K task scores. Combining the K task scores can obtain the sorting score of the candidate content, where this sorting score is used to characterize the degree of interest of the recommended object in the candidate content. Generally, the higher the sorting score of the candidate content, the more suitable the candidate content is for pushing to the recommended object. If the sorting score of a certain candidate content is greater than the sorting scores of other candidate content, the candidate content can be used as the content to be pushed.

[0138] Specifically, the basic structure of the deep learning model will be introduced below in conjunction with the diagrams.

[0139] Exemplarily, please refer to Figure 4 , Figure 4It is a schematic structural diagram of a single-task single-expert network in an embodiment of the present application. As shown in the figure, the deep learning model includes a task network, assuming that the comprehensive feature vector is 128-dimensional. Thus, the comprehensive feature vector is used as the input of the task network, and a 128-dimensional expert vector can be output through the expert network. The activity feature vector is used as the input of the activity gating network, and a 1-dimensional or 128-dimensional activity weight vector can be output through the activity gating network. The activity weight vector and the expert vector are multiplied element by element to obtain a 128-dimensional weighted expert vector. The weighted expert vector is used as the input of the prediction network, and a task score is output through the prediction network.

[0140] Exemplarily, please refer to Figure 5 , Figure 5 It is a schematic structural diagram of a multi-task single-expert network in an embodiment of the present application. As shown in the figure, the deep learning model includes two task networks, assuming that the comprehensive feature vector is 128-dimensional. Thus, the comprehensive feature vector is used as the input of the task network, and a 128-dimensional expert vector 1 can be output through expert network 1, and a 128-dimensional expert vector 2 can be output through expert network 2. The activity feature vector is used as the input of the activity gating network, and a 1-dimensional or 128-dimensional activity weight vector can be output through the activity gating network. The activity weight vector is multiplied element by element with expert vector 1 and expert vector 2 respectively to obtain a 128-dimensional weighted expert vector 1 and a 128-dimensional weighted expert vector 2. The weighted expert vector 1 is used as the input of prediction network 1, and a task score 1 is output through prediction network 1, and the weighted expert vector 2 is used as the input of prediction network 2, and a task score 2 is output through prediction network 2.

[0141] Exemplarily, please refer to Figure 6 , Figure 6This is a schematic diagram of a multi - task and multi - expert network in an embodiment of the present application. As shown in the figure, the deep - learning model includes two task networks. Assume that the comprehensive feature vector is 128 - dimensional. Thus, taking the comprehensive feature vector as the input in one task network, through expert network 1, a 128 - dimensional expert vector 1 can be output, and through expert network 2, a 128 - dimensional expert vector 2 can be output. Similarly, taking the comprehensive feature vector as the input in another task network, through expert network 3, a 128 - dimensional expert vector 3 can be output, and through expert network 4, a 128 - dimensional expert vector 4 can be output. Taking the activity feature vector as the input of the activity gating network, through the activity gating network, a 2 - dimensional or 128 - dimensional activity weight vector can be output. Element - wise multiply the activity weight vector with expert vector 1 and expert vector 2 respectively. Thus, a 128 - dimensional weighted expert vector 1 and a 128 - dimensional weighted expert vector 2 are obtained. Element - wise multiply the activity weight vector with expert vector 3 and expert vector 4 respectively. Thus, a 128 - dimensional weighted expert vector 3 and a 128 - dimensional weighted expert vector 4 are obtained. Taking the weighted expert vector 1 and the weighted expert vector 2 as the input of prediction network 1, through prediction network 1, a task score 1 is output, and taking the weighted expert vector 3 and the weighted expert vector 4 as the input of prediction network 2, through prediction network 2, a task score 2 is output.

[0142] Assume that the 2 - dimensional activity weight vector is (0.2, 0.1), and the expert vector 1 is (0.5, 0.4, …, 0.8). Based on this, multiply 0.2 with each element in expert vector 1 respectively, and the obtained weighted expert vector 1 can be expressed as (0.1, 0.08, …, 0.16). The expert vector 2 is (0.9, 0.5, …, 0.6). Based on this, multiply 0.1 with each element in expert vector 2 respectively, and the obtained weighted expert vector 2 can be expressed as (0.05, 0.04, …, 0.08).

[0143] In an embodiment of the present application, a content recommendation method based on a deep - learning model is provided. Through the above - mentioned method, the historical behavior data of the object is used to construct its corresponding activity feature vector as the input of the activity gating network, and then the activity gating network is used to weight each expert network, so that the deep - learning model can better learn the data distribution differences of objects with different activities. Thus, the recommendation accuracy and efficiency of the recommendation system for objects with different activities are improved.

[0144] Optionally, based on the above - mentioned Figure 3 corresponding respective embodiments, in another optional embodiment provided by the embodiment of the present application, in response to a content recommendation request for a recommended object, obtaining a comprehensive feature vector and an activity feature vector may specifically include:

[0145] In response to a content recommendation request for a recommended object, obtain the object data of the recommended object, where the object data includes object activity data and object association data, and the object association data includes one or more of object basic data, object consumption sequence data, and object description data;

[0146] Obtain candidate content from the candidate content set corresponding to the recommended object, where the candidate content set includes Q candidate contents, Q≥1, and Q is an integer;

[0147] Obtain the content data of the candidate content, where the content data includes one or more of content basic data and content statistical data;

[0148] Perform feature extraction on the object association data and the content data to obtain a comprehensive feature vector;

[0149] Perform feature extraction on the object activity data to obtain an activity feature vector.

[0150] In one or more embodiments, a method for constructing a comprehensive feature vector and an activity feature vector is introduced. As can be seen from the foregoing embodiments, the content recommendation request may carry the object identifier of the recommended object. Based on this, the relevant data of the recommended object can be extracted from the log. In addition, the recalled content needs to be roughly sorted according to the relevant data of the recommended object to obtain a candidate content set. For the sake of convenience of description, the following will take one candidate content in a candidate content set as an example for introduction.

[0151] (1) Object activity data;

[0152] Specifically, the object activity data represents the number of contents consumed by the object in the historical period. Among them, the object activity data includes but is not limited to the number of contents liked by the object in the historical period, the number of contents commented by the object in the historical period, the number of contents shared by the object in the historical period, the number of contents coin-voted by the object in the historical period, etc.

[0153] Exemplarily, for example, if the recommended object liked 5 contents in the past month, the corresponding feature value can be "5". Exemplarily, for example, if the recommended object shared 20 contents in the past month, the corresponding feature value can be "20". It can be seen that the above feature values form continuous features.

[0154] (2) Object basic data;

[0155] Specifically, the object description data represents the basic data related to the object. Among them, the object basic data includes but is not limited to the object identifier, object gender, object's age range, object's city type, the viewing duration of the object in the last week, and the viewing duration of the object in the last month, etc.

[0156] Exemplarily, for example, if the object's gender is "male", the corresponding feature value can be "1". Exemplarily, for example, if the age range to which the object belongs is "18 to 30", the corresponding feature value can be "1". Also for example, if the age range to which the object belongs is "31 to 50", the corresponding feature value can be "2". Exemplarily, for example, if the type of city where the object is located is "first-tier city", the corresponding feature value can be "1". Exemplarily, for example, if the consumption duration of the object last week is "13 hours", the corresponding feature value can be "13". It can be seen that the above feature values constitute fixed-length discrete features (i.e., discrete sequence features with a fixed length).

[0157] (3) Object consumption sequence data;

[0158] Specifically, the object consumption sequence data represents the sequence of the object's consumption content within the historical period. Among them, the object consumption sequence data includes but is not limited to the sequence of the object's liked content within the historical period, the sequence of the object's commented content within the historical period, the sequence of the object's shared content within the historical period, the sequence of the object's coin-voted content within the historical period, etc.

[0159] Exemplarily, the recommended object liked 30 pieces of content in the past month. Based on this, obtain the identifiers of these 30 pieces of content in chronological order, and use each content identifier as a feature value. It can be seen that the above feature values constitute non-fixed-length discrete features (i.e., discrete sequence features with an unfixed length).

[0160] (4) Object description data;

[0161] Specifically, the object description data represents the context data related to the object. Among them, the object description data includes but is not limited to the object's network category, the terminal system used by the object, the object's current location, the object's current time, the object's primary portrait, and the object's secondary portrait, etc.

[0162] Exemplarily, for example, if the object's network category is "4G", the corresponding feature value can be "1". Exemplarily, for example, if the terminal system used by the object is "Android", the corresponding feature value can be "2". Exemplarily, for example, if the object's current location is "Guangdong Province", the corresponding feature value can be "3". Exemplarily, for example, if the object's current time is "2022-07-17 20:41:24", the corresponding feature value can be "1658061684". Exemplarily, the object's primary portrait can be "sports", the corresponding feature value can be "23", and the object's secondary portrait can be "basketball", the corresponding feature value can be "18". It can be seen that the above feature values can constitute continuous features and fixed-length discrete features.

[0163] (5) Content basic data;

[0164] Specifically, the object description data represents the basic data related to the content. Among them, the content basic data includes but is not limited to video identifier, video category, video duration interval, video owner account, etc.

[0165] Exemplarily, for example, if the video category is "long video", its corresponding feature value can be "1". Again, for example, if the video category is "short video", its corresponding feature value can be "0". Exemplarily, for example, if the video duration interval is "1 to 2 hours", its corresponding feature value can be "1". Exemplarily, for example, if the video owner account feature is "publisher account", its corresponding feature value can be "11134". Exemplarily, the first-level portrait of the content can be "sports", its corresponding feature value can be "23", and the second-level portrait of the content can be "football", its corresponding feature value can be "15". It can be seen that the above feature values constitute fixed-length discrete features.

[0166] (6) Content statistical data;

[0167] Specifically, the content statistical data represents the statistical posterior data related to the content within the historical period. Among them, the content statistical data includes but is not limited to the exposure count of the content, the click count of the content, the click-through rate of the content, the share count of the content, the forward count of the content, or the video duration of the content, etc.

[0168] Exemplarily, for example, the exposure count of the content in the past month is 2000, and its corresponding feature value can be "2000". Exemplarily, for example, the share count of the recommended object in the past month is 1000, and its corresponding feature value can be "1000". It can be seen that the above feature values constitute continuous features.

[0169] Based on this, the comprehensive feature vector and the activity feature vector can be constructed according to the feature type. The feature type includes fixed-length discrete features, non-fixed-length discrete features, and continuous features. The embedding layer obtains the feature vectors of different feature types through the following embedding methods for the features of different feature types respectively, and these feature vectors can be concatenated to obtain the comprehensive feature vector. Among them, for fixed-length discrete features, the corresponding feature vector is obtained through the discrete feature index. For non-fixed-length discrete features, the feature vector corresponding to each element is obtained through the discrete feature index, and then the corresponding feature vector is obtained through the pooling method. For continuous features, the feature values are directly concatenated, or concatenated after normalization.

[0170] Secondly, in the embodiment of the present application, a method for constructing a comprehensive feature vector and an activity feature vector is provided. Through the above method, multiple activity features are selected to represent the object, thereby characterizing the activity level of the object from multiple angles such as duration and interaction. At the same time, multiple other features are selected to represent the object and candidate content, thereby incorporating the influence of activity features on other features in the model training process.

[0171] Optionally, in the above Figure 3 On the basis of the corresponding embodiments, in another optional embodiment provided by the embodiment of the present application, the deep learning model further includes a feature weighted network;

[0172] In response to a content recommendation request for a recommendation object, after obtaining a comprehensive feature vector and an activity feature vector, the following may also be included:

[0173] Based on the comprehensive feature vector, a target comprehensive feature vector is obtained through a feature weighted network;

[0174] For each task network, based on the comprehensive feature vector, N expert vectors are obtained through N expert networks, which may include:

[0175] For each task network, based on the target comprehensive feature vector, N expert vectors are obtained through N expert networks.

[0176] In one or more embodiments, a method of weighting the features of the comprehensive feature vector is introduced. As can be seen from the above embodiments, after the comprehensive feature vector is obtained through the embedding layer, the feature importance of the comprehensive feature vector can also be selected.

[0177] Specifically, for ease of understanding, see Figure 7 , Figure 7 This is a structural diagram of a feature weighted network introduced in an embodiment of the present application. As shown in the figure, the comprehensive feature vector is used as a feature weighted network, and the target comprehensive feature vector is output through the feature weighted network. Exemplarily, the comprehensive feature vector and the target comprehensive feature vector usually have the same dimension. The target comprehensive feature vector is then used as the input of each expert network, and the corresponding expert vector is output through the expert network.

[0178] It should be noted that the feature weighted network can be a squeeze and excitation network (Squeeze-and-Excitation Network, SENet), or a parameter personalized network (parameter personalized net, PPNet), or an attention mechanism network for information extraction and filtering.

[0179] Secondly, in the embodiments of the present application, a method for feature weighting of the comprehensive feature vector is provided. Through the above method, a feature weighting network is further introduced in the deep learning model to implement the selection of feature importance. For different object groups, features are weighted to reduce the noise of invalid information and enhance the weights of important features, strengthening the distinction between different object groups from the feature side.

[0180] Optionally, on the basis of the above Figure 3 corresponding embodiments, in another optional embodiment provided by the embodiments of the present application, based on the comprehensive feature vector, the target comprehensive feature vector is obtained through the feature weighting network, which may specifically include:

[0181] Perform pooling processing on the feature vectors of each feature domain in the comprehensive feature vector to obtain compressed feature vectors. Among them, the comprehensive feature vector includes the feature vectors of T feature domains, and the compressed feature vector includes T elements, where T>1 and T is an integer;

[0182] Based on the compressed feature vector, obtain the feature weight vector through the excitation network, where the feature weight vector includes T elements;

[0183] Multiply the feature vectors of the T feature domains element by element using the feature weight vector to obtain the target comprehensive feature vector.

[0184] In one or more embodiments, a method for feature weighting based on SENet is introduced. As can be seen from the foregoing embodiments, the feature weighting network can be SENet. Among them, SENet mainly includes three stages, namely, the squeeze stage, the excitation stage, and the reweight stage.

[0185] Specifically, for ease of understanding, please refer to Figure 8 , Figure 8 is a schematic diagram for obtaining the target comprehensive feature vector based on the feature weighting network in the embodiments of the present application. As shown in the figure, the comprehensive feature vector converges the feature vectors from different feature domains. Among them, the feature domain represents the category to which the feature belongs. For example, the object gender, the age range to which the object belongs, and the object network category are three different feature domains. In the squeeze stage, the feature vectors of each feature domain can be compressed and represented through average pooling (mean-pooling) or max pooling (max-pooling), that is, T feature weight values are obtained. The T feature weight values constitute the compressed feature vector. The compressed feature vector can be expressed as [z1,…,z T .

[0186] Exemplarily, the mean-pooling can be used to calculate the feature weight value:

[0187]

[0188] where z i represents the feature weight value corresponding to the i-th feature domain, and 1 ≤ i ≤ T. e i represents the feature vector corresponding to the i-th feature domain in the comprehensive feature vector. F sq (·) represents compression calculation. meanpooling(e i ) represents performing feature compression calculation on e i . d i represents the dimension of the feature vector corresponding to the i-th feature domain. t represents the feature value of the t-th dimension.

[0189] Next, the compressed feature vector is used as the input of the excitation network, and a feature weight vector with the same dimension is output through the excitation network. The feature weight vector includes T elements, and each element represents the weight value of a feature domain. It should be noted that the excitation network usually includes one or two fully connected networks. In the excitation stage, the importance weights of features can be learned based on the compressed feature vector obtained in the previous step. The feature weight vector can be calculated in the following way:

[0190] A = F ex (Z) = σ2(W2σ1(W1Z)) = (a1,…,a T ); Equation (2)

[0191] where A represents the feature weight vector. Z represents the compressed feature vector. F ex (·) represents excitation calculation. σ1 and σ2 represent activation functions. W1 and W2 represent parameter matrices to be learned.

[0192] Finally, each element in the feature weight vector is multiplied element-wise with the feature vector corresponding to each feature domain in the comprehensive feature vector. Thus, a weight update is performed on the original input feature vector. That is, the target comprehensive feature vector is calculated in the following way:

[0193] V = F re (A, E) = [a1·e1, a2·e2,…,a T ·e T = [v1,…,v T ; Equation (3)

[0194] where V represents the target comprehensive feature vector. F re (·) represents weight update calculation. E represents the comprehensive feature vector. a i represents the weight value corresponding to the i-th feature domain, that is, the i-th dimension of the feature weight vector. e i represents the feature vector corresponding to the i-th feature domain. vi It represents the i-th weighted feature vector in the target comprehensive feature vector.

[0195] Again, in the embodiments of the present application, a method for feature weighting based on SENet is provided. Through the above method, SENet can be used to learn the importance of different features personalizedly, effectively filter out invalid noise, and thus improve the prediction accuracy of the deep learning model.

[0196] Optionally, based on the above Figure 3 In another optional embodiment provided by the embodiments of the present application on the basis of the corresponding various embodiments, the feature weighting network includes X gating networks and X multi-layer perceptrons, where X≥1 and X is an integer;

[0197] Based on the comprehensive feature vector, obtaining the target comprehensive feature vector through the feature weighting network may specifically include:

[0198] Obtaining the screened feature vector corresponding to the screened data, where the screened data is the data screened from the object association data and the content data;

[0199] Based on the screened feature vector, obtaining X gating feature vectors through X gating networks, where the input of each gating network is the screened feature vector;

[0200] Based on the comprehensive feature vector and the X gating feature vectors, obtaining the target comprehensive feature vector through X multi-layer perceptrons, where the input of each multi-layer perceptron is different.

[0201] In one or more embodiments, a method for feature weighting based on PPNet is introduced. As can be seen from the foregoing embodiments, the feature weighting network can be PPNet), where PPNet mainly includes X gating networks and X multi-layer perceptrons, and X is an integer greater than or equal to 1. Before passing through PPNet, it is necessary to screen the data from the business scenario object association data and the content data (that is, obtain the screened data). Then, based on these screened data, the corresponding screened feature vector can be constructed. Among them, the feature dimension of the screened feature vector is smaller than the feature dimension of the comprehensive feature vector.

[0202] Specifically, for the sake of understanding, please refer to Figure 9 , Figure 9Another schematic diagram for implementing feature weighting based on a feature weighting network in an embodiment of the present application is shown in the figure. Assume X is 3. The screened feature vectors are respectively used as the inputs of gating network 1, gating network 2, and gating network 3. Through gating network 1, gating feature vector 1 is obtained. Through gating network 2, gating feature vector 2 is obtained. Through gating network 3, gating feature vector 3 is obtained. Among them, the comprehensive feature vector has the same dimension as each gating feature vector. Based on this, first, the comprehensive feature vector is multiplied element-wise with gating feature vector 1 to obtain feature vector 1. Feature vector 1 is used as the input of MLP1. The feature vector output by MLP1 is multiplied element-wise with gating feature vector 2 to obtain feature vector 2. Feature vector 2 is used as the input of MLP2. The feature vector output by MLP2 is multiplied element-wise with gating feature vector 3 to obtain feature vector 3. Feature vector 3 is used as the input of MLP3, and the target comprehensive feature vector is output through MLP3.

[0203] Again, in an embodiment of the present application, a method for feature weighting based on PPNet is provided. Through the above method, PPNet supports the personalized ability of MLP network parameters through a gating network, thereby improving the prediction ability for the target.

[0204] Optionally, based on the corresponding various embodiments above, in another optional embodiment provided by the embodiments of the present application, for each task network, an activity weight vector is used to multiply at least one of the N expert vectors element-wise to obtain at least one weighted expert vector, which specifically may include: Figure 3 For each task network, each weight value in the activity weight vector is used to multiply the corresponding expert vector among the N expert vectors element-wise to obtain N weighted expert vectors, where the activity weight vector includes N weight values, or the activity weight vector has the same dimension as the expert vector.

[0205] In one or more embodiments, a method for implementing activity weighting based on one or more layers of MLP is introduced. As can be seen from the foregoing embodiments, the activity gating network can adopt one or more layers of MLP. After obtaining the N expert vectors, the activity feature vector can be input into one or more layers of MLP, and the activity weight vector is output through the MLP. The N expert vectors can be weighted using the activity weight vector, where the activity weight vector includes N weight values, or the activity weight vector has the same dimension as the expert vector. Hereinafter, an example of a task network will be used for introduction, and other task networks are also processed in a similar manner, so details will not be elaborated here.

[0206] Specifically, for ease of understanding, please refer to

[0207] For Figure 10 ,Figure 10 This is a schematic diagram of implementing feature weighting based on an activity gating network in an embodiment of the present application. As shown in the figure, taking a task network as an example, assume that the task network includes 3 expert networks (for example, 2 shared expert networks and 1 exclusive expert network). Based on this, the following method can be used to calculate N weighted expert vectors:

[0208]

[0209] Among them, \(x\) t represents the N weighted expert vectors corresponding to the \(t\)-th task network. \(t\) represents the \(t\)-th task network. \(N\) represents the number of expert networks included in a task network. \(x\) pc represents the activity feature vector. \(g(\cdot)\) represents the activity gating network (for example, a three-layer MLP can be selected). \(e\) (i) represents the \(i\)-th expert network. \(x\) represents the comprehensive feature vector or the target comprehensive feature vector.

[0210] In one case, taking \(N = 1\) as an example, based on this, the activity weight vector is 1-dimensional, and assume that the activity weight vector is 0.2, and the expert vector is \((0.5, 0.4, \ldots, 0.8)\). Based on this, multiply 0.2 by each element in the expert vector respectively, and the obtained weighted expert vector can be expressed as \((0.1, 0.08, \ldots, 0.16)\).

[0211] In another case, exemplarily, taking \(N = 1\) as an example, based on this, the activity weight vector can be 128-dimensional. Assume that the 128-dimensional activity weight vector is \((0.2, 0.7, \ldots, 0.1)\), and the expert vector is \((0.5, 0.4, \ldots, 0.8)\). Based on this, multiply each element of the activity weight vector by the corresponding element in the expert vector respectively, and the obtained weighted expert vector can be expressed as \((0.1, 0.28, \ldots, 0.08)\).

[0212] Assume that the 1-dimensional activity weight vector is 0.2, and the expert vector is \((0.5, 0.4, \ldots, 0.8)\). Based on this, multiply 0.2 by each element in the expert vector respectively, and the obtained weighted expert vector can be expressed as \((0.1, 0.08, \ldots, 0.16)\).

[0213] Assume that the 128-dimensional activity weight vector is \((0.2, 0.7, \ldots, 0.1)\), and the expert vector is \((0.5, 0.4, \ldots, 0.8)\). Based on this, multiply each element of the activity weight vector by the corresponding element in the expert vector respectively, and the obtained weighted expert vector can be expressed as \((0.1, 0.28, \ldots, 0.08)\).

[0214] Secondly, in the embodiments of the present application, an activity weighting method based on one or more multi-layer perceptrons (MLPs) is provided. Through the above method, in order to fully explore the data distribution differences of objects with different activities, an activity gating network can be learned based on activity features, and the output of the activity gating network is used to weight each expert network, thereby effectively strengthening the personalized features related to object activity and alleviating the phenomenon that the model is dominated by high-activity objects.

[0215] Optionally, on the basis of the above Figure 3 corresponding embodiments, in another optional embodiment provided by the embodiments of the present application, based on the activity feature vector, an activity weight vector is obtained through the activity gating network, which may specifically include:

[0216] Based on N expert vectors and the activity feature vector, an activity weight vector is obtained through the activity gating network, where the activity gating network is a network using the attention mechanism, and the activity weight vector includes N weight values;

[0217] For each task network, at least one expert vector among the N expert vectors is multiplied element-wise by the activity weight vector, which may specifically include:

[0218] For each task network, each weight value in the activity weight vector is used to multiply the corresponding expert vectors among the N expert vectors element-wise to obtain N weighted expert vectors.

[0219] In one or more embodiments, an activity weighting method based on the attention mechanism network is introduced. As can be seen from the foregoing embodiments, the activity gating network can adopt the attention mechanism network. After obtaining the N expert vectors, the N expert vectors and the activity feature vector can be jointly input into the attention mechanism network, and the activity weight vector is output through the attention mechanism network. Among them, the activity weight vector includes N weight values, and each weight value acts on a corresponding expert vector. Hereinafter, an example of a task network will be used for introduction, and other task networks are processed in a similar manner, so details will not be described here.

[0220] Specifically, for the convenience of understanding, please refer to Figure 11 , Figure 11 is another schematic diagram of feature weighting based on the activity gating network in the embodiments of the present application. As shown in the figure, an example of a task network including 3 expert networks (i.e., N = 3) is used for introduction. At the same time, the activity feature vector is used as the query (Q), and the N expert vectors are used as the key (K) and value (V). First, Q is multiplied by KT (i.e., the transpose of K) in matrix multiplication (matmul) to generate a similarity matrix. Then the similarity matrix is divided by Denote the dimension of K. Then, perform normalization using the exponential normalization (softmax) function to obtain the activity weight vector, where the activity weight vector includes N weight values. Then, perform weighted summation of the activity weight vector and V, that is, use each weight value in the activity weight vector to perform element-wise multiplication on each expert vector to obtain N weighted expert vectors. For example, if the first element in the activity weight vector is 0.2 and the expert vector 1 is (0.5, 0.4, …, 0.8), then multiply 0.2 by each element in the expert vector 1 respectively, and the weighted expert vector 1 can be expressed as (0.1, 0.08, …, 0.16).

[0221] Based on this, the following method can be used to calculate N weighted expert vectors:

[0222]

[0223] Among them, attention(Q, K, V) represents N weighted expert vectors. Q represents query, K represents key, and V represents value. Denote the dimension of K.

[0224] Secondly, in the embodiments of the present application, an activity weighting method based on an attention mechanism network is provided. Through the above method, for feature vectors with more dimensions, important features can be strengthened, weighted for each expert network, so as to effectively strengthen the personalized features related to object activity and alleviate the phenomenon that the model is dominated by high-activity objects.

[0225] Optionally, on the basis of the above Figure 3 corresponding embodiments, in another optional embodiment provided by the embodiments of the present application, after obtaining the activity weight vector through the activity gating network based on the activity feature vector, it may further include:

[0226] For each task network, determine the largest P weight values from the activity weight vector, where the activity weight vector includes N weight values, each weight value corresponds to an expert network, N > P ≥ 1, and P is an integer;

[0227] For each task network, determine the corresponding P expert networks from the N expert networks according to the P weight values;

[0228] For each task network, perform element-wise multiplication on at least one expert vector among the N expert vectors using the activity weight vector to obtain at least one weighted expert vector, which may specifically include:

[0229] For each task network, each of the P weight values is used to perform an element-wise multiplication on the corresponding expert vectors among the P expert vectors to obtain P weighted expert vectors, where the P expert vectors are the output results of P expert networks.

[0230] In one or more embodiments, an activity weighting method based on a sparse gating network is introduced. As can be seen from the foregoing embodiments, the activity gating network can adopt a sparse gating network, where the sparse gating network can be an MLP. After obtaining N expert vectors, the activity feature vector can be input into the sparse gating network, and the activity weight vector is output through the sparse gating network. Among them, the activity weight vector includes N weight values, and each weight value represents the importance of an expert network. Generally, the higher the weight value, the more important the corresponding expert network is. Hereinafter, a task network will be taken as an example for introduction, and other task networks are processed in a similar manner, so details will not be elaborated here.

[0231] Specifically, for the sake of easy understanding, please refer to Figure 12 , Figure 12 FIG. is another schematic diagram of implementing feature weighting based on an activity gating network in the embodiments of the present application. As shown in the figure, a task network including 3 expert networks (i.e., N = 3) is taken as an example for introduction. The activity feature vector is used as the input of the activity gating network, and an N-dimensional activity weight vector is output through the activity gating network. Assume that the activity weight vector is (0.2, 0.7, 0.1), where "0.2" corresponds to expert network 1, "0.7" corresponds to expert network 2, and "0.1" corresponds to expert network 3. Taking P = 1 as an example, that is, determining the largest weight value from the activity weight vector, that is, "0.7". Then, according to this weight value, the corresponding expert network 2 is determined from the N expert networks, and the expert vector 2 is multiplied element-wise by this weight value to obtain the corresponding weighted expert vector 2.

[0232] It should be noted that P can take other values, and the dot product results between the P weight values and the corresponding P expert vectors are calculated in a similar manner to obtain P weighted expert vectors.

[0233] Secondly, in the embodiments of the present application, an activity weighting method based on a sparse gating network is provided. Through the above method, considering that a task network may include a large number of expert networks, therefore, based on the activity weight vector output by the sparse gating network, it is possible to further select which expert vectors output by the expert networks are weighted. Thus, the computational amount can be saved, the model capacity can be improved, and the calculation speed is fast.

[0234] Optionally, in the above Figure 3Based on the corresponding embodiments, in another alternative embodiment provided by the embodiments of the present application, K>1, and the N expert networks include a shared expert network and exclusive expert networks, and each task network further includes a task gating network;

[0235] For each task network, based on at least one weighted expert vector, obtain the task score of the candidate content through the prediction network, which may specifically include:

[0236] For each task network, based on the comprehensive feature vector, obtain the task feature vector through the task gating network;

[0237] For each task network, perform element-wise multiplication on the task feature vector and at least one weighted expert vector to obtain the target feature vector;

[0238] For each task network, based on the target feature vector, obtain the task score of the candidate content through the prediction network.

[0239] In one or more embodiments, a deep learning model based on a progressive layered extraction (PLE) architecture is provided. As can be seen from the foregoing embodiments, the multi-object ranking framework based on PLE uses the historical behavior lengths of multiple types of objects as activity features, introducing the personalized differences of different groups. At the same time, based on the feature weighting network, importance selection is performed on the feature inputs to enhance effective information and filter out invalid noise. To further strengthen the role of the activity features, an activity gating network is further introduced, greatly improving the online recommendation accuracy and efficiency of different activity groups, and enhancing the user experience.

[0240] It can be understood that the multi-object ranking framework based on PLE also introduces a task gating network corresponding to each task. Exemplarily, the task gating network can adopt one or more layers of MLP. That is, the comprehensive feature vector can be input into the MLP, and the task feature vector is output through the MLP. Then, the task feature vector is used to weight each weighted expert vector. Exemplarily, the task gating network can adopt an attention mechanism network. That is, the comprehensive feature vector and each weighted expert vector can be input into the attention mechanism network, and the task feature vector is output through the attention mechanism network. Then, the task feature vector is used to weight each weighted expert vector. Exemplarily, the task gating network can adopt a sparse gating network. That is, the comprehensive feature vector can be input into the sparse gating network, and the task feature vector is output through the sparse gating network. The task feature vector can be used to weight several weighted expert vectors.

[0241] Specifically, for ease of understanding, please refer to Figure 13 , Figure 13This is a schematic diagram of the structure of a deep learning model in an embodiment of the present application. As shown in the figure, N expert networks include a shared expert network and exclusive expert networks. For example, it includes 2 exclusive expert networks and 1 shared expert network. In a task network, the comprehensive feature vector is used as the input of task gating network 1 to obtain task feature vector 1. After passing through exclusive expert network 1 and the shared expert network, 2 expert vectors can be obtained. Then, the activity weight vector is used to perform weighted calculations on these 2 expert vectors respectively to obtain 2 weighted expert vectors. Based on this, task feature vector 1 and these 2 weighted expert vectors are multiplied element by element to obtain target feature vector 1. Finally, target feature vector 1 is used as the input of prediction network 1 to obtain task score 1.

[0242] Similarly, in another task network, the comprehensive feature vector is used as the input of task gating network 2 to obtain task feature vector 2. After passing through exclusive expert network 2 and the shared expert network, 2 expert vectors can be obtained. Then, the activity weight vector is used to perform weighted calculations on these 2 expert vectors respectively to obtain 2 weighted expert vectors. Based on this, task feature vector 2 and these 2 weighted expert vectors are multiplied element by element to obtain target feature vector 2. Finally, target feature vector 2 is used as the input of prediction network 2 to obtain task score 2.

[0243] It can be understood that in the Figure 13 shown framework, assuming that the weighted expert vector is 128-dimensional, the task feature vector can be 2-dimensional or 128 * 2-dimensional, which is not limited here.

[0244] Secondly, in the embodiment of the present application, a deep learning model based on PLE is provided. In the above manner, based on the PLE multi-object ranking model, a feature weighting network and an activity gating network are added thereto. In order to simultaneously model multiple objectives, the PLE model can be selected as the basic model. The PLE model includes a shared expert network for multiple objectives and also provides exclusive expert networks unique to different objectives, effectively alleviating the "seesaw" effect that may occur between different objectives. In addition, the gating mechanism acts not only on the shared expert network but also on the exclusive expert networks unique to each objective, effectively helping each objective learn different activity weights.

[0245] Optionally, based on the above Figure 3 corresponding respective embodiments, in another optional embodiment provided by the embodiment of the present application, K > 1, and the N expert networks are N shared expert networks, and each task network further includes a task gating network;

[0246] For each task network, based on at least one weighted expert vector, the task score of the candidate content is obtained through the prediction network, which may specifically include:

[0247] For each task network, based on the comprehensive feature vector, obtain the task feature vector through the task gating network;

[0248] For each task network, perform element-wise multiplication of the task feature vector and at least one weighted expert vector to obtain the target feature vector;

[0249] For each task network, based on the target feature vector, obtain the task score of the candidate content through the prediction network.

[0250] In one or more embodiments, a deep learning model based on the multi-gate mixture-of-experts (MMoE) architecture is introduced. As can be seen from the foregoing embodiments, the multi-objective ranking framework based on MMoE uses the historical behavior lengths of multiple types of objects as the activity features, introducing the personalized differences of different groups. At the same time, based on the feature weighting network, importance selection is performed on the feature inputs to enhance the effective information and filter out the invalid noise. To further strengthen the role of the activity features, an activity gating network is further introduced, greatly improving the online recommendation accuracy and efficiency of different activity groups, and enhancing the user experience.

[0251] It can be understood that the multi-objective ranking framework based on MMoE also introduces the task gating network corresponding to each task. Exemplarily, the task gating network can adopt one or more layers of MLP, or, the attention mechanism network, or, the sparse gating network, which is not limited herein.

[0252] Specifically, for ease of understanding, please refer to Figure 14 , Figure 14 which is another structural schematic diagram of the deep learning model in the embodiments of the present application. As shown in the figure, the N expert networks all belong to the shared expert networks. For example, it includes 3 shared expert networks. In a task network, the comprehensive feature vector is used as the input of the task gating network 1 to obtain the task feature vector 1. Three expert vectors can be obtained through the 3 shared expert networks, and then the activity weight vector is used to perform weighted calculations on these 3 expert vectors respectively to obtain 3 weighted expert vectors. Based on this, the task feature vector 1 and these 3 weighted expert vectors are subjected to element-wise multiplication to obtain the target feature vector 1. Finally, the target feature vector 1 is used as the input of the prediction network 1 to obtain the task score 1.

[0253] Similarly, in another task network, the comprehensive feature vector is used as the input of the task gating network 2 to obtain the task feature vector 2. Three expert vectors can be obtained through three shared expert networks, and then the activity weight vector is used to perform weighted calculations on these three expert vectors respectively to obtain three weighted expert vectors. Based on this, the task feature vector 2 is element-wise multiplied with these three weighted expert vectors to obtain the target feature vector 2. Finally, the target feature vector 2 is used as the input of the prediction network 2 to obtain the task score 2.

[0254] It can be understood that in the Figure 14 shown framework, assuming that the weighted expert vector is 128-dimensional, the task feature vector can be 3-dimensional or 128 * 3-dimensional, which is not limited here.

[0255] Secondly, in the embodiments of the present application, a deep learning model based on MMoE is provided. Through the above method, based on the MMoE multi-objective ranking model, a feature weighting network and an activity gating network are added thereto. In order to simultaneously model multiple objectives, the MMoE model can be selected as the basic model. The MMoE model includes shared expert networks for multiple objectives, which are used to learn the correlation between subtasks, thereby facilitating the improvement of the prediction accuracy.

[0256] Optionally, based on the above Figure 3 corresponding embodiments, in another optional embodiment provided by the embodiments of the present application, K > 1, and the N expert networks are N shared expert networks;

[0257] For each task network, based on at least one weighted expert vector, the task score of the candidate content is obtained through the prediction network, which may specifically include:

[0258] For each task network, matrix transformation is performed on at least one weighted expert vector through a matrix set to obtain a target feature vector, where the matrix set includes a transformation matrix or an identity matrix;

[0259] For each task network, based on the target feature vector, the task score of the candidate content is obtained through the prediction network.

[0260] In one or more embodiments, a deep learning model based on sub-network routing (SNR) is introduced. As can be seen from the foregoing embodiments, the multi-object ranking framework based on SNR uses the historical behavior lengths of multiple types of objects as activity features, introducing the personalized differences of different groups. At the same time, based on the feature weighting network, importance selection is performed on the feature inputs to enhance the effective information and filter out the invalid noise. In order to further strengthen the role of the activity features, an activity gating network is further introduced, greatly improving the online recommendation accuracy and efficiency of different activity groups, and enhancing the user experience.

[0261] Specifically, for ease of understanding, please refer to Figure 15 , Figure 15 , which is another structural schematic diagram of the deep learning model in the embodiments of the present application. As shown in the figure, N expert networks all belong to shared expert networks. For example, it includes 3 shared expert networks. Three expert vectors can be obtained through 3 shared expert networks, and then these 3 expert vectors are respectively weighted and calculated using the activity weight vector to obtain 3 weighted expert vectors (weighted expert vector 1, weighted expert vector 2, and weighted expert vector 3 respectively). Based on this, matrix transformation can be performed on the 3 weighted expert vectors through a matrix set to obtain 2 target feature vectors (target feature vector 1 and target feature vector 2 respectively). Finally, target feature vector 1 is used as the input of prediction network 1 to obtain task score 1, and target feature vector 2 is used as the input of prediction network 2 to obtain task score 2. Two conversion methods will be introduced below in combination with formulas.

[0262] (1) The target feature vectors can be calculated in the following manner:

[0263]

[0264] Among them, V1 represents target feature vector 1, and V2 represents target feature vector 2. U1 represents weighted expert vector 1, U2 represents weighted expert vector 2, and U3 represents weighted expert vector 3. W ij represents the transformation matrix between U j and V i . Z ij represents a binary coding variable, that is, 0 or 1. By setting it to 0, a certain shared expert network is discarded, and by setting it to 1, a certain shared expert network is selected.

[0265] (2) The target feature vectors can be calculated in the following manner:

[0266]

[0267] Among them, V1 represents the target feature vector 1, and V2 represents the target feature vector 2. U1 represents the weighted expert vector 1, U2 represents the weighted expert vector 2, and U3 represents the weighted expert vector 3. I ij represents the identity matrix of U j and V i . Z ij represents the binary coding variable, which is either 0 or 1. By setting it to 0, a certain shared expert network is discarded, and by setting it to 1, a certain shared expert network is selected.

[0268] It should be noted that in addition to using PLE, MMoE, or SNR as the basic model, the deep learning model provided in this application can also use a multiple-level sparse sharing model (MSSM) or the like as the basic model, and no exhaustive listing is made here.

[0269] Secondly, in the embodiments of this application, a deep learning model based on SNR is provided. Through the above method, based on the SNR multi-objective ranking model, a feature weighting network and an activity gating network are added. To simultaneously model multiple objectives, the SNR model can be selected as the basic model. The SNR model includes shared expert networks for multiple objectives and is used to learn the correlation between subtasks. In addition, the SNR model can achieve more flexible parameter sharing, while maintaining the computational advantages of the classical multi-task model. The shared low-level hidden layer is modularized into multiple sub-networks, and the connection between the sub-networks and the learnable latent variables is controlled to achieve flexible parameter sharing.

[0270] Optionally, based on the above Figure 3 corresponding embodiments, in another optional embodiment provided by the embodiments of this application, for each task network, after obtaining the task score of the candidate content through the prediction network based on at least one weighted expert vector, it may further include:

[0271] Determine the ranking score of the candidate content according to the K task scores;

[0272] Obtain the ranking scores of each candidate content among other candidate contents, where the other candidate contents are from the candidate content set of the recommended object;

[0273] Determine the score ranking result of the candidate content according to the ranking score of the candidate content and the ranking scores of each candidate content among other candidate contents;

[0274] If the score ranking result of the candidate content is within the recommended ranking range, determine that the candidate content belongs to the content to be pushed.

[0275] In one or more embodiments, a method for determining whether the content push condition is satisfied is introduced. As can be seen from the foregoing embodiments, each task network can predict a task score. Therefore, for a recommended object, the sorting scores for each candidate content can be obtained. At the same time, it is also necessary to obtain the sorting scores of other candidate contents in the candidate content set. Based on this, these sorting scores are sorted to obtain the score sorting results of each candidate content. Generally, the larger the sorting score, the higher the position in the score sorting result. Exemplarily, assume that the score sorting result of a certain candidate content (for example, candidate content A) is the fifth in the overall sorting, and the recommended sorting range is from the first to the tenth. Then this candidate content (for example, candidate content A) belongs to the content to be pushed.

[0276] After the deep learning model is trained with a large number of samples, the model parameters can well fit the data of different object groups. Thus, the deep learning model can be exported and applied to the online service. The deep learning model provided in this application is mainly applicable to the fine ranking stage of the recommendation system. For ease of understanding, please refer to Figure 16 , Figure 16 which is a schematic flowchart of applying the deep learning model in the embodiments of this application. As shown in the figure, specifically:

[0277] In step B1, the recommendation system receives a content recommendation request triggered by the recommended object through the terminal. The content recommendation request may carry an object identifier. Based on this, the recommendation system can extract object-related data based on the object identifier.

[0278] In step B2, according to the object-related data of the recommended object, a part of the candidate contents that the recommended object may be interested in is quickly recalled from the massive content. The characteristic of recall is relatively fast speed.

[0279] In step B3, through the object data and the candidate content data, the recalled candidate contents are roughly sorted to further reduce the number of selected candidate contents on the premise of ensuring a certain degree of accuracy. The rough ranking process can be selected according to the scenario. The purpose is to prevent the number of candidate contents returned by the recall link from being too large, resulting in the speed of the ranking link not being able to keep up. Therefore, a rough ranking stage can be added between the recall stage and the fine ranking stage.

[0280] In step B4, for each candidate content returned by the rough ranking, the relevant data of the current request recommended object is pulled online and spliced. Based on this, the spliced feature vector can be used as the input of the deep learning model, and the deep learning model scores each candidate content after rough ranking to obtain the sorting score of each candidate content.

[0281] In step B5, all candidate contents are re-sorted in descending order of the sorting scores.

[0282] In step B6, the candidate content with a high push ranking (or ranking first) is used as the content to be pushed.

[0283] It should be noted that the deep learning model provided in this application can be applied not only to recommendation scenarios of instant messaging apps, but also to recommendation scenarios of live broadcast apps, recommendation scenarios of question-and-answer apps, and recommendation scenarios of video accounts, etc.

[0284] Secondly, in the embodiment of the present application, a method for determining whether the content push condition is met is provided. Through the above method, according to the score ranking results of each candidate content, the candidate content with a higher ranking is pushed to the recommendation object first, thereby improving the accuracy of content recommendation.

[0285] Optionally, in the above Figure 3 On the basis of the corresponding embodiments, in another optional embodiment provided by the embodiment of the present application, determining the ranking scores of the candidate contents according to the K task scores may specifically include:

[0286] If K = 1, the K task scores are used as the ranking scores of the candidate contents;

[0287] If K>1, the ranking score of the candidate content is calculated according to the K task scores and the task weight of each task score in the K task scores.

[0288] In one or more embodiments, a method for calculating the ranking score of candidate content is introduced. As can be seen from the above embodiments, the deep learning model includes K task networks, each of which is used to output a task score. Based on this, if K=1, the task score is the ranking score of the candidate content.

[0289] Specifically, if K>1, in one case, the ranking score of the candidate content can be calculated based on the K task scores. In another case, the ranking score of the candidate content can be calculated based on the K task scores and the task weight of each task score in the K task scores.

[0290] Again, in the embodiment of the present application, a method for calculating the ranking score of candidate content is provided. Through the above method, one or more task types are set according to the business scenario, and then the ranking score is calculated in combination with the task score corresponding to each task type, which is used as the basis for evaluating whether the candidate content is suitable for recommendation, thereby increasing the adaptability and flexibility of the solution.

[0291] Optionally, in the above Figure 3 On the basis of the corresponding embodiments, in another optional embodiment provided by the embodiment of the present application, the K task scores include a duration score, an interaction score, a completion rate score, and a fast scrolling rate score;

[0292] According to the K task scores and the task weights of each task score among the K task scores, the sorting score of the candidate content is calculated, which may specifically include:

[0293] The duration score is weighted by the duration weight to obtain the first score;

[0294] The interaction score is weighted by the interaction weight to obtain the second score;

[0295] The completion rate score is weighted by the completion rate weight to obtain the third score;

[0296] The fast-forward rate score is weighted by the fast-forward rate weight to obtain the fourth score;

[0297] According to the first score, the second score, the third score, and the fourth score, the sorting score of the candidate content is calculated. Among them, the sorting score of the candidate content is positively correlated with the duration score, the interaction score, and the completion rate score, and the sorting score of the candidate content is negatively correlated with the fast-forward rate score.

[0298] In one or more embodiments, a method for content push based on four types of tasks is introduced. As can be seen from the foregoing embodiments, the K tasks may include a duration task, an interaction task, a completion rate task, and a fast-forward rate task. Based on this, the K task scores output by the deep learning model include a duration score, an interaction score, a completion rate score, and a fast-forward rate score. The K prediction networks are respectively a duration prediction network, an interaction prediction network, a completion rate prediction network, and a fast-forward rate prediction network.

[0299] Specifically, for ease of understanding, please refer to Figure 17 , Figure 17 is a schematic structural diagram of a multi-task deep learning model in an embodiment of the present application. As shown in the figure, taking the duration task as an example, the task feature vector 1 output by the task gating network 1 is element-wise multiplied by the weighted expert vector corresponding to the exclusive expert network 1 and the weighted expert vector 3 corresponding to the shared expert network to obtain the target feature vector 1. The target feature vector 1 is used as the input of the duration prediction network, and the duration score is output through the duration prediction network.

[0300] It should be noted that other tasks are predicted to obtain the corresponding task scores in a similar manner, which will not be elaborated here.

[0301] Exemplarily, the sorting score of the candidate content can be calculated in the following manner:

[0302] S total = w1 * S1 + w2 * S2 + w3 * S3 + w4 * (1 - S4); Equation (8)

[0303] where, S totalRepresents the sorting score of the candidate content. w1 represents the duration weight (e.g., 0.7). S1 represents the duration score, and w1*S1 represents the first score. w2 represents the interaction weight (e.g., 0.8). S2 represents the interaction score, and w2*S2 represents the second score. w3 represents the completion rate weight (e.g., 0.3). S3 represents the completion rate score, and w3*S3 represents the third score. w4 represents the fast swipe rate weight (e.g., 0.5). S4 represents the fast swipe rate score, and w4*S4 represents the fourth score.

[0304] It can be seen that the greater the duration score, interaction score, or completion rate score, the higher the sorting score of the candidate content. And the greater the fast swipe rate score, the lower the sorting score of the candidate content.

[0305] Furthermore, in the embodiments of the present application, a content push method based on four types of tasks is provided. Through the above method, the recommendation situation of the content is jointly predicted from the perspectives of satisfaction rate, interaction rate, completion rate, and fast swipe rate, thereby improving the prediction accuracy and being applicable to specific services.

[0306] Optionally, based on the corresponding respective embodiments above, in another optional embodiment provided by the embodiments of the present application, it may further include: Figure 3 Obtain a set of data to be trained, where each data to be trained includes object data of a sample object, content data of a sample content, and K task labels;

[0307] For each data to be trained, perform feature processing on the object data of the sample object and the content data of the sample content to obtain a sample feature vector of the data to be trained, where the sample feature vector includes a comprehensive sample feature vector and an activity sample feature vector;

[0308] Based on the sample feature vector of each data to be trained, obtain K prediction scores for each data to be trained through a deep learning model;

[0309] According to the K prediction scores of each data to be trained and the corresponding K task labels, determine the total loss result;

[0310] Update the model parameters of the deep learning model according to the total loss result.

[0311] In one or more embodiments, a training method of a deep learning model is introduced. As can be seen from the foregoing embodiments, before the deep learning model goes online for service, a large amount of data to be trained is required to train the deep learning model so that the model parameters can be learned sufficiently to fit the data distribution of the data to be trained.

[0312] Specifically, for ease of understanding, please refer to

[0313] Figure 18 ​, Figure 18 This is a schematic diagram of a process for training a deep learning model in an embodiment of the present application. As shown in the figure, model training is roughly divided into four steps, which will be introduced separately below.

[0314] In step C1, data extraction is performed to obtain the object data of the sample object and the content data of the sample content. That is, according to the data embedding service of the online system, the historical behavior logs of the sample object are extracted, mainly including various behaviors related to content viewing.

[0315] In step C2, the extracted sample data is preprocessed. That is, after obtaining the preliminary sample data, the sample data also needs to be filled with null values (i.e., filled with the mean value), abnormal samples are removed (i.e., feature values higher or lower than certain thresholds are removed), abnormal feature values are corrected (i.e., feature values higher or lower than certain thresholds are corrected according to experience), weight settings are performed (i.e., according to business requirements and the ratio of positive and negative samples. For example, if the number of positive samples for liking is very small, in order to ensure the training effect, the positive samples for liking are weighted 1 to 2 times), etc., so as to obtain effective sample data.

[0316] In step C3, samples are constructed based on the effective sample data, that is, feature construction and label construction are performed. For feature construction, through the feature reporting system, the sample feature vectors required for model training are extracted, including the comprehensive sample feature vector and the activity sample feature vector mentioned in the foregoing embodiments. For label construction, it includes but is not limited to duration task labels, interaction task labels, completion rate task labels, and fast sliding rate task labels. By splicing the feature and label data, a large amount of data to be trained can be obtained.

[0317] Exemplarily, when the viewing duration is greater than or equal to the duration threshold (for example, 3 minutes), the duration task label can be set to "1", otherwise, the duration task label is set to "0". Exemplarily, when the sample object has interacted with the sample content (such as liking or commenting, etc.), the interaction task label can be set to "1", otherwise, the interaction task label is set to "0". Exemplarily, when the sample object has completely played the sample content, the completion rate task label can be set to "1", otherwise, the completion rate task label is set to "0". Exemplarily, when the viewing time of the sample object for the sample content is less than the duration threshold (for example, 5 seconds), the fast sliding rate task label can be set to "1", otherwise, the fast sliding rate task label is set to "0".

[0318] In step C4, train the model, that is, after obtaining valid data to be trained, input the sample feature vectors of the data to be trained into the deep learning model for sufficient training. During the training process, the loss function of the model needs to be set. K predicted scores of the data to be trained are output through the deep learning model. Calculate the total loss result of the K predicted scores and the K task labels using the loss function. Finally, use the Adaptive Moment Estimation (Adam) optimizer to solve for the optimal value.

[0319] In step C5, output the trained deep learning model.

[0320] Secondly, in the embodiments of the present application, a training method for a deep learning model is provided. Through the above method, during the model training process, there is no need to perform threshold division according to the object activity, and the normalized eigenvalue can be directly used for model training, avoiding the negative impact on the model caused by improper threshold selection. In addition, on the basis of the same modeling, the object activity is explicitly introduced for modeling, so as to take into account the data distribution differences of different activity object groups. The same amount of machine resources as the unified modeling method can be used to serve online inference, thus effectively improving the accuracy and efficiency of the recommendation system.

[0321] Optionally, on the basis of the above Figure 3 In another optional embodiment provided by the embodiments of the present application corresponding to each of the above embodiments, according to the K predicted scores of each data to be trained and the corresponding K task labels, determine the total loss result, which may specifically include:

[0322] For each task among the K tasks, calculate the task loss result using the first loss function according to the predicted score of each data to be trained and the corresponding task label;

[0323] If K = 1, then use the task loss result as the total loss result;

[0324] If K > 1, then sum up the task loss results of each task to obtain the total loss result;

[0325] Or,

[0326] According to the K predicted scores of each data to be trained and the corresponding K task labels, determine the total loss result, which may specifically include:

[0327] For each task among the K tasks, calculate the task loss result using the second loss function according to the sample weight, the corresponding predicted score, and the corresponding task label of each data to be trained, where the sample weight is related to the historical behavior data of the sample object for the sample content;

[0328] If K = 1, then use the task loss result as the total loss result;

[0329] If K > 1, then the task loss results for each task are summed to obtain the total loss result.

[0330] In one or more embodiments, two ways of training a deep learning model based on a loss function are introduced. As can be seen from the foregoing embodiments, if there is only one task, the task loss result corresponding to this task is the total loss result. If there are multiple tasks, the task loss results corresponding to each task are calculated separately, and then the task loss results for each task are summed to obtain the total loss result.

[0331] Specifically, after obtaining the effective data to be trained, the data to be trained can be input into the deep learning model for sufficient training, and the loss function of the model needs to be set during the training process. For example, the cross-entropy loss function can be used as the first loss function to train the model, or the weighted cross-entropy loss function can be used as the second loss function to train the model.

[0332] Exemplarily, the first loss function can be used to calculate the task loss result:

[0333]

[0334] where L represents the task loss result corresponding to the task. M is the total number of samples. y i represents the task label of the i-th data to be trained for this task. represents the predicted score of the i-th data to be trained for this task.

[0335] Exemplarily, the second loss function can be used to calculate the task loss result:

[0336]

[0337] where L represents the task loss result corresponding to the task. M is the total number of samples. w i represents the sample weight of the i-th data to be trained. y i represents the task label of the i-th data to be trained for this task. represents the predicted score of the i-th data to be trained for this task. It can be understood that the sample weight of the data to be trained is related to the historical behavior data of the sample object on the sample content. Assuming that the sample weight of the sample content is "1" initially, the sample weight can be increased when the sample object clicks, likes, comments, follows, or forwards. For example, if the sample object clicks and likes a certain sample content, the sample weight of this sample content is "3".

[0338] Again, in the embodiments of the present application, two ways of training a deep learning model based on a loss function are provided. Through the above methods, the weighted cross-entropy loss function is used to train the model, which is beneficial to strengthening the influence of object activity on the prediction accuracy of the model.

[0339] The method provided by the present application has been experimentally tested in immersive video services. Compared with traditional methods, the offline metrics of the present application have been significantly improved, that is, the area under the curve (AUC) has been significantly increased. Small traffic tests have found that compared with traditional methods, the online metrics of the present application have also been greatly improved, that is:

[0340] (1) Overall video service metrics: The number of content likes increased by 1.84%, the number of shares increased by 3.65%, the number of follows increased by 3.09%, and the video view (VV) increased by 1.05%.

[0341] (2) Metrics for the low-activity object group of videos: The number of content likes increased by 2.31%, the number of shares increased by 4.67%, the number of follows increased by 4.80%, and the video VV increased by 2.27%.

[0342] It can be seen that the present application not only significantly improves the overall important metrics of the service, but also has a more significant improvement in the metrics of the low-activity object group. This also shows that the present application fully exploits the data distribution differences of different groups, improving key metrics such as the accuracy of recommendations, the number of daily active users (DAU), the consumption duration, VV, and interactions, thereby enhancing the user experience.

[0343] The content recommendation device based on the deep learning model in the present application will be described in detail below. Please refer to Figure 19 The content recommendation device 30 based on the deep learning model includes:

[0344] An acquisition module 310, configured to respond to a content recommendation request for a recommended object, and acquire a comprehensive feature vector and an activity feature vector, where the comprehensive feature vector includes the feature vector of the recommended object and the feature vector of the candidate content, and the activity feature vector characterizes the activity of the recommended object in a historical period;

[0345] The acquisition module 310 is further configured to, for each task network, obtain N expert vectors through N expert networks based on the comprehensive feature vector;

[0346] The acquisition module 310 is further configured to obtain an activity weight vector through an activity gating network based on the activity feature vector;

[0347] The obtaining module 310 is further configured to, for each task network, perform element-wise multiplication on at least one of the N expert vectors by using an activity weight vector to obtain at least one weighted expert vector;

[0348] The obtaining module 310 is further configured to, for each task network, based on at least one weighted expert vector, obtain a task score of the candidate content through a prediction network;

[0349] The recommendation module 320 is configured to, if it is determined according to the K task scores that the candidate content belongs to the content to be pushed, push the candidate content to the recommendation object, where the K task scores are the output results of the K task networks.

[0350] In the embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, the historical behavior data of the user is used to construct its corresponding activity feature vector as the input of the activity gating network, and then the activity gating network is used to weight each expert network, so that the deep learning model can better learn the data distribution differences of users with different activity levels. Therefore, the recommendation accuracy and efficiency of the recommendation system for users with different activity levels are improved.

[0351] Optionally, on the basis of the corresponding embodiment above, Figure 19 In another embodiment of the content recommendation device 30 of the deep learning model provided in the embodiment of the present application,

[0352] The obtaining module 310 is specifically configured to respond to a content recommendation request for a recommendation object, and obtain object data of the recommendation object, where the object data includes object activity data and object association data, and the object association data includes one or more of object basic data, object consumption sequence data, and object description data;

[0353] Obtain candidate content from the candidate content set corresponding to the recommendation object, where the candidate content set includes Q candidate contents, Q≥1, and Q is an integer;

[0354] Obtain content data of the candidate content, where the content data includes one or more of content basic data and content statistical data;

[0355] Perform feature processing on the object association data and the content data to obtain a comprehensive feature vector;

[0356] Perform feature processing on the object activity data to obtain an activity feature vector.

[0357] In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, multiple activity features for representing an object are selected. Thus, the activity degree of the object is characterized from multiple perspectives such as duration and interaction. At the same time, multiple other features for representing the object and candidate content are selected. Thus, the influence of the activity features on other features can be incorporated during the model training process.

[0358] Optionally, based on the corresponding embodiment above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiment of the present application, the deep learning model further includes a feature weighting network; Figure 19 The obtaining module 310 is further configured to, in response to a content recommendation request for a recommended object, after obtaining a comprehensive feature vector and an activity feature vector, obtain a target comprehensive feature vector based on the comprehensive feature vector through the feature weighting network;

[0359] Specifically, for each task network, the obtaining module 310 is configured to obtain N expert vectors through N expert networks based on the target comprehensive feature vector.

[0360] In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, a feature weighting network is further introduced into the deep learning model to implement the selection of feature importance. For different object groups, the features are weighted to reduce the noise of invalid information, enhance the weights of important features, and strengthen the differentiation of different object groups from the feature side.

[0361] Optionally, based on the corresponding embodiment above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiment of the present application,

[0362] The obtaining module 310 is specifically configured to perform pooling processing on the feature vectors of each feature domain in the comprehensive feature vector to obtain a compressed feature vector, where the comprehensive feature vector includes feature vectors of T feature domains, the compressed feature vector includes T elements, T>1, and T is an integer; Figure 19 Based on the compressed feature vector, obtain a feature weight vector through an excitation network, where the feature weight vector includes T elements;

[0363] Multiply the feature vectors of the T feature domains element by element using the feature weight vector to obtain a target comprehensive feature vector.

[0364]

[0365]

[0366] ​​In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, SENet can learn the importance of different features in a personalized manner, effectively filtering out invalid noise, and thus improving the prediction accuracy of the deep learning model.

[0367] Optionally, based on the corresponding embodiment above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiment of the present application, the feature weighting network includes X gating networks and X multi-layer perceptrons, where X≥1 and X is an integer; Figure 19 The obtaining module 310 is specifically configured to obtain the screened feature vectors corresponding to the screened data, where the screened data is the data screened from the object-associated data and the content data;

[0368] Based on the screened feature vectors, X gating feature vectors are obtained through X gating networks, where the input of each gating network is the screened feature vector;

[0369] Based on the comprehensive feature vector and the X gating feature vectors, a target comprehensive feature vector is obtained through X multi-layer perceptrons, where the input of each multi-layer perceptron is different.

[0370] In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, PPNet supports the personalized ability of the MLP network parameters through the gating network, thereby improving the estimation ability for the target.

[0371] Optionally, based on the corresponding embodiment above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiment of the present application,

[0372] The obtaining module 310 is specifically configured to, for each task network, perform element-wise multiplication on each corresponding expert vector among the N expert vectors by using each weight value in the activity weight vector, where the activity weight vector includes N weight values, or the activity weight vector and the expert vector have the same dimension. Figure 19 In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, in order to fully explore the data distribution differences of different activity objects, an activity gating network can be learned based on the activity features, and the output of the activity gating network is used to weight each expert network, thereby effectively strengthening the personalized features related to the object activity and alleviating the phenomenon that the model is dominated by high-activity objects.

[0373] Optionally, based on the corresponding embodiment above,

[0374]

[0375] Figure 19 ​​Based on the corresponding embodiments, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiments of the present application,

[0376] An acquisition module 310, specifically configured to obtain an activity weight vector through an activity gating network based on N expert vectors and an activity feature vector, where the activity gating network is a network using an attention mechanism, and the activity weight vector includes N weight values;

[0377] The acquisition module 310 is specifically configured to, for each task network, multiply each of the N expert vectors corresponding to each weight value in the activity weight vector element - by - element to obtain N weighted expert vectors.

[0378] In the embodiments of the present application, a content recommendation device based on a deep learning model is provided. By using the above - mentioned device, for feature vectors with a large number of dimensions, important features can be strengthened, weighted for each expert network, thereby effectively strengthening personalized features related to object activity and alleviating the phenomenon that the model is dominated by high - activity objects.

[0379] Optionally, based on the corresponding embodiments above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiments of the present application, the content recommendation device 30 based on a deep learning model further includes a determination module 330; Figure 19 Based on the corresponding embodiments above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiments of the present application, the content recommendation device 30 based on a deep learning model further includes a determination module 330;

[0380] The determination module 330 is configured to, after obtaining the activity weight vector through the activity gating network based on the activity feature vector, for each task network, determine the largest P weight values from the activity weight vector, where the activity weight vector includes N weight values, each weight value corresponds to an expert network, N > P ≥ 1, and P is an integer;

[0381] For each task network, determine the corresponding P expert networks from the N expert networks according to the P weight values;

[0382] The acquisition module 310 is specifically configured to, for each task network, multiply each of the P expert vectors corresponding to each of the P weight values element - by - element to obtain P weighted expert vectors, where the P expert vectors are the output results of the P expert networks.

[0383] In the embodiments of the present application, a content recommendation device based on a deep learning model is provided. By using the above - mentioned device, considering that a task network may include a large number of expert networks, therefore, based on the activity weight vector output by the sparse gating network, it is possible to further select which expert vectors output by the expert networks are weighted. Thus, the amount of calculation can be saved, the model capacity can be improved, and the calculation speed is fast.

[0384] Optionally, based on the corresponding embodiments described above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiments of the present application, K>1, and the N expert networks include a shared expert network and exclusive expert networks, and each task network further includes a task gating network; Figure 19 The obtaining module 310 is specifically configured to, for each task network, obtain a task feature vector through the task gating network based on the comprehensive feature vector;

[0385] For each task network, perform element-wise multiplication of the task feature vector and at least one weighted expert vector to obtain a target feature vector;

[0386] For each task network, obtain a task score of the candidate content through the prediction network based on the target feature vector.

[0387] In the embodiments of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, based on the PLE multi-objective ranking model, a feature weighting network and an activity gating network are added thereto. In order to simultaneously model multiple objectives, the PLE model can be selected as the basic model. The PLE model includes a shared expert network for multiple objectives and also provides exclusive expert networks unique to different objectives, effectively alleviating the "seesaw" effect that may occur between different objectives.

[0388] Optionally, based on the corresponding embodiments described above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiments of the present application, K>1, and the N expert networks are N shared expert networks, and each task network further includes a task gating network;

[0389] The obtaining module 310 is specifically configured to, for each task network, obtain a task feature vector through the task gating network based on the comprehensive feature vector; Figure 19

[0390] For each task network, perform element-wise multiplication of the task feature vector and at least one weighted expert vector to obtain a target feature vector;

[0391] For each task network, obtain a task score of the candidate content through the prediction network based on the target feature vector.

[0392]

[0393] ​​In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, based on the MMoE multi-object ranking model, a feature weighting network and an activity gating network are added thereto. In order to model multiple objectives simultaneously, the MMoE model can be selected as the basic model. The MMoE model includes a shared expert network for multiple objectives, which is used to learn the correlation between subtasks, thereby facilitating the improvement of the prediction accuracy.

[0394] Optionally, on the basis of the corresponding embodiment above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiment of the present application, K>1, and the N expert networks are N shared expert networks; Figure 19 The obtaining module 310 is specifically configured to, for each task network, perform matrix transformation on at least one weighted expert vector through a matrix set to obtain a target feature vector, where the matrix set includes a transformation matrix or an identity matrix;

[0395] For each task network, based on the target feature vector, obtain the task score of the candidate content through the prediction network.

[0396] In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, based on the SNR multi-object ranking model, a feature weighting network and an activity gating network are added thereto. In order to model multiple objectives simultaneously, the SNR model can be selected as the basic model. The SNR model includes a shared expert network for multiple objectives, which is used to learn the correlation between subtasks. In addition, the SNR model can achieve more flexible parameter sharing while maintaining the computational advantages of the classical multi-task model. The shared low-level hidden layer is modularized into multiple sub-networks, and the connection between the sub-networks and the learnable latent variables is controlled to achieve flexible parameter sharing.

[0397] Optionally, on the basis of the corresponding embodiment above, in another embodiment of the content recommendation device 30 for a deep learning model provided by the embodiment of the present application,

[0398] The determining module 330 is further configured to, after obtaining the task scores of the candidate content through the prediction network based on at least one weighted expert vector, determine the ranking score of the candidate content according to the K task scores; Figure 19 The obtaining module 310 is further configured to obtain the ranking scores of each candidate content among other candidate contents, where the other candidate contents are from the candidate content set of the recommended object;

[0399]

[0400]

[0401] ​​The determination module 330 is further configured to determine the score ranking result of the candidate content according to the ranking score of the candidate content and the ranking scores of each candidate content among other candidate contents;

[0402] The determination module 330 is further configured to determine that the candidate content belongs to the content to be pushed if the score ranking result of the candidate content is within the recommended ranking range.

[0403] In the embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, according to the score ranking result of each candidate content, the candidate content with a more forward ranking is preferentially pushed to the recommended object, thereby improving the accuracy of content recommendation.

[0404] Optionally, on the basis of the corresponding embodiment above, Figure 19 In another embodiment of the content recommendation device 30 based on a deep learning model provided in the embodiment of the present application,

[0405] The determination module 330 is specifically configured to use the K task scores as the ranking scores of the candidate content if K = 1;

[0406] If K > 1, the ranking score of the candidate content is calculated according to the K task scores and the task weights of each task score among the K task scores.

[0407] In the embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, one or more task types are set according to the business scenario, and then the ranking score is calculated by combining the task scores corresponding to each task type, and this is used as the basis for evaluating whether the candidate content is suitable for recommendation, thereby increasing the adaptability and flexibility of the solution.

[0408] Optionally, on the basis of the corresponding embodiment above, Figure 19 In another embodiment of the content recommendation device 30 based on a deep learning model provided in the embodiment of the present application, the K task scores include a duration score, an interaction score, a completion rate score, and a fast swipe rate score;

[0409] The determination module 330 is specifically configured to weight the duration score with the duration weight to obtain a first score;

[0410] Weight the interaction score with the interaction weight to obtain a second score;

[0411] Weight the completion rate score with the completion rate weight to obtain a third score;

[0412] Weight the fast swipe rate score with the fast swipe rate weight to obtain a fourth score;

[0413] According to the first score, the second score, the third score, and the fourth score, the ranking score of the candidate content is calculated. Among them, the ranking score of the candidate content is positively correlated with the duration score, the interaction score, and the completion rate score, and the ranking score of the candidate content is negatively correlated with the fast skip rate score.

[0414] In the embodiments of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, the recommendation situation of the content is jointly predicted from the perspectives of the satisfaction rate, the interaction rate, the completion rate, and the fast skip rate, thereby improving the prediction accuracy and being applicable to specific services.

[0415] Optionally, based on the corresponding embodiments above Figure 19 In another embodiment of the content recommendation device 30 based on the deep learning model provided in the embodiments of the present application, the content recommendation device 30 based on the deep learning model further includes a processing module 340 and a training module 350;

[0416] The acquisition module 310 is further configured to acquire a set of data to be trained, where each piece of data to be trained includes object data of a sample object, content data of a sample content, and K task labels;

[0417] The processing module 340 is configured to perform feature extraction on the object data of the sample object and the content data of the sample content for each piece of data to be trained, to obtain a sample feature vector of the data to be trained, where the sample feature vector includes a comprehensive sample feature vector and an activity sample feature vector;

[0418] The acquisition module 310 is further configured to obtain K prediction scores of each piece of data to be trained through the deep learning model based on the sample feature vector of each piece of data to be trained;

[0419] The determination module 330 is further configured to determine a total loss result according to the K prediction scores of each piece of data to be trained and the corresponding K task labels;

[0420] The training module 350 is configured to update the model parameters of the deep learning model according to the total loss result.

[0421] In the embodiments of the present application, a content recommendation device based on a deep learning model is provided. By using the above device, during the model training process, there is no need to perform threshold division according to the object activity, and the normalized feature values can be directly used for model training, avoiding the negative impact on the model caused by improper threshold selection. In addition, on the basis of the same modeling, the object activity is explicitly introduced for modeling, so as to take into account the data distribution differences of different activity object groups, and the same amount of machine resources as the unified modeling method can be used to serve online inference, thereby effectively improving the accuracy and efficiency of the recommendation system.

[0422] Optionally, based on the above Figure 19 In another embodiment of the content recommendation device 30 for a deep learning model provided in an embodiment of the present application,

[0423] The determination module 330 is specifically configured to, for each of the K tasks, calculate a task loss result by using a first loss function according to the prediction score of each training data to be trained and the corresponding task label;

[0424] If K = 1, the task loss result is used as the total loss result;

[0425] If K > 1, the task loss results of each task are summed to obtain the total loss result;

[0426] Or,

[0427] The determination module 330 is specifically configured to, for each of the K tasks, calculate a task loss result by using a second loss function according to the sample weight, the corresponding prediction score, and the corresponding task label of each training data to be trained, where the sample weight is related to the historical behavior data of the sample object on the sample content;

[0428] If K = 1, the task loss result is used as the total loss result;

[0429] If K > 1, the task loss results of each task are summed to obtain the total loss result.

[0430] In an embodiment of the present application, a content recommendation device based on a deep learning model is provided. By using the above device and training the model with a weighted cross-entropy loss function, it is beneficial to strengthen the influence of object activity on the prediction accuracy of the model.

[0431] Figure 20 FIG. is a schematic structural diagram of a computer device provided in an embodiment of the present application. The computer device 400 may vary greatly due to configuration or performance differences, and may include one or more central processing units (CPUs) 422 (for example, one or more processors) and a memory 432, and one or more storage media 430 (for example, one or more mass storage devices) for storing application programs 442 or data 444. Among them, the memory 432 and the storage media 430 may be transient storage or persistent storage. The program stored in the storage media 430 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the computer device. Further, the central processing unit 422 may be configured to communicate with the storage media 430 and execute a series of instruction operations in the storage media 430 on the computer device 400.

[0432] The computer device 400 may also include one or more power supplies 426, one or more wired or wireless network interfaces 450, one or more input / output interfaces 458, and / or one or more operating systems 441, such as Windows Server TM , Mac OS X TM , Uni TM , Linux TM , FreeBSD TM and so on.

[0433] The steps performed by the computer device in the foregoing embodiments may be based on the Figure 20 computer device structure shown.

[0434] In an embodiment of the present application, there is also provided a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the methods described in the foregoing embodiments are implemented.

[0435] In an embodiment of the present application, there is also provided a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the methods described in the foregoing embodiments are implemented.

[0436] It can be understood that in the specific implementation of the present application, for relevant data such as object data of the recommended object, when the foregoing embodiments of the present 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 relevant laws, regulations, and standards of relevant countries and regions.

[0437] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above may refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0438] In several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods may be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed couplings, direct couplings, or communication connections to each other may be through some interfaces, and the indirect couplings or communication connections of devices or units may be in electrical, mechanical, or other forms.

[0439] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0440] In addition, each functional unit in various embodiments of the present application can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0441] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store computer programs.

[0442] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of the present application.

Claims

1. A content recommendation method based on a deep learning model, characterized in that, The deep learning model includes an activity gating network and K task networks. Each task network includes N expert networks and a prediction network, where K ≥ 1, N ≥ 1, and both K and N are integers. The content recommendation method includes: In response to a content recommendation request for a recommended object, obtaining a comprehensive feature vector and an activity feature vector. Among them, the comprehensive feature vector includes the feature vector of the recommended object and the feature vector of the candidate content. The activity feature vector characterizes the activity of the recommended object during a historical period, and the activity feature vector is obtained based on object activity data, where the object activity data includes any one or a combination of the number of liked contents by the object during the historical period, the number of commented contents by the object during the historical period, the number of shared contents by the object during the historical period, and the number of coin - voted contents by the object during the historical period; For each task network, based on the comprehensive feature vector, obtaining N expert vectors through the N expert networks; Based on the activity feature vector, obtaining an activity weight vector through the activity gating network, which includes: based on the N expert vectors and the activity feature vector, obtaining the activity weight vector through the activity gating network. Among them, the activity gating network is a network using an attention mechanism, and the activity weight vector includes N weight values; For each task network, performing element - wise multiplication on at least one of the N expert vectors using the activity weight vector, which includes: for each task network, performing element - wise multiplication on each of the N expert vectors corresponding to the N weight values in the activity weight vector to obtain N weighted expert vectors; For each task network, based on the at least one weighted expert vector, obtaining a task score of the candidate content through the prediction network; If it is determined that the candidate content belongs to the content to be pushed according to the K task scores, then pushing the candidate content to the recommended object, where the K task scores are the output results of the K task networks.

2. The content recommendation method according to claim 1, wherein The step of "In response to a content recommendation request for a recommended object, obtaining a comprehensive feature vector and an activity feature vector" includes: In response to a content recommendation request for a recommended object, obtaining the object data of the recommended object, where the object data includes object activity data and object association data, and the object association data includes one or more of object basic data, object consumption sequence data, and object description data; Obtaining candidate content from the candidate content set corresponding to the recommended object, where the candidate content set includes Q candidate contents, Q ≥ 1, and Q is an integer; Obtaining the content data of the candidate content, where the content data includes one or more of content basic data and content statistical data; Performing feature processing on the object association data and the content data to obtain the comprehensive feature vector; Performing feature processing on the object activity data to obtain the activity feature vector.

3. The content recommendation method according to claim 1, wherein The deep learning model further includes a feature weighting network; After the response to the content recommendation request for the recommended object and obtaining the comprehensive feature vector and the activity feature vector, the method further includes: Based on the comprehensive feature vector, obtaining a target comprehensive feature vector through the feature weighting network; For each of the task networks, based on the comprehensive feature vector, obtaining N expert vectors through the N expert networks, including: For each of the task networks, based on the target comprehensive feature vector, obtaining N expert vectors through the N expert networks.

4. The content recommendation method according to claim 3, wherein The obtaining of the target comprehensive feature vector through the feature weighting network based on the comprehensive feature vector includes: Performing pooling processing on the feature vectors of each feature domain in the comprehensive feature vector to obtain a compressed feature vector, where the comprehensive feature vector includes feature vectors of T feature domains, the compressed feature vector includes T elements, T>1, and T is an integer; Based on the compressed feature vector, obtaining a feature weight vector through an activation network, where the feature weight vector includes T elements; Using the feature weight vector to perform element-wise multiplication on the feature vectors of the T feature domains to obtain the target comprehensive feature vector.

5. The content recommendation method according to claim 3, wherein The feature weighting network includes X gating networks and X multi-layer perceptrons, X≥1, and X is an integer; The obtaining of the target comprehensive feature vector through the feature weighting network based on the comprehensive feature vector includes: Obtaining the screened feature vector corresponding to the screened data, where the screened data is data screened from the object-associated data and the content data; Based on the screened feature vector, obtaining X gating feature vectors through the X gating networks, where the input of each gating network is the screened feature vector; Based on the comprehensive feature vector and the X gating feature vectors, obtaining the target comprehensive feature vector through the X multi-layer perceptrons, where the inputs of each multi-layer perceptron are different.

6. The content recommendation method according to any one of claims 1 to 5, characterized in that For each of the task networks, performing element-wise multiplication on at least one of the N expert vectors using the activity weight vector to obtain at least one weighted expert vector, including: For each of the task networks, performing element-wise multiplication on each of the N expert vectors using each weight value in the activity weight vector to obtain N weighted expert vectors, where the activity weight vector includes N weight values, or the activity weight vector has the same dimension as the expert vector.

7. The content recommendation method according to any one of claims 1 to 5, characterized in that, After obtaining the activity weight vector through the activity gating network based on the activity feature vector, the method further includes: For each of the task networks, determining the largest P weight values from the activity weight vector, where the activity weight vector includes N weight values, each weight value corresponds to an expert network, N>P≥1, and P is an integer; For each of the task networks, determining the corresponding P expert networks from the N expert networks according to the P weight values; For each of the task networks, multiplying at least one of the N expert vectors by the activity weight vector to obtain at least one weighted expert vector, includes: For each of the task networks, multiplying each of the P weight values by the corresponding expert vectors among the P expert vectors to obtain P weighted expert vectors, where the P expert vectors are the output results of the P expert networks.

8. The content recommendation method according to claim 1, characterized in that, K > 1, and the N expert networks include a shared expert network and exclusive expert networks, and each of the task networks further includes a task gating network; For each of the task networks, obtaining the task score of the candidate content through the prediction network based on the at least one weighted expert vector, includes: For each of the task networks, obtaining a task feature vector through the task gating network based on the comprehensive feature vector; For each of the task networks, multiplying the task feature vector by the at least one weighted expert vector to obtain a target feature vector; For each of the task networks, obtaining the task score of the candidate content through the prediction network based on the target feature vector.

9. The content recommendation method according to claim 1, characterized in that, K > 1, and the N expert networks are N shared expert networks, and each of the task networks further includes a task gating network; For each of the task networks, obtaining the task score of the candidate content through the prediction network based on the at least one weighted expert vector, includes: For each of the task networks, obtaining a task feature vector through the task gating network based on the comprehensive feature vector; For each of the task networks, multiplying the task feature vector by the at least one weighted expert vector to obtain a target feature vector; For each of the task networks, obtaining the task score of the candidate content through the prediction network based on the target feature vector.

10. The content recommendation method according to claim 1, wherein K > 1, and the N expert networks are N shared expert networks; For each of the task networks, obtaining the task score of the candidate content through the prediction network based on the at least one weighted expert vector, includes: For each of the task networks, performing a matrix transformation on the at least one weighted expert vector through a matrix set to obtain a target feature vector, where the matrix set includes a transformation matrix or an identity matrix; For each of the task networks, obtaining the task score of the candidate content through the prediction network based on the target feature vector.

11. The content recommendation method according to claim 1, wherein After obtaining the task score of the candidate content through the prediction network based on the at least one weighted expert vector for each of the task networks, the method further includes: Determining a ranking score of the candidate content according to the K task scores; Obtaining the ranking scores of each of the other candidate contents, where the other candidate contents are from the candidate content set of the recommended object; Determining a score ranking result of the candidate content according to the ranking score of the candidate content and the ranking scores of each of the other candidate contents; If the score ranking result of the candidate content is within the recommended ranking range, it is determined that the candidate content belongs to the content to be pushed.

12. The content recommendation method according to claim 11, wherein Determining the ranking score of the candidate content according to the K task scores includes: If K = 1, use the K task scores as the ranking score of the candidate content; If K > 1, calculate the ranking score of the candidate content according to the K task scores and the task weights of each task score among the K task scores.

13. The content recommendation method according to claim 12, wherein The K task scores include duration score, interaction score, completion rate score, and fast swipe rate score; Calculating the ranking score of the candidate content according to the K task scores and the task weights of each task score among the K task scores includes: Weight the duration score with the duration weight to obtain the first score; Weight the interaction score with the interaction weight to obtain the second score; Weight the completion rate score with the completion rate weight to obtain the third score; Weight the fast swipe rate score with the fast swipe rate weight to obtain the fourth score; Calculate the ranking score of the candidate content according to the first score, the second score, the third score, and the fourth score, where the ranking score of the candidate content is positively correlated with the duration score, the interaction score, and the completion rate score, and the ranking score of the candidate content is negatively correlated with the fast swipe rate score.

14. The content recommendation method according to any one of claims 1 to 13, characterized in that The method further includes: Obtain a set of data to be trained, where each piece of data to be trained includes object data of a sample object, content data of a sample content, and K task labels; For each piece of data to be trained, perform feature extraction on the object data of the sample object and the content data of the sample content to obtain a sample feature vector of the data to be trained, where the sample feature vector includes a comprehensive sample feature vector and an activity sample feature vector; Based on the sample feature vectors of each piece of data to be trained, obtain the K predicted scores of each piece of data to be trained through the deep learning model; Determine the total loss result according to the K predicted scores of each piece of data to be trained and the corresponding K task labels; Update the model parameters of the deep learning model according to the total loss result.

15. The content recommendation method according to claim 14, wherein Determining the total loss result according to the K predicted scores of each piece of data to be trained and the corresponding K task labels includes: For each task among the K tasks, calculate the task loss result by using the first loss function according to the predicted score of each piece of data to be trained and the corresponding task label; If K = 1, use the task loss result as the total loss result; If K > 1, sum the task loss results of each task to obtain the total loss result; Or, Determining the total loss result according to the K predicted scores of each piece of data to be trained and the corresponding K task labels includes: For each of the K tasks, according to the sample weights, corresponding prediction scores, and corresponding task labels of each piece of data to be trained, a task loss result is calculated using a second loss function, where the sample weights are related to the historical behavior data of the sample object with respect to the sample content; If K = 1, then the task loss result is used as the total loss result; If K > 1, then the task loss results of each task are summed to obtain the total loss result.

16. A content recommendation device based on a deep learning model, characterized in that, The deep learning model includes an activity gating network and K task networks. Each task network includes N expert networks and a prediction network, where K ≥ 1, N ≥ 1, and both K and N are integers. The content recommendation device includes: An acquisition module, configured to respond to a content recommendation request for a recommended object, and acquire a comprehensive feature vector and an activity feature vector, where the comprehensive feature vector includes the feature vector of the recommended object and the feature vector of the candidate content, the activity feature vector characterizes the activity of the recommended object in a historical period, the activity feature vector is obtained based on object activity data, and the object activity data includes any one or a combination of the number of content liked by the object in the historical period, the number of content commented by the object in the historical period, the number of content shared by the object in the historical period, and the number of content coin - voted by the object in the historical period; The acquisition module is further configured to, for each task network, based on the comprehensive feature vector, obtain N expert vectors through the N expert networks; The acquisition module is further configured to, based on the activity feature vector, obtain an activity weight vector through the activity gating network, including: obtaining the activity weight vector based on the N expert vectors and the activity feature vector through the activity gating network, where the activity gating network is a network using an attention mechanism, and the activity weight vector includes N weight values; The acquisition module is further configured to, for each task network, perform element - wise multiplication on at least one of the N expert vectors using the activity weight vector, including: for each task network, performing element - wise multiplication on each of the N expert vectors corresponding to each weight value in the activity weight vector to obtain N weighted expert vectors; The acquisition module is further configured to, for each task network, obtain the task score of the candidate content through the prediction network based on the at least one weighted expert vector; A recommendation module, configured to, if it is determined according to the K task scores that the candidate content belongs to the content to be pushed, push the candidate content to the recommended object, where the K task scores are the output results of the K task networks.

17. The device according to claim 16, characterized in that, The acquisition module is specifically configured to: In response to a content recommendation request for a recommended object, obtain object data of the recommended object, where the object data includes object activity data and object association data, and the object association data includes one or more of object basic data, object consumption sequence data, and object description data; Obtain candidate content from the candidate content set corresponding to the recommended object, where the candidate content set includes Q candidate contents, Q≥1, and Q is an integer; Obtain content data of the candidate content, where the content data includes one or more of content basic data and content statistical data; Perform feature extraction on the object association data and the content data to obtain the comprehensive feature vector; Perform feature extraction on the object activity data to obtain the activity feature vector.

18. The device according to claim 16, characterized in that, The deep learning model further includes a feature weighting network; The obtaining module is further configured to, after obtaining the comprehensive feature vector and the activity feature vector in response to a content recommendation request for a recommended object, obtain a target comprehensive feature vector based on the comprehensive feature vector through the feature weighting network; The obtaining module is specifically configured to: For each task network, obtain N expert vectors through the N expert networks based on the target comprehensive feature vector.

19. The device according to claim 18, characterized in that, The obtaining module is specifically configured to: Perform pooling processing on the feature vectors of each feature domain in the comprehensive feature vector to obtain a compressed feature vector, where the comprehensive feature vector includes feature vectors of T feature domains, the compressed feature vector includes T elements, T>1, and T is an integer; Based on the compressed feature vector, obtain a feature weight vector through an excitation network, where the feature weight vector includes T elements; Multiply the feature vectors of the T feature domains element by element using the feature weight vector to obtain the target comprehensive feature vector.

20. The device according to claim 18, characterized in that, The feature weighting network includes X gating networks and X multi-layer perceptrons, X≥1, and X is an integer; The obtaining module is specifically configured to: Obtain a screened feature vector corresponding to the screened data, where the screened data is data screened from the object association data and the content data; Based on the screened feature vector, obtain X gating feature vectors through the X gating networks, where the input of each gating network is the screened feature vector; Based on the comprehensive feature vector and the X gating feature vectors, obtain the target comprehensive feature vector through the X multi-layer perceptrons, where the inputs of each multi-layer perceptron are different.

21. The device according to any one of claims 16 to 20, characterized in that, The obtaining module is specifically configured to: For each task network, multiply each expert vector in the N expert vectors element by element using each weight value in the activity weight vector, where the activity weight vector includes N weight values, or the activity weight vector has the same dimension as the expert vector.

22. The device according to any one of claims 16 to 20, characterized in that, The device further includes a determination module; The determining module is configured to, after obtaining an activity weight vector through the activity gating network based on the activity feature vector, for each of the task networks, determine the largest P weight values from the activity weight vector, where the activity weight vector includes N weight values, each weight value corresponding to an expert network, N > P ≥ 1, and P is an integer; for each of the task networks, determine the corresponding P expert networks from the N expert networks according to the P weight values; The obtaining module is specifically configured to: For each of the task networks, use each of the P weight values to perform element-wise multiplication on the corresponding expert vectors among the P expert vectors to obtain P weighted expert vectors, where the P expert vectors are the output results of the P expert networks.

23. The device according to claim 16, wherein K > 1, and the N expert networks include shared expert networks and exclusive expert networks, and each of the task networks further includes a task gating network; The obtaining module is specifically configured to: For each of the task networks, obtain a task feature vector through the task gating network based on the comprehensive feature vector; For each of the task networks, perform element-wise multiplication on the task feature vector and the at least one weighted expert vector to obtain a target feature vector; For each of the task networks, obtain a task score of the candidate content through the prediction network based on the target feature vector.

24. The device according to claim 16, characterized in that, K > 1, and the N expert networks are N shared expert networks, and each of the task networks further includes a task gating network; The obtaining module is specifically configured to: For each of the task networks, obtain a task feature vector through the task gating network based on the comprehensive feature vector; For each of the task networks, perform element-wise multiplication on the task feature vector and the at least one weighted expert vector to obtain a target feature vector; For each of the task networks, obtain a task score of the candidate content through the prediction network based on the target feature vector.

25. The device according to claim 16, characterized in that, K > 1, and the N expert networks are N shared expert networks; The obtaining module is specifically configured to: For each of the task networks, perform matrix transformation on the at least one weighted expert vector through a matrix set to obtain a target feature vector, where the matrix set includes a transformation matrix or an identity matrix; For each of the task networks, obtain a task score of the candidate content through the prediction network based on the target feature vector.

26. The device according to claim 16, characterized in that, The apparatus further includes a determining module; The determining module is configured to, for each of the task networks, after obtaining a task score of the candidate content through the prediction network based on the at least one weighted expert vector, determine a ranking score of the candidate content according to the K task scores; The obtaining module is further configured to obtain ranking scores of each of the other candidate contents among the other candidate contents, where the other candidate contents are from the candidate content set of the recommended object; The determining module is further configured to determine a score ranking result of the candidate content according to the ranking score of the candidate content and the ranking scores of each candidate content in the other candidate contents; if the score ranking result of the candidate content is within the recommended ranking range, it is determined that the candidate content belongs to the content to be pushed.

27. The device according to claim 26, wherein The determining module is specifically configured to: If K = 1, use the K task scores as the ranking score of the candidate content; If K > 1, calculate the ranking score of the candidate content according to the K task scores and the task weights of each task score in the K task scores.

28. The device according to claim 27, characterized in that, The K task scores include a duration score, an interaction score, a completion rate score, and a fast skip rate score; The determining module is specifically configured to: Weight the duration score with a duration weight to obtain a first score; Weight the interaction score with an interaction weight to obtain a second score; Weight the completion rate score with a completion rate weight to obtain a third score; Weight the fast skip rate score with a fast skip rate weight to obtain a fourth score; Calculate the ranking score of the candidate content according to the first score, the second score, the third score, and the fourth score, where the ranking score of the candidate content is positively correlated with the duration score, the interaction score, and the completion rate score, and the ranking score of the candidate content is negatively correlated with the fast skip rate score.

29. The device according to any one of claims 16 to 28, characterized in that, The device further includes a processing module, a determining module, and a training module; The obtaining module is further configured to obtain a set of data to be trained, where each piece of data to be trained includes object data of a sample object, content data of a sample content, and K task labels; The processing module is configured to perform feature extraction on the object data of the sample object and the content data of the sample content for each piece of data to be trained to obtain a sample feature vector of the data to be trained, where the sample feature vector includes a comprehensive sample feature vector and an activity sample feature vector; The obtaining module is further configured to obtain K prediction scores of each piece of data to be trained through the deep learning model based on the sample feature vector of each piece of data to be trained; The determining module is configured to determine a total loss result according to the K prediction scores of each piece of data to be trained and the corresponding K task labels; The training module is configured to update the model parameters of the deep learning model according to the total loss result.

30. The device according to claim 29, characterized in that, The determining module is specifically configured to: For each of the K tasks, calculate a task loss result by using a first loss function according to the prediction score of each piece of data to be trained and the corresponding task label; If K = 1, use the task loss result as the total loss result; If K > 1, sum up the task loss results of each task to obtain the total loss result; Or, The determining module is specifically configured to: For each of the K tasks, according to the sample weights, corresponding prediction scores, and corresponding task labels of each piece of training data to be trained, a task loss result is calculated using a second loss function, where the sample weights are related to the historical behavior data of the sample object with respect to the sample content; If K = 1, then the task loss result is used as the total loss result; If K > 1, then the task loss results of each task are summed to obtain the total loss result.

31. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the content recommendation method described in any one of claims 1 to 15 are implemented.

32. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the content recommendation method described in any one of claims 1 to 15 are implemented.

33. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, the steps of the content recommendation method described in any one of claims 1 to 15 are implemented.

Images