WO2023061087A1 - 信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品 - Google Patents

信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品 Download PDF

Info

Publication number
WO2023061087A1
WO2023061087A1 PCT/CN2022/116402 CN2022116402W WO2023061087A1 WO 2023061087 A1 WO2023061087 A1 WO 2023061087A1 CN 2022116402 W CN2022116402 W CN 2022116402W WO 2023061087 A1 WO2023061087 A1 WO 2023061087A1
Authority
WO
WIPO (PCT)
Prior art keywords
features
recommendation
recommended
information
feature
Prior art date
Application number
PCT/CN2022/116402
Other languages
English (en)
French (fr)
Inventor
马骊
赵忠
梁瀚明
赵光耀
傅妍玫
户维波
何新昇
吴铭津
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2023061087A1 publication Critical patent/WO2023061087A1/zh
Priority to US18/196,373 priority Critical patent/US20230281448A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Definitions

  • the present application relates to the field of Internet of Vehicles and artificial intelligence technology, and in particular to an information recommendation method, device, electronic equipment, computer readable storage medium and computer program product.
  • AI Artificial Intelligence
  • Artificial intelligence technology is widely used in recommendation systems. For example, the information that users are interested in is recommended to appropriate users through the multi-recommendation target ranking model of the recommendation system.
  • the recommended dimension also known as the target
  • the fusion scheme in related technologies cannot accurately predict the scores of different users for information, making it unsuitable for personalized recommendation, and thus it is difficult to improve the recommendation accuracy of the recommendation system.
  • Embodiments of the present application provide an information recommendation method, device, electronic device, computer-readable storage medium, and computer program product, which can accurately predict a user's recommendation score for information to be recommended, so as to improve the recommendation accuracy of a recommendation system.
  • An embodiment of the present application provides an information recommendation method, including:
  • An embodiment of the present application provides an information recommendation device, including:
  • a feature encoding module configured to encode a plurality of reference features to obtain an encoding feature of each of the reference features
  • the first prediction module is configured to perform a first mapping process on the plurality of encoded features to obtain a plurality of first recommendation scores corresponding to a plurality of recommendation dimensions one-to-one, wherein the first recommendation scores represent the target object The recommendation score for the information to be recommended in the corresponding recommendation dimension;
  • the feature mapping module is configured to perform the following processing for each of the recommended dimensions: performing a second mapping process on the plurality of encoded features in the recommended dimensions to obtain the mapped features of the recommended dimensions;
  • the second prediction module is configured to perform fusion processing on the first recommendation scores of the plurality of recommendation dimensions based on the mapping features of each of the recommended dimensions to obtain fusion features, and to process the information to be recommended based on the fusion features. Perform recommendation score prediction processing to obtain a second recommendation score of the target object for the information to be recommended;
  • the information recommending module is configured to perform a recommending operation in which the information to be recommended corresponds to the target object based on the second recommendation score.
  • An embodiment of the present application provides an electronic device, including:
  • memory for storing computer-executable instructions
  • the processor is configured to implement the information recommendation method provided in the embodiment of the present application when executing the computer-executable instructions stored in the memory.
  • An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions for implementing the information recommendation method provided in the embodiment of the present application when executed by a processor.
  • An embodiment of the present application provides a computer program product, including a computer program or a computer-executable instruction.
  • the computer program or computer-executable instruction is executed by a processor, the information recommendation method provided in the embodiment of the present application is implemented.
  • the first recommendation score of the target object in multiple recommendation dimensions (such as click, duration, interaction, etc.) for the information to be recommended is predicted, and then the The encoding features are mapped on each recommendation dimension to obtain the mapping features characterized by the first recommendation score in the corresponding recommendation dimension, and the first recommendation score of each recommendation dimension is fused based on the mapping features, and based on the fusion result, the target object is predicted to be suitable for the information to be recommended.
  • the final second recommendation score in this way, according to the reference characteristics of different target objects, the fusion method suitable for the corresponding target objects can be used to fuse the scores of each recommendation dimension, so as to realize automatic accuracy according to the tendency of the target object on different recommendation dimensions.
  • the purpose of the final recommendation score can improve the prediction accuracy of the final recommendation score, provide accurate recommendation reference data for the recommendation system, and then improve the recommendation accuracy and user experience.
  • FIG. 1 is a schematic diagram of the architecture of an information recommendation system 10 provided by an embodiment of the present application.
  • FIG. 2 is a schematic structural diagram of an electronic device 500 for information recommendation provided by an embodiment of the present application
  • FIG. 3A is a schematic flowchart of an information recommendation method provided by an embodiment of the present application.
  • Fig. 3B is a schematic diagram of determining the first recommendation score provided by the embodiment of the present application.
  • Fig. 3C is a schematic diagram of determining the fitting features provided by the embodiment of the present application.
  • FIG. 4 is a schematic diagram 4 of information recommendation provided by the embodiment of the present application.
  • FIG. 5A is a schematic flow chart of the model training method provided by the embodiment of the present application.
  • FIG. 5B is a schematic flowchart of a method for updating model parameters provided in the embodiment of the present application.
  • FIG. 6 is a schematic diagram 5 of information recommendation provided by the embodiment of the present application.
  • FIG. 7A is a first schematic diagram of the information recommendation effect provided by the embodiment of the present application.
  • FIG. 7B is a second schematic diagram of the information recommendation effect provided by the embodiment of the present application.
  • FIG. 7C is a third schematic diagram of the information recommendation effect provided by the embodiment of the present application.
  • first ⁇ second ⁇ third is only used to distinguish similar objects, and does not represent a specific order for objects. Understandably, “first ⁇ second ⁇ third” is used in Where permitted, the specific order or sequence may be interchanged such that the embodiments of the application described herein can be practiced in other sequences than illustrated or described herein.
  • CTR Click through rate
  • Duration refers to the duration of the user's consumption of information, such as the duration of the user's reading information.
  • Interaction including but not limited to operations such as likes, sharing, favorites, reposts, and attention of information by users.
  • Sorting Score the recalled candidate information, and select the top-ranked multiple information from the recalled candidate information in order of scores from high to low as the recommendation result.
  • Embedded representation From a mathematical point of view, Embedding is a function for mapping. The original data is mapped or embedded into another numerical vector space through the function. It uses continuous vectors to represent discrete variables, so it is called Embedding is because this representation method is often accompanied by dimensionality reduction, just as high-dimensional data is squeezed and embedded into a low-dimensional space.
  • Multi-layer perceptron It is a feed-forward artificial neural network model that maps multiple input data sets to a single output data set. Multi-layer perceptron can handle nonlinear separable problem.
  • Multi-gate Mixture-of-Experts Network A common network structure used for multi-recommendation target learning, which consists of multiple expert networks and multiple gated networks. It is a DNN network structure, and the expert network is used to extract different features from the input data used for multi-recommendation target learning, which is equivalent to dividing the information included in the input data into multiple regions, and each region corresponds to an expert network. An expert network extracts features of different dimensions from the input data, and the gating network is used to assign the weight of each expert network. There will be multiple gating networks for multiple tasks.
  • task A corresponds to The gating network outputs the selected probability of each expert network, and uses the output selected probability as the weight of the corresponding expert network, so that the features output by multiple expert networks can be weighted and summed to obtain the comprehensive result of the corresponding task A feature.
  • Personalized features predict the user's needs and preferences based on the user's previous click data, interaction data, similar user click data, and interaction data, etc., and then recommend items that the user may like.
  • Self-adaptation Automatically adjust the processing method and parameter weight according to the data characteristics of the processed data.
  • Self-adaptation is a process in which a mathematical model continuously approaches the target.
  • Multi-recommendation target fusion Learn to obtain the estimated scores of multiple targets, and each estimated score is added or multiplied according to the importance of each target, business index requirements and other strategies.
  • the multi-recommendation target ranking model is generally used to estimate the scores (ie, recommendation scores) of multiple recommendation targets (ie, recommendation dimensions).
  • scores ie, recommendation scores
  • recommendation dimensions ie, recommendation dimensions
  • a formula fusion method is adopted to fuse multiple scores into a comprehensive score for sorting. Specifically, the prediction models of each target are trained separately, and then the predicted scores of different targets are fused through a formula. Then fusion is performed by addition, multiplication or more complex formulas. Parameters are used in the fusion process. In order to find relatively good parameters, different parameter groups need to be searched offline. Commonly used methods are grid search (grid -search) or heuristic methods (such as genetic algorithm, particle swarm algorithm, etc.).
  • the above method has at least the following disadvantages: if the prediction models of multiple targets are trained separately, the cost of this method is high, and multiple prediction models cannot share parameters, so they cannot be trained together, nor can they accelerate feature learning, and the online service load pressure is relatively high. Large, the number of loaded prediction models is large, the calculation amount is relatively large, the resource consumption is high, and the stability is poor; if the new target data added is relatively sparse, it is difficult to carry out effective model training and iteration. It can be seen that whether multiple prediction models are trained individually or multiple recommended target networks are jointly trained, this method relies too much on manual rules, and there are differences in data distribution between offline and online. The offline search parameter verification effect depends on the collection of online data and the specification of performance indicators.
  • the embodiments of the present application provide an information recommendation method, device, electronic device, computer-readable storage medium, and computer program product, which can accurately predict a user's recommendation score for information, so as to improve the recommendation accuracy and user experience of the recommendation system.
  • the information recommendation method provided by the embodiment of the present application can be implemented by various electronic devices, for example, it can be implemented by a terminal alone, it can also be implemented by a server alone, or it can be implemented jointly by a terminal and a server.
  • the terminal independently executes the information recommendation method described below, or the terminal sends a recommendation request to the server, and the server executes the information recommendation method according to the received recommendation request.
  • the electronic device used for information recommendation provided by the embodiment of the present application may be various types of terminal devices or servers, wherein the server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides cloud computing services; the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, a vehicle terminal, etc., but is not limited thereto.
  • the terminal and the server may be connected directly or indirectly through wired or wireless communication, which is not limited in this embodiment of the present application.
  • the server can be a server cluster deployed in the cloud, and open artificial intelligence cloud services (Aiaas, AI as a Service) to users.
  • AIaaS platform will split several types of common AI services and provide independent services in the cloud. Or packaged services, this service model is similar to an AI-themed mall, all users can access one or more artificial intelligence services provided by the AIaaS platform through the application programming interface.
  • one of the artificial intelligence cloud services may be an information recommendation service, that is, the server in the cloud encapsulates the information recommendation program provided by the embodiment of the present application.
  • the user calls the information recommendation service in the cloud service through a terminal (running a client, such as an instant messaging client, a live broadcast client, a short video client, a social client, etc.), so that the server deployed in the cloud calls the encapsulated information recommendation service
  • the program determines the recommendation score of the target object for the information to be recommended, and executes the recommendation operation of the target object corresponding to the information to be recommended based on the recommendation score.
  • the server alone implements the information recommendation method provided by the embodiment of the present application as an example for illustration.
  • the server performs encoding processing on multiple reference features respectively to obtain the encoding features of each reference feature, and the reference features include at least one of the following: object features of the target object, information features of the information to be recommended; performing a first mapping process on the multiple encoding features , to obtain a plurality of first recommendation scores in one-to-one correspondence with a plurality of recommendation dimensions, wherein the first recommendation score represents the recommendation score of the target object in the corresponding recommendation dimension for the information to be recommended; the following processing is performed for each recommendation dimension: A coded feature is subjected to the second mapping process in the recommended dimension to obtain the mapped feature of the recommended dimension; based on the mapped feature of each recommended dimension, the first recommendation scores of multiple recommended dimensions are fused to obtain the fused feature, and based on the fused feature Perform recommendation score prediction processing on the information to be recommended to obtain the second recommendation score of the target object for the information to be recommended; based
  • FIG. 1 is a schematic diagram of an information recommendation system 10 provided by an embodiment of the present application.
  • the terminal 400 is connected to the server 200 through the network 300, and the network 300 may be a wide area network or a local area network, or a combination of both.
  • the terminal 400 (running a client, such as an instant messaging client, a live broadcast client, a short video client, a social client, etc.) can be used to obtain the user's information recommendation request.
  • the target object For example, after the target object opens the news client running on the terminal, the terminal automatically obtains a news recommendation request for the target object.
  • the terminal after the terminal obtains the information recommendation request, it calls the information recommendation interface of the server 200 (which can be provided as a cloud service, that is, the information recommendation service), and the server 200 obtains multiple reference features of the target object based on the information recommendation request.
  • the target object is a user who needs to recommend information
  • the reference features include at least one of the following: object features of the target object (object features are user features, such as user age, user gender, etc.), information about the information to be recommended Features: Recall the information to be recommended from the information base to be recommended as the candidate information for ranking.
  • multiple reference features are encoded to obtain the encoding features of each reference feature.
  • the reference features include at least one of the following: object features of the target object, information features of recalled information to be recommended; multiple encoding features Perform the first mapping process to obtain a plurality of first recommendation scores corresponding to multiple recommendation dimensions one by one; perform the following processing for each recommendation dimension: perform the second mapping process on multiple coding features in the recommendation dimension to obtain the recommended dimension Mapping feature: Based on the mapping feature of each recommendation dimension, the first recommendation score of multiple recommendation dimensions is fused to obtain the fusion feature, and based on the fusion feature, the recommendation score prediction process is performed on the recommended information to obtain the target object for the information to be recommended based on the second recommendation score of the information to be recommended; rearrange the recalled information to be recommended based on the second recommendation score of the information to be recommended, and select a number of top-ranked information (that is, a plurality of information to be recommended starting from the first place) to be recommended
  • the recommendation information is pushed to the terminal 400 for display.
  • the target object involved in the embodiment of this application is the recipient of the information recommended by the information recommendation system.
  • the target object is the recipient of the news recommended by the news recommendation system.
  • the object characteristics of the involved target objects are obtained with the consent of the target objects.
  • the information recommendation method provided by the embodiment of the present application can also be applied to information recommendation scenarios related to Internet of Vehicles services (such as refueling, navigation, parking, maintenance, etc.), such as when recommending information to a vehicle terminal , execute the information recommendation method provided by the embodiment of the present application on the target object of the vehicle terminal, determine the final recommendation score of the target object for the information to be recommended, and perform the recommendation operation of the target object corresponding to the information to be recommended based on the final recommendation score; for example, for the final
  • the information to be recommended with a recommendation score lower than the score threshold applies the corresponding masking mode, and the information to be recommended with a final recommendation score exceeding the score threshold is recommended to the vehicle terminal, thereby avoiding the widespread dissemination of low-quality information, indirectly improving the overall information quality, and improving user experience.
  • FIG. 2 is a schematic structural diagram of the electronic device 500 for information recommendation provided by the embodiment of the present application.
  • the electronic device 500 is a server as an example.
  • the electronic device 500 for information recommendation shown in FIG. 2 includes: at least one processor 510 , a memory 550 , at least one network interface 520 and a user interface 530 .
  • Various components in the electronic device 500 are coupled together through the bus system 540 .
  • the bus system 540 is used to realize connection and communication between these components.
  • the bus system 540 also includes a power bus, a control bus and a status signal bus.
  • the various buses are labeled as bus system 540 in FIG. 2 .
  • Processor 510 can be a kind of integrated circuit chip, has signal processing capability, such as general-purpose processor, digital signal processor (DSP, Digital Signal Processor), or other programmable logic device, discrete gate or transistor logic device, discrete hardware Components, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor, etc.
  • DSP digital signal processor
  • DSP Digital Signal Processor
  • Memory 550 includes volatile memory or nonvolatile memory, and may include both volatile and nonvolatile memory.
  • the non-volatile memory may be a read-only memory (ROM, Read Only Memory), and the volatile memory may be a random access memory (RAM, Random Access Memory).
  • the memory 550 described in the embodiment of the present application is intended to include any suitable type of memory.
  • Memory 550 optionally includes one or more storage devices located physically remote from processor 510 .
  • memory 550 is capable of storing data to support various operations, examples of which include programs, modules, and data structures, or subsets or supersets thereof, as exemplified below.
  • Operating system 551 including system programs for processing various basic system services and performing hardware-related tasks, such as framework layer, core library layer, driver layer, etc., for implementing various basic services and processing hardware-based tasks;
  • the network communication module 552 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 520.
  • Exemplary network interfaces 520 include: Bluetooth, Wireless Compatibility Authentication (WiFi), and Universal Serial Bus ( USB, Universal Serial Bus), etc.;
  • Presentation module 553 for enabling presentation of information via one or more output devices 531 (e.g., display screen, speakers, etc.) associated with user interface 530 (e.g., a user interface for operating peripherals and displaying content and information );
  • output devices 531 e.g., display screen, speakers, etc.
  • user interface 530 e.g., a user interface for operating peripherals and displaying content and information
  • the input processing module 554 is configured to detect one or more user inputs or interactions from one or more of the input devices 532 and translate the detected inputs or interactions.
  • the information recommendation apparatus provided in the embodiments of the present application may be implemented in software, for example, it may be the above-mentioned information recommendation service in the server, or it may be the above-mentioned information recommendation plug-in in the terminal.
  • the information recommendation apparatus provided in the embodiments of the present application may be provided in various software embodiments, including various forms of application programs, software, software modules, scripts or codes.
  • the information recommending device provided by the embodiment of the present application can be realized by software.
  • FIG. 2 shows the information recommending device 555 stored in the memory 550, which can be software in the form of programs and plug-ins, including the following Software modules: feature encoding module 5551, first prediction module 5552, feature mapping module 5553, second prediction module 5554 and information recommendation module 5555, these modules are logical, so any combination or further split. The function of each module will be explained below.
  • the information recommendation method provided by the embodiment of the present application will be described below in conjunction with the accompanying drawings.
  • the execution subject of the following information recommendation method may be a server, specifically, the server may be implemented by running various computer programs above; of course, according to the From the understanding below, it is not difficult to see that the information recommendation method provided by the embodiment of the present application can also be implemented cooperatively by the terminal and the server.
  • FIG. 3A is a schematic flowchart of an information recommendation method provided by an embodiment of the present application, which will be described in conjunction with the steps shown in FIG. 3A .
  • step 101 the server encodes multiple reference features to obtain the coded features of each reference feature.
  • the reference features include at least one of the following: object features of the target object, information features of the information to be recommended, and the object features are the basic attribute features of the target object (such as age, gender, occupation, education level, consumption level, etc.), portrait features (such as hobbies, browsing, clicking, favorites, purchases, etc.), and the contextual characteristics of the recommendation (environmental characteristics, such as recommendation time, recommendation scene, etc.), the information characteristics are the information labels, information categories, and so on of the information to be recommended.
  • Information sources, interaction features, and interaction features are features related to information content and object features. Specifically, interaction features are the intersection of information features and object features.
  • the consumption characteristics of recommended information by users with the same age and gender are the interaction characteristics between the information to be recommended and users.
  • any one or more features can be selected and freely combined as a reference feature, that is, the scale dimension of the reference feature can be set, such as using all or part of the object features of the target object as a reference feature, or will be recommended All or part of the information features of the information are used as reference features, or all object features of the target object and all information features of the information to be recommended are used as reference features, or part of the object features of the target object and all information features of the information to be recommended are used as reference features , using part of the object features of the target object and part of the information features of the information to be recommended as reference features, and so on.
  • the reference features include at least one of continuous features and discrete features.
  • Step 101 can be implemented in the following manner: for each reference feature, the following processing is performed: when the reference feature is a continuous feature, the continuous The features are discretized to obtain the discrete features of the continuous features, and the discrete features of the continuous features are encoded to obtain the encoding features of the continuous features; when the reference features are discrete features, the discrete features are Encoding processing, to obtain the encoding features of discrete features, through the embodiment of the present application, different forms of features can be uniformly encoded, which is equivalent to mapping to the same encoding space, so that subsequent data processing can be performed based on unified reference features, and data processing can be improved efficiency and accuracy.
  • the data form of continuous features is continuous data, which can be any value in the interval, for example, video duration and release failure are all continuous features
  • the data form of discrete features is discrete data, generally in integer
  • the discrete data 0 is used to indicate that the user's gender is male
  • the discrete data 1 is used to indicate that the user's gender is 1.
  • discretize continuous features for example, adopt card room test, for example, discretize by adopting decision tree, and discretize by dividing intervals.
  • the encoding process of different reference features is different. Refer to FIG. 4.
  • FIG. 4 is a schematic diagram of information recommendation provided by the embodiment of this application.
  • the information to be recommended is video as an example for illustration.
  • the continuous features in the reference features can be For video duration, release effectiveness, etc., discrete features can be video ID, user account level, user gender, etc., and different features need to be encoded in a targeted manner.
  • discretize continuous features such as video duration and release effectiveness, or normalize or standardize continuous features to obtain discrete features of continuous features (that is, discrete values, also known as sparse features)
  • discrete features of continuous features that is, discrete values, also known as sparse features
  • the discrete features of the continuous features are encoded, for example, the discrete features are mapped to another vector space to obtain the encoded features of the continuous features (embedded representation Embedding, also known as dense features); for video identification, user account level,
  • the discrete features such as user gender are directly encoded, for example, the discrete features are mapped to another vector space to obtain the encoded features of the discrete features (embedded representation Embedding, also known as dense features).
  • a first mapping process is performed on a plurality of encoding features to obtain a plurality of first recommendation scores corresponding to a plurality of recommendation dimensions one-to-one.
  • FIG. 3B is a schematic diagram of determining the first recommendation score provided by the embodiment of the present application. Step 102 can be implemented through steps 1021 to 1024 shown in FIG. 3B:
  • step 1021 feature intersection processing is performed on a plurality of encoded features to obtain at least one intersection feature.
  • multiple reference features are linearly processed to obtain first-order features
  • the server can perform feature cross processing on multiple coded features in the following manner to obtain at least one cross feature: Perform second-order feature intersection processing on any two encoding features among the plurality of encoding features to obtain second-order intersection features of any two encoding features; concatenate first-order features and each second-order intersection feature to obtain at least one intersection feature.
  • the cross feature can be memorized, and the non-linear capability of subsequent data processing can be improved, thereby improving recommendation accuracy.
  • multiple reference features are linearly calculated. For example, based on the weight of each reference feature, multiple reference features are weighted and summed to obtain first-order features; any two encoded features are used for second-order features.
  • cross features of any specified order can also be obtained, such as performing i on any i (2 ⁇ i ⁇ m, m is the number of reference features) coding features among the coding features corresponding to multiple reference features First-order crossover processing, to obtain the i-order crossover feature of any i coded features, and splicing the first-order feature and the i-order crossover feature to obtain the crossover feature, so that high-order information can be considered in the subsequent prediction of the first recommendation score, and the improvement Recommended accuracy.
  • step 1022 based on a plurality of coding features, predict the fit of the information to be recommended in each recommended dimension, and obtain the fitting features corresponding to each recommended dimension.
  • the fitting feature is a feature used to characterize the fitting degree of freedom between any two recommended dimensions in all recommended dimensions.
  • the fitting feature is usually a high-order feature.
  • a high-order feature is a feature whose order is greater than the set threshold.
  • the recommended dimension refers to metrics used to evaluate recommender systems, for example, recommender dimensions include click-through rate, number of interactions, viewing time, and more.
  • FIG. 3C is a schematic diagram of determining the fitting features provided by the embodiment of the present application.
  • Step 1022 can be implemented by executing steps 10221 to 10223 shown in FIG. 3C for each recommended dimension: Step 10221 Among them, through each expert network in the multi-gated mixed expert network, the first fully connected processing is performed on multiple coding features to obtain the first hidden layer features, and the fourth mapping process is performed on the first hidden layer features to obtain the corresponding The mapping features of an expert network; in step 10222, through the gated network corresponding to the recommended dimension in the multi-gated mixed expert network, the second full-connection processing is performed on multiple coding features to obtain the second hidden layer feature, and the second hidden layer feature is obtained.
  • step 10223 Perform the fifth mapping process on the hidden layer features to obtain the weight features corresponding to each expert network; in step 10223, based on the weight features of each expert network, perform weighted summation processing on the mapping features of each expert network to obtain the corresponding recommendation dimension fitting features.
  • the multi-gated mixed expert network is composed of multiple expert networks and multiple gating networks.
  • the expert network is used to extract different mapping features.
  • the structure of the expert network can be a fully connected neural network structure.
  • the weight characteristics of each expert network, each gating network is equivalent to a classifier, and the gating network of each recommendation dimension will judge which expert network will fit it better according to the current input coding features, so as to estimate The weight characteristics of each expert network are obtained.
  • the number of gated networks is consistent with the number of recommended dimensions, and the number of expert networks can be consistent with or inconsistent with the number of recommended dimensions, that is, each recommended dimension corresponds to a gated network.
  • the recommended dimensions are click-through rate, viewing duration, and interaction times
  • the recommended dimension of "click-through rate” corresponds to gated network 1
  • the recommended dimension of "watching time” corresponds to gated network 2
  • the recommended dimension of "interaction times” corresponds to gated network 3.
  • the coding features corresponding to multiple reference features are input into each expert network.
  • the first full connection processing is performed on multiple coding features to obtain the corresponding first One hidden layer feature, and perform linear or nonlinear mapping processing on the first hidden layer feature through the activation function to obtain the mapping feature corresponding to each expert network; then input multiple encoding features and multiple expert network output mapping features to
  • the second fully connected processing is performed on multiple encoded features through the gating network, such as multi-layer perception processing, to obtain the second hidden layer features, and then linear or nonlinear mapping is performed on the second hidden layer features through the activation function Processing to obtain the weight feature corresponding to each expert network, and use the weight feature to perform weighted summation processing on the mapping features output by each expert network, and obtain the fitting of the corresponding recommended dimension (recommended dimension corresponding to the gated network) feature (the output of the gating network), and the output of the gating network corresponding to each recommended dimension is used
  • step 1023 the following processing is performed for each of the recommended dimensions: concatenating the intersection features and the fitting features of the recommended dimensions to obtain concatenated features corresponding to the recommended dimensions.
  • the subsequent score prediction based on the concatenated features can improve the score prediction accuracy, thereby improving the recommendation accuracy.
  • the cross features and the fitting features corresponding to each recommended dimension are spliced to obtain the splicing corresponding to each recommended dimension feature. Still taking the three recommendation dimensions of click-through rate, viewing duration, and number of interactions as an example, combine the cross features obtained above with the fitting features of the recommendation dimension of "click-through rate" to obtain the recommendation dimension of "click-through rate”.
  • step 1024 the following processing is performed for each recommended dimension: a third mapping process is performed on the spliced features of the recommended dimension to obtain the first recommendation score of the recommended dimension corresponding to the information to be recommended.
  • the third mapping process is performed on the corresponding splicing features to obtain the mapping features corresponding to the splicing features, and the mapping features of the splicing features are biased by the activation function to obtain corresponding to each recommended dimension.
  • the first recommendation score represents the recommendation score of the target object in the corresponding recommendation dimension for the information to be recommended.
  • linear logistic regression processing can be performed on the concatenated features through a logistic regression function.
  • the linear logistic regression processing here can be linear sum processing, and the obtained linear sum results can be used as projection features.
  • the linear summation result can be substituted into the logistic regression function to obtain the logistic regression feature as the projection feature, and then the recommendation score prediction process is performed on the projection feature through the activation function to obtain the first recommendation score representing the level of the recommendation score.
  • step 103 the following processing is performed for each recommended dimension: a second mapping process is performed on the multiple encoded features in the recommended dimension to obtain the mapped features of the recommended dimension.
  • step 103 can be implemented in the following manner: perform horizontal concatenation processing on the first recommendation scores of multiple recommendation dimensions to obtain a tiling vector; perform third full connection processing on multiple coding features to obtain a third latent layer feature; the sixth mapping process is performed on the third hidden layer feature to obtain a mapping feature with the same dimension as the tiling vector.
  • step 104 based on the mapping features of each recommended dimension, the first recommendation scores of multiple recommended dimensions are fused to obtain the fused features, and based on the fused features, the recommendation score prediction process is performed on the to-be-recommended information to obtain the Second recommendation score for recommendation information.
  • step 104 based on the mapping features of each recommendation dimension, fusion processing is performed on the first recommendation scores of multiple recommendation dimensions to obtain fusion features, which is equivalent to using the mapping features of each recommendation dimension as each recommendation
  • the weight of the first recommendation score of each dimension can be weighted and summed for the first recommendation scores of multiple recommendation dimensions based on the weight of the first recommendation score of each recommendation dimension.
  • the score matrix composed of the first recommendation score of the dimension, and the mapping matrix composed of the mapping features corresponding to each recommended dimension is obtained; the element product calculation is performed between the score matrix and the mapping matrix to obtain the fusion feature.
  • the score matrix is the tiling vector x in obtained by horizontally concatenating the vector representations of the first recommendation scores above.
  • the mapping matrix is the mapping feature with the same dimension as the tiling vector, denoted as w iu .
  • step 104 based on the fusion feature, the recommendation score prediction processing is performed on the information to be recommended, and the second recommendation score of the target object for the information to be recommended is obtained, which can be realized in the following manner: the seventh mapping process is performed on the fusion feature to obtain The mapping feature corresponding to the fusion feature; based on the mapping feature corresponding to the fusion feature, the recommendation score prediction process is performed on the information to be recommended, and the second recommendation score of the target object for the information to be recommended is obtained.
  • the fusion features are mapped. For example, the fusion features are linearly projected through the logistic regression function, and then the obtained projection values are passed through the activation function to predict the second recommendation score of the target object for the information to be recommended.
  • step 105 based on the second recommendation score, a recommendation operation operation of the target object corresponding to the information to be recommended is performed.
  • the second recommendation score is the final score of multiple recommendation dimensions, which is used to represent the overall evaluation of the target object on the information to be recommended.
  • the second recommendation score exceeds the score threshold, the information to be recommended is recommended to the target object.
  • the information recommendation method provided by the embodiment of the present application can be applied in the recall phase of the recommendation system.
  • the recommendation system includes a recall stage, a rough sorting stage, a fine sorting stage, and a rearrangement stage.
  • the recall stage is: select and candidate information from the candidate pool, and send it to the subsequent sorting stage for scoring.
  • the candidate pool is a pool of candidate information available for recommendation.
  • the rough sorting stage is to sort the thousands or hundreds of candidate information recalled and selected.
  • the stage of fine sorting is to precisely sort the results of rough sorting.
  • the rearrangement stage is: to make small adjustments to the results of fine sorting.
  • the evaluation level of the target object for the information to be recommended can also be determined according to the second recommendation score, and then different recommendation operations are performed according to the evaluation level. For example, when the evaluation level includes the first level, the second level, and the third level with successively higher levels (the user is more and more interested), when the evaluation level for the information to be recommended is the first level, in the ranking stage of the recommendation system , carry out down-weight recommendation on the recommended information to reduce the number of recommendations or frequency of recommendation. For example, before the down-weight ranking is adopted, the information may be recommended to 100 people within a week. After the down-weight sorting is adopted, the information may be recommended to Only recommend the information to 20 people.
  • the degree of power reduction is negatively correlated with the final score of the information to be recommended, that is, the lower the final score of the information to be recommended, the greater the degree of power reduction.
  • the evaluation level of the information to be recommended is the second level
  • the information to be recommended is freely recommended, and the free recommendation means that the recommended information is not biased to recommend, neither weighted nor downgraded, so that it is based on user needs and Recommendations based on the quality of the information itself.
  • the evaluation level of the information to be recommended is the third level
  • the weighted recommendation is carried out on the information to be recommended, so that the information to be recommended that the target object is interested in can be recommended to more other users who are similar to the target object, increasing the value of the information to be recommended Exposure and click-through rates.
  • the information recommendation method described above is implemented by calling a score prediction model.
  • the score prediction model includes: a feature encoding layer, a first recommendation score prediction layer, a feature mapping layer, and a second recommendation score prediction Layers; wherein, the first recommendation score prediction layer includes a first feature extraction layer, a second feature extraction layer, a feature splicing layer and a sub-score prediction layer; the second recommendation score prediction layer includes a feature fusion layer and a total score prediction layer.
  • FIG. 5A is a schematic flow chart of the model training method provided by the embodiment of the present application.
  • the score prediction model can be trained in the following manner:
  • the server uses the feature coding layer to respectively
  • the multiple sample reference features of the training samples are encoded to obtain the sample encoding features of each sample reference feature.
  • the training samples carry the first label of the object sample for the information sample in multiple recommended dimensions, and the second label of the object sample for the information sample.
  • step 202 through the first recommendation score prediction layer, the first mapping process is performed on a plurality of sample coding features, and a plurality of first prediction results corresponding to a plurality of recommended dimensions are obtained, wherein the first prediction result Characterize the recommendation score of the object sample for the information sample in the corresponding recommendation dimension; in step 203, perform the following processing for each recommendation dimension through the feature mapping layer: perform the second mapping process on each multiple coding features in the recommendation dimension to obtain the recommendation Dimension sample mapping features; in step 204, the following processing is performed through the second recommendation score prediction layer, based on the sample mapping features of each recommended dimension, the first prediction results of multiple recommended dimensions are fused to obtain the sample fusion feature , and based on the sample fusion features, carry out recommendation score prediction processing on information samples, and obtain the second prediction results of object samples for information samples; in step 205, based on the first prediction results of each recommendation dimension, corresponding to each recommendation dimension The first label, the second prediction result and the second label, update the model parameters of the score prediction model.
  • the training samples are input into the score prediction model.
  • multiple sample reference features of the training samples are encoded through the feature coding layer, and the sparse features of multiple reference features are converted into dense features.
  • the first feature extraction layer in the first recommendation score prediction layer perform feature cross processing on multiple sample coding features of training samples to obtain sample cross features;
  • the second feature extraction layer treat multiple sample coding features based on Predict the fit of recommendation information in at least two recommended dimensions, and obtain the sample fitting features corresponding to each recommended dimension;
  • through the feature splicing layer respectively splicing the sample cross features and the sample fitting features corresponding to each recommended dimension processing to obtain the sample splicing features corresponding to each recommendation dimension;
  • the sub-score prediction layer based on the sample splicing features, the recommendation scores of the information to be recommended in at least two recommendation dimensions are predicted, and the object samples are obtained for the information samples in at least two recommendation Dimension's first prediction result.
  • a second mapping process is performed on multiple coding features of the training samples in each recommended dimension to obtain sample mapping features.
  • the object samples are fused with the first prediction results of the information samples in at least two recommended dimensions to obtain the corresponding sample fusion Features:
  • the recommendation score prediction processing is performed on the information to be recommended, and the second prediction result of the target object for the information to be recommended is obtained.
  • the encoding process of different sample reference features is different.
  • the discrete features of the continuous features are encoded, for example, the discrete features are mapped to another vector space, and Coding features of continuous features (embedded representation Embedding, also known as dense features);
  • the sample reference features are discrete features, the discrete features such as video identification, user account level, and user gender are directly encoded, for example, discrete Type features are mapped to another vector space to obtain the encoding features of discrete features (embedded representation Embedding, also known as dense features).
  • FIG. 5B is a schematic flow chart of the method for updating model parameters provided by the embodiment of the present application.
  • Step 205 can be implemented through steps 2051 to 2054 shown in FIG. 5B: in step 2051, for each A recommendation dimension, based on the first prediction result and the first label of the recommended dimension, construct the first loss function corresponding to the first recommendation score prediction layer; in step 2052, based on the second prediction result and the second label, construct the corresponding second Recommend the second loss function of the score prediction layer; in step 2053, carry out weighted summation of the second loss function and the first loss function to obtain the third loss function of the score prediction model; in step 2054, based on the third loss function Update the model parameters of the score prediction model.
  • the above step 2051 can be implemented in the following manner: based on the first prediction result corresponding to each recommended dimension and the corresponding first label, construct a sub-loss function corresponding to each recommended dimension; determine the recommendation weight corresponding to each recommended dimension , based on each recommendation weight, perform weighted summation on the sub-loss functions corresponding to each recommendation dimension to obtain a first loss function corresponding to the first recommendation score prediction layer.
  • a corresponding sub-loss function can be constructed based on the first predicted result and the first label of the object sample in the corresponding recommended dimension for the information sample, and multiple Add the sub-loss functions of the recommendation dimension to get the first loss function of the first recommendation score prediction layer Among them, n represents the number of recommended dimensions, loss j represents the sub-loss function corresponding to the jth recommended dimension, 1 ⁇ j ⁇ n.
  • the second loss function L( ⁇ ) of the second recommendation score prediction layer can be constructed and expressed as :
  • ⁇ ) ⁇ (f(w i,um
  • ⁇ ( ) is a sigmoid function
  • ⁇ ) is the second prediction result
  • the first Two prediction results are converted into estimated probability p(w i,um
  • n is the total number of training samples
  • ⁇ is the model parameter
  • yi is the second label
  • a i is different weights set according to the recommended dimension, such as a training
  • the sample has two recommendation dimensions of click and interaction, and a i can be set to 2, which is larger than the weight (less than 2) of the training sample with only one recommendation dimension of the click dimension, and the model is more inclined to learn training samples with the interaction dimension.
  • the third loss function of the overall score prediction model is the sum of the first loss function (that is, the sum of the individual sub-loss functions of each recommendation dimension among multiple recommendation dimensions) and the second loss function:
  • the third loss function After constructing the third loss function, judge whether the value of the third loss function exceeds the preset threshold according to the value of the third loss function, and determine the score prediction model based on the third loss function when the value of the third loss function exceeds the preset threshold
  • the error signal of the error information is backpropagated in the score prediction model, and the model parameters of each layer are updated during the propagation process.
  • the reference features of the training samples are input to the input layer of the neural network model, through the hidden layer, and finally reach the output layer and output the result.
  • This is the forward propagation process of the neural network model. Since the neural network If there is an error between the output result of the model and the actual result, the error between the output result and the actual value is calculated, and the error is backpropagated from the output layer to the hidden layer until it is propagated to the input layer.
  • the process of backpropagation Adjust the value of the model parameters according to the error; continuously iterate the above process until convergence, wherein the score prediction model belongs to the neural network model.
  • the information recommendation method provided by the embodiment of the present application can be applied to all recommendation systems that use a multi-recommendation target ranking model, such as client recommendation, browser information flow scenarios, news, express recommendation and other information flow products, and can also be applied to In other recommendation scenarios such as e-commerce, advertising recommendation scenarios, etc.
  • the information recommendation method provided by the embodiment of the present application will be described by taking the fusion of multi-recommendation target scores of the three recommendation dimensions of click, duration, and interaction as an example.
  • FIG. 6 is a schematic diagram of information recommendation provided by the embodiment of the present application.
  • Information recommendation is performed through a multi-recommendation target ranking model.
  • the model includes: a sparse feature layer, a feature extraction layer, a sub-score prediction layer, a feature mapping layer, and a fusion part , and then the training and application of the score prediction model will be described in conjunction with FIG. 6 .
  • the user-side features of the training samples ie, the above-mentioned reference features
  • it can be selected from the user features of the object samples (ie, the above-mentioned object features), and the information features of the information to be recommended
  • the user features are the basis of the target object Attribute characteristics (such as age, gender, occupation, education level, consumption level, etc.), portrait characteristics (such as hobbies, browsing, clicks, favorites, purchases, etc.), and contextual characteristics of the recommendation (environmental characteristics, Such as recommended time, recommended scene, etc.)
  • information features are the interactive features of information labels, information categories, information sources, information content and user features of information samples
  • interaction features refer to the intersection of information features and user features, such as statistical information to be recommended In the consumption situation of users of different ages and genders, through the specific age and gender of the users, the consumption characteristics of the users under the age and gender of the user for the recommended information are obtained, and the consumption characteristics are the interaction characteristics between the information to be recommended and the user.
  • the discrete features are sparse features
  • the discrete features need to be encoded through the sparse feature layer, such as through the Embedding vector conversion process, to obtain the corresponding encoded features (also known as dense features); the obtained encoded features and the user side itself are dense features.
  • the feature extraction layer comprises a cross feature extraction layer and a fitting feature extraction layer, wherein the cross feature extraction layer can be a factorization machine (FM, Factorization Machine) model, and the fitting feature extraction layer can be an MMoE model, by crossing the feature extraction layer, Perform second-order feature crossing on each coding feature of user-side features to obtain corresponding second-order crossing features, and splicing the first-order features and second-order crossing features of each coding feature to obtain low-order crossing features with memory; MMoE model It consists of multiple expert networks and multiple gating networks.
  • the expert network is used to extract different features. It can be a DNN network structure.
  • the gating network is used to assign the weight of each expert network.
  • Each gating network is equivalent to a classification
  • the gating network of each recommended dimension will judge which expert network will fit it better according to the current input coding features, so the weight of each expert network is estimated.
  • the sub-score prediction layer includes three models for score prediction of click, duration, and interaction.
  • the three models are independent of each other, and the cross-features and the output corresponding to each target output by the MMoE model are spliced and then input into the corresponding model.
  • Score prediction to obtain the corresponding score that is, the above-mentioned first recommendation score).
  • the feature mapping layer is essentially the user's lightweight network, and the features input to the feature mapping layer can come from the coding features output by the sparse feature layer, that is, the features input to the feature mapping layer can be the coded output of the sparse feature layer. Some or all of the features can also be other new features, such as obtaining user-side features that are different from the user-side features input to the sparse feature layer, and the newly acquired user-side features can even include information features, and the newly acquired The user-side features are encoded and then input to the feature mapping layer.
  • the introduction of user personalized features can give the optimal fusion method of all target scores according to different users, which is equivalent to automatically giving the final score according to the user's tendency on different goals, and achieve a relatively better business performance. Effect.
  • the role of the fusion part is to predict the final score of the target object for the information to be recommended.
  • the corresponding sub-loss function can be constructed based on the score and the object sample carried by the training sample for the label of the information sample in the corresponding target, and the sub-loss functions of each target can be added to obtain the sub-loss function Loss function for the score prediction layer Among them, n represents the number of targets, loss j represents the sub-loss function corresponding to the jth target, 1 ⁇ j ⁇ n.
  • the loss function L( ⁇ ) of the fusion part can be constructed based on the final score and the label of the object sample carried by the training sample for the information sample, expressed as:
  • ⁇ ) ⁇ (f(wi ,um
  • ⁇ ( ) is the sigmoid function
  • ⁇ ) is the final score
  • n is the total number of training samples
  • y i is the label
  • a i is the different weights set for the target.
  • a training sample has two targets of click and interaction, a i can be If it is set to 2, the weight (less than 2) of the training sample that only clicks on the target is larger, and the model is more inclined to learn the training sample with interaction.
  • the overall loss function of the multi-recommendation target ranking model is the sum of the loss function of the sub-score prediction layer and the loss function of the fusion part, expressed as:
  • the overall loss function of the multi-recommendation target ranking model After constructing the overall loss function of the multi-recommendation target ranking model, judge whether it exceeds the preset threshold according to the value of the overall loss function (such as the gradient value) of the multi-recommendation target ranking model. When the preset threshold is exceeded, based on the multi-recommendation target ranking model The overall loss function determines the error signal of the model, backpropagates the error signal in the fractional prediction model, and updates the model parameters of each layer during the propagation process.
  • the overall loss function determines the error signal of the model, backpropagates the error signal in the fractional prediction model, and updates the model parameters of each layer during the propagation process.
  • the multi-recommendation target ranking model provided by the embodiment of this application is an end-to-end model, which does not need to consider the influence of data distribution in offline and online scenarios, and the loss function of the fusion part is jointly trained with other loss functions of the multi-recommendation target.
  • online prediction only needs to load one model, which improves the convenience and stability of service deployment.
  • the user-side features including user features, information features of candidate articles, cross features and context features, etc.
  • the estimated click-through rate, estimated duration or duration probability of a candidate article is converted into a score to form a multi-recommendation target score vector x im .
  • a feature mapping layer is constructed according to the required user-side features.
  • Figure 7A- Figure 7C is a schematic diagram of the information recommendation effect provided by the embodiment of this application, taking the application of the information recommendation method provided by the embodiment of this application to the scene of point-of-view picture-text recommendation as an example, using this application
  • the information recommendation method provided by the embodiment has a relative increase in the three goals of click-through rate, total reading time, and number of likes. For example, the average click-through rate has been relatively improved.
  • the empty running period uses the general formula fusion and grid search method
  • the experimental period is the improvement effect of the information recommendation method provided by the embodiment of the application compared with the general fusion and grid search method.
  • the embodiment of the present application provides an end-to-end multi-recommendation target score fusion model based on user personalized features, aiming at solving the problem of merging the scores of multiple recommendation target models on different targets into one score for ranking.
  • the information recommendation method provided in the embodiment of the application is based on the MMoE-based multi-recommendation target model, and introduces the MLP network to adaptively learn multi-recommendation target fusion scoring, that is, by introducing user-side features, it can adaptively learn each user to a different target
  • the personalized weight of the score, the best fusion score is obtained by integrating the characteristics of the user side, and then the optimization of each goal is achieved, no longer relying on manual formulas and search parameters, saving time and effort.
  • the software modules stored in the information recommendation device 555 of the memory 550 in FIG. 2 may include: feature coding
  • the module 5551 is configured to perform encoding processing on multiple reference features of the target object respectively to obtain the encoding features of each reference feature;
  • the first prediction module 5552 is configured to perform the first mapping process on the multiple encoding features to obtain the corresponding A plurality of first recommendation scores corresponding to a plurality of recommendation dimensions one-to-one, wherein the first recommendation score represents the recommendation score of the target object in the corresponding recommendation dimension for the information to be recommended;
  • the feature mapping module 5553 is configured for each The recommended dimension performs the following processing: perform a second mapping process on the multiple encoded features in the recommended dimension to obtain the mapped features of the recommended dimension;
  • the second prediction module 5554 is configured to be based on each of the recommended dimensions
  • the first recommendation scores of the multiple recommendation dimensions are fused to obtain the fused features
  • the reference features include at least one of continuous features and discrete features
  • the feature encoding module 5551 is further configured to perform the following processing for each of the reference features: when the reference features are the continuous features In the feature, discretize the continuous feature to obtain the discrete feature of the continuous feature, and encode the discrete feature of the continuous feature to obtain the encoded feature of the continuous feature; When the reference feature is the discrete feature, encoding processing is performed on the discrete feature to obtain an encoded feature of the discrete feature.
  • the first prediction module 5552 is further configured to perform feature intersection processing on the multiple encoding features to obtain at least one intersection feature; based on the multiple encoding features, the information to be recommended is processed in each The fit of the recommended dimension is predicted to obtain the fitting feature corresponding to each of the recommended dimensions; the following processing is performed for each of the recommended dimensions: the cross feature is compared with the fitting feature of the recommended dimension Splicing processing to obtain the splicing features corresponding to the recommended dimensions; perform the following processing for each of the recommended dimensions: perform a third mapping process on the splicing features of the recommended dimensions to obtain the information to be recommended corresponding to the recommended dimensions First recommendation score.
  • the first prediction module 5552 is further configured to perform linear processing on multiple reference features to obtain first-order features; perform second-order feature intersection processing on any two encoded features among the multiple encoded features to obtain any two A second-order intersection feature of a coded feature; the first-order feature and the second-order intersection feature are spliced to obtain at least one intersection feature.
  • the first prediction module 5552 is further configured to perform the following processing for each of the recommended dimensions: through each expert network in the multi-gated mixed expert network, perform the first comprehensive connection processing to obtain the first hidden layer features, and perform fourth mapping processing on the first hidden layer features to obtain the mapping features corresponding to each of the expert networks; through the multi-gated mixed expert network corresponding to the A gating network of the recommended dimension performs a second full-connection process on the plurality of encoded features to obtain second hidden layer features, and performs a fifth mapping process on the second hidden layer features to obtain The weight feature of the network: based on the weight feature of each of the expert networks, weighted summation is performed on the mapping features of each of the expert networks to obtain fitting features corresponding to the recommended dimensions.
  • the feature mapping module 5553 is further configured to perform horizontal concatenation processing on the first recommendation scores of multiple recommended dimensions to obtain a tiling vector; to perform a third full connection processing on multiple coding features to obtain a third latent layer feature; the sixth mapping process is performed on the third hidden layer feature to obtain a mapping feature with the same dimension as the tiling vector.
  • the second prediction module 5554 is further configured to obtain a score matrix composed of the first recommendation scores of each of the recommended dimensions, and obtain a mapping matrix composed of mapping features corresponding to each of the recommended dimensions ; Computing the element-wise product of the score matrix and the mapping matrix to obtain the fusion feature.
  • the second prediction module 5554 is further configured to perform a seventh mapping process on the fusion feature to obtain a mapping feature corresponding to the fusion feature; based on the mapping feature corresponding to the fusion feature, the to-be The recommended information performs recommendation score prediction processing to obtain a second recommendation score of the target object for the information to be recommended.
  • the information recommendation method is implemented by calling a score prediction model
  • the score prediction model includes: a feature encoding layer, a first recommendation score prediction layer, a feature mapping layer, and a second recommendation score prediction layer
  • the device also includes: a model
  • the training module is configured to respectively encode a plurality of sample reference features of the training samples through the feature encoding layer to obtain a sample encoding feature of each of the sample reference features, and the training samples carry object samples for information samples in The first labels of multiple recommended dimensions, and the second label of the object sample for the information sample
  • the first mapping process is performed on the coding features of the multiple samples to obtain the A plurality of first prediction results corresponding to a plurality of recommendation dimensions one-to-one, wherein the first prediction result represents the recommendation score of the object sample with respect to the information sample in the corresponding recommendation dimension
  • the feature mapping layer Perform the following processing for each of the recommended dimensions: perform a second mapping process on each of the plurality of encoded features in the recommended dimension to
  • the recommendation score prediction process is to obtain the second prediction result of the object sample for the information sample; based on the first prediction result of each of the recommended dimensions and the first label corresponding to each of the recommended dimensions, the second The second prediction result is associated with the second label, and the model parameters of the score prediction model are updated.
  • the model training module is further configured to, for each recommended dimension, construct a first loss function corresponding to the first recommendation score prediction layer based on the first prediction result and the first label of the recommended dimension; based on the first The second prediction result and the second label are used to construct the second loss function corresponding to the second recommended score prediction layer; the second loss function and the first loss function are weighted and summed to obtain the third loss function of the score prediction model; based on the third The loss function updates the model parameters of the score prediction model.
  • the model training module is further configured to perform the following processing for each of the recommended dimensions: based on the first prediction result corresponding to the recommended dimension and the first label of the recommended dimension, construct the corresponding recommended Dimension sub-loss function; determine the recommendation weight corresponding to each of the recommended dimensions, based on the recommendation weight corresponding to each of the recommended dimensions, perform weighted summation on the sub-loss functions of multiple recommended dimensions, and obtain the corresponding A first loss function of the recommendation score prediction layer.
  • An embodiment of the present application provides a computer program product or computer program, where the computer program product or computer program includes computer-executable instructions, and the computer-executable instructions are stored in a computer-readable storage medium.
  • the processor of the electronic device reads the computer-executable instructions from the computer-readable storage medium, and the processor executes the computer-executable instructions, so that the electronic device executes the information recommendation method described above in the embodiment of the present application.
  • the embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are stored.
  • the processor When the computer-executable instructions are executed by a processor, the processor will execute the method provided by the embodiment of the present application.
  • the information recommendation method is, for example, the information recommendation method shown in FIG. 3A .
  • the computer-readable storage medium can be memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; Various equipment.
  • executable instructions may take the form of programs, software, software modules, scripts, or code written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and its Can be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.
  • executable instructions may, but do not necessarily correspond to files in a file system, may be stored as part of a file that holds other programs or data, for example, in a Hyper Text Markup Language (HTML) document in one or more scripts, in a single file dedicated to the program in question, or in multiple cooperating files (for example, files that store one or more modules, subroutines, or sections of code).
  • HTML Hyper Text Markup Language
  • executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one location, or, alternatively, on multiple electronic devices distributed across multiple locations and interconnected by a communication network. to execute.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请提供了一种信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品，应用于车联网领域以及人工智能技术领域；方法包括：对多个参考特征进行编码处理，得到每个所述参考特征的编码特征，其中，所述参考特征包括以下至少之一：目标对象的对象特征、待推荐信息的信息特征；对多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，其中，第一推荐分数表征目标对象针对待推荐信息在对应推荐维度的推荐分数；对多个编码特征进行第二映射处理，得到映射特征；基于映射特征，对多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于融合特征对待推荐信息进行推荐分数预测处理，得到目标对象针对待推荐信息的第二推荐分数；基于第二推荐分数，执行待推荐信息对应目标对象的推荐操作。

Description

信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品
相关申请的交叉引用
本申请基于申请号为202111184748.8、申请日为2021年10月12日的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此引入本申请作为参考。
技术领域
本申请涉及车联网领域以及人工智能技术，尤其涉及一种信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品。
背景技术
人工智能(AI，Artificial Intelligence)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法和技术及应用系统。
人工智能技术广泛应用于推荐系统中，例如，通过推荐系统的多推荐目标排序模型将用户感兴趣的信息推荐给合适的用户，多推荐目标排序模型从用户对信息的点击、消费时长及互动行为等推荐维度(又称目标)对信息进行预估得分，在得到针对每个目标的得分后，如何将多个得分融合将影响推荐系统的精度和用户体验。
相关技术中的融合方案无法准确预测不同用户针对信息的得分，使得无法适用于个性化推荐，进而难以提升推荐系统的推荐精度。
发明内容
本申请实施例提供一种信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品，能够准确预测用户针对待推荐信息的推荐分数，以提升推荐系统的推荐精度。
本申请实施例的技术方案是这样实现的：
本申请实施例提供一种信息推荐方法，包括：
对多个参考特征进行编码处理，得到每个所述参考特征的编码特征；
对所述多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，其中，所述第一推荐分数表征所述目标对象针对待推荐信息在对应推荐维度的推荐分数；
针对每个所述推荐维度执行以下处理：对所述多个编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的映射特征；
基于每个所述推荐维度的映射特征，对所述多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于所述融合特征对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数；
基于所述第二推荐分数，执行所述待推荐信息对应所述目标对象的推荐操作。
本申请实施例提供一种信息推荐装置，包括：
特征编码模块，配置为对多个参考特征进行编码处理，得到每个所述参考特征的编码特征；
第一预测模块，配置为对所述多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，其中，所述第一推荐分数表征所述目标对象针对待推荐信息在对应推荐维度的推荐分数；
特征映射模块，配置为针对每个所述推荐维度执行以下处理：对所述多个编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的映射特征；
第二预测模块，配置为基于每个所述推荐维度的映射特征，对所述多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于所述融合特征对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数；
信息推荐模块，配置为基于所述第二推荐分数，执行所述待推荐信息对应所述目标对象的推荐操作。
本申请实施例提供一种电子设备，包括：
存储器，用于存储计算机可执行指令；
处理器，用于执行所述存储器中存储的计算机可执行指令时，实现本申请实施例提供的信息推荐方法。
本申请实施例提供一种计算机可读存储介质，存储有计算机可执行指令，用于被处理器执行时，实现本申请实施例提供的信息推荐方法。
本申请实施例提供一种计算机程序产品，包括计算机程序或计算机可执行指令，所述计算机程序或计算机可执行指令被处理器执行时实现本申请实施例提供的信息推荐方法。
本申请实施例具有以下有益效果：
先通过对目标对象的多个参考特征的编码特征，预测得到目标对象针对待推荐信息在多个推荐维度(如点击、时长、互动等维度)的第一推荐分数，再通过特征映射的方式将编码特征在各推荐维度进行映射，得到表征在相应的推荐维度第一推荐分数的映射特征，将各推荐维度的第一推荐分数基于映射特征进行融合处理，基于融合结果预测目标对象针对待推荐信息最终的第二推荐分数；如此，可根据不同目标对象的参考特征，采用适合相应目标对象的融合方式进行各个推荐维度的分数的融合，实现自动根据目标对象在不同推荐维度上的倾向性得到精确的最终推荐分数的目的，能够提高最终推荐分数的预测准确性，为推荐系统提供准确的推荐参考数据，进而提升推荐精度和用户体验。
附图说明
图1是本申请实施例提供的信息推荐系统10的架构示意图；
图2是本申请实施例提供的用于信息推荐的电子设备500的结构示意图；
图3A是本申请实施例提供的信息推荐方法的流程示意图；
图3B为本申请实施例提供的第一推荐分数的确定示意图；
图3C为本申请实施例提供的拟合特征的确定示意图；
图4为本申请实施例提供的信息推荐示意图四；
图5A为本申请实施例提供的模型训练方法的流程示意图；
图5B为本申请实施例提供的模型参数更新方法的流程示意图；
图6为本申请实施例提供的信息推荐示意图五；
图7A为本申请实施例提供的信息推荐效果示意图一；
图7B为本申请实施例提供的信息推荐效果示意图二；
图7C为本申请实施例提供的信息推荐效果示意图三。
具体实施方式
为了使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请作进一步地详细描述，所描述的实施例不应视为对本申请的限制，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例，都属于本申请保护的范围。
在以下的描述中，涉及到“一些实施例”，其描述了所有可能实施例的子集，但是可以理解，“一些实施例”可以是所有可能实施例的相同子集或不同子集，并且可以在不冲突的 情况下相互结合。
在以下的描述中，所涉及的术语“第一\第二\第三”仅仅是区别类似的对象，不代表针对对象的特定排序，可以理解地，“第一\第二\第三”在允许的情况下可以互换特定的顺序或先后次序，以使这里描述的本申请实施例能够以除了在这里图示或描述的以外的顺序实施。
除非另有定义，本文所使用的所有的技术和科学术语与属于本申请的技术领域的技术人员通常理解的含义相同。本文中所使用的术语只是为了描述本申请实施例的目的，不是旨在限制本申请。
对本申请实施例进行进一步详细说明之前，对本申请实施例中涉及的名词和术语进行说明，本申请实施例中涉及的名词和术语适用于如下的解释。
1)点击率(CTR，Click Through Rate)：是指网站页面上某一信息被点击的次数与被显示次数之比。
2)时长(Duration)：是指用户对信息的消费时长，如用户阅读信息的时长。
3)互动：包括但不限于用户对信息的点赞、分享、收藏、转发、关注等操作。
4)排序：对召回的候选信息进行打分，按照分数从高到低的顺序从召回的候选信息中选取排序靠前的多个信息作为推荐结果。
5)嵌入式表示(Embedding)：从数学角度看，Embedding是一个用于映射的函数，通过函数将原始数据映射或者嵌入到另一个数值向量空间，是用连续向量表示离散变量，之所以称之为Embedding，是因为这种表示方法往往伴随着降维，就像高维数据被挤压嵌入到低维空间中一样。
6)多层感知器(MLP，Multiple Layer Perceptron)：是一种前馈人工神经网络模型，其将输入的多个数据集映射到单一的输出的数据集上，多层感知器能够处理非线性可分离的问题。
7)多门控混合专家网络(MMoE，Multi-gate Mixture-of-Experts)：用于多推荐目标学习的常用网络结构，由多个专家网络和多个门控网络构成，其中，专家网络多为DNN网络结构，专家网络用于从用于进行多推荐目标学习的输入数据中提取不同的特征，相当于将输入数据所包括的信息划分到多个区域，每个区域对应一个专家网络，每个专家网络从输入数据中提取不同维度的特征，门控网络用于分配每个专家网络的权重，针对多个任务会具有多个门控网络，以任务A为例进行说明，任务A对应的门控网络输出每个专家网络的被选择的概率，将输出的被选择的概率作为对应专家网络的权重，从而可以对多个专家网络输出的特征进行加权求和处理，得到对应任务A的综合特征。
8)个性化特征：根据用户之前的点击数据、互动数据、相似用户的点击数据以及互动数据等等预测用户的需求和偏好，进而给用户推荐可能喜欢的物品。
9)自适应：根据所处理的数据的数据特征自动调整处理方法和参数权重，自适应是是一种数学模型不断逼近目标的过程。
10)多推荐目标融合：学习得到多个目标的预估打分，各预估打分根据每个目标的重要性、业务指标需求等策略进行相加或相乘。
在推荐系统中，多推荐目标排序模型一般用于预估多个推荐目标(即推荐维度)的得分(即推荐分数)，相关技术中难以把多个得分融合成用于进行排序的综合得分，并且在业务上实现最佳的效果。
相关技术中会采取公式融合方法把多个得分融合成用于进行排序的综合得分，具体而言，对各个目标的预测模型进行单独训练，然后将预测出的不同目标的得分通过公式进行融合，然后通过相加、相乘或更复杂的公式进行融合，融合的过程中会使用到参数，为了找到相对较好的参数，还需要离线搜索不同的参数组，常用的方法有网格搜索(grid-search)或者是启发式方法(如遗传算法、粒子群算法等)。
上述方式至少存在以下缺点：如果是多个目标的预测模型单独训练，此种方式成本较高，多个预测模型无法共享参数，从而无法共同训练，也无法加速特征学习，而且在线服务负载 压力较大，加载预测模型个数较多，计算量相对较大，资源消耗多，稳定性较差；如果增加的新目标数据比较稀疏，难以进行有效的模型训练和迭代。可见，不论是多个预测模型单独训练还是多推荐目标网络共同训练，这种方法过于依赖人工规则，存在离线和在线的数据分布差异，离线搜参验证效果依赖在线数据的收集和效果指标的指定，多个目标的重要性难以量化；调参需要遍历很多组参数组合，耗时耗力，难以适应业务数据的实时变化，成本高，缺少个性和场景化；当目标不断增多，公式排序能力受限，无法寻到最优参数组合，可能导致业务指标变差。故上述方法不适用于全量用户，没有考虑用户的个性化水平差异，由于每个用户对不同目标的倾向性不同，限制模型在全部用户上实现最优效果。
为此，本申请实施例提供一种信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品，能够准确预测用户针对信息的推荐分数，以提升推荐系统的推荐精度和用户体验。
本申请实施例提供的信息推荐方法可以由各种电子设备实施，例如，可以由终端单独实施，也可以由服务器单独实施，也可以由终端和服务器协同实施。例如终端独自执行下文所述的信息推荐方法，或者，终端向服务器发送推荐请求，服务器根据接收的推荐请求执行信息推荐方法。
本申请实施例提供的用于信息推荐的电子设备可以是各种类型的终端设备或服务器，其中，服务器可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云计算服务的云服务器；终端可以是智能手机、平板电脑、笔记本电脑、台式计算机、智能音箱、智能手表、车载终端等，但并不局限于此。终端以及服务器可以通过有线或无线通信方式进行直接或间接地连接，本申请实施例在此不做限制。
以服务器为例，例如可以是部署在云端的服务器集群，向用户开放人工智能云服务(AiaaS，AI as a Service)，AIaaS平台会把几类常见的AI服务进行拆分，并在云端提供独立或者打包的服务，这种服务模式类似于一个AI主题商城，所有的用户都可以通过应用程序编程接口的方式来接入使用AIaaS平台提供的一种或者多种人工智能服务。
例如，其中的一种人工智能云服务可以为信息推荐服务，即云端的服务器封装有本申请实施例提供的信息推荐程序。用户通过终端(运行有客户端，例如即时通信客户端、直播客户端、短视频客户端、社交客户端等)调用云服务中的信息推荐服务，以使部署在云端的服务器调用封装的信息推荐程序，确定目标对象针对待推荐信息的推荐分数，并基于推荐分数执行待推荐信息对应目标对象的推荐操作。
在一些实施例中，以服务器单独实施本申请实施例提供的信息推荐方法为例进行说明。服务器分别对多个参考特征进行编码处理，得到各参考特征的编码特征，参考特征包括以下至少之一：目标对象的对象特征、待推荐信息的信息特征；对多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，其中，第一推荐分数表征目标对象针对待推荐信息在对应推荐维度的推荐分数；针对每个推荐维度执行以下处理：对多个编码特征在推荐维度进行第二映射处理，得到推荐维度的映射特征；基于每个推荐维度的映射特征，对多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于融合特征对待推荐信息进行推荐分数预测处理，得到目标对象针对待推荐信息的第二推荐分数；基于待推荐信息的第二推荐分数，执行待推荐信息对应目标对象的推荐操作。
在一些实施例中，以服务器和终端协同实施本申请实施例提供的信息推荐方法为例进行说明。参见图1，图1是本申请实施例提供的信息推荐系统10的架构示意图。终端400通过网络300连接服务器200，网络300可以是广域网或者局域网，又或者是二者的组合。终端400(运行有客户端，例如即时通信客户端、直播客户端、短视频客户端、社交客户端等)可以被用来获取用户的信息推荐请求，以终端400的用户是目标对象为例，例如，当目标对象打开终端上运行的新闻客户端后，终端自动获取针对目标对象的新闻推荐请求。
在一些实施例中，终端获取信息推荐请求后，调用服务器200的信息推荐接口(可以提供为云服务的形式，即信息推荐服务)，服务器200基于信息推荐请求，获取目标对象的多 个参考特征，目标对象即为需要推荐信息的某个用户，参考特征包括以下至少之一：目标对象的对象特征(对象特征即为用户特征，例如，用户年龄、用户性别等数据)、待推荐信息的信息特征；从待推荐信息库中召回符合上述参考特征的待推荐信息作为候选信息进行排序。
可以理解的是，在本申请实施例中，涉及到用户特征等相关的数据，当本申请实施例运用到具体产品或技术中时，需要获得用户许可或者同意，且相关数据的收集、使用和处理需要遵守相关国家和地区的相关法律法规和标准。
在排序阶段，对多个参考特征进行编码处理，得到每个参考特征的编码特征，参考特征包括以下至少之一：目标对象的对象特征、召回的待推荐信息的信息特征；对多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数；针对每个推荐维度执行以下处理：对多个编码特征在推荐维度进行第二映射处理，得到推荐维度的映射特征；基于每个推荐维度的映射特征，对多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于融合特征对待推荐信息进行推荐分数预测处理，得到目标对象针对待推荐信息的第二推荐分数；基于待推荐信息的第二推荐分数对召回的待推荐信息进行重排，并选择排名靠前的多个(即，从排名第一位开始的多个待推荐信息)待推荐信息推送至终端400显示。
需要说明的是，本申请实施例中涉及的目标对象是信息推荐系统推荐信息的接收者，如目标对象打开新闻客户端时，目标对象即为新闻推荐系统推荐新闻的接收者，本申请实施例所涉及的目标对象的对象特征均是在征得目标对象同意的情况下获取的。
在一些实施例中，本申请实施例提供的信息推荐方法还可应用于与车联网服务(如加油、导航、停车、维修等)相关的信息推荐场景中，如在对车载终端进行信息推荐时，对车载终端的目标对象执行本申请实施例提供的信息推荐方法，确定目标对象针对待推荐信息的最终推荐分数，并基于最终推荐分数执行待推荐信息对应目标对象的推荐操作；例如，对最终推荐分数低于分数阈值的待推荐信息应用相应的屏蔽模式，将最终推荐分数超过分数阈值的待推荐信息推荐给车载终端，从而避免质量低下的信息的广泛传播，间接提高整体信息质量，提高了用户体验。
下面说明本申请实施例提供的用于信息推荐的电子设备的结构，参见图2，图2是本申请实施例提供的用于信息推荐的电子设备500的结构示意图，以电子设备500是服务器为例说明，图2所示的用于信息推荐的电子设备500包括：至少一个处理器510、存储器550、至少一个网络接口520和用户接口530。电子设备500中的各个组件通过总线系统540耦合在一起。可理解，总线系统540用于实现这些组件之间的连接通信。总线系统540除包括数据总线之外，还包括电源总线、控制总线和状态信号总线。但是为了清楚说明起见，在图2中将各种总线都标为总线系统540。
处理器510可以是一种集成电路芯片，具有信号的处理能力，例如通用处理器、数字信号处理器(DSP，Digital Signal Processor)，或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等，其中，通用处理器可以是微处理器或者任何常规的处理器等。
存储器550包括易失性存储器或非易失性存储器，也可包括易失性和非易失性存储器两者。其中，非易失性存储器可以是只读存储器(ROM，Read Only Memory)，易失性存储器可以是随机存取存储器(RAM，Random Access Memory)。本申请实施例描述的存储器550旨在包括任意适合类型的存储器。存储器550可选地包括在物理位置上远离处理器510的一个或多个存储设备。
在一些实施例中，存储器550能够存储数据以支持各种操作，这些数据的示例包括程序、模块和数据结构或者其子集或超集，下面示例性说明。
操作系统551，包括用于处理各种基本系统服务和执行硬件相关任务的系统程序，例如框架层、核心库层、驱动层等，用于实现各种基础业务以及处理基于硬件的任务；
网络通信模块552，用于经由一个或多个(有线或无线)网络接口520到达其他电子设备，示例性的网络接口520包括：蓝牙、无线相容性认证(WiFi)、和通用串行总线(USB，Universal Serial Bus)等；
呈现模块553，用于经由一个或多个与用户接口530相关联的输出装置531(例如，显示屏、扬声器等)使得能够呈现信息(例如，用于操作外围设备和显示内容和信息的用户接口)；
输入处理模块554，用于对一个或多个来自一个或多个输入装置532之一的一个或多个用户输入或互动进行检测以及翻译所检测的输入或互动。
在一些实施例中，本申请实施例提供的信息推荐装置可以采用软件方式实现，例如，可以是上文所述的服务器中信息推荐服务，还可以是上文所述的终端中信息推荐插件。当然，不局限于此，本申请实施例提供的信息推荐装置可以提供为各种软件实施例，包括应用程序、软件、软件模块、脚本或代码在内的各种形式。
在一些实施例中，本申请实施例提供的信息推荐装置可以采用软件方式实现，图2示出了存储在存储器550中的信息推荐装置555，其可以是程序和插件等形式的软件，包括以下软件模块：特征编码模块5551、第一预测模块5552、特征映射模块5553、第二预测模块5554和信息推荐模块5555，这些模块是逻辑上的，因此根据所实现的功能可以进行任意的组合或进一步拆分。将在下文中说明每个模块的功能。
下面将结合附图对本申请实施例提供的信息推荐方法进行说明，下述信息推荐方法的执行主体可以为服务器，具体可以是服务器通过运行上文的各种计算机程序来实现的；当然，根据对下文的理解，不难看出也可以由终端和服务器协同实施本申请实施例提供的信息推荐方法。
参见图3A，图3A是本申请实施例提供的信息推荐方法的流程示意图，将结合图3A示出的步骤进行说明。
在步骤101中，服务器对多个参考特征进行编码处理，得到每个参考特征的编码特征。
参考特征包括以下至少之一：目标对象的对象特征、待推荐信息的信息特征，对象特征是目标对象的基础属性特征(如年龄、性别、职业、受教育程度、消费水平等)、画像特征(如兴趣爱好、浏览、点击、收藏、购买等行为数据等)、以及推荐所处的上下文特征(环境特征，如推荐时间、推荐场景等)，信息特征是待推荐信息的信息标签、信息类别、信息来源、交互特征，交互特征即为信息内容与对象特征相关的特征，交互特征具体是信息特征与对象特征的交集，如统计待推荐信息在不同年龄、性别用户的消费情况，得到与该用户具有相同年龄、性别的用户对待推荐信息的消费特征，该消费特征即为待推荐信息与用户之间的交互特征。
可以理解的是，在本申请实施例中，涉及到基础属性特征、画像特征等相关的数据，当本申请实施例运用到具体产品或技术中时，需要获得用户许可或者同意，且相关数据的收集、使用和处理需要遵守相关国家和地区的相关法律法规和标准。
在实际实施时，可从中选择任意一个特征或多个特征进行自由组合作为参考特征，也即参考特征的规模维度可设置，如将目标对象的所有或部分对象特征作为参考特征，或将待推荐信息的所有或部分信息特征作为参考特征，或将目标对象的所有对象特征和待推荐信息的所有信息特征作为参考特征，或将目标对象的部分对象特征和待推荐信息的所有信息特征作为参考特征，将目标对象的部分对象特征和待推荐信息的部分信息特征作为参考特征，等等。
在一些实施例中，参考特征包括连续型特征及离散型特征中至少之一，步骤101可通过如下方式实现：针对每个参考特征执行以下处理：当参考特征为连续型特征时，对连续型特征进行离散化处理，得到连续型特征的离散型特征，并对连续型特征的离散型特征进行编码处理，得到连续型特征的编码特征；当参考特征为离散型特征时，对离散型特征进行编码处理，得到离散型特征的编码特征，通过本申请实施例可以将不同形式的特征进行统一编码，相当于映射到相同的编码空间，从而可以基于统一的参考特征进行后续数据处理，提升数据处理效率以及准确度。
作为示例，连续型特征的数据形式是连续数据，它可以是区间内的任意取值，例如，视 频时长、发布失效均属于连续型特征，离散型特征的数据形式是离散数据，一般都以整数的形式表现，例如，用户的年龄、用户的性别，利用离散数据0表征用户性别为男，利用离散数据1表征用户性别为1。对连续型特征进行离散化的方式有多种，例如，采取卡房检验的方式，例如采取决策树的方式进行离散化，还可以利用划分区间的方式进行离散化。作为示例，不同的参考特征的编码过程是不同的，参见图4，图4为本申请实施例提供的信息推荐示意图，以待推荐信息为视频为例进行说明，参考特征中的连续型特征可为视频时长、发布实效等，离散型特征可为视频标识、用户账号等级、用户性别等，需要对不同的特征进行针对性的编码处理。例如，对视频时长、发布实效等连续型特征进行离散化，或对连续型特征进行归一化或标准化处理，得到连续型特征的离散型特征(即离散数值，又称稀疏特征)，再对连续型特征的离散型特征进行编码处理，例如将离散型特征映射到另一个向量空间，得到连续型特征的编码特征(嵌入式表示Embedding，又称稠密特征)；对于视频标识、用户账号等级、用户性别等离散型特征直接进行编码处理，例如将离散型特征映射到另一个向量空间，得到离散型特征的编码特征(嵌入式表示Embedding，又称稠密特征)。
在步骤102中，对多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数。
在一些实施例中，参见图3B，图3B为本申请实施例提供的第一推荐分数的确定示意图，步骤102可通过图3B示出的步骤1021至步骤1024实现：
在步骤1021中，对多个编码特征进行特征交叉处理，得到至少一个交叉特征。
在一些实施例中，在执行步骤1021之前，对多个参考特征进行线性处理，得到一阶特征，之后，服务器可通过如下方式对多个编码特征进行特征交叉处理，得到至少一个交叉特征：对多个编码特征中任意两个编码特征进行二阶特征交叉处理，得到任意两个编码特征的二阶交叉特征；将一阶特征与每个二阶交叉特征进行拼接处理，得到至少一个交叉特征。通过本申请实施例可以使得交叉特征具有记忆性，并且可以提升后续数据处理的非线性能力，从而提高推荐准确度。
如图4所示，对多个参考特征进行线性计算，如基于每个参考特征的权重，将多个参考特征进行加权求和处理，得到一阶特征；对任意两个编码特征进行二阶特征交叉处理，得到任意两个编码特征的二阶交叉特征，以参考特征是对象特征为例进行说明，对象特征的编码特征x iu＝[x iu1,x iu2,...,x iun]，x iu表征用户i的对象特征，m为对象特征的数目，x iu1表征用户i的第1个对象特征，则任意两个编码特征的二阶交叉特征为p i,j＝＜x iua,x iub＞,a＝1,....,m,b＝1,...,m，对所有任意两个编码特征的二阶交叉特征p＝∑(p i,j)与参考特征的一阶特征进行拼接处理，得到低阶且具有记忆性的交叉特征，交叉特征的数目与二阶交叉特征的数目相同，低阶特征指的是阶数低于设定阈值的特征。
在一些实施例中，还可获取任意指定阶的交叉特征，如对多个参考特征分别对应的编码特征中任意i(2≤i≤m，m为参考特征的个数)个编码特征进行i阶交叉处理，得到任意i个编码特征的i阶交叉特征，将一阶特征及i阶交叉特征进行拼接处理，得到交叉特征，以便后续进行第一推荐评分的预测时考虑到高阶信息，提升推荐准确度。
在一些实施例中，为了提高处理速度，还可将一阶特征与i阶交叉特征进行拼接处理，得到的拼接特征，再对得到的拼接特征进行矩阵分解处理，如当i＝2时，对所有任意两个编码特征的二阶交叉特征与参考特征的一阶特征进行拼接得到的拼接特征后，对得到的拼接特征进行矩阵分解处理，得到分解特征，通过激活函数对分解特征进行非线性映射处理，得到对应的交叉特征。
通过上述方式，通过对各编码特征进行特征交叉处理，捕捉不同编码特征之间的交叉信息，增强编码特征表征能力，避免遗漏特征边界，以便后续基于准确的交叉特征进行后续的预测处理。
在步骤1022中，基于多个编码特征，对待推荐信息在每个推荐维度的拟合性进行预测， 得到对应每个推荐维度的拟合特征。
拟合特征为用于表征所有推荐维度中任意两个推荐维度之间的拟合自由度的特征，拟合特征通常为高阶特征，高阶特征是阶数大于设定阈值的特征，推荐维度指的是用于评估推荐系统的指标，例如，推荐维度包括点击率、互动次数、观看时长等等。
在一些实施例中，参见图3C，图3C为本申请实施例提供的拟合特征的确定示意图，步骤1022可通过针对每个推荐维度执行图3C示出的步骤10221至步骤10223实现：步骤10221中，通过多门控混合专家网络中每个专家网络，对多个编码特征进行第一全连接处理，得到第一隐层特征，并对第一隐层特征进行第四映射处理，得到对应每个专家网络的映射特征；步骤10222中，通过多门控混合专家网络中对应推荐维度的门控网络，对多个编码特征进行第二全连接处理，得到第二隐层特征，并对第二隐层特征进行第五映射处理，得到对应每个专家网络的权重特征；步骤10223中，基于每个专家网络的权重特征，对每个专家网络的映射特征进行加权求和处理，得到对应推荐维度的拟合特征。通过本申请实施例同时学习针对多个推荐维度的任务，使这些任务取得比单独训练一个推荐维度的任务更好的效果，可以缓解数据处理过程中的过拟合现象。
多门控混合专家网络由多个专家网络和多个门控网络构成，专家网络用于提取不同的映射特征，专家网络的结构可为全连接神经网络结构，门控网络用于输出分配给每个专家网络的权重特征，每个门控网络相当于一个分类器，每个推荐维度的门控网络会根据当前输入的编码特征，判别它由哪些专家网络来拟合会更好，从而预估出每个专家网络的权重特征。门控网络的数量与推荐维度的数量一致，专家网络的数量可以与推荐维度的数量一致或者不一致，即每个推荐维度对应一个门控网络，如推荐维度为点击率、观看时长和互动次数三个维度时，“点击率”这一推荐维度对应门控网络1，“观看时长”这一推荐维度对应门控网络2，“互动次数”这一推荐维度对应门控网络3。
如图4所示，将多个参考特征对应的编码特征均输入至每个专家网络中，首先通过对应各推荐维度的专家网络，对多个编码特征进行第一全连接处理，得到对应的第一隐层特征，并通过激活函数对第一隐层特征进行线性或非线性映射处理，得到对应每个专家网络的映射特征；然后将多个编码特征以及多个专家网络输出的映射特征输入至门控网络中，通过门控网络对多个编码特征进行第二全连接处理，例如多层感知处理，得到第二隐层特征，然后通过激活函数对第二隐层特征进行线性或非线性映射处理，得到对应每个专家网络的权重特征，并以此权重特征，对每个专家网络输出的映射特征进行加权求和处理，得到对应推荐维度(与门控网络对应的推荐维度)的拟合特征(门控网络的输出)，并将每个推荐维度对应的门控网络的输出作为整个多门控混合专家网络的输出。
在步骤1023中，针对每个所述推荐维度执行以下处理：将所述交叉特征与所述推荐维度的拟合特征进行拼接处理，得到对应所述推荐维度的拼接特征。
这里，将低阶且具有记忆性的交叉特征与对应各推荐维度的高阶特征进行拼接后，基于拼接特征进行后续的分数预测，可以提高分数预测准确度，从而提高推荐准确度。
如图4所示，在获得每个推荐维度对应的门控网络输出的拟合特征后，分别将交叉特征及对应每个推荐维度的拟合特征进行拼接处理，得到对应每个推荐维度的拼接特征。仍以点击率、观看时长和互动次数这三个推荐维度为例，将上述得到的交叉特征与“点击率”这一推荐维度的拟合特征进行拼接，得到“点击率”这一推荐维度的拼接特征；将上述得到的交叉特征与“观看时长”这一推荐维度的拟合特征进行拼接，得到“观看时长”这一推荐维度的拼接特征；将上述得到的交叉特征与“互动次数”这一推荐维度的拟合特征进行拼接，得到“互动次数”这一推荐维度的拼接特征。
在步骤1024中，针对每个推荐维度执行以下处理：对推荐维度的拼接特征进行第三映射处理，得到待推荐信息对应推荐维度的第一推荐分数。
在一些实施例中，针对各推荐维度，对相应的拼接特征进行第三映射处理，得到对应拼接特征的映射特征，通过激活函数对拼接特征的映射特征进行偏置处理，得到对应每个推荐 维度的第一推荐分数，第一推荐分数表征目标对象针对待推荐信息在对应推荐维度的推荐分数。
在一些实施例中，针对各推荐维度，可以通过逻辑回归函数对拼接特征进行线性逻辑回归处理，这里的线性逻辑回归处理可以是线性加和处理，将得到的线性加和结果作为投影特征，也可以是将线性加和结果代入逻辑回归函数，得到逻辑回归特征作为投影特征，然后经过激活函数对投影特征进行推荐分数预测处理，得到表征推荐分数高低的第一推荐分数。
仍以上述示例为例，在分别得到对应点击率、观看时长和互动次数这三个推荐维度的拼接特征后，分别预测得到“点击率”这一推荐维度的第一推荐分数、“观看时长”这一推荐维度的第一推荐分数，以及“互动次数”这一推荐维度的第一推荐分数。
在步骤103中，针对每个推荐维度执行以下处理：对多个编码特征在推荐维度进行第二映射处理，得到推荐维度的映射特征。
在一些实施例中，步骤103可通过如下方式实现：对多个推荐维度的第一推荐分数进行横向拼接处理，得到平铺向量；对多个编码特征进行第三全连接处理，得到第三隐层特征；对第三隐层特征进行第六映射处理，得到与平铺向量维度相同的映射特征。
这里，在实际实施时，在得到对应每个推荐维度的第一推荐分数之后，对每个第一推荐分数的向量表示进行横向拼接处理，得到第一推荐分数对应的平铺向量，记为x in＝[x i1,x i2,...,x in]，其中，x in表示第n个推荐维度的第一推荐分数的向量表示，n表示推荐维度的个数；然后，将多个编码特征降维至与平铺向量维度规模相同的映射特征，如将多个编码特征通过第三全连接处理，得到对应的第三隐层特征，并通过激活函数对隐层特征进行非线性映射处理(第六映射处理)，得到与平铺向量维度规模相同的映射特征，以便于后续映射特征与各推荐维度的第一推荐分数的关联计算。
在步骤104中，基于每个推荐维度的映射特征，对多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于融合特征对待推荐信息进行推荐分数预测处理，得到目标对象针对待推荐信息的第二推荐分数。
在一些实施例中，步骤104中基于每个推荐维度的映射特征，对多个推荐维度的第一推荐分数进行融合处理，得到融合特征，相当于是将每个推荐维度的映射特征作为每个推荐维度的第一推荐分数的权重，可以基于每个推荐维度的第一推荐分数的权重对多个推荐维度的第一推荐分数进行加权求和处理，具体可通过如下方式实现：获取由每个推荐维度的第一推荐分数构成的分数矩阵，并获取由每个推荐维度对应的映射特征构成的映射矩阵；将分数矩阵与映射矩阵进行元素积计算，得到融合特征。
分数矩阵即为上述对各第一推荐分数的向量表示进行横向拼接处理得到的平铺向量x in，映射矩阵为与平铺向量维度规模相同的映射特征，记为w iu，将分数矩阵与映射矩阵进行元素积计算得到的融合特征记为：w i,um＝w iu·x im，表示将多推荐维度的分数矩阵和映射矩阵做矩阵点乘，以衡量目标对象在不同推荐维度下的倾向大小。
在一些实施例中，步骤104中基于融合特征对待推荐信息进行推荐分数预测处理，得到目标对象针对待推荐信息的第二推荐分数，可通过如下方式实现：对融合特征进行第七映射处理，得到对应融合特征的映射特征；基于对应融合特征的映射特征，对待推荐信息进行推荐分数预测处理，得到目标对象针对待推荐信息的第二推荐分数。
在进行推荐分数预测时，对融合特征进行映射处理，如将融合特征通过逻辑回归函数做线性投影，然后将得到的投影值经过激活函数，预测得到目标对象针对待推荐信息的第二推荐分数。
在步骤105中，基于第二推荐分数，执行待推荐信息对应目标对象的推荐操作操作。
第二推荐分数是综合多个推荐维度的最终得分，用于表征目标对象针对待推荐信息的整体评价，当第二推荐分数超过分数阈值时，将待推荐信息推荐给目标对象。
在一些实施例中，本申请实施例提供的信息推荐方法可适用于在推荐系统的召回阶段。
下面介绍推荐系统的结构，推荐系统包括召回阶段、粗排阶段、精排阶段以及重排阶段， 召回阶段是：从候选池中选出和候选信息，交给后面的排序阶段进行打分，候选池是可供推荐的候选信息池。粗排阶段是：对召回选择的几千或者几百候选信息进行排序。精排阶段是：对粗排的结果进行精准排序。重排阶段是：对精排结果做小幅调整。
在获取召回的每个候选信息的第二推荐分数后，按照第二推荐分数由高到低的顺序，对召回的候选信息进行排序，并选择排名靠前的多个的候选信息(即，从排名第一位开始的多个候选信息)推送至终端显示。
在一些实施例中，还可根据第二推荐分数可确定目标对象针对待推荐信息的评价等级，进而根据评价等级进行不同的推荐操作。例如，评价等级包括等级依次增高(用户越来越感兴趣)的第一等级、第二等级和第三等级时，当针对待推荐信息的评价等级为第一等级时，在推荐系统的排序阶段，对待推荐信息进行降权推荐，以减少推荐次数或推荐频率，例如，在未采取降权排序之前，在一周内可能会向100个人推荐该信息，在采取降权排序之后，在一周内可能只向20个人推荐该信息，另外，降权的幅度与待推荐信息的最终得分数呈负相关关系，即待推荐信息的最终得分越低，降权幅度越大，降权排序后在一定时间内针对该信息的推荐次数或推荐频率就越低；在推荐系统的召回阶段，将包含待推荐信息的召回结果中对待推荐信息进行暂时过滤或永久过滤，然后对过滤后的信息(召回得到的信息)进行粗排处理、精排处理以及重排处理，最后基于重排结果进行推荐，以避免将用户不感兴趣的信息推荐给目标对象或与目标对象相似的其他用户。
当针对待推荐信息的评价等级为第二等级时，对待推荐信息进行自由推荐，自由推荐即不对待推荐信息进行偏向性推荐，既不加权推荐，也不降权推荐，使其基于用户需求和信息自身质量进行推荐。当针对待推荐信息的评价等级为第三等级时，对待推荐信息进行加权推荐，从而使目标对象感兴趣的待推荐信息可以被推荐给更多与目标对象相似的其他用户，增加待推荐信息的曝光率和点击率。
在一些实施例中，上述信息推荐方法是通过调用分数预测模型实现的，如图4所示，分数预测模型包括：特征编码层、第一推荐分数预测层、特征映射层和第二推荐分数预测层；其中，第一推荐分数预测层包括第一特征提取层、第二特征提取层、特征拼接层和子分数预测层；第二推荐分数预测层包括特征融合层和总分数预测层。
在一些实施例中，参见图5A，图5A为本申请实施例提供的模型训练方法的流程示意图，可通过如下方式对分数预测模型进行训练：在步骤201中，服务器通过特征编码层，分别对训练样本的多个样本参考特征进行编码处理，得到每个样本参考特征的样本编码特征，训练样本携带对象样本针对信息样本在多个推荐维度的第一标签，以及对象样本针对信息样本的第二标签；在步骤202中，通过第一推荐分数预测层，对多个样本编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一预测结果，其中，第一预测结果表征对象样本针对信息样本在对应推荐维度的推荐分数；在步骤203中，通过特征映射层，针对每个推荐维度执行以下处理：对各多个编码特征在推荐维度进行第二映射处理，得到推荐维度的样本映射特征；在步骤204中，通过第二推荐分数预测层执行以下处理，基于每个推荐维度的样本映射特征，对多个推荐维度的第一预测结果进行融合处理，得到样本融合特征，并基于样本融合特征，对信息样本进行推荐分数预测处理，得到对象样本针对信息样本的第二预测结果；在步骤205中，基于每个推荐维度的第一预测结果、对应每个的推荐维度的第一标签，第二预测结果与第二标签，更新分数预测模型的模型参数。
在实际实施时，将训练样本输入至分数预测模型中，首先，通过特征编码层将训练样本的多个样本参考特征进行编码处理，将多个参考特征的稀疏特征转换为稠密特征。其次，通过第一推荐分数预测层中的第一特征提取层，对训练样本的多个样本编码特征进行特征交叉处理，得到样本交叉特征；通过第二特征提取层，基于多个样本编码特征对待推荐信息在至少两个推荐维度的拟合性进行预测，得到对应每个推荐维度的样本拟合特征；通过特征拼接层，分别将样本交叉特征及对应每个推荐维度的样本拟合特征进行拼接处理，得到对应每个推荐维度的样本拼接特征；通过子分数预测层，基于样本拼接特征针对待推荐信息在至少两 个推荐维度的推荐分数进行预测，得到对象样本针对信息样本在至少两个推荐维度的第一预测结果。再次，通过特征映射层，对训练样本的多个编码特征在每个推荐维度进行第二映射处理，得到样本映射特征。最后，通过第二分数预测层中的特征融合层，基于每个推荐维度的样本映射特征，将对象样本针对信息样本在至少两个推荐维度的第一预测结果进行融合处理，得到对应的样本融合特征；通过总分数预测层，基于样本融合特征，对待推荐信息进行推荐分数预测处理，得到目标对象针对待推荐信息的第二预测结果。
作为示例，不同的样本参考特征的编码过程是不同的，当样本参考特征是连续型特征时，对连续型特征的离散型特征进行编码处理，例如将离散型特征映射到另一个向量空间，得到连续型特征的编码特征(嵌入式表示Embedding，又称稠密特征)；当样本参考特征是离散型特征时，对于视频标识、用户账号等级、用户性别等离散型特征直接进行编码处理，例如将离散型特征映射到另一个向量空间，得到离散型特征的编码特征(嵌入式表示Embedding，又称稠密特征)。
在一些实施例中，参见图5B，图5B为本申请实施例提供的模型参数更新方法的流程示意图，步骤205可通过图5B示出的步骤2051至步骤2054实现：在步骤2051中，针对每个推荐维度，基于第一预测结果及推荐维度的第一标签，构造对应第一推荐分数预测层的第一损失函数；在步骤2052中，基于第二预测结果及第二标签，构造对应第二推荐分数预测层的第二损失函数；在步骤2053中，将第二损失函数及第一损失函数进行加权求和，得到分数预测模型的第三损失函数；在步骤2054中，基于第三损失函数更新分数预测模型的模型参数。
在一些实施例中，上述步骤2051可通过如下方式实现：基于各推荐维度对应的第一预测结果及相应的第一标签，构造对应各推荐维度的子损失函数；确定各推荐维度对应的推荐权重，基于各推荐权重对对应各推荐维度的子损失函数进行加权求和，得到对应第一推荐分数预测层的第一损失函数。
这里，对于每个推荐维度，在得到对应的第一预测结果后，可基于第一预测结果与对象样本针对信息样本在相应推荐维度的第一标签，构建对应的子损失函数，并将多个推荐维度的子损失函数相加得到第一推荐分数预测层的第一损失函数
Figure PCTCN2022116402-appb-000001
其中，n表示推荐维度的个数，loss j表示第j个推荐维度对应的子损失函数，1≤j≤n。
在得到综合每个推荐维度最终的第二预测结果后，可基于第二预测结果与对象样本针对信息样本的第二标签，构建第二推荐分数预测层的第二损失函数L(θ)表示为：
Figure PCTCN2022116402-appb-000002
其中，p(w i,um|θ)＝σ(f(w i,um|θ))，σ(·)为sigmoid函数，f(w i,um|θ)为第二预测结果，将第二预测结果转换为预估概率p(w i,um|θ)，n为训练样本的总数，θ为模型参数，yi为第二标签，a i为根据推荐维度设置的不同权重，如一个训练样本有点击和互动两个推荐维度，a i可设置为2，比单纯仅有点击维度这一个推荐维度的训练样本的权重(小于2)大，模型更倾向学习带互动维度的训练样本。
分数预测模型整体的第三损失函数为第一损失函数(即多个推荐维度中每个推荐维度单独的子损失函数之和)和第二损失函数相加表示：
Figure PCTCN2022116402-appb-000003
在构建第三损失函数后，根据第三损失函数的值判断第三损失函数的值是否超出预设阈值，当第三损失函数的值超出预设阈值时，基于第三损失函数确定分数预测模型的误差信号，将误差信息在分数预测模型中反向传播，并在传播的过程中更新各个层的模型参数。
这里，对反向传播进行说明，将训练样本的参考特征输入到神经网络模型的输入层，经 过隐藏层，最后达到输出层并输出结果，这是神经网络模型的前向传播过程，由于神经网络模型的输出结果与实际结果有误差，则计算输出结果与实际值之间的误差，并将该误差从输出层向隐藏层反向传播，直至传播到输入层，在反向传播的过程中，根据误差调整模型参数的值；不断迭代上述过程，直至收敛，其中，分数预测模型属于神经网络模型。
下面，将说明本申请实施例在一个实际的应用场景中的示例性应用。本申请实施例提供的信息推荐方法可应用于一切使用多推荐目标排序模型的推荐系统上，如可应用于客户端推荐、浏览器信息流场景、新闻、快报推荐等信息流产品，还可以应用于如电商领域、广告推荐场景等其他推荐场景。接下来以点击、时长、互动这三个推荐维度的多推荐目标分数融合为例，对本申请实施例提供的信息推荐方法进行说明。
参见图6，图6为本申请实施例提供的信息推荐示意图，通过多推荐目标排序模型进行信息推荐，该模型包括：稀疏特征层、特征提取层、子得分预测层、特征映射层、融合部分，接下来将结合图6对分数预测模型的训练和应用进行说明。
1、稀疏特征层
在选择训练样本的用户侧特征(即上述的参考特征)时，可从对象样本的用户特征(即上述的对象特征)、待推荐信息的信息特征中选择，其中，用户特征是目标对象的基础属性特征(如年龄、性别、职业、受教育程度、消费水平等)、画像特征(如兴趣爱好、浏览、点击、收藏、购买等行为数据等)、以及推荐所处的上下文特征(环境特征，如推荐时间、推荐场景等)，信息特征是信息样本的信息标签、信息类别、信息来源、信息内容与用户特征的交互特征，交互特征是指信息特征与用户特征的交集，如统计待推荐信息在不同年龄、性别用户的消费情况，通过具体的用户年龄、性别，得到该用户年龄、性别下的用户对待推荐信息的消费特征，该消费特征即为待推荐信息与用户之间的交互特征。
当用户侧特征中存在连续型特征时，需先对连续型特征进行离散化处理，或对连线型特征进行归一化或标准化处理得到离散型特征；通常情况下，离散型特征为稀疏特征，需通过稀疏特征层对离散型特征进行编码处理，如通过Embedding向量转化处理，得到对应的编码特征(又称稠密特征)；将得到的编码特征及用户侧本身即为稠密特征的用户侧特征进行拼接，得到用户侧特征向量(即编码特征)为x iu＝[x iu1,x iu2,...,x iun]，其中，m为用户侧特征的个数。
2、特征提取层
特征提取层包括交叉特征提取层和拟合特征提取层，其中，交叉特征提取层可为因子分解机(FM，Factorization Machine)模型，拟合特征提取层可为MMoE模型，通过交叉特征提取层，对用户侧特征的各编码特征进行二阶特征交叉，得到对应的二阶交叉特征，将各编码特征的一阶特征与二阶交叉特征进行拼接，得到低阶具有记忆性的交叉特征；MMoE模型由多个专家网络和多个门控网络构成，专家网络用于提取不同的特征，可为DNN网络结构，门控网络用于分配每个专家网络的权重，每个门控网络相当于一个分类器，每个推荐维度的门控网络会根据当前输入的编码特征，判别它由哪些专家网络来拟合会更好，故预估出每个专家网络的权重。最后，将低阶具有记忆性的交叉特征与MMoE模型输出的对应各个目标的高阶特征进行拼接后，输入至子分数预测层中进行分数预测。
3、子分数预测层
其中，子分数预测层包括对点击、时长、互动进行分数预测的三个模型，三个模型相互独立，将交叉特征与MMoE模型输出的对应各个目标的输出进行拼接后输入至相应的模型中进行分数预测，得到相应的得分(即上述的第一推荐分数)。
4、特征映射层
在得到对应每个目标的得分之后，对各得分的向量表示进行横向拼接处理，得到多推荐目标得分向量，记为x im＝[x i1,x i2,...,x in]，其中，x in表示第n个目标的得分的向量表示，n表示目标的个数；然后，通过特征映射层将用户侧特征的编码特征降维至与多推荐目标得分向量维度规模相同的矩阵，记为w iu，其中，特征映射层可为MLP网络，如DNN网络；然 后将多推荐目标得分向量与降维后用户侧特征的矩阵进行元素积计算得到的融合特征记为：w i,um＝w iu⊙x im，以衡量目标对象在不同目标下的倾向大小。
需要说明的是，特征映射层实质为用户的轻量网络，输入到特征映射层的特征可以来源于稀疏特征层输出的编码特征，即输入到特征映射层的特征可以是稀疏特征层输出的编码特征的部分或全部，也可以是其他新的特征，如获取与输入到稀疏特征层的用户侧特征不同的用户侧特征，新获取的用户侧特征中甚至可以包括信息特征，并对新获取的用户侧特征进行编码处理后输入至特征映射层。
通过上述方式，用户个性化特征的引入可以根据不同用户给出所有目标分数的最优融合方式，相当于自动根据用户在不同目标上的倾向给出最终得分，在业务表现上达到相对更优的效果。
5、融合部分
融合部分的作用在于预测目标对象针对待推荐信息的最终得分，在实际实施时，将上述融合特征w i,um＝w iu⊙x im通过DNN预测得到目标对象针对待推荐信息的最终得分(即上述的第二推荐分数)：final score＝f(w i,um|θ)。
6、损失函数
这里，在得到对应每个目标的得分后，可基于得分与训练样本携带的对象样本针对信息样本在相应目标的标签，构建对应的子损失函数，并将各个目标的子损失函数相加得到子分数预测层的损失函数
Figure PCTCN2022116402-appb-000004
其中，n表示目标的个数，loss j表示第j个目标对应的子损失函数，1≤j≤n。
在得到最终得分后，可基于最终得分与训练样本携带的对象样本针对信息样本的标签，构建融合部分的损失函数L(θ)表示为：
Figure PCTCN2022116402-appb-000005
其中，p(w i,um|θ)＝σ(f(w i,um|θ))，σ(·)为sigmoid函数，f(w i,um|θ)为最终得分，将最终得分转换为预估概率p(w i,um|θ)，n为训练样本的总数，y i为标签，a i为目标设置的不同权重，如一个训练样本有点击和互动两个目标，a i可设置为2，比单纯仅有点击这一个目标的训练样本的权重(小于2)大，模型更倾向学习带互动的训练样本。
多推荐目标排序模型整体的损失函数为子分数预测层的损失函数和融合部分的损失函数之和，表示为：
Figure PCTCN2022116402-appb-000006
在构建多推荐目标排序模型整体的损失函数后，根据多推荐目标排序模型整体的损失函数的值(如梯度值)判断是否超出预设阈值，当超出预设阈值时，基于多推荐目标排序模型整体的损失函数确定模型的误差信号，将误差信号在分数预测模型中反向传播，并在传播的过程中更新各个层的模型参数。
可见，本申请实施例提供的多推荐目标排序模型是一种端到端的模型，不需要考虑离线和在线两种场景下数据的分布影响，融合部分的损失函数和多推荐目标其他损失函数联合训练，在线预测使用只需要加载一个模型，提高服务部署的便利性和稳定性。
8、预测阶段
以待推荐信息为文章为例，当用户请求时，将用户侧特征(包括用户特征、候选文章的信息特征、交叉特征和上下文特征等)输入到多推荐目标排序模型中，得到当前用户对每篇候选文章的预估点击率，预估时长或时长概率，转换为得分构成多推荐目标得分向量x im，同时根据需要的用户侧特征构建特征映射层，经过MLP后输出w iu，将x im和w iu进行点乘， 得到点乘结果w i,um＝w iu⊙x im，最后把w i,um输入到融合部分，得到用户对候选文章的最终得分final score＝f(w i,um|θ)，根据最终得分从大到小的顺序，对所有候选文章进行排序，返回前面K篇文章作为结果呈现给用户。
参见图7A-图7C，图7A-图7C为本申请实施例提供的信息推荐效果示意图，以将本申请实施例提供的信息推荐方法应用于看点图文推荐场景下为例，使用本申请实施例提供的信息推荐方法相较于，一般的公式融合和网格搜参方法，在点击率、阅读总时长和点赞人数这三个目标上的相对提升幅度，如点击率平均相对提升了1.16％，最高相对提升1.62％(图7A)，时长平均相对提升了1.17％，最高相对提升1.38％(图7B)、点赞平均相对提升了2.76％，最高相对提升3.77％(图7C)，其中，空跑期是用一般的公式融合和网格搜参方法，实验期为本申请实施例提供的信息推荐方法相对对比一般融合和网格搜参方法的提升效果。
通过上述方式，本申请实施例提供一种端到端的基于用户个性化特征的多推荐目标分数融合模型，旨在解决多推荐目标模型在不同目标上的打分融合成一个得分进行排序的问题，本申请实施例提供的信息推荐方法在基于MMoE的多推荐目标模型的基础上，引入MLP网络自适应学习多推荐目标融合打分，即通过引入用户侧特征，可以自适应的学习每个用户到不同目标得分的个性化权重，综合用户侧特征得到最佳的融合得分，进而实现各个目标上的最优，不再依赖人工公式和搜参，省时省力。
下面继续说明本申请实施例提供的信息推荐装置555的实施为软件模块的示例性结构，在一些实施例中，存储在图2中存储器550的信息推荐装置555中的软件模块可以包括：特征编码模块5551，配置为分别对目标对象的多个参考特征进行编码处理，得到每个参考特征的编码特征；第一预测模块5552，配置为对所述多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，其中，所述第一推荐分数表征所述目标对象针对待推荐信息在对应推荐维度的推荐分数；特征映射模块5553，配置为针对每个所述推荐维度执行以下处理：对所述多个编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的映射特征；第二预测模块5554，配置为基于每个所述推荐维度的映射特征，对所述多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于所述融合特征对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数；信息推荐模块5555，配置为基于第二推荐分数，执行待推荐信息对应目标对象的推荐操作。
在一些实施例中，参考特征包括连续型特征及离散型特征中至少之一，特征编码模块5551，还配置为针对每个所述参考特征执行以下处理：当所述参考特征为所述连续型特征时，对所述连续型特征进行离散化处理，得到所述连续型特征的离散型特征，并对所述连续型特征的离散型特征进行编码处理，得到所述连续型特征的编码特征；当所述参考特征为所述离散型特征时，对所述离散型特征进行编码处理，得到所述离散型特征的编码特征。
在一些实施例中，第一预测模块5552，还配置为对所述多个编码特征进行特征交叉处理，得到至少一个交叉特征；基于所述多个编码特征，对所述待推荐信息在每个所述推荐维度的拟合性进行预测，得到对应每个所述推荐维度的拟合特征；针对每个所述推荐维度执行以下处理：将所述交叉特征与所述推荐维度的拟合特征进行拼接处理，得到对应所述推荐维度的拼接特征；针对每个所述推荐维度执行以下处理：对所述推荐维度的拼接特征进行第三映射处理，得到所述待推荐信息对应所述推荐维度的第一推荐分数。
在一些实施例中，第一预测模块5552，还配置为对多个参考特征进行线性处理，得到一阶特征；对多个编码特征中任意两个编码特征进行二阶特征交叉处理，得到任意两个编码特征的二阶交叉特征；将一阶特征及二阶交叉特征进行拼接处理，得到至少一个交叉特征。
在一些实施例中，第一预测模块5552，还配置为针对每个所述推荐维度执行以下处理：通过多门控混合专家网络中每个专家网络，对所述多个编码特征进行第一全连接处理，得到第一隐层特征，并对所述第一隐层特征进行第四映射处理，得到对应每个所述专家网络的映射特征；通过所述多门控混合专家网络中对应所述推荐维度的门控网络，对所述多个编码特征进行第二全连接处理，得到第二隐层特征，并对所述第二隐层特征进行第五映射处理，得 到对应每个所述专家网络的权重特征；基于每个所述专家网络的权重特征，对每个所述专家网络的映射特征进行加权求和处理，得到对应所述推荐维度的拟合特征。
在一些实施例中，特征映射模块5553，还配置为对多个推荐维度的第一推荐分数进行横向拼接处理，得到平铺向量；对多个编码特征进行第三全连接处理，得到第三隐层特征；对第三隐层特征进行第六映射处理，得到与平铺向量维度相同的映射特征。
在一些实施例中，第二预测模块5554，还配置为获取由每个所述推荐维度的第一推荐分数构成的分数矩阵，并获取由每个所述推荐维度对应的映射特征构成的映射矩阵；将所述分数矩阵与所述映射矩阵进行元素积计算，得到所述融合特征。
在一些实施例中，第二预测模块5554，还配置为对所述融合特征进行第七映射处理，得到对应所述融合特征的映射特征；基于对应所述融合特征的映射特征，对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数。
在一些实施例中，信息推荐方法是通过调用分数预测模型实现的，分数预测模型包括：特征编码层、第一推荐分数预测层、特征映射层和第二推荐分数预测层；装置还包括：模型训练模块，配置为通过所述特征编码层，分别对训练样本的多个样本参考特征进行编码处理，得到每个所述样本参考特征的样本编码特征，所述训练样本携带对象样本针对信息样本在多个推荐维度的第一标签，以及所述对象样本针对所述信息样本的第二标签；通过所述第一推荐分数预测层，对所述多个样本编码特征进行第一映射处理，得到与多个所述推荐维度一一对应的多个第一预测结果，其中，所述第一预测结果表征所述对象样本针对所述信息样本在对应推荐维度的推荐分数；通过所述特征映射层，针对每个所述推荐维度执行以下处理：对各所述多个编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的样本映射特征；通过所述第二推荐分数预测层执行以下处理，基于每个所述推荐维度的样本映射特征，对所述多个推荐维度的第一预测结果进行融合处理，得到样本融合特征，并基于所述样本融合特征，对所述信息样本进行推荐分数预测处理，得到所述对象样本针对所述信息样本的第二预测结果；基于每个所述推荐维度的第一预测结果、对应每个所述的推荐维度的第一标签，所述第二预测结果与所述第二标签，更新所述分数预测模型的模型参数。
在一些实施例中，模型训练模块，还配置为针对每个推荐维度，基于第一预测结果及所述推荐维度的第一标签，构造对应第一推荐分数预测层的第一损失函数；基于第二预测结果及第二标签，构造对应第二推荐分数预测层的第二损失函数；将第二损失函数及第一损失函数进行加权求和，得到分数预测模型的第三损失函数；基于第三损失函数更新分数预测模型的模型参数。
在一些实施例中，模型训练模块，还配置为针对每个所述推荐维度执行以下处理：基于所述推荐维度对应的第一预测结果及所述推荐维度的第一标签，构造对应所述推荐维度的子损失函数；确定每个所述推荐维度对应的推荐权重，基于每个所述推荐维度对应的推荐权重对多个所述推荐维度的子损失函数进行加权求和，得到对应所述第一推荐分数预测层的第一损失函数。
本申请实施例提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机可执行指令，该计算机可执行指令存储在计算机可读存储介质中。电子设备的处理器从计算机可读存储介质读取该计算机可执行指令，处理器执行该计算机可执行指令，使得该电子设备执行本申请实施例上述的信息推荐方法。
本申请实施例提供一种存储有计算机可执行指令的计算机可读存储介质，其中存储有计算机可执行指令，当计算机可执行指令被处理器执行时，将被处理器执行本申请实施例提供的信息推荐方法，例如，如图3A示出的信息推荐方法。
在一些实施例中，计算机可读存储介质可以是FRAM、ROM、PROM、EPROM、EEPROM、闪存、磁表面存储器、光盘、或CD-ROM等存储器；也可以是包括上述存储器之一或任意组合的各种设备。
在一些实施例中，可执行指令可以采用程序、软件、软件模块、脚本或代码的形式， 按任意形式的编程语言(包括编译或解释语言，或者声明性或过程性语言)来编写，并且其可按任意形式部署，包括被部署为独立的程序或者被部署为模块、组件、子例程或者适合在计算环境中使用的其它单元。
作为示例，可执行指令可以但不一定对应于文件系统中的文件，可以可被存储在保存其它程序或数据的文件的一部分，例如，存储在超文本标记语言(HTML，Hyper Text Markup Language)文档中的一个或多个脚本中，存储在专用于所讨论的程序的单个文件中，或者，存储在多个协同文件(例如，存储一个或多个模块、子程序或代码部分的文件)中。
作为示例，可执行指令可被部署为在一个电子设备上执行，或者在位于一个地点的多个电子设备上执行，又或者，在分布在多个地点且通过通信网络互连的多个电子设备上执行。
以上所述，仅为本申请的实施例而已，并非用于限定本申请的保护范围。凡在本申请的精神和范围之内所作的任何修改、等同替换和改进等，均包含在本申请的保护范围之内。

Claims (15)

  1. 一种信息推荐方法，所述方法由电子设备执行，所述方法包括：
    对多个参考特征进行编码处理，得到每个所述参考特征的编码特征，其中，所述参考特征包括以下至少之一：目标对象的对象特征、待推荐信息的信息特征；
    对多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，其中，所述第一推荐分数表征所述目标对象针对所述待推荐信息在对应推荐维度的推荐分数；
    针对每个所述推荐维度执行以下处理：对所述多个编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的映射特征；
    基于每个所述推荐维度的映射特征，对所述多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于所述融合特征对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数；
    基于所述待推荐信息的第二推荐分数，执行所述待推荐信息对应所述目标对象的推荐操作。
  2. 如权利要求1所述的方法，其中，所述参考特征包括连续型特征及离散型特征中至少之一，所述对多个参考特征进行编码处理，得到每个所述参考特征的编码特征，包括：
    针对每个所述参考特征执行以下处理：
    当所述参考特征为所述连续型特征时，对所述连续型特征进行离散化处理，得到所述连续型特征的离散型特征，并对所述连续型特征的离散型特征进行编码处理，得到所述连续型特征的编码特征；
    当所述参考特征为所述离散型特征时，对所述离散型特征进行编码处理，得到所述离散型特征的编码特征。
  3. 如权利要求1所述的方法，其中，所述对所述多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，包括：
    对所述多个编码特征进行特征交叉处理，得到至少一个交叉特征；
    基于所述多个编码特征，对所述待推荐信息在每个所述推荐维度的拟合性进行预测，得到对应每个所述推荐维度的拟合特征；
    针对每个所述推荐维度执行以下处理：将所述交叉特征与所述推荐维度的拟合特征进行拼接处理，得到对应所述推荐维度的拼接特征；
    针对每个所述推荐维度执行以下处理：对所述推荐维度的拼接特征进行第三映射处理，得到所述待推荐信息对应所述推荐维度的第一推荐分数。
  4. 如权利要求3所述的方法，其中，对所述多个编码特征进行特征交叉处理，得到至少一个交叉特征之前，所述方法还包括：
    对所述多个参考特征进行线性处理，得到一阶特征；
    所述对所述多个编码特征进行特征交叉处理，得到至少一个交叉特征，包括：
    对所述多个编码特征中任意两个编码特征进行二阶特征交叉处理，得到所述任意两个编码特征的二阶交叉特征；
    将所述一阶特征与每个所述二阶交叉特征进行拼接处理，得到至少一个所述交叉特征。
  5. 如权利要求3所述的方法，其中，所述基于所述多个编码特征，对所述待推荐信息在每个所述推荐维度的拟合性进行预测，得到对应每个所述推荐维度的拟合特征，包括：
    针对每个所述推荐维度执行以下处理：
    通过多门控混合专家网络中每个专家网络，对所述多个编码特征进行第一全连接处理，得到第一隐层特征，并对所述第一隐层特征进行第四映射处理，得到对应每个所述专家网络的映射特征；
    通过所述多门控混合专家网络中对应所述推荐维度的门控网络，对所述多个编码特征进行第二全连接处理，得到第二隐层特征，并对所述第二隐层特征进行第五映射处理，得到对 应每个所述专家网络的权重特征；
    基于每个所述专家网络的权重特征，对每个所述专家网络的映射特征进行加权求和处理，得到对应所述推荐维度的拟合特征。
  6. 如权利要求1所述的方法，其中，所述对所述多个编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的映射特征，包括：
    对多个所述推荐维度的第一推荐分数进行横向拼接处理，得到平铺向量；
    对所述多个编码特征进行第三全连接处理，得到第三隐层特征；
    对所述第三隐层特征进行第六映射处理，得到与所述平铺向量维度相同的映射特征。
  7. 如权利要求1所述的方法，其中，所述基于每个所述推荐维度的映射特征，对所述多个推荐维度的第一推荐分数进行融合处理，得到融合特征，包括：
    获取由每个所述推荐维度的第一推荐分数构成的分数矩阵，并获取由每个所述推荐维度对应的映射特征构成的映射矩阵；
    将所述分数矩阵与所述映射矩阵进行元素积计算，得到所述融合特征。
  8. 如权利要求1所述的方法，其中，所述基于所述融合特征对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数，包括：
    对所述融合特征进行第七映射处理，得到对应所述融合特征的映射特征；
    基于对应所述融合特征的映射特征，对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数。
  9. 如权利要求1至8任一项所述的方法，其中，所述信息推荐方法是通过调用分数预测模型实现的，所述分数预测模型包括：特征编码层、第一推荐分数预测层、特征映射层和第二推荐分数预测层；所述方法还包括：
    通过所述特征编码层，分别对训练样本的多个样本参考特征进行编码处理，得到每个所述样本参考特征的样本编码特征，所述训练样本携带对象样本针对信息样本在多个推荐维度的第一标签，以及所述对象样本针对所述信息样本的第二标签；
    通过所述第一推荐分数预测层，对所述多个样本编码特征进行第一映射处理，得到与多个所述推荐维度一一对应的多个第一预测结果，其中，所述第一预测结果表征所述对象样本针对所述信息样本在对应推荐维度的推荐分数；
    通过所述特征映射层，针对每个所述推荐维度执行以下处理：对各所述多个样本编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的样本映射特征；
    通过所述第二推荐分数预测层执行以下处理，基于每个所述推荐维度的样本映射特征，对所述多个推荐维度的第一预测结果进行融合处理，得到样本融合特征，并基于所述样本融合特征，对所述信息样本进行推荐分数预测处理，得到所述对象样本针对所述信息样本的第二预测结果；
    基于每个所述推荐维度的第一预测结果、对应每个所述的推荐维度的第一标签，所述第二预测结果与所述第二标签，更新所述分数预测模型的模型参数。
  10. 如权利要求9所述的方法，其中，所述基于每个所述推荐维度的第一预测结果、对应每个所述的推荐维度的第一标签，所述第二预测结果与所述第二标签，更新所述分数预测模型的模型参数，包括：
    针对每个所述推荐维度，基于所述第一预测结果及所述推荐维度的第一标签，构造对应所述第一推荐分数预测层的第一损失函数；
    基于所述第二预测结果及所述第二标签，构造对应所述第二推荐分数预测层的第二损失函数；
    将所述第二损失函数及所述第一损失函数进行加权求和，得到所述分数预测模型的第三损失函数；
    基于所述第三损失函数更新所述分数预测模型的模型参数。
  11. 如权利要求10所述的方法，其中，所述针对每个所述推荐维度，基于所述第一预 测结果及所述推荐维度的第一标签，构造对应所述第一推荐分数预测层的第一损失函数，包括：
    针对每个所述推荐维度执行以下处理：基于所述推荐维度对应的第一预测结果及所述推荐维度的第一标签，构造对应所述推荐维度的子损失函数；
    确定每个所述推荐维度对应的推荐权重，基于每个所述推荐维度对应的推荐权重对多个所述推荐维度的子损失函数进行加权求和，得到对应所述第一推荐分数预测层的第一损失函数。
  12. 一种信息推荐装置，所述装置包括：
    特征编码模块，配置为对多个参考特征进行编码处理，得到每个所述参考特征的编码特征，其中，所述参考特征包括以下至少之一：目标对象的对象特征、待推荐信息的信息特征；
    第一预测模块，配置为对多个编码特征进行第一映射处理，得到与多个推荐维度一一对应的多个第一推荐分数，其中，所述第一推荐分数表征所述目标对象针对待推荐信息在对应推荐维度的推荐分数；
    特征映射模块，配置为针对每个所述推荐维度执行以下处理：对所述多个编码特征在所述推荐维度进行第二映射处理，得到所述推荐维度的映射特征；
    第二预测模块，配置为基于每个所述推荐维度的映射特征，对所述多个推荐维度的第一推荐分数进行融合处理，得到融合特征，并基于所述融合特征对所述待推荐信息进行推荐分数预测处理，得到所述目标对象针对所述待推荐信息的第二推荐分数；
    信息推荐模块，配置为基于所述第二推荐分数，执行所述待推荐信息对应所述目标对象的推荐操作。
  13. 一种电子设备，所述电子设备包括：
    存储器，用于存储计算机可执行指令；
    处理器，用于执行所述存储器中存储的计算机可执行指令时，实现权利要求1至11任一项所述的信息推荐方法。
  14. 一种计算机可读存储介质，存储有计算机可执行指令，用于被处理器执行时，实现权利要求1至11任一项所述的信息推荐方法。
  15. 一种计算机程序产品，包括计算机程序或计算机可执行指令，所述计算机程序或计算机可执行指令被处理器执行时实现权利要求1至11任一项所述的信息推荐方法。
PCT/CN2022/116402 2021-10-12 2022-09-01 信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品 WO2023061087A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/196,373 US20230281448A1 (en) 2021-10-12 2023-05-11 Method and apparatus for information recommendation, electronic device, computer readable storage medium and computer program product

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111184748.8 2021-10-12
CN202111184748.8A CN113626719B (zh) 2021-10-12 2021-10-12 信息推荐方法、装置、设备、存储介质及计算机程序产品

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/196,373 Continuation US20230281448A1 (en) 2021-10-12 2023-05-11 Method and apparatus for information recommendation, electronic device, computer readable storage medium and computer program product

Publications (1)

Publication Number Publication Date
WO2023061087A1 true WO2023061087A1 (zh) 2023-04-20

Family

ID=78391020

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/116402 WO2023061087A1 (zh) 2021-10-12 2022-09-01 信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品

Country Status (3)

Country Link
US (1) US20230281448A1 (zh)
CN (1) CN113626719B (zh)
WO (1) WO2023061087A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116226789A (zh) * 2023-05-08 2023-06-06 锋睿领创(珠海)科技有限公司 基于人工智能的数据同分布判别方法、装置、设备及介质
CN116662814A (zh) * 2023-07-28 2023-08-29 腾讯科技（深圳）有限公司 对象意向预测方法、装置、计算机设备及存储介质
CN117495142A (zh) * 2023-11-18 2024-02-02 北京连华永兴科技发展有限公司 一种企业水处理方案推荐方法及系统
CN117765378A (zh) * 2024-02-22 2024-03-26 成都信息工程大学 多尺度特征融合的复杂环境下违禁物品检测方法和装置
CN118118565A (zh) * 2024-04-30 2024-05-31 成都鸿钰网络科技有限公司 一种森林防火数据通信协同共享方法及系统

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US12014398B2 (en) * 2021-07-07 2024-06-18 Baidu Usa Llc Systems and methods for gating-enhanced multi-task neural networks with feature interaction learning
CN113626719B (zh) * 2021-10-12 2022-02-08 腾讯科技（深圳）有限公司 信息推荐方法、装置、设备、存储介质及计算机程序产品
CN114265979B (zh) * 2021-12-17 2022-11-18 北京百度网讯科技有限公司 确定融合参数的方法、信息推荐方法和模型训练方法
CN114491093B (zh) * 2021-12-22 2023-03-28 北京达佳互联信息技术有限公司 多媒体资源推荐、对象表征网络的生成方法及装置
CN114528482B (zh) * 2022-01-25 2022-12-27 北京三快在线科技有限公司 一种推荐对象确定方法、装置、电子设备及存储介质
CN114661994B (zh) * 2022-03-28 2022-10-14 中软数智信息技术(武汉)有限公司 基于人工智能的用户兴趣数据处理方法、系统及云平台
CN114707041B (zh) * 2022-04-11 2023-12-01 中国电信股份有限公司 消息推荐方法、装置、计算机可读介质及电子设备
US20230334514A1 (en) * 2022-04-18 2023-10-19 Microsoft Technology Licensing, Llc Estimating and promoting future user engagement of applications
CN114817751B (zh) * 2022-06-24 2022-09-23 腾讯科技（深圳）有限公司 数据处理方法、装置、电子设备、存储介质及程序产品
CN115392365B (zh) * 2022-08-18 2024-04-26 腾讯科技（深圳）有限公司 多模态特征的获取方法、装置及电子设备
CN115544385B (zh) * 2022-11-22 2023-04-04 浙江大华技术股份有限公司 平台推荐方法、电子设备和计算机可读存储介质
CN117252665B (zh) * 2023-11-14 2024-02-20 苏州元脑智能科技有限公司 业务推荐方法、装置、电子设备及存储介质
CN117349458B (zh) * 2023-12-05 2024-04-09 北京搜狐新媒体信息技术有限公司 一种多媒体推荐方法、装置、设备和存储介质
CN117876038A (zh) * 2024-03-12 2024-04-12 云筑信息科技(成都)有限公司 一种ctr预估模型推荐方法
CN118013060B (zh) * 2024-03-19 2024-06-14 腾讯科技（深圳）有限公司 数据处理方法、装置、设备、存储介质及产品
CN118070139A (zh) * 2024-04-24 2024-05-24 中国科学技术大学 一种跨领域推荐方法、装置、电子设备及存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111291266A (zh) * 2020-02-13 2020-06-16 腾讯科技（北京）有限公司 基于人工智能的推荐方法、装置、电子设备及存储介质
CN112163165A (zh) * 2020-10-21 2021-01-01 腾讯科技（深圳）有限公司 信息推荐方法、装置、设备及计算机可读存储介质
CN112232546A (zh) * 2020-09-09 2021-01-15 北京三快在线科技有限公司 一种推荐概率预估方法、装置、电子设备及存储介质
US20210110306A1 (en) * 2019-10-14 2021-04-15 Visa International Service Association Meta-transfer learning via contextual invariants for cross-domain recommendation
CN112699305A (zh) * 2021-01-08 2021-04-23 网易传媒科技（北京）有限公司 多目标推荐方法、装置、计算设备以及介质
CN113342868A (zh) * 2021-08-05 2021-09-03 腾讯科技（深圳）有限公司 一种信息推荐方法、装置、设备及计算机可读存储介质
CN113626719A (zh) * 2021-10-12 2021-11-09 腾讯科技（深圳）有限公司 信息推荐方法、装置、设备、存储介质及计算机程序产品

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107909433A (zh) * 2017-11-14 2018-04-13 重庆邮电大学 一种基于大数据移动电子商务的商品推荐方法
CN108763362B (zh) * 2018-05-17 2020-10-09 浙江工业大学 基于随机锚点对选择的局部模型加权融合Top-N电影推荐方法
US10475105B1 (en) * 2018-07-13 2019-11-12 Capital One Services, Llc Systems and methods for providing improved recommendations
US11392593B2 (en) * 2019-01-18 2022-07-19 Macy's IP Holdings, LLC Systems and methods for calculating recommendation scores based on combined signals from multiple recommendation systems

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210110306A1 (en) * 2019-10-14 2021-04-15 Visa International Service Association Meta-transfer learning via contextual invariants for cross-domain recommendation
CN111291266A (zh) * 2020-02-13 2020-06-16 腾讯科技（北京）有限公司 基于人工智能的推荐方法、装置、电子设备及存储介质
CN112232546A (zh) * 2020-09-09 2021-01-15 北京三快在线科技有限公司 一种推荐概率预估方法、装置、电子设备及存储介质
CN112163165A (zh) * 2020-10-21 2021-01-01 腾讯科技（深圳）有限公司 信息推荐方法、装置、设备及计算机可读存储介质
CN112699305A (zh) * 2021-01-08 2021-04-23 网易传媒科技（北京）有限公司 多目标推荐方法、装置、计算设备以及介质
CN113342868A (zh) * 2021-08-05 2021-09-03 腾讯科技（深圳）有限公司 一种信息推荐方法、装置、设备及计算机可读存储介质
CN113626719A (zh) * 2021-10-12 2021-11-09 腾讯科技（深圳）有限公司 信息推荐方法、装置、设备、存储介质及计算机程序产品

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116226789A (zh) * 2023-05-08 2023-06-06 锋睿领创(珠海)科技有限公司 基于人工智能的数据同分布判别方法、装置、设备及介质
CN116226789B (zh) * 2023-05-08 2023-08-18 锋睿领创(珠海)科技有限公司 基于人工智能的数据同分布判别方法、装置、设备及介质
CN116662814A (zh) * 2023-07-28 2023-08-29 腾讯科技（深圳）有限公司 对象意向预测方法、装置、计算机设备及存储介质
CN116662814B (zh) * 2023-07-28 2023-10-31 腾讯科技（深圳）有限公司 对象意向预测方法、装置、计算机设备及存储介质
CN117495142A (zh) * 2023-11-18 2024-02-02 北京连华永兴科技发展有限公司 一种企业水处理方案推荐方法及系统
CN117765378A (zh) * 2024-02-22 2024-03-26 成都信息工程大学 多尺度特征融合的复杂环境下违禁物品检测方法和装置
CN117765378B (zh) * 2024-02-22 2024-04-26 成都信息工程大学 多尺度特征融合的复杂环境下违禁物品检测方法和装置
CN118118565A (zh) * 2024-04-30 2024-05-31 成都鸿钰网络科技有限公司 一种森林防火数据通信协同共享方法及系统

Also Published As

Publication number Publication date
US20230281448A1 (en) 2023-09-07
CN113626719A (zh) 2021-11-09
CN113626719B (zh) 2022-02-08

Similar Documents

Publication Publication Date Title
WO2023061087A1 (zh) 信息推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品
WO2021203819A1 (zh) 一种内容推荐方法、装置、电子设备和存储介质
CN110781321B (zh) 一种多媒体内容推荐方法及装置
US20220215032A1 (en) Ai-based recommendation method and apparatus, electronic device, and storage medium
CN111008332B (zh) 内容项推荐方法、装置、服务器以及存储介质
CN111382361B (zh) 信息推送方法、装置、存储介质和计算机设备
CN111444428A (zh) 基于人工智能的信息推荐方法、装置、电子设备及存储介质
WO2022016522A1 (zh) 推荐模型的训练方法、推荐方法、装置及计算机可读介质
CN110941764A (zh) 对象推荐方法、装置、计算机设备和存储介质
CN114265979A (zh) 确定融合参数的方法、信息推荐方法和模型训练方法
CN111798280A (zh) 多媒体信息推荐方法、装置和设备及存储介质
CN111400613A (zh) 物品推荐方法、装置、介质及计算机设备
CN113569129A (zh) 点击率预测模型处理方法、内容推荐方法、装置及设备
CN110717597A (zh) 利用机器学习模型获取时序特征的方法和装置
CN112150184A (zh) 点击率预估方法及其系统、计算机系统及计算机可读介质
US20230316106A1 (en) Method and apparatus for training content recommendation model, device, and storage medium
CN114358807A (zh) 基于可预测用户特征属性的用户画像方法及系统
CN113342868B (zh) 一种信息推荐方法、装置、设备及计算机可读存储介质
CN115222112A (zh) 一种行为预测方法、行为预测模型的生成方法及电子设备
CN112785328A (zh) 一种内容推送方法、装置及计算机存储介质
CN113538030B (zh) 一种内容推送方法、装置及计算机存储介质
CN117648576B (zh) 数据增强模型训练及数据处理方法、装置、设备、介质
CN114996561B (zh) 一种基于人工智能的信息推荐方法及装置
CN113496304B (zh) 网络媒介信息的投放控制方法、装置、设备及存储介质
CN116976943A (zh) 媒体价值预测方法、装置、设备以及可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22880032

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE