Recommendation method and device, equipment, storage medium and program product

Abstract

The invention provides a recommendation method and device, equipment, a storage medium and a product, and the method comprises the steps: determining a group of weight parameters corresponding to a target user through a gating network in an MTL model based on user data; for any first network layer in the MTL model, adjusting first output data of the first network layer based on a group of weight parameters output by a gating network corresponding to the first network layer to obtain second output data of the first network layer after being adjusted by the gating network; adjusting first output data of at least one first network layer in the MTL model through at least one gating network to obtain an interested degree of the adjusted to-be-recommended material output by the MTL model by a target user; and recommending materials to the target user based on the adjusted interested degree. Through the application, the problem that the recommendation result of the low-frequency user is dominated by the high-frequency user can be solved.

Description

Technical Field

[0001] This application relates to the technical field of information flow recommendation, and particularly to a recommendation method, apparatus, device, storage medium and program product. Background Art

[0002] The recommendation model captures users' interests by learning a large number of interactions between users and items, and recommends items that users are interested in. It is widely used in application software such as e-commerce, video or music. However, for low-frequency users who have no interaction behavior or few interaction behaviors with items, it is difficult for the recommendation model to capture their true preferences, and it is extremely easy to be dominated by the data of high-frequency users during the training process, resulting in the inability to accurately recommend items that low-frequency users are interested in. Summary of the Invention

[0003] This application provides a recommendation method, apparatus, device, storage medium and program product, which can effectively solve the problem that the recommendation results of low-frequency users are dominated by high-frequency users, and improve the recommendation accuracy of the recommendation model for low-frequency users.

[0004] In a first aspect, this application provides a recommendation method, including: in response to a recommendation request, obtaining user data of a target user and item data of an item to be recommended; based on the user data, determining a set of weight parameters corresponding to the target user through a gating network set in a multi-task learning (MTL) model; wherein, the gating network includes at least one, and each gating network corresponds to a first network layer in the MTL model, and each gating network is trained based on at least the user data of low-frequency sample users; the MTL model outputs the degree of interest of the target user in the item to be recommended based on the input user data and item data; for any first network layer in the MTL model, based on a set of weight parameters output by the gating network corresponding to the first network layer, adjusting the first output data of the first network layer to obtain second output data of the first network layer after being adjusted by the gating network; after at least one gating network adjusts the first output data of at least one first network layer in the MTL model, obtaining the adjusted degree of interest of the target user in the item to be recommended output by the MTL model; and recommending an item to the target user based on the adjusted degree of interest.

[0005] In some possible implementation manners, determining a set of weight parameters corresponding to the target user through a gating network set in the MTL model based on the user data includes: based on the user data, obtaining third output data of the last activation layer in the gating network through multiple activation layers in the gating network; multiplying the scaling coefficient of the gating network by the third output data to obtain a set of weight parameters output by the gating network; wherein, the scaling coefficient is obtained after training the gating network based on at least the user data of low-frequency sample users.

[0006] In some possible embodiments, when the number of gating networks is one, the gating network is a first type of gating network, and the first network layer corresponding to the first type of gating network is used to extract features from user data and material data; based on the user data, a set of weight parameters corresponding to the target user is determined through the gating network set in the MTL model, including: based on the user data and the consumption data of the target user, a set of weight parameters output by the first type of gating network is obtained through the first type of gating network; wherein, the consumption data is used to represent the consumption behavior of the target user on materials at different times in different recommendation scenarios.

[0007] In some possible embodiments, based on a set of weight parameters output by the gating network corresponding to the first network layer, the first output data of the first network layer is adjusted to obtain the second output data of the first network layer after being adjusted by the gating network, including: through the first network layer corresponding to the first type of gating network, features are extracted from user data and material data to obtain the first output data output by the first network layer; the first output data is multiplied element-wise by a set of weight parameters output by the first type of gating network to obtain the second output data.

[0008] In some possible embodiments, when the number of gating networks is multiple, the multiple gating networks include a first type of gating network and a second type of gating network. The first network layer corresponding to the first type of gating network is used to extract features from user data and material data, and the first network layer corresponding to the second type of gating network is used to perform data transformation processing on the input data; based on the user data, a set of weight parameters corresponding to the target user is determined through the gating network set in the MTL model, including: based on the user data and the consumption data of the target user, a set of weight parameters output by the first type of gating network is obtained through the first type of gating network; wherein, the consumption data is used to represent the consumption behavior of the target user on materials at different times in different recommendation scenarios; based on the user data and the material data, a set of weight parameters output by the second type of gating network is obtained through the second type of gating network.

[0009] In some possible embodiments, based on the user data and the material data, a set of weight parameters output by the second type of gating network is obtained through the second type of gating network, including: the user data, the material data, and the second output data of the first network layer corresponding to the first type of gating network are concatenated to obtain concatenated data; by inputting the concatenated data into the second type of gating network, a set of weight parameters output by the second type of gating network is obtained.

[0010] In some possible embodiments, each gating network includes multiple activation layers connected in series, the number of activation layers included in different gating networks is different, and / or, the activation functions of the activation layers included in different gating networks are different.

[0011] Second aspect, the present application provides a recommendation device, which includes: a data acquisition module, configured to acquire user data of a target user and material data of materials to be recommended in response to a recommendation request; a first data processing module, configured to determine a set of weight parameters corresponding to the target user based on the user data through a gating network set in a multi-task learning (MTL) model; wherein, the gating network includes at least one, and each gating network corresponds to a first network layer in the MTL model, and each gating network is trained based on at least user data of low-frequency sample users; the MTL model outputs the degree of interest of the target user in the materials to be recommended based on the input user data and material data; a second data processing module, configured to, for any first network layer in the MTL model, adjust the first output data of the first network layer based on a set of weight parameters output by the gating network corresponding to the first network layer to obtain second output data of the first network layer adjusted by the gating network; after at least one gating network adjusts the first output data of at least one first network layer in the MTL model, obtain the adjusted degree of interest of the target user in the materials to be recommended output by the MTL model; a recommendation module, configured to recommend materials to the target user based on the adjusted degree of interest.

[0012] In some possible implementation manners, the first data processing module is configured to, based on the user data, obtain third output data of the last activation layer in the gating network through multiple activation layers in the gating network; multiply the scaling coefficient of the gating network by the third output data to obtain a set of weight parameters output by the gating network; wherein, the scaling coefficient is obtained by training the gating network based on at least user data of low-frequency sample users.

[0013] In some possible implementation manners, when the number of gating networks is one, the gating network is a first type of gating network, and the first network layer corresponding to the first type of gating network is used to extract features from the user data and the material data. The first data processing module is configured to obtain a set of weight parameters output by the first type of gating network based on the user data and the consumption data of the target user through the first type of gating network; wherein, the consumption data is used to represent the consumption behavior of the target user on materials at different times in different recommendation scenarios.

[0014] In some possible implementation manners, the second data processing module is configured to extract features from the user data and the material data through the first network layer corresponding to the first type of gating network to obtain first output data output by the first network layer; perform a dot product of the set of weight parameters output by the first type of gating network and the first output data to obtain second output data.

[0015] In some possible embodiments, when the number of gating networks is multiple, the multiple gating networks include a first type of gating network and a second type of gating network. The first network layer corresponding to the first type of gating network is used to extract features from user data and material data, and the first network layer corresponding to the second type of gating network is used to perform data transformation processing on input data. A first data processing module is configured to obtain a set of weight parameters output by the first type of gating network based on user data and the consumption data of the target user through the first type of gating network, where the consumption data is used to represent the consumption behavior of the target user for materials at different times in different recommendation scenarios; and obtain a set of weight parameters output by the second type of gating network based on user data and material data through the second type of gating network.

[0016] In some possible embodiments, the first data processing module is configured to splice user data, material data, and the second output data of the first network layer corresponding to the first type of gating network to obtain spliced data; and obtain a set of weight parameters output by the second type of gating network by inputting the spliced data into the second type of gating network.

[0017] In some possible embodiments, each gating network includes multiple activation layers connected in series, and the number of activation layers included in different gating networks is different, and / or the activation functions of the activation layers included in different gating networks are different.

[0018] In a third aspect, the present application provides an electronic device, including: a memory for storing executable instructions; and a processor for implementing the method provided by the present application when executing the executable instructions stored in the memory.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium storing executable instructions for implementing the method provided by the present application when the executable instructions are executed by a processor.

[0020] In a fifth aspect, the present application provides a computer program product including a computer program or instructions for implementing the method provided by the present application when the computer program or instructions are executed by a processor.

[0021] The present application has the following beneficial effects:

[0022] In the present application, by obtaining the weight parameters corresponding to the target user through the gating network trained with the user data of low-frequency users, and adjusting the output of the MTL model with the weight parameters corresponding to the target user, the problem that the recommendation results of low-frequency users are dominated by high-frequency users can be effectively solved, and the recommendation accuracy of the recommendation model for low-frequency users can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a schematic diagram of the architecture of a recommendation system provided by an embodiment of the present application;

[0024] Figure 2 is a schematic structural diagram of an electronic device provided by an embodiment of the present application;

[0025] Figure 3 is an optional flowchart of a recommendation method provided by an embodiment of the present application;

[0026] Figure 4 is a schematic structural diagram of a gating network provided by an embodiment of the present application;

[0027] Figure 5 is a schematic structural diagram of another gating network provided by an embodiment of the present application;

[0028] Figure 6 is a schematic structural diagram of an MTL model provided by an embodiment of the present application;

[0029] Figure 7 is a schematic structural diagram of a recommendation model provided by an embodiment of the present application. Detailed implementation manners

[0030] In order to make the objectives, technical solutions, and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be construed as limiting the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present application.

[0031] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments. However, it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0032] If similar descriptions such as "first / second" appear in the application documents, the following explanation is added. In the following description, the terms "first / second / third" are only used to distinguish similar objects and do not represent a specific order for the objects. It can be understood that "first / second / third" can be interchanged with a specific order or sequence when permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0034] In a recommendation system, after a user browses the materials recommended by the recommendation system, the user will interact with the materials, generating various interaction behaviors such as clicks, forwards, likes, comments, collections, etc. The recommendation system often collects the user's interaction behaviors with the materials as samples to model recommendation models for different objectives such as click-through rate and interaction rate. Thus, when the user makes the next request, the recommendation models for different objectives are used to make a real-time prediction of the user's preferences, determine which materials in the material library the user is more interested in, and then recommend them to the user.

[0035] Recommendation models often require a large amount of user data for training in order to fully learn the user's preferences and thus achieve accurate prediction and recommendation. However, for low-frequency users or new users, on the one hand, the historical data of these users is scarce, resulting in less training data for the recommendation model, making it difficult for the recommendation model to learn the interest preferences of low-frequency users or new users. On the other hand, since the historical data of high-frequency users is much more than that of low-frequency users, the recommendation model is dominated by the data of high-frequency users during the training process, leading to poor recommendation results for low-frequency users.

[0036] The embodiments of the present application provide a recommendation method, device, equipment, storage medium, and program product, which solve the problem that the recommendation results of low-frequency users are dominated by high-frequency users, enabling better recommendation results for different types of users and effectively improving the robustness and stability of the recommendation model.

[0037] The following describes the exemplary applications of the electronic device provided in the embodiments of the present application. The electronic device provided in the embodiments of the present application can be various types of user terminals such as laptop computers, tablet computers, desktop computers, mobile devices (e.g., mobile phones, wearable smart watches, dedicated messaging devices), or can also be implemented as a server. Below, the exemplary application when the electronic device is implemented as a server will be described.

[0038] See Figure 1 , Figure 1 is a schematic diagram of the architecture of a recommendation system 100 provided in the embodiments of the present application. To implement the recommendation method of the embodiments of the present application, electronic devices 400-1 and 400-2 ( Figure 1 only two electronic devices are shown here, but it does not limit the number of electronic devices included in the recommendation system of the embodiments of the present application) are connected to the server 200 through the network 300. The network 300 can be a wide area network, a local area network, or a combination of the two.

[0039] In one embodiment, the server 200 is configured to obtain user data of a target user and material data of a target material from the database 500, and obtain at least one trained gating network and a multi-task learning (MTL) model from the database 500, wherein each gating network is trained based on user data of low-frequency users at least.

[0040] In one embodiment, the server 200 determines at least one set of weight parameters corresponding to the target user based on the user data through at least one gating network, wherein each gating network outputs a set of weight parameters, each gating network corresponds to each of at least one first network layer of the MTL model, through the MTL model, first output data of each first network layer is obtained based on the user data and the material data, and the first output data of each first network layer is adjusted respectively based on the at least one set of weight parameters to obtain second output data of each first network layer; based on the second output data, output data of the MTL model is determined, wherein the output data is used to represent the degree of interest of the target user in the target material; based on the output data and the materials to be recommended, materials are recommended to the target user.

[0041] In one embodiment, the terminal 400 is configured to display the materials recommended by the server on a graphical interface.

[0042] In some embodiments, the server 200 may be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and big data and artificial intelligence platforms. The electronic device 400 may be an electronic device with a display screen such as a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart watch, etc., but is not limited thereto. The electronic device and the server may be directly or indirectly connected through wired or wireless communication methods, which are not limited in the embodiments of the present invention.

[0043] Figure 2 It is a schematic structural diagram of the electronic device 400 provided by the embodiments of the present application. Figure 2 The illustrated electronic device 400 includes: at least one processor 410, a memory 450, at least one network interface 420, and a user interface 430. Each component in the electronic device 400 is coupled together through a bus system 440. It can be understood that the bus system 440 is used to realize the connection and communication between these components. The bus system 440 includes not only a data bus, but also a power bus, a control bus, and a status signal bus. However, for the sake of clarity, in Figure 2 all kinds of buses are labeled as the bus system 440.

[0044] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0045] The user interface 430 includes one or more output devices 431 that enable the presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, a mouse, a microphone, a touch screen display, a camera, and other input buttons and controls.

[0046] The memory 450 can be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disc drives, etc. The memory 450 optionally includes one or more storage devices that are physically remote from the processor 410.

[0047] The memory 450 includes volatile memory or non-volatile memory, and can also include both volatile and non-volatile memory. The non-volatile memory can be read only memory (ROM), and the volatile memory can be random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0048] In some embodiments, the memory 450 is capable of storing data to support various operations. Examples of such data include programs, modules, and data structures, or subsets or supersets thereof, which are described below by way of example.

[0049] The operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0050] The network communication module 452 is used to reach other computing devices via one or more (wired or wireless) network interfaces 420. Exemplary network interfaces 420 include: Bluetooth, Wireless Fidelity (WiFi), and Universal Serial Bus (USB), etc.;

[0051] A presentation module 453 for enabling presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 431 associated with the user interface 430 (e.g., a display screen, a speaker, etc.).

[0052] An input processing module 454 for detecting one or more user inputs or interactions from one of one or more input devices 432 and translating the detected inputs or interactions.

[0053] In some embodiments, the recommendation device 455 provided by the embodiments of the present application may be implemented in software. Figure 2 Shown is the recommendation device 455 stored in the memory 450, which may be software in the form of a program and a plug-in, etc., including the following software modules: a data acquisition module 4551, a first data processing module 4552, a second data processing module 4553, and a recommendation module 4554. These modules are logical, and thus can be combined arbitrarily or further split according to the functions implemented. The functions of each module will be described below.

[0054] In other embodiments, the recommendation device 455 provided by the embodiments of the present application may be implemented in hardware. As an example, the recommendation device 455 provided by the embodiments of the present application may be a processor in the form of a hardware decoding processor, which is programmed to execute the recommendation method provided by the embodiments of the present application. For example, a processor in the form of a hardware decoding processor may employ one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0055] Next, the recommendation method provided by the embodiments of the present application will be described in combination with the exemplary applications and implementations of the electronic device provided by the embodiments of the present application.

[0056] See Figure 3 , Figure 3 which is an optional flowchart of a recommendation method provided by the embodiments of the present application. Next, the recommendation method provided by the embodiments of the present application will be described in combination with Figure 3 the steps shown.

[0057] In the embodiments of the present disclosure, the recommendation method is executed by an electronic device, which may be a server or a terminal. The terminal includes, for example, at least one of a mobile phone, a smart watch, a wearable device, an Internet of Things device, a vehicle with communication function, a smart vehicle, a tablet computer (Pad), a computer with wireless transceiver function, a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal device in industrial control, a wireless terminal device in self-driving, a wireless terminal device in remote medical surgery, a wireless terminal device in smart grid, a wireless terminal device in transportation safety, a wireless terminal device in smart city, and a wireless terminal device in smart home, but is not limited thereto.

[0058] Step S301: In response to a recommendation request, obtain user data of a target user and material data of materials to be recommended.

[0059] In some embodiments, the electronic device may obtain the user data of the target user and the material data of the materials to be recommended from a database. The user data is data representing the user identity and user interest preferences. The material data is data representing the material type, the material interaction information, and the material publisher.

[0060] In some embodiments, the target user may be a low-frequency user, which is a user with historical behavior data less than a first threshold, and the low-frequency user may also be referred to as a cold start user. The target user may also be a high-frequency user, which is a user with historical behavior data greater than the first threshold, and the high-frequency user may also be referred to as a hot start user.

[0061] In some embodiments, the user data may include at least one of the following: identity data, preference data. Among them, the identity data includes at least one of the following: age data, gender data, height data, weight data, appearance data, education data, work data. The preference data includes the degree of interest of the user in different types of materials. Different types of materials may include text materials, image materials, graphic materials, video materials. Different types of materials may also include materials of multiple different topic types. For example, different types of materials include: materials associated with entertainment star topics, materials associated with current affairs and politics topics, materials associated with sports event topics, materials associated with technology and digital topics.

[0062] In some embodiments, the material data includes at least one of the following: material type data, click-through rate of the material at different times, interaction rate of the material at different times, and material publisher data. Among them, the click-through rate is the ratio of the number of times the material is clicked to the number of times the material is exposed. The interaction rate is the ratio of the number of times the user has an interaction behavior with the material to the number of times the material is exposed. The interaction behavior includes at least one of the following: commenting on the material, forwarding the material, liking the material, and favoriting the material.

[0063] Step S302: Based on the user data, determine a set of weight parameters corresponding to the target user through the gating network set in the MTL model.

[0064] In some embodiments, the gating network includes at least one, and each gating network corresponds to a first network layer in the MTL model. Each gating network is trained at least based on the user data of low-frequency sample users; the MTL model outputs the degree of interest of the target user in the material to be recommended based on the input user data and material data.

[0065] In some embodiments, the electronic device can obtain at least one trained gating network, input the user data of the target user into the at least one gating network, and obtain at least one set of weight parameters output by the at least one gating network, where each gating network outputs a set of weight parameters.

[0066] In some embodiments, each gating network can be trained at least based on the user data of low-frequency users. In some embodiments, the training data of different gating networks can be the same or different. In some embodiments, the training data can obtain the historical data of users and the historical data of materials from the database as training data, or can generate training data through a flink real-time task. For example, generate the consumption data of all users in various recommendation scenarios on Weibo in the past 3 days, 7 days, or 30 days through a flink real-time task, and use this consumption data as training data.

[0067] In some embodiments, when the training data of different gating networks is the same, each gating network can be trained based on the consumption data of low-frequency users. In some embodiments, each gating network can be trained based on the user data of low-frequency users and the user data of high-frequency users, or can be trained based on the user data of low-frequency users and the consumption data of low-frequency users, or can also be trained based on the user data of low-frequency users, the consumption data of low-frequency users, the user data of high-frequency users, and the consumption data of high-frequency users.

[0068] In some embodiments, when the training data of different gating networks is different, some of the multiple gating networks can be trained based on the user data of low-frequency users, and some of the other gating networks can be trained based on the user data of low-frequency users and the material data. In some embodiments, some of the multiple gating networks can be trained based on the user data of low-frequency users and the consumption data of low-frequency users, and some of the other gating networks can be trained based on the user data of low-frequency users and the material data. In some embodiments, some of the multiple gating networks can be trained based on the user data of low-frequency users and the user data of high-frequency users, and some of the other gating networks can be trained based on the user data of low-frequency users, the user data of high-frequency users, and the material data. In some embodiments, some of the multiple gating networks can be trained based on the user data of low-frequency users, the consumption data of low-frequency users, the user data of high-frequency users, and the consumption data of high-frequency users, and some of the other gating networks can be trained based on the user data of low-frequency users, the user data of high-frequency users, and the material data.

[0069] In some embodiments, the consumption data represents the consumption behavior of the target user for materials at different times in different recommendation scenarios, and the consumption behavior includes at least one of the following: browsing materials, clicking on materials, clicking on the comment area of materials, clicking on the author's homepage of materials, liking materials, commenting on materials, collecting materials, forwarding materials, paying to view materials, etc. The consumption data includes at least one of the following: the number of consumption days, the number of exposed materials, the number of refreshes, the number of clicks, the number of interactions, and the total consumption duration.

[0070] In one example, if the consumption data represents the consumption data of the user within 3 days, 7 days, and 30 days in different recommendation scenarios, then the number of consumption days includes the number of days the user consumes materials within 3 days, within 7 days, and within 30 days. The number of exposed materials includes the total number of materials the user browses within 3 days, within 7 days, and within 30 days. The number of refreshes includes the total number of times the user refreshes materials within 3 days, within 7 days, and within 30 days. The number of clicks includes the number of times the user clicks on materials within 3 days, within 7 days, and within 30 days. The number of interactions includes the number of interactions the user has with materials within 3 days, within 7 days, and within 30 days. The total consumption duration includes the total duration the user consumes materials within 3 days, within 7 days, and within 30 days.

[0071] Understandably, consumption data can accurately reflect the behavioral differences of different users (such as low-frequency users and high-frequency users) compared to other data. Therefore, by training the gating network with consumption data, users can be refined into different types, so that users with the same behavior habits are classified into the same type, thereby achieving the purpose of personalized recommendation for different types of users.

[0072] In some embodiments, each gating network includes multiple activation layers. The activation functions of the multiple activation layers can be the same or different. The activation function of each activation layer can be one of the following: sigmoid function, tanh function, rectified linear unit relu function, exponential linear unit ELU function, maxout function, swish function, mish function.

[0073] In some embodiments, each gating network includes multiple activation layers connected in series. The number of activation layers included in different gating networks is different, and / or the activation functions of the activation layers included in different gating networks are different.

[0074] In some embodiments, the network structures of different gating networks can be the same or different. In some embodiments, the number of activation layers included in different gating networks can be the same or different. In some embodiments, the activation functions of the activation layers included in different gating networks can be the same or different. In some embodiments, the input data of different gating networks can be the same or different.

[0075] In an example, it is assumed that there are a total of 3 gating networks, including: gating network 1 (gate1), gating network 2 (gate2), and gating network 3 (gate3). Gate1 includes 2 activation layers. The activation function of the first activation layer is the relu function, and the activation function of the second activation layer is the sigmoid function. Gate2 includes 2 activation layers. The activation function of the first activation layer is the ELU function, and the activation function of the second activation layer is the sigmoid function. Gate3 includes 3 activation layers. The activation function of the first activation layer is the swish function, the activation function of the second activation layer is the tanh function, and the activation function of the third activation layer is the mish function. That is to say, gate1 and gate2 include the same number of activation layers, but the activation functions of the activation layers are partially the same and partially different. Gate3 has a different number of activation layers from gate1 and gate2, and the activation functions are also different.

[0076] In one example, it is assumed that there are a total of 2 gating networks, including: gate1 and gate2. Gate1 includes 3 activation layers connected in series, and gate2 includes 3 activation layers. The first activation layer is connected in parallel with the second activation layer, the first activation layer is connected in series with the third activation layer, and the second activation layer is connected in series with the third activation layer.

[0077] In some embodiments, before inputting user data into the gating network, the user data can also be preprocessed. For example, the preprocessing parameters include w and b. If the input data of the gating network is represented by x, the preprocessing process can be expressed by the formula f(x) = wx + b.

[0078] In some embodiments, Figure 4 is a schematic structural diagram of a gating network provided by an embodiment of the present application. As Figure 4 shown, taking the gating network including two activation layers connected in series, where the activation function of the first activation layer is the relu function and the activation function of the second activation layer is the sigmoid function as an example, the input data x (for example, the input data x includes user data and consumption data, or the input data x includes user data and material data) passes through the first activation layer. The first activation layer performs cross-learning on the user data to obtain the output data y1. The second activation layer maps the output data y1 of the first activation layer to the interval (0, 1) to obtain the weight parameter g.

[0079] In some embodiments, step S302 further includes: based on the user data, obtaining the third output data of the last activation layer in the gating network through multiple activation layers in the gating network; multiplying the scaling coefficient of the gating network by the third output data to obtain a set of weight parameters output by the gating network. The scaling coefficient is obtained by training the gating network at least based on the user data of low-frequency sample users.

[0080] It can be understood that the gating network further includes a multiplier connected in series with the last activation layer. The electronic device inputs the user data into the gating network, processes the user data through multiple activation layers of the gating network to obtain the third output data of the last activation layer, and the third output data is multiplied by the scaling coefficient through the multiplier to obtain the weight output by the gating network.

[0081] Figure 5 is another schematic structural diagram of a gating network provided by an embodiment of the present application. As Figure 5As shown, taking the gating network including two activation layers connected in series and a multiplier connected to the last activation layer as an example, where the activation function of the first activation layer is the relu function and the activation function of the second activation layer is the sigmoid function, the input data x (for example, the input data x includes user data and consumption data, or the input data x includes user data and material data) passes through the first activation layer. The first activation layer performs cross-learning on the user data to obtain the output data y1. The second activation layer maps the output data y1 of the first activation layer to the interval (0, 1) to obtain the output data y2. y2 is multiplied by the scaling coefficient γ through the multiplier to obtain the weight parameter g.

[0082] In some embodiments, when the number of gating networks is one, the gating network is a first type of gating network, and the first network layer corresponding to the first type of gating network is used to extract features from user data and material data. The above step S302 includes: based on the user data and the consumption data of the target user, through the first type of gating network, obtaining a set of weight parameters output by the first type of gating network. Among them, the consumption data is used to represent the consumption behavior of the target user on materials at different times in different recommendation scenarios. That is to say, through the first type of gating network, based on the user data and the consumption data, the weight parameters output by the first type of gating network are obtained.

[0083] In some embodiments, the first type of gating network corresponds to the first network layer in the MTL model. That is to say, the first network layer corresponding to the first type of gating network is the first network layer in the MTL model. In some embodiments, the first network layer corresponding to the first type of gating network is used to extract features from the input data. That is to say, the weight parameters output by the first type of gating network are used to adjust the output data obtained after extracting features from the input data in the MTL model.

[0084] In some embodiments, if the number of gating networks is multiple, the gating network includes a first type of gating network and a second type of gating network. In some embodiments, the input data of the first type of gating network and the second type of gating network can be the same or different. In some embodiments, the network layers corresponding to the first type of gating network and the second type of network in the MTL model are different.

[0085] In some embodiments, when the number of gating networks is multiple, the multiple gating networks include a first type of gating network and a second type of gating network. The first network layer corresponding to the first type of gating network is used to extract features from user data and material data, and the first network layer corresponding to the second type of gating network is used to perform data transformation processing on input data. The above step S302 includes: based on user data and the consumption data of the target user, through the first type of gating network, obtaining a set of weight parameters output by the first type of gating network. Among them, the consumption data is used to represent the consumption behavior of the target user on materials at different times in different recommendation scenarios; based on user data and material data, through the second type of gating network, obtaining a set of weight parameters output by the second type of gating network. That is to say, the input data of the first type of gating network includes user data and consumption data, and the input data of the second type of gating network includes user data and material data, that is, the input data of the first type of gating network and the second type of gating network are different.

[0086] Step S303: For any first network layer in the MTL model, based on a set of weight parameters output by the gating network corresponding to the first network layer, adjust the first output data of the first network layer to obtain the second output data of the first network layer after being adjusted by the gating network.

[0087] In some embodiments, through the MTL model, based on user data and material data, obtain the first output data of each first network layer.

[0088] In some embodiments, the MTL model simultaneously learns multiple targets and tasks through a single model. The framework of the MTL model mainly adopts a shared-bottom structure, that is, different tasks share the underlying hidden layer, and then different target towers are established at the top to correspond to different tasks. Specifically, the MTL model can be a PLE (Progressive Layered Extraction) model, an SNR (Sub-Network Routing) model, etc.

[0089] In some embodiments, each gating network corresponds to each of at least one first network layer in the MTL model. It can be understood that the MTL model includes multiple network layers. The network layer in the multiple network layers corresponding to the gating network is the first network layer. The number of first network layers is equal to the number of gating networks, and one first network layer corresponds to one gating network. The output data of the first network layer corresponding to the gating network is adjusted by the weight parameters output by the gating network.

[0090] That is to say, the electronic device inputs user data and material data into the MTL model, and processes the user data and material data through multiple network layers in the MTL model until the output data of the last network layer in the MTL model is obtained. During the process of processing the user data and material data, for the first network layer corresponding to the gating network, the output data of the first network layer can be adjusted by the weight parameters output by the corresponding gating network, while the output data of the network layer not corresponding to the gating network is directly input into the next network layer for continued data processing.

[0091] In one example, Figure 6 is a schematic structural diagram of an MTL model provided by an embodiment of the present application. As Figure 6 shown, taking the MTL model as the SRN model as an example, the SNR model includes 5 network layers. The first network layer is the embedding layer, the second network layer includes expert layer 11, expert layer 21, and expert layer 31, the third network layer includes expert layer 12, expert layer 22, and expert layer 32, the fourth network layer includes target tower 11, target tower 12, and target tower 31, and the fifth network layer includes target tower 12, target tower 22, and target tower 32. The electronic device inputs user data and material data into the SNR model, and processes the user data and material data through 5 network layers in sequence, and finally obtains the output data of the SNR model.

[0092] During the process of the SNR model processing the user data and material data, if the first network layer (i.e., the embedding layer) and the third network layer (i.e., expert layer 12, expert layer 22, and expert layer 32) in the SNR model are the first network layers, then the output data of the embedding layer, the output data of expert layer 12, the output data of expert layer 22, and the output data of expert layer 32 are obtained. The output data of these network layers can be adjusted by the weight parameters output by the corresponding gating network, while the output data of the remaining network layers is directly input into the next network layer for continued processing.

[0093] In some embodiments, each gating network outputs a set of weight parameters. This set of weight parameters corresponds to one user, and the number of weight parameters included in this set of weight parameters is equal to the dimension of the first output data of the first network layer. For example, if the dimension of the first output data of the first network layer corresponding to gating network 1 is 256 dimensions, then gating network 1 outputs 256 weight parameters, and these 256 weight parameters are obtained based on the user data of the same user.

[0094] In some embodiments, when the number of gating networks is one, the first output data of a first network layer corresponding to the gating network is adjusted based on a set of weight parameters to obtain the adjusted second output data.

[0095] In some embodiments, when the number of gating networks is multiple, based on multiple sets of weight parameters, the first output data of multiple first network layers corresponding to the multiple gating networks is adjusted to obtain adjusted second output data. Wherein, each set of weight parameters adjusts one first output data.

[0096] In some embodiments, the above-mentioned adjusting the first output data of the first network layer based on a set of weight parameters output by the gating network corresponding to the first network layer to obtain the second output data of the first network layer after being adjusted by the gating network includes: extracting features from the user data and the material data through the first network layer corresponding to the first type of gating network to obtain the first output data output by the first network layer; multiplying a set of weight parameters output by the first type of gating network by the first output data to obtain the second output data.

[0097] That is to say, the second output data is obtained by multiplying the weight parameters by the first output data. For example, features are extracted from the user data and the material data through the first network layer corresponding to the first type of gating network to obtain the first output data, and the weight parameters output by the first type of gating network are multiplied by the output data of this first network layer to obtain the second output data.

[0098] In some embodiments, based on the user data and the material data, a set of weight parameters output by the second type of gating network is obtained through the second type of gating network, including: concatenating the user data, the material data, and the second output data of the first network layer corresponding to the first type of gating network to obtain concatenated data; inputting the concatenated data into the second type of gating network to obtain a set of weight parameters output by the second type of gating network.

[0099] In some embodiments, the input data of the second type of gating network may further include the second output data of the first network layer corresponding to the first type of gating network. In some embodiments, for the second type of gating network, the user data, the material data, and the second output data of the second type of gating network are concatenated to obtain concatenated data; through the second type of gating network, based on the concatenated data, a set of weight parameters output by the second type of gating network is obtained.

[0100] In some embodiments, the first output data of the first network layer corresponding to the second type of gating network is obtained, and the weight parameters output by the second type of gating network are multiplied by the first output data to obtain the second output data.

[0101] Understandably, the MTL model inputs model samples of multiple targets into the same model. Through learning in one model, it outputs predicted values for different targets. However, in a multi-task model, the shared layer parameters of all users and the parameters of the independent target towers are the same, and the training results of low-frequency users will be dominated by high-frequency users. Through the gate network, different weight parameters can be realized for different users, achieving a thousand people with a thousand parameters, and effectively solving the problem that the recommendation results of low-frequency users are dominated by high-frequency users.

[0102] Step S304: Adjust the first output data of at least one first network layer in the MTL model through at least one gate network to obtain the adjusted degree of interest of the target user in the to-be-recommended material output by the MTL model; based on the adjusted degree of interest, recommend the material to the target user.

[0103] In some embodiments, the output data of the MTL model is determined based on the second output data. In some embodiments, the second output data is used as the input data of the next network layer in the MTL model, that is, the second output data is input into the next network layer for further data processing until the output data of the last network layer in the MTL model is obtained. The output data of the MTL model is used to represent the degree of interest of the target user in the to-be-recommended material.

[0104] In some embodiments, the output data of the MTL model may include at least one of the following: predicted click-through rate, predicted interaction rate, predicted consumption duration, predicted forwarding rate, predicted collection rate. Among them, the predicted click-through rate is the probability that the to-be-recommended material is clicked by the target user. The predicted interaction rate is the probability that the target user generates an interaction behavior with the to-be-recommended material. The predicted consumption duration is the predicted duration for which the target user consumes the to-be-recommended material. The predicted forwarding rate is the probability that the to-be-recommended material is forwarded by the target user. The predicted collection rate is the probability that the to-be-recommended material is collected by the target user.

[0105] In some embodiments, based on the output data of the MTL model and the to-be-recommended material, the material is recommended to the target user. In some embodiments, the electronic device recommends the material to the target user based on the output data. For example, when the predicted click-through rate is greater than the first threshold, the to-be-recommended material is recommended to the target user. For example, when the predicted interaction rate is greater than the second threshold, the to-be-recommended material is recommended to the target user. For example, when the predicted consumption duration is greater than the third threshold, the to-be-recommended material is recommended to the target user. For example, when the predicted click-through rate is greater than the first threshold, the predicted interaction rate is greater than the second threshold, and the predicted consumption duration is greater than the third threshold, the to-be-recommended material is recommended to the target user.

[0106] In some embodiments, the recommendation model used in the above recommendation method includes at least one gate network and an MTL model.Figure 7 This is a schematic structural diagram of a recommendation model provided by an embodiment of the present application. As Figure 7 shown, the gating network 1 is a first type of gating network. The input data x1 of the gating network 1 may include user data, item data, and consumption data. The weight parameter g1 output by the gating network 1 is used to adjust the output data of the embedding layer, and the adjusted output data is used as the input data of the expert layer 11, the expert layer 21, and the expert layer 31.

[0107] The gating networks 2, 3, and 4 are all second type of gating networks. The input data of the gating networks 2, 3, and 4 includes the input data x2 and the output data of the embedding layer adjusted by the weight parameter g1. The input data x2 includes user data and item data. The weight parameter g2 output by the gating network 2 is used to adjust the output data of the expert layer 11, the output data of the expert layer 21, and the output data of the expert layer 31 respectively. The weight parameter g3 output by the gating network 3 is used to adjust the output data of the expert layer 12, the output data of the expert layer 22, and the output data of the expert layer 32 respectively. The weight parameter g4 output by the gating network 4 is used to adjust the output data of the target tower 11, the output data of the target tower 21, and the output data of the target tower 31 respectively.

[0108] In the recommendation method of the embodiment of the present disclosure, the weight parameter corresponding to the target user is obtained through the gating network trained by the user data of the low-frequency user, and the output of the MTL model is adjusted through the weight parameter corresponding to the target user, which can effectively solve the problem that the recommendation result of the low-frequency user is dominated by the high-frequency user and improve the recommendation accuracy of the recommendation model for the low-frequency user.

[0109] Next, the exemplary structure of the recommendation device 455 implemented as a software module provided by the embodiment of the present application will be continued. In some embodiments, as Figure 2 shown, the software module in the recommendation device 455 stored in the memory 450 may include:

[0110] A data acquisition module 4551, configured to acquire user data of a target user and material data of a material to be recommended in response to a recommendation request; a first data processing module 4552, configured to determine a set of weight parameters corresponding to the target user based on the user data through a gating network set in a multi-task learning (MTL) model; wherein, the gating network includes at least one, and each gating network corresponds to a first network layer in the MTL model, and each gating network is trained based on at least the user data of low-frequency sample users; the MTL model outputs the degree of interest of the target user in the material to be recommended based on the input user data and material data; a second data processing module 4553, configured to, for any first network layer in the MTL model, adjust the first output data of the first network layer based on a set of weight parameters output by the gating network corresponding to the first network layer to obtain second output data of the first network layer after being adjusted by the gating network; after at least one gating network adjusts the first output data of at least one first network layer in the MTL model, obtain the adjusted degree of interest of the target user in the material to be recommended output by the MTL model; a recommendation module 4554, configured to recommend materials to the target user based on the adjusted degree of interest.

[0111] In some possible implementation manners, the first data processing module 4552 is configured to obtain third output data of the last activation layer in the gating network based on the user data through multiple activation layers in the gating network; multiply the scaling coefficient of the gating network by the third output data to obtain a set of weight parameters output by the gating network; wherein, the scaling coefficient is obtained by training the gating network based on at least the user data of low-frequency sample users.

[0112] In some possible implementation manners, when the number of gating networks is one, the gating network is a first type of gating network, and the first network layer corresponding to the first type of gating network is used to extract features from the user data and the material data. The first data processing module 4552 is configured to obtain a set of weight parameters output by the first type of gating network based on the user data and the consumption data of the target user through the first type of gating network; wherein, the consumption data is used to represent the consumption behavior of the target user for materials at different times in different recommendation scenarios.

[0113] In some possible implementation manners, the second data processing module 4553 is configured to extract features from the user data and the material data through the first network layer corresponding to the first type of gating network to obtain first output data output by the first network layer; perform dot multiplication on the first output data with a set of weight parameters output by the first type of gating network to obtain second output data.

[0114] In some possible embodiments, when the number of gating networks is multiple, the multiple gating networks include a first type of gating network and a second type of gating network. The first network layer corresponding to the first type of gating network is used to extract features from user data and material data, and the first network layer corresponding to the second type of gating network is used to perform data transformation processing on the input data. The first data processing module 4552 is configured to obtain a set of weight parameters output by the first type of gating network based on the user data and the consumption data of the target user through the first type of gating network, where the consumption data is used to represent the consumption behavior of the target user for materials at different times in different recommendation scenarios; and obtain a set of weight parameters output by the second type of gating network based on the user data and the material data through the second type of gating network.

[0115] In some possible embodiments, the first data processing module 4552 is configured to splice the user data, the material data, and the second output data of the first network layer corresponding to the first type of gating network to obtain spliced data; and obtain a set of weight parameters output by the second type of gating network by inputting the spliced data into the second type of gating network.

[0116] In some possible embodiments, each gating network includes multiple activation layers connected in series, the number of activation layers included in different gating networks is different, and / or the activation functions of the activation layers included in different gating networks are different.

[0117] The embodiments of the present application provide a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the recommendation method described above in the embodiments of the present application.

[0118] The embodiments of the present application provide a computer-readable storage medium storing executable instructions, where the executable instructions are stored. When the executable instructions are executed by a processor, the processor will be caused to execute the recommendation method provided in the embodiments of the present application. For example, as Figure 3 shown in the recommendation method.

[0119] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disc, or CD-ROM; or may be various devices including one or any combination of the above memories.

[0120] In some embodiments, the executable instructions may take the form of a program, software, software module, script, or code, and may be written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including being deployed as a stand-alone program or being deployed as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0121] As an example, the executable instructions may or may not correspond to a file in a file system, and may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, stored in a single file dedicated to the program being discussed, or, stored in multiple cooperating files (e.g., files that store one or more modules, subroutines, or portions of code).

[0122] As an example, the executable instructions may be deployed to execute on one computing device, or on multiple computing devices located at one location, or, on multiple computing devices distributed across multiple locations and interconnected by a communication network. As described above, the foregoing are only embodiments of the present application and are not intended to limit the protection scope of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included within the protection scope of the present application.

Claims

1. A recommendation method, characterized in that: The method comprises: In response to the recommendation request, obtaining user data of the target user and material data of the material to be recommended; Based on the user data, a set of weight parameters corresponding to the target user is determined by a gating network set in a multi-task learning MTL model; wherein the gating network includes at least one, each gating network corresponds to a first network layer in the MTL model, and each gating network is trained based on at least user data of low-frequency sample users; the MTL model outputs the interest degree of the target user in the material to be recommended based on the input user data and the material data; For any first network layer in the MTL model, based on a set of weight parameters output by a gating network corresponding to the first network layer, adjusting the first output data of the first network layer to obtain second output data of the first network layer adjusted by the gating network; The at least one gating network adjusts the first output data of at least one first network layer in the MTL model to obtain the adjusted interest level of the target user in the material to be recommended output by the MTL model; and based on the adjusted interest level, recommend materials to the target user.

2. The method according to claim 1, characterized in that The determining, based on the user data, a set of weight parameters corresponding to the target user through a gating network set in the MTL model includes: Based on the user data, obtaining third output data of a last activation layer in the gating network through a plurality of activation layers in the gating network; The scaling factor of the gating network is multiplied by the third output data to obtain a set of weight parameters output by the gating network; wherein the scaling factor is obtained after the gating network is trained based on at least the user data of the low-frequency sample user.

3. The method according to claim 1, characterized in that In the case where the number of the gated networks is one, the gated network is a first type of gated network, and the first network layer corresponding to the first type of gated network is used to extract features from the user data and the material data; The determining, based on the user data, a set of weight parameters corresponding to the target user through a gating network set in the MTL model includes: Based on the user data and the consumption data of the target user, a set of weight parameters output by the first type of gating network is obtained through the first type of gating network; wherein the consumption data is used to represent the consumption behavior of the target user on the material at different times in different recommendation scenarios.

4. The method according to claim 3, characterized in that The adjusting the first output data of the first network layer based on a set of weight parameters output by the gating network corresponding to the first network layer to obtain the second output data of the first network layer adjusted by the gating network includes: Performing feature extraction on the user data and the material data through a first network layer corresponding to the first type of gating network to obtain first output data output by the first network layer; A set of weight parameters output by the first type of gating network is point-multiplied with the first output data to obtain the second output data.

5. The method according to claim 1, characterized in that In the case where there are multiple gated networks, the multiple gated networks include a first type of gated network and a second type of gated network, the first network layer corresponding to the first type of gated network is used to extract features from the user data and the material data, and the first network layer corresponding to the second type of gated network is used to perform data transformation processing on the input data; Based on the user data, the gating network set in the MTL model determines a set of weight parameters corresponding to the target user, including: Based on the user data and the consumption data of the target user, a set of weight parameters output by the first type of gating network is obtained through the first type of gating network; wherein the consumption data is used to represent the consumption behavior of the target user on the material at different times in different recommendation scenarios; Based on the user data and the material data, a set of weight parameters output by the second type of gating network is obtained through the second type of gating network.

6. The method according to claim 5, characterized in that The obtaining, based on the user data and the material data, a set of weight parameters output by the second type of gating network through the second type of gating network comprises: splicing the user data, the material data and the second output data of the first network layer corresponding to the first type of gating network to obtain spliced ​​data; By inputting the spliced ​​data into the second type of gating network, a set of weight parameters output by the second type of gating network is obtained.

7. The method according to any one of claims 1 to 6, characterized in that: Each of the gating networks includes a plurality of activation layers connected in series, and different gating networks include different numbers of activation layers, and / or different gating networks include different activation functions of the activation layers.

8. A recommendation device, characterized in that: include: A data acquisition module, used to acquire user data of a target user and material data of a material to be recommended in response to a recommendation request; A first data processing module is used to determine a set of weight parameters corresponding to the target user based on the user data through a gating network set in a multi-task learning MTL model; wherein the gating network includes at least one, each gating network corresponds to a first network layer in the MTL model, and each gating network is trained based on at least user data of low-frequency sample users; the MTL model outputs the interest level of the target user in the material to be recommended based on the input user data and the material data; A second data processing module is used to adjust the first output data of any first network layer in the MTL model based on a set of weight parameters output by the gating network corresponding to the first network layer, so as to obtain the second output data of the first network layer adjusted by the gating network; adjust the first output data of at least one first network layer in the MTL model through the at least one gating network, so as to obtain the adjusted interest level of the target user in the material to be recommended output by the MTL model; A recommendation module is used to recommend materials to the target user based on the adjusted interest level.

9. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; A processor, configured to implement the method according to any one of claims 1 to 7 when executing the executable instructions or computer programs stored in the memory.

10. A computer-readable storage medium storing executable instructions or a computer program, characterized in that: When the executable instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images