WO2023155306A1 - 基于图神经网络的数据推荐方法及装置、电子设备 - Google Patents

基于图神经网络的数据推荐方法及装置、电子设备 Download PDF

Info

Publication number
WO2023155306A1
WO2023155306A1 PCT/CN2022/090754 CN2022090754W WO2023155306A1 WO 2023155306 A1 WO2023155306 A1 WO 2023155306A1 CN 2022090754 W CN2022090754 W CN 2022090754W WO 2023155306 A1 WO2023155306 A1 WO 2023155306A1
Authority
WO
WIPO (PCT)
Prior art keywords
user
data
information
social relationship
graph
Prior art date
Application number
PCT/CN2022/090754
Other languages
English (en)
French (fr)
Inventor
吴粤敏
舒畅
陈又新
Original Assignee
平安科技(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 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2023155306A1 publication Critical patent/WO2023155306A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/953Querying, e.g. by the use of web search engines
    • G06F16/9536Search customisation based on social or collaborative filtering
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/01Social networking

Definitions

  • the present application relates to the technical field of artificial intelligence, and specifically relates to a graph neural network-based data recommendation method, device, and electronic equipment.
  • the recommendation system when a user enters keywords into a search engine for retrieval, the recommendation system usually recalls and ranks candidate data sets according to the keywords entered by the user, so as to recommend data that may be of interest to the user.
  • the existing data recommendation algorithms such as those based on machine learning, need to sort the recommended data according to the user's historical behavior data. The lower problem affects the experience of new users on the search engine.
  • the embodiment of the present application provides a data recommendation method based on a graph neural network, the method includes: obtaining target search information of a new user, and generating a candidate data set according to the target search information; the associated user associated with the new user; obtain the first search data of the associated user; generate a social relationship graph of the new user according to the new user, the associated user and the first search data;
  • the social relationship graph is input into the preset graph neural network model for feature prediction to obtain the target feature information of the new user; according to the target feature information, the candidate data sets are sorted to obtain the recommended data set.
  • the embodiment of the present application also proposes a graph neural network-based data recommendation device, the device includes: an acquisition module, used to acquire target search information of new users; a first generation module, used to The target search information is used to generate a candidate data set; the determination module is used to determine the associated user associated with the new user; the acquisition module is also used to obtain the first search data of the associated user; the second generation module, It is used to generate a social relationship graph of the new user according to the new user, the associated user and the first search data; a prediction module is used to input the social relationship graph into a preset graph neural network model Perform feature prediction to obtain target feature information of the new user; a sorting module is used to sort the candidate data sets according to the target feature information to obtain recommended data sets.
  • the embodiment of the present application also proposes an electronic device, the electronic device includes a memory and a processor, the memory stores a program, and when the program is executed by the processor, a graph neural network based A network data recommendation method, the method comprising: acquiring target search information of a new user, and generating a candidate data set according to the target search information; determining associated users associated with the new user; acquiring the associated user's First search data; generate a social relationship graph of the new user according to the new user, the associated user, and the first search data; input the social relationship graph into a preset graph neural network model for characterization Prediction, to obtain the target feature information of the new user; according to the target feature information, sort the candidate data sets to obtain the recommended data set.
  • the present application provides a storage medium for computer-readable storage.
  • the storage medium stores one or more programs, and one or more programs can be executed by one or more processors to implement a system based on A data recommendation method of a graph neural network, the method comprising: obtaining target search information of a new user, and generating a candidate data set according to the target search information; determining associated users associated with the new user; obtaining the association The user's first search data; according to the new user, the associated user and the first search data, generate a social relationship graph of the new user; input the social relationship graph into a preset graph neural network model Perform feature prediction to obtain target feature information of the new user; sort the candidate data sets according to the target feature information to obtain recommended data sets.
  • the graph neural network-based data recommendation method, device, and electronic device proposed in this application can determine the associated users associated with the new user, and obtain the first search data of the associated users, so that according to the new user, the associated user, and the first search data Data to generate a social relationship graph of a new user, so the social relationship graph combines the user's social relationship and the first search data of different users at the same time. Based on this, the social relationship graph is input into the neural network model of the preset graph, so that the neural network model of the preset graph can identify the common data needs and evaluation mode characteristics among associated users along the deep propagation of the social relationship chain, so as to provide new users with Determine more interpretable target feature information.
  • the candidate data sets for the new user are sorted to obtain the recommended data set for the new user, which can alleviate the cold start problem on the user side and improve the accuracy of data recommendation for new users. This in turn improves the experience of new users with the data search function.
  • FIG. 1 is a structural block diagram of an electronic device applied in an embodiment of the present application
  • FIG. 2 is a flow chart of a data recommendation method based on a graph neural network provided in Embodiment 1 of the present application;
  • FIG. 3 is a schematic diagram of a social relationship graph in an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of a preset graph neural network model in an embodiment of the present application.
  • FIG. 5 is a flow chart of a data recommendation method based on a graph neural network provided in Embodiment 2 of the present application;
  • Fig. 6 is a structural block diagram of a graph neural network-based data recommendation device applied in the embodiment of the present application.
  • Artificial intelligence is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
  • artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics
  • artificial intelligence software technologies mainly include computer vision technology, Robotics, biometrics, speech processing, natural language processing, and machine learning/deep learning.
  • FIG. 1 is a structural block diagram of an electronic device applied in an embodiment of the present application.
  • the electronic device may be a server, a smart phone, a tablet computer, a portable computer, a desktop computer, and other terminal devices with computing functions.
  • the electronic device includes: a memory 11 , a processor 12 , a network interface 13 and a data bus 14 .
  • the memory 11 includes at least one type of readable storage medium, and the at least one type of readable storage medium may be a non-volatile storage medium such as a flash memory, a hard disk, a multimedia card, or a card-type memory.
  • the readable storage medium may be an internal storage unit of the electronic device, such as a hard disk of the electronic device.
  • the readable storage medium can also be an external memory of the electronic device, such as a plug-in hard disk equipped on the electronic device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash card (Flash Card), etc.
  • the readable storage medium of the memory 11 is generally used to store a data recommendation program based on a graph neural network installed in an electronic device, various sample sets, pre-trained models, and the like.
  • the memory 11 can also be used to temporarily store data that has been output or will be output.
  • Processor 12 can be a central processing unit (Central Processing Unit, CPU) in some embodiments, microprocessor or other data processing chip, is used for running the program code stored in memory 11 or processing data, for example executes data recommendation program wait.
  • CPU Central Processing Unit
  • microprocessor or other data processing chip is used for running the program code stored in memory 11 or processing data, for example executes data recommendation program wait.
  • the network interface 13 may optionally include a standard wired interface or wireless interface (such as a WI-FI interface), which is generally used to establish a communication connection between the electronic device and other electronic devices.
  • a standard wired interface or wireless interface such as a WI-FI interface
  • the data bus 14 is used to realize connection communication between these components.
  • the electronic device may also include a user interface
  • the user interface may include an input unit such as a keyboard (Keyboard), a voice input device such as a microphone (microphone) and other devices with voice recognition functions, a voice output device such as a stereo, earphones, etc.
  • the user interface may also include a standard wired interface and a wireless interface.
  • the electronic device may further include a display, and the display may also be called a display screen or a display unit.
  • the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an organic light-emitting diode (Organic Light-Emitting Diode, OLED) touch device, and the like.
  • the display is used to display information processed in the electronic device and to display a visualized user interface.
  • the electronic device further includes a touch sensor.
  • the area provided by the touch sensor for the user to perform a touch operation is called a touch area.
  • the touch sensor here may be a resistive touch sensor, a capacitive touch sensor, and the like.
  • the touch sensor includes not only a contact-type touch sensor, but also a proximity-type touch sensor and the like.
  • the touch sensor may be a single sensor, or a plurality of sensors arranged in an array, for example.
  • the area of the display of the electronic device may be the same as or different from the area of the touch sensor.
  • the display and the touch sensor are stacked to form a touch display screen. The device detects a touch operation triggered by a user based on a touch screen.
  • a data recommendation method based on a graph neural network disclosed in an embodiment of the present application will be described in detail below.
  • FIG. 2 is a flowchart of a data recommendation method based on a graph neural network provided in Embodiment 1 of the present application.
  • the processor 12 Based on the electronic device embodiment shown in FIG. 1, the processor 12 implements the following steps when executing the program stored in the memory 11:
  • Step S200 Obtain target search information of a new user.
  • a new user may be a user who enters search information for the first time, and the target search information includes but is not limited to: search sentences or keywords input by the new user's voice; text information or pictures input by the new user through a search engine information etc.
  • the user information corresponding to the search request may be obtained first, and the user information includes but is not limited to user ID, user IP address, user device code, and other information used to identify the user. After that, it is confirmed whether the user information has been entered in the database. If the user information has not been entered, a new user is created according to the user information and entered into the database.
  • the target user corresponding to the user information is determined from the database, and the target feature information of the target user is directly obtained, and step S260 is executed.
  • Step S210 Generate a candidate data set according to the target search information.
  • web crawlers can be used to collect big data information such as Internet web page links and specific access resources (such as a certain document, a certain product on a shopping page, etc.), and then extract and organize the big data information Build an index library.
  • the searcher can split the target search information input by the user into words or words with retrieval significance, and then quickly detect candidate data in the index library according to the split words or words to generate candidate data sets .
  • a new user enters the search keyword "XX Technology” in the search engine, a set of web page links containing "XX Technology” is called to add the candidate data set.
  • Step S220 Determine associated users associated with the new user.
  • associated users may include, but are not limited to: friend users associated with new users when using social software, work users (including customers, colleagues at the same level, colleagues at the same level, etc.), temporary chat users (such as commenting or forwarding users), follow users, like users, dislike users, bookmark users, etc.; and, the device address book contacts of new users.
  • the associated user associated with the new user can be determined from various information channels such as social application software, office application software, or multimedia application software accessed by the new user by obtaining the account information access authority or device information access authority of the new user.
  • Step S230 Obtain the first search data of the associated user.
  • the first search data may include historical search information of the associated user, such as texts, pictures, etc. that the associated user has searched using a search engine. Further, the first search data may also specifically include evaluation data of associated users on historical search information, such as evaluation content (including likes, dislikes, evaluation types, etc.) Recommended data for historical search information, such as data sent when associated users use the software sharing function.
  • Step S240 Generate a social relationship graph of the new user according to the new user, associated users and the first search data.
  • the social relationship information between the new user and the associated user can be determined, such as colleagues, friends, follow, likes, or dislikes.
  • the associated relationship information between the associated user and the first search data can be determined, such as likes, dislikes, recommendations, positive reviews, or negative reviews.
  • the new user, the associated user, and the first search data are respectively used as different nodes, and then according to the relationship information between different nodes, a connection relationship is established between different nodes to obtain a social relationship graph of the new user.
  • FIG. 3 is a schematic diagram of a social relationship graph in an embodiment of the present application.
  • the social relationship graph includes four user nodes and two data nodes, and the four user nodes include a new user node A, a first associated user node B, a second associated user node C and a third associated user Node D, the two data nodes include item node 1 and item node 2.
  • the social relationship information related to the new user includes: the new user node A is connected to the first associated user node B, and its connection relationship is "dislike"; the new user node A is connected to the second associated user node C, and its connection relationship is a like relationship; the new user node A is connected to the third associated user node D, and the connection relationship is a follow relationship.
  • the relationship information related to the second associated user node C includes: the second associated user node C is connected to the item node 1, and its connection relationship is "dislike”; the second associated user node C is connected to the item node 2, and its connection relationship is "like". It can be seen that the social relationship graph adopts a graph data structure, which is convenient for combing the relationship between different users and between users and resources.
  • Step S250 input the social relationship graph into the preset graph neural network model for feature prediction, and obtain the target feature information of the new user.
  • the training steps of the preset graph neural network model may include:
  • the training samples include the first social relationship sample graph of the first user and the second social relationship sample graph of the second user, the first user and the second user are any two users,
  • the annotation data is used to indicate whether the first user is similar to the second user.
  • the second social relationship sample graph and the first social relationship sample graph reference may be made to the above description of the social relationship graph, and details are not repeated here.
  • the first social relationship sample graph of the first user and the second social relationship sample graph of the second user are taken as training samples, and then node search is performed on the first social relationship sample graph and the second social relationship sample graph .
  • the label data of the training sample is determined for the first user is similar to the second user. If the second node is not found in the first social relationship sample graph and the first node is not found in the second social relationship sample graph, the labeled data of the training sample is determined to be that the first user is not similar to the second user.
  • the first social relationship sample graph and the second social relationship sample graph are input into the preset graph neural network model. Generate the first feature information of the first social relationship sample graph and the second feature information of the second social relationship sample graph through the preset graph neural network model, and fuse the first feature information and the second feature information to obtain the fusion result.
  • the fusion result may represent a similarity value S between the first feature information and the second feature information.
  • the preset accuracy rate can be set and adjusted manually, such as 90% or 95%, without any specific limitation.
  • the number of training samples used to train the neural network model of the preset graph each time may be multiple.
  • the accurate value R of each training sample can be calculated, and then the accurate value R of multiple training samples can be averaged to obtain the verification result.
  • the calculation method of the accurate value R is:
  • s1 is a standard value (such as 0.9) set when the labeled data is similar
  • s2 is a standard value (such as 0.1) set when the labeled data is dissimilar.
  • the first value range (such as 0.8 to 1) when the labeled data is similar. If the similarity value S corresponding to the training sample belongs to the first value range, the sample count The value N is cumulatively incremented by one. And, when the labeled data is dissimilar, set a second value range (for example, 0 to 0.2), and if the similarity value S corresponding to the training sample belongs to the second value range, the sample count value N is cumulatively increased by one. Based on this, after traversing all the training samples, take the ratio between the finally determined sample count value N and the number of training samples as the verification result.
  • the first value range such as 0.8 to 1
  • the preset graph neural network model may include a first feature generation module, a second feature generation module, and a fusion module, then in the training step of the preset graph neural network model, the The first social relationship sample graph is input to the first feature generation module to obtain first feature information, and the second social relationship sample graph is input to the second feature generation module to obtain second feature information. Then input the first feature information and the second feature information into the fusion module to obtain a fusion result.
  • the fusion module may use a dot product operation unit, and then input the first feature information and the second feature information into the fusion module to perform a dot product operation to obtain the similarity value S between the first feature information and the second feature information.
  • the loss value can be calculated for the fusion result and the labeled data output by the neural network model of the preset graph through the root mean square error loss function L, and through the backpropagation of the loss value Adjust the parameters of the preset graph neural network model until the end of training.
  • the root mean square error loss function L satisfies:
  • m is the number of training samples
  • y i is the fusion result corresponding to the i-th training sample
  • y i is the fusion result corresponding to the i-th training sample
  • input the social relationship graph into the preset graph neural network model for feature prediction, and obtain the target feature information of the new user specifically: input the social relationship graph into the first feature generation module or the second feature generation module for Feature prediction, to obtain target feature information of new users.
  • the trained preset graph neural network model can intelligently analyze user social relations and associated data features, generate reliable feature information for different users, and satisfy similar comparison results between different users at the same time. Therefore, in practical applications, only It is necessary to input the new user's social relationship graph into the feature generation module in the preset graph neural network model to obtain the target feature information of the new user, which has better model application performance and reliability.
  • both the first feature generation module and the second feature generation module include a sequentially connected double graph convolutional layer (double graph convolutional layer, DGCL), pooling layer and The fully connected layer forms a similarity bi-directional graph convolutional network (similarity bi-directional graph convolutional networks, S-BIGCN).
  • DGCL double graph convolutional layer
  • S-BIGCN similarity bi-directional graph convolutional networks
  • the parameter sharing of the double-layer graph convolutional layer is suitable for social relationship graphs with undirected graph structures, and the pooling layer can use the average pooling layer, which not only prevents overfitting, but also makes the feature generation module have a better performance.
  • the training speed and inference speed do not need to consume too much computer GPU resources, so the preset graph neural network model is easier to implement.
  • the preset graph neural network model may also use a graph attention network model, a graph autoencoder network model, or a graph generation network model.
  • Step S260 According to the target feature information, the candidate data sets are sorted to obtain the recommended data set.
  • the candidate data set includes multiple candidate data.
  • the social relationship graph related to each candidate data can be obtained, and the social relationship graph related to the candidate data can be input into the preset
  • the feature information of the candidate data is obtained.
  • the target feature information and the feature information of each candidate data are substituted into the similarity calculation formula to obtain the ranking value corresponding to each candidate data.
  • the sorting process of the candidate data set can be realized.
  • the similarity calculation formula includes but is not limited to the cosine similarity calculation method, the Euclidean distance calculation method, the Manhattan distance calculation formula or the Jaccard similarity coefficient calculation formula, etc.
  • the social relationship graph that combines the user's social relationship and the first search data of different users is input into the preset graph neural network model, so that the preset graph neural network model follows the social relationship chain.
  • Deep dissemination identifies the common data needs and evaluation mode characteristics among associated users, so as to determine more interpretable target feature information for new users.
  • the candidate data sets for the new user are sorted to obtain the recommended data set for the new user, which can alleviate the cold start problem on the user side and improve the accuracy of data recommendation for new users. This in turn improves the experience of new users with the data search function.
  • FIG. 5 is a flowchart of a data recommendation method based on a graph neural network provided in Embodiment 2 of the present application.
  • the processor 12 Based on the electronic device embodiment shown in FIG. 1, the processor 12 implements the following steps when executing the program stored in the memory 11:
  • Step S500 Obtain target search information of a new user.
  • Step S510 Generate a candidate data set according to the target search information.
  • Step S520 Determine associated users associated with the new user.
  • Step S530 Obtain the first search data of the associated user.
  • Step S540 Generate a social relationship graph of the new user according to the new user, associated users and the first search data.
  • Step S550 input the social relationship graph into the preset graph neural network model for feature prediction, and obtain target feature information of the new user.
  • Step S560 Obtain multiple similar users of the new user and the similarity between the similar users and the new user according to the target feature information.
  • step S560 is specifically: acquiring third characteristic information of multiple known users. According to the target feature information and the third feature information of each known user, the similarity between the new user and the known users is calculated, and the known users whose similarity is within a preset threshold range are taken as similar users. Wherein, the value of the similarity belongs to [0, 1], and the larger the value of the similarity, the higher the similarity between the new user and the known user.
  • the preset threshold range can be artificially specified, such as [0.8, 1]; or, the similarity between new users and different known users can also be sorted from large to small, and the similarity of the k-th ranking is taken as the target
  • the thresholds d and k are positive integers and are manually specified, and then the preset threshold range is determined according to the target threshold d (for example, the preset threshold range is [d, 1]), which is not limited.
  • a plurality of registered known users may be obtained from the database.
  • the target feature information and the third feature information of each known user are substituted into a preset similarity calculation formula to obtain the similarity between the new user and the known users.
  • the similarity calculation formula can be a cosine similarity calculation method, namely:
  • sim(u, v) is the similarity between the new user and the known user
  • u is the target characteristic information
  • v is the third characteristic information of the known user.
  • acquiring the third characteristic information of multiple known users may specifically include the following steps:
  • the historical search information of a plurality of known users and the operation data of the known users on the historical search information are acquired, and the historical search information of the known users includes a plurality of second search data.
  • the second search data may refer to historical data recommended to known users after responding to a known user's search request.
  • the operation data includes operation records for each second search data, and the operation types corresponding to the operation records include but are not limited to at least one of the following: number of clicks, collection operations, sharing times, specific ratings, like operations, and dislike operations, It is used to indicate the known user's feedback on different second search data.
  • a scoring matrix is generated.
  • a row vector corresponding to each known user is obtained from the scoring matrix, the row vector includes a plurality of first scoring elements, and the plurality of first scoring elements are in one-to-one correspondence with a plurality of second search data.
  • the third characteristic information of the known user can be determined.
  • the rows of the scoring matrix represent different known users
  • the columns of the scoring matrix represent different second search data
  • the matrix elements of the i-th row and j-th column represent the ratings of the i-th known users on the j-th second search data element.
  • the corresponding scoring element is determined according to the operation record of the operation type. It can be understood that if the operation record is a numerical value, such as the number of clicks, the number of shares, or the specific scoring situation, the operation record can be directly taken as the scoring element. If the operation record is not a numerical value, the operation record can be quantified first to obtain the score element. For example, for the favorite operation, quantize the favorite operation as "1", quantize the non-collection operation as "0"; quantize the like operation as "1”; and quantize the click operation as "-1", etc.
  • the click times of the known user E on the second search data 1, the second search data 2 and the second search data 3 are 1, 1 and 0 respectively, and the corresponding click times of the known user F are 1, 1 and 0 respectively.
  • the number of clicks corresponding to user G is known to be 2, 1 and 0 respectively, and the scoring matrix shown in Table 1 below can be generated.
  • each row vector (or column vector) in the scoring matrix can also be normalized to obtain a normalized row vector (or column vector), so that the normalized row vector (or a column vector) where each scoring element is distributed in [0, 1].
  • the third characteristic information of the known user is determined according to the row vector corresponding to the known user, the third characteristic information can be stored, which is convenient for directly calling the third characteristic information of the known user in practical applications. Further, after the preset graph neural network model is trained (or updated each time), and the social graph of the known user is obtained, the social graph of the known user can be input to the neural network model of the preset graph. In the network model, the new third feature information is obtained, and the new feature information is used to replace the original third feature information for storage, so as to realize the function of feature update.
  • Step S570 Obtain the first rating information of multiple similar users on the candidate data set.
  • the candidate data set may include multiple candidate data.
  • a column vector corresponding to each second search data may also be obtained from the scoring matrix, and the column vector includes a plurality of second scoring elements, A plurality of second rating elements correspond to a plurality of known users one by one.
  • the fourth feature information of the second search data is determined according to the column vector corresponding to the second search data.
  • the column vector [1, 1, 2] can be taken as the fourth feature information of the second search data 1
  • the column vector [1, 0, 1] can be taken as the fourth feature information of the second search data 2 Information, take the column vector [0, 1, 0] as the fourth feature information of the second search data 3 .
  • step S570 may be: determining the fourth characteristic information of each candidate data from the fourth characteristic information of a plurality of second search data by matching the candidate data with the second search data. According to the third characteristic information of the similar users and the fourth characteristic information of the candidate data, the first score value of the similar users to the candidate data is calculated.
  • the third feature information w of similar users and the fourth feature information c of candidate data can be substituted into the similarity calculation formula to obtain the first scoring value score(u,c) of similar users on candidate data, for example, similar
  • the degree calculation formula adopts the cosine similarity calculation method, namely:
  • the fourth characteristic information of the candidate data may also be generated based on the label of the candidate data, or other neural network-based learning methods may be used to generate the fourth characteristic information of the candidate data.
  • Step S580 According to the first scoring information and in combination with the similarity between each similar user and the new user, calculate the second scoring information of the new user on the candidate data set.
  • the second scoring information includes a second scoring value of each candidate data by the new user.
  • the specific manner of calculating the second scoring information may be:
  • Like(u,cj) is the second scoring value of the new user on the jth candidate data cj
  • sim(u,wi) is the similarity between the new user and the i-th similar user
  • score(wi,cj) is The first rating value of the i-th similar user to the j-th candidate data cj.
  • Step S590 According to the second scoring information, sort the candidate data sets to obtain the recommended data sets.
  • the plurality of candidate data in the candidate data set may be processed in descending order according to the descending order of the second scoring value to obtain the recommended data set.
  • the results of the descending processing can be divided into p groups of recommended data (p is a positive integer, and p ⁇ 2), and then each group of recommended data can be randomized Processing to obtain p groups of recommendation data after disordering to form a recommendation data set.
  • FIG. 6 is a structural block diagram of a graph neural network-based data recommendation device applied in the embodiment of the present application.
  • the graph neural network-based data recommendation device 600 includes:
  • An acquisition module 610 configured to acquire target search information of a new user.
  • the first generation module 620 is configured to generate candidate data sets according to the target search information.
  • a determining module 630 configured to determine associated users associated with the new user.
  • the obtaining module 610 is further configured to obtain the first search data of the associated user.
  • the second generating module 640 is configured to generate a social relationship graph of the new user according to the new user, associated users and the first search data.
  • the prediction module 650 is used to input the social relationship graph into the preset graph neural network model to perform feature prediction and obtain the target feature information of the new user.
  • the sorting module 660 is configured to sort the candidate data sets according to the target feature information to obtain the recommended data set.
  • the embodiment of the present application also provides an electronic device, the electronic device includes a memory and a processor, a program is stored in the memory, and when the program is executed by the processor, a data recommendation method based on a graph neural network is implemented, the method includes: obtaining The target search information of the new user, and generate a candidate data set according to the target search information; determine the associated users associated with the new user; obtain the first search data of the associated user; generate The social relationship graph of the new user; the social relationship graph is input into the preset graph neural network model for feature prediction, and the target feature information of the new user is obtained; according to the target feature information, the candidate data sets are sorted to obtain the recommended data set.
  • the embodiment of the present application also provides a storage medium for computer-readable storage.
  • the storage medium stores one or more programs, and one or more programs can be executed by one or more processors to implement a graph-based A neural network data recommendation method, the method comprising: obtaining target search information of a new user, and generating a candidate data set according to the target search information; determining associated users associated with the new user; obtaining the first search data of the associated user; New users, associated users and the first search data, generate a social relationship graph of the new user; input the social relationship graph into the preset graph neural network model for feature prediction, and obtain the target feature information of the new user; according to the target feature information, the The candidate data sets are sorted to obtain the recommended data set.
  • the storage medium may be non-volatile or volatile, which is not specifically limited.
  • the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be composed of several physical components. Components cooperate to execute.
  • Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit .
  • a processor such as a central processing unit, digital signal processor, or microprocessor
  • Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media).
  • computer storage media includes both volatile and nonvolatile media implemented in any method or technology for storage of information, such as computer readable instructions, data structures, program modules, or other data. permanent, removable and non-removable media.
  • Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cartridges, tape, magnetic disk storage or other magnetic storage devices, or can Any other medium used to store desired information and which can be accessed by a computer.
  • communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media .

Abstract

本申请公开了一种基于图神经网络的数据推荐方法及装置、电子设备,具体涉及人工智能技术领域。该方法包括:获取新用户的目标搜索信息,并根据目标搜索信息,生成候选数据集。确定与新用户相关联的关联用户,并获取关联用户的第一搜索数据。根据新用户、关联用户和第一搜索数据,生成新用户的社交关系图。将社交关系图输入到预设图神经网络模型中进行特征预测,得到新用户的目标特征信息。根据目标特征信息,对候选数据集进行排序处理,得到推荐数据集,能够缓解用户侧冷启动问题,提升对新用户进行数据推荐的准确性,进而改善了新用户对数据搜索功能的使用体验。

Description

基于图神经网络的数据推荐方法及装置、电子设备
本申请要求于2022年2月16日提交中国专利局、申请号为202210143649.3,发明名称为“基于图神经网络的数据推荐方法及装置、电子设备”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域,具体涉及一种基于图神经网络的数据推荐方法及装置、电子设备。
背景技术
现有技术中,当用户向搜索引擎输入关键词进行检索时,推荐系统通常会根据用户输入的关键词进行候选数据集的召回和排序,从而向用户推荐可能感兴趣的数据。然而,发明人发现,现有的数据推荐算法,比如基于机器学习的推荐算法,需要根据用户的历史行为数据进行推荐数据排序,因此,容易因新用户的历史行为数据不足而出现数据推荐准确性较低的问题,影响了新用户对搜索引擎的使用体验。
技术问题
以下是发明人意识到的现有技术的技术问题:现有的数据推荐算法,比如基于机器学习的推荐算法,需要根据用户的历史行为数据进行推荐数据排序,容易因新用户的历史行为数据不足而出现数据推荐准确性较低的问题,影响了新用户对搜索引擎的使用体验。
技术解决方案
第一方面,本申请实施例提供了一种基于图神经网络的数据推荐方法,所述方法包括:获取新用户的目标搜索信息,并根据所述目标搜索信息,生成候选数据集;确定与所述新用户相关联的关联用户;获取所述关联用户的第一搜索数据;根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;将所述社交关系图输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
第二方面,本申请实施例还提出了一种基于图神经网络的数据推荐装置,所述装置包括:获取模块,用于获取新用户的目标搜索信息;第一生成模块,用于根据所述目标搜索信息,生成候选数据集;确定模块,用于确定与所述新用户相关联的关联用户;所述获取模块,还用于获取所述关联用户的第一搜索数据;第二生成模块,用于根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;预测模块,用于将所述社交关系图 输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;排序模块,用于根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
第三方面,本申请实施例还提出了一种电子设备,所述电子设备包括存储器和处理器,所述存储器上存储有程序,所述程序被所述处理器执行时实现一种基于图神经网络的数据推荐方法,所述方法包括:获取新用户的目标搜索信息,并根据所述目标搜索信息,生成候选数据集;确定与所述新用户相关联的关联用户;获取所述关联用户的第一搜索数据;根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;将所述社交关系图输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
第四方面,本申请提供了一种存储介质,用于计算机可读存储,存储介质存储有一个或者多个程序,一个或者多个程序可被一个或者多个处理器执行,以实现一种基于图神经网络的数据推荐方法,所述方法包括:获取新用户的目标搜索信息,并根据所述目标搜索信息,生成候选数据集;确定与所述新用户相关联的关联用户;获取所述关联用户的第一搜索数据;根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;将所述社交关系图输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
有益效果
本申请提出的基于图神经网络的数据推荐方法及装置、电子设备,可以确定与新用户相关联的关联用户,并获取关联用户的第一搜索数据,从而根据新用户、关联用户和第一搜索数据,生成新用户的社交关系图,故社交关系图同时融合了用户社交关系以及不同用户的第一搜索数据。基于此,将社交关系图输入到预设图神经网络模型中,使得预设图神经网络模型沿着社交关系链的深层传播识别关联用户之间的共同数据需求和评价模式特征,从而为新用户确定更具备可解释性的目标特征信息。之后,再基于新用户的目标特征信息,对新用户的候选数据集进行排序,得到为新用户推送的推荐数据集,能够缓解用户侧冷启动问题,提升对新用户进行数据推荐的准确性,进而改善了新用户对数据搜索功能的使用体验。
附图说明
附图用来提供对本申请技术方案的进一步理解,并且构成说明书的一部分,与本申请的实施例一起用于解释本申请的技术方案,并不构成对本申请技术方案的限制。
图1是本申请实施例所应用的一种电子设备的结构框图;
图2是本申请实施例一提供的基于图神经网络的数据推荐方法的流程图;
图3是本申请实施例中一种社交关系图的示意图;
图4是本申请实施例中一种预设图神经网络模型的结构示意图;
图5是本申请实施例二提供的基于图神经网络的数据推荐方法的流程图;
图6是本申请实施例所应用的一种基于图神经网络的数据推荐装置的结构框图。
本发明的实施方式
应当理解,此处所描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
在后续的描述中,使用用于表示元件的诸如“模块”、“部件”或“单元”的后缀仅为了有利于本申请的说明,其本身没有特有的意义。因此,“模块”、“部件”或“单元”可以混合地使用。
本申请实施例可以基于人工智能技术对相关的数据进行获取和处理。人工智能(artificial intelligence,AI)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、技术及应用系统。其中,人工智能基础技术一般包括如传感器、专用人工智能芯片、云计算、分布式存储、大数据处理技术、操作/交互系统、机电一体化等技术,而人工智能软件技术主要包括计算机视觉技术、机器人技术、生物识别技术、语音处理技术、自然语言处理技术以及机器学习/深度学习等几大方向。
本申请提供一种基于图神经网络(graph neural networks,GNN)的数据推荐方法,应用于一种电子设备。参照图1所示,图1是本申请实施例所应用的一种电子设备的结构框图。
在本实施例中,电子设备可以是服务器、智能手机、平板电脑、便携计算机、桌上型计算机等具有运算功能的终端设备。
该电子设备包括:存储器11、处理器12、网络接口13及数据总线14。
存储器11包括至少一种类型的可读存储介质,至少一种类型的可读存储介质可为如闪存、硬盘、多媒体卡、卡型存储器等的非易失性存储介质。在一些实施例中,可读存储介质可以是电子设备的内部存储单元,例如该电子设备的硬盘。在另一些实施例中,可读存储介质也可以是电子设备的外部存储器,例如电子设备上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。
在本实施例中,存储器11的可读存储介质通常用于存储安装于电子设备的基于图神经网络的数据推荐程序、多种样本集及预先训练好的模型等。存储器11还可以用于暂时地存储已经输出或者将要输出的数据。
处理器12在一些实施例中可以是一中央处理器(Central Processing Unit,CPU),微处理器或其他数据处理芯片,用于运行存储器11中存储的程序代码或处理数据,例如执行数据推荐程序等。
网络接口13可选地可以包括标准的有线接口、无线接口(如WI-FI接口),通常用于在该电子设备与其他电子设备之间建立通信连接。
数据总线14用于实现这些组件之间的连接通信。
可选的,该电子设备还可以包括用户接口,用户接口可以包括输入单元比如键盘(Keyboard)、语音输入装置比如麦克风(microphone)等具有语音识别功能的设备、语音输出装置比如音响、耳机等,可选地用户接口还可以包括标准的有线接口、无线接口。
可选的,该电子设备还可以包括显示器,显示器也可以称为显示屏或显示单元。在一些 实施例中可以是LED显示器、液晶显示器、触控式液晶显示器以及有机发光二极管(Organic Light-Emitting Diode,OLED)触摸器等。显示器用于显示在电子设备中处理的信息以及用于显示可视化的用户界面。
可选的,该电子设备还包括触摸传感器。触摸传感器所提供的供用户进行触摸操作的区域称为触控区域。此外,这里的触摸传感器可以为电阻式触摸传感器、电容式触摸传感器等。而且,触摸传感器不仅包括接触式的触摸传感器,也可包括接近式的触摸传感器等。此外,触摸传感器可以为单个传感器,也可以为例如阵列布置的多个传感器。
此外,该电子设备的显示器的面积可以与触摸传感器的面积相同,也可以不同。可选地,将显示器与触摸传感器层叠设置,以形成触摸显示屏。该装置基于触摸显示屏侦测用户触发的触控操作。
下面对本申请实施例公开的一种基于图神经网络的数据推荐方法进行具体说明。
实施例一
如图2所示,图2是本申请实施例一提供的基于图神经网络的数据推荐方法的流程图。基于图1所示的电子设备实施例,处理器12执行存储器11中存储的程序时实现如下步骤:
步骤S200:获取新用户的目标搜索信息。
在本申请实施例中,新用户可以是第一次输入搜索信息的用户,目标搜索信息包括但不限于:新用户语音输入的搜索语句或关键词;新用户通过搜索引擎输入的文字信息或图片信息等。具体的,每次接收到搜索请求时,可以先获取搜索请求对应的用户信息,用户信息包括但不限于用户ID、用户IP地址和用户设备码等用于标识用户身份的信息。之后,确认数据库中是否已录入用户信息,若未录入用户信息,则根据用户信息新建新用户并录入数据库。相应的,若已录入用户信息,则从数据库中确定用户信息对应的目标用户,并直接获取目标用户的目标特征信息,执行步骤S260。
步骤S210:根据目标搜索信息,生成候选数据集。
在本申请实施例中,可以预先通过网络爬虫搜集互联网的网页链接以及具体访问资源(比如某一文档、购物页的某一商品等)等大数据信息,再对这些大数据信息进行提取并组织建立索引库。实际应用中,可以由检索器将用户输入的目标搜索信息拆分成具有检索意义的字或词,再根据拆分得到的字或词,在索引库中快速检出候选数据以生成候选数据集。示例性的,当新用户在搜索引擎中输入搜索关键词“XX科技”时,调取包含“XX科技”的网页链接集合以加入候选数据集。
步骤S220:确定与新用户相关联的关联用户。
在本申请实施例中,关联用户可以包括但不限于:新用户使用社交软件时关联的好友用户、工作用户(包括客户、同级同事、上下级同事等)、临时对话用户(比如评论或转发用户)、关注用户、点赞用户、点踩用户和收藏用户等;以及,新用户的设备通讯录联系人。相应的,可以通过获取新用户的账号信息访问权限或使用设备信息访问权限,从新用户访问的社交应用软件、办公应用软件或多媒体应用软件等多种信息渠道确定与新用户相关联的关联用户。
步骤S230:获取关联用户的第一搜索数据。
其中,第一搜索数据可以包括关联用户的历史搜索信息,比如关联用户利用搜索引擎搜索过的文本、图片等。进一步的,第一搜索数据具体还可以包括关联用户对历史搜索信息的评价数据,比如关联用户对任一网页链接或具体访问资源的评价内容(包括喜欢、不喜欢和评价类型等);关联用户对历史搜索信息的推荐数据,比如关联用户使用软件分享功能时发送的数据。
步骤S240:根据新用户、关联用户和第一搜索数据,生成新用户的社交关系图。
在本申请实施例中,根据新用户和关联用户,可以确定新用户与关联用户之间的社交关系信息,比如同事、好友、关注、点赞或不喜欢等。根据关联用户和第一搜索数据,可以确定关联用户和第一搜索数据之间的关联关系信息,比如喜欢、不喜欢、推荐、好评或差评等。基于此,可选的,分别以新用户、关联用户以及第一搜索数据为不同节点,再根据不同节点之间的关系信息,为不同节点之间建立连接关系,得到新用户的社交关系图。
请参阅图3,图3是本申请实施例中一种社交关系图的示意图。如图3所示,该社交关系图中包括四个用户节点和两个数据节点,四个用户节点包括新用户节点A、第一关联用户节点B、第二关联用户节点C和第三关联用户节点D,两个数据节点包括物品节点1和物品节点2。其中,与新用户相关的社交关系信息包括:新用户节点A与第一关联用户节点B连接,其连接关系为“不喜欢”;新用户节点A与第二关联用户节点C连接,其连接关系为点赞关系;新用户节点A与第三关联用户节点D连接,其连接关系为关注关系。与第二关联用户节点C相关的关系信息包括:第二关联用户节点C与物品节点1连接,其连接关系为“不喜欢”;第二关联用户节点C与物品节点2连接,其连接关系为“喜欢”。可见,社交关系图采用图数据结构,便于梳理不同用户之间以及用户与资源之间的关系。
步骤S250:将社交关系图输入到预设图神经网络模型中进行特征预测,得到新用户的目标特征信息。
作为一种可选的实施方式,预设图神经网络模型的训练步骤可以包括:
首先,获取训练样本以及训练样本的标注数据,训练样本包括第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图,第一用户和第二用户为任意两位用户,标注数据用于表示第一用户与第二用户是否相似。其中,第二社交关系样本图和第一社交关系样本图可参照上述对社交关系图的说明,不再赘述。一种实现方式中,先取第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图作为训练样本,再对第一社交关系样本图和第二社交关系样本图进行节点搜索。若在第一社交关系样本图搜索到第二用户对应的第二节点,和/或,在第二社交关系样本图中搜索到第一用户对应的第一节点,则将训练样本的标注数据确定为第一用户与第二用户相似。若未在第一社交关系样本图搜索到第二节点,并且未在第二社交关系样本图中搜索到第一节点,将训练样本的标注数据确定为第一用户与第二用户不相似。
之后,将第一社交关系样本图和第二社交关系样本图输入预设图神经网络模型。通过预设图神经网络模型生成第一社交关系样本图的第一特征信息以及第二社交关系样本图的第二 特征信息,并对第一特征信息和第二特征信息进行融合,得到融合结果。其中,融合结果可以表示第一特征信息与第二特征信息之间的相似值S。
最后,利用标注数据对融合结果进行验证,得到验证结果,若验证结果未达到预设准确率,则重新执行获取训练样本以及训练样本的标注数据的步骤,若验证结果达到预设准确率,训练结束。预设准确率可以由人为设定及调整,比如取90%或95%,不做具体限定。
可以理解,每次用于训练预设图神经网络模型的训练样本数量可以是多个。一种实现方式中,可以计算每个训练样本的准确值R,再对多个训练样本的准确值R进行平均计算,得到验证结果,准确值R的计算方式为:
Figure PCTCN2022090754-appb-000001
其中,s1是标注数据为相似时设定的标准值(比如0.9),s2为标注数据为不相似时设定的标准值(比如0.1)。
另一种实现方式中,也可以在标注数据为相似的情况下,设定第一取值范围(比如0.8至1),若训练样本对应的相似值S属于第一取值范围,则样本计数值N累计加一。以及,在标注数据为不相似的情况下,设定第二取值范围(比如0至0.2),若训练样本对应的相似值S属于第二取值范围,则样本计数值N累计加一。基于此,遍历所有训练样本后,取最终确定的样本计数值N与训练样本数量之间的比值作为验证结果。
进一步的,作为一种可选的实施方式,预设图神经网络模型可以包括第一特征生成模块、第二特征生成模块以及融合模块,则在预设图神经网络模型的训练步骤中,可以将第一社交关系样本图输入第一特征生成模块,得到第一特征信息,以及将第二社交关系样本图输入第二特征生成模块,得到第二特征信息。再将第一特征信息和第二特征信息输入融合模块,得到融合结果。具体的,融合模块可以采用点积运算单元,则将第一特征信息和第二特征信息输入融合模块中进行点积运算,得到第一特征信息与第二特征信息之间的相似值S。
此外,在训练预设图神经网络模型的过程中,可以通过均方根误差损失函数L,对预设图神经网络模型输出的融合结果与标注数据计算损失值,并通过损失值的反向传播进行预设图神经网络模型的参数调整,直至训练结束。均方根误差损失函数L满足:
Figure PCTCN2022090754-appb-000002
其中,m为训练样本的数量,y i为第i个训练样本对应的融合结果,
Figure PCTCN2022090754-appb-000003
为第i个训练样本对应的标注数据。
相应的,将社交关系图输入到预设图神经网络模型中进行特征预测,得到新用户的目标特征信息,具体为:将社交关系图输入到第一特征生成模块或第二特征生成模块中进行特征预测,得到新用户的目标特征信息。可见,训练好的预设图神经网络模型能够智能分析用户社交关系及关联数据特征,为不同用户生成可靠的特征信息,并同时满足不同用户之间的相 似对比结果,因此在实际应用中,只需将新用户的社交关系图输入预设图神经网络模型中的特征生成模块,即可得到新用户的目标特征信息,具备更优的模型应用性能和可靠性。
请参与图4,图4是本申请实施例中一种预设图神经网络模型的结构示意图。如图4所示,作为一种可选的实施方式,第一特征生成模块和第二特征生成模块均包括依次连接的双层图卷积层(double graph convolutional layer,DGCL)、池化层和全连接层,形成相似度-双向图卷积网络(similarity bi-directional graph convolutional networks,S-BIGCN)。具体的,双层图卷积层的参数共享,适用于无向图结构的社会关系图,而池化层可以采用平均池化层,既防止过拟合,又使得特征生成模块具备更优的训练速度以及推理速度,不需要耗费过多的计算机GPU资源,故预设图神经网络模型更易落地。
其他可选的实施方式中,预设图神经网络模型也可以采用图注意力网络模型、图自编码网络模型或图生成网络模型等。
步骤S260:根据目标特征信息,对候选数据集进行排序处理,得到推荐数据集。
在本申请实施例中,候选数据集包括多个候选数据,一种可选的实施方式中,可以获取每个候选数据相关的社交关系图,并将候选数据相关的社交关系图输入到预设图神经网络模型中,得到候选数据的特征信息。将目标特征信息与每个候选数据的特征信息代入相似度计算公式中,得到每个候选数据对应的排序值。根据每个候选数据对应的排序值大小,即可实现对候选数据集的排序处理。其中,相似度计算公式包括但不限于余弦相似度计算方式、欧氏距离计算方式、曼哈顿距离计算公式或Jaccard相似系数计算公式等。
可见,实施上述方法实施例,将同时融合了用户社交关系以及不同用户的第一搜索数据的社交关系图输入到预设图神经网络模型中,使得预设图神经网络模型沿着社交关系链的深层传播识别关联用户之间的共同数据需求和评价模式特征,从而为新用户确定更具备可解释性的目标特征信息。之后,再基于新用户的目标特征信息,对新用户的候选数据集进行排序,得到为新用户推送的推荐数据集,能够缓解用户侧冷启动问题,提升对新用户进行数据推荐的准确性,进而改善了新用户对数据搜索功能的使用体验。
实施例二
如图5所示,图5是本申请实施例二提供的基于图神经网络的数据推荐方法的流程图。基于图1所示的电子设备实施例,处理器12执行存储器11中存储的程序时实现如下步骤:
步骤S500:获取新用户的目标搜索信息。
步骤S510:根据目标搜索信息,生成候选数据集。
步骤S520:确定与新用户相关联的关联用户。
步骤S530:获取关联用户的第一搜索数据。
步骤S540:根据新用户、关联用户和第一搜索数据,生成新用户的社交关系图。
步骤S550:将社交关系图输入到预设图神经网络模型中进行特征预测,得到新用户的目标特征信息。
可以理解的是,本实施例中步骤S500至S550的具体实现方式还可以参照上述实施例一中对步骤S200至S250的描述,在此不再赘述。
步骤S560:根据目标特征信息,获取新用户的多位相似用户以及相似用户与新用户的相似度。
作为一种可选的实施方式,步骤S560具体为:获取多位已知用户的第三特征信息。根据目标特征信息以及每位已知用户的第三特征信息,计算新用户与已知用户之间的相似度,并取相似度在预设阈值范围内的已知用户作为相似用户。其中,相似度的取值属于[0,1],相似度的取值越大,表示新用户与已知用户的相似程度越高。预设阈值范围可以是人为指定,比如[0.8,1];或者,也可以对新用户与不同已知用户之间的相似度从大到小进行排序,取排序第k位的相似度为目标阈值d,k为正整数并由人为指定,再根据目标阈值d确定预设阈值范围(比如预设阈值范围为[d,1]),对此不做限定。
具体的,可以从数据库中获取已录入的多位已知用户。一种实现方式中,将目标特征信息和每位已知用户的第三特征信息代入预设的相似度计算公式中,得到新用户与已知用户之间的相似度。比如,相似度计算公式可以是余弦相似度计算方式,即:
Figure PCTCN2022090754-appb-000004
其中,sim(u,v)为新用户与已知用户之间的相似度,u为目标特征信息,v为已知用户的第三特征信息。
可见,通过挖掘新用户与已知用户的相似性,筛选出新用户的相似用户,以便根据相似用户的行为数据进行推荐参考。
进一步的,作为一种可选的实施方式,获取多位已知用户的第三特征信息,具体可以包括以下步骤:
获取多位已知用户的历史搜索信息以及已知用户对历史搜索信息的操作数据,已知用户的历史搜索信息包括多个第二搜索数据。第二搜索数据可以指响应于已知用户的搜索请求后,向已知用户推荐过的历史数据。操作数据包括对每个第二搜索数据的操作记录,操作记录对应的操作类型包括但不限于以下至少一种:点击次数、收藏操作、分享次数、具体评分情况、点赞操作和点踩操作,用于表示已知用户对不同第二搜索数据的反馈情况。
之后,根据历史搜索信息和操作数据,生成评分矩阵。从评分矩阵中获取每位已知用户对应的行向量,行向量包括多个第一评分元素,多个第一评分元素与多个第二搜索数据一一对应。根据已知用户对应的行向量,即可确定已知用户的第三特征信息。
其中,评分矩阵的行表示不同已知用户,评分矩阵的列表示不同第二搜索数据,则第i行第j列的矩阵元素表示第i位已知用户对第j个第二搜索数据的评分元素。
一种实现方式中,如果采用的操作类型为一种,则根据该操作类型的操作记录,确定相应的评分元素。可以理解,如果操作记录为数值,比如点击次数、分享次数或具体评分情况等,则可以直接取操作记录为评分元素。如果操作记录并非数值,可以先对操作记录进行量化处理,得到评分元素。比如,针对收藏操作,将已收藏操作量化为“1”,将未收藏操作量化为“0”;将点赞操作量化为“1”;以及,将点踩操作量化为“-1”等。
示例性的,假设已知用户E对第二搜索数据1、第二搜索数据2和第二搜索数据3的点 击次数分别为1、1和0,已知用户F对应的点击次数分别为1、0和0,已知用户G对应的点击次数分别为2、1和0,可以生成如下表1所示的评分矩阵。
表1评分矩阵示例表
  第二搜索数据1 第二搜索数据2 第二搜索数据3
已知用户E 1 1 0
已知用户F 1 0 1
已知用户G 2 1 0
基于表1,取行向量[1,1,0]作为已知用户E的第三特征信息,取行向量[1,0,1]作为已知用户F的第三特征信息,取行向量[2,1,0]为已知用户G的第三特征信息。
另一种实现方式中,如果采用的操作类型为两种及以上,则可以根据不同操作类型的加权权重,对所有操作类型的操作记录进行加权求和计算,得到相应的评分元素。比如,假设点击次数和分享次数各自的加权权重为0.5,已知用户E分别对第二搜索数据1、第二搜索数据2和第二搜索数据3的分享次数为1、0和0,则已知用户E的行向量=[1*0.5+1*0.5,1*0.5,0]=[1,0.5,0]。
此外,还可选的,还可以将评分矩阵中每个行向量(或列向量)进行归一化处理,得到归一化后的行向量(或列向量),使得归一化后的行向量(或列向量)中每个评分元素分布在[0,1]中。
基于此,根据已知用户对应的行向量,确定已知用户的第三特征信息之后,可以对第三特征信息进行存储,便于实际应用中直接调取已知用户的第三特征信息。进一步的,也可以在训练好(或每次更新完)预设图神经网络模型后,并获得已知用户的社交关系图的情况下,将已知用户的社交关系图输入到预设图神经网络模型中,得到新的第三特征信息,并利用新的特征信息替代原有的第三特征信息进行存储,实现特征更新作用。
步骤S570:获取多位相似用户对候选数据集的第一评分信息。
在本申请实施例中,候选数据集可以包括多个候选数据。一种可选的实施方式中,根据历史搜索信息和操作数据,生成评分矩阵之后,还可以从评分矩阵中获取每个第二搜索数据对应的列向量,列向量包括多个第二评分元素,多个第二评分元素与多位已知用户一一对应。根据第二搜索数据对应的列向量,确定第二搜索数据的第四特征信息。
仍以表1为例,可以取列向量[1,1,2]为第二搜索数据1的第四特征信息,取列向量[1,0,1]为第二搜索数据2的第四特征信息,取列向量[0,1,0]为第二搜索数据3的第四特征信息。
相应的,步骤S570可以是:通过将候选数据与第二搜索数据进行匹配,从多个第二搜索数据的第四特征信息中,确定每个候选数据的第四特征信息。根据相似用户的第三特征信息以及候选数据的第四特征信息,计算相似用户对候选数据的第一评分值。
具体的,可以将相似用户的第三特征信息w以及候选数据的第四特征信息c代入相似度计算公式中,得到相似用户对候选数据的第一评分值score(u,c),比如,相似度计算公式采用 余弦相似度计算方式,即:
Figure PCTCN2022090754-appb-000005
其他实现方式中,也可以基于候选数据的标签生成候选数据的第四特征信息,或者采用其他基于神经网络学习的方法生成候选数据的第四特征信息。
步骤S580:根据第一评分信息,并结合每位相似用户与新用户的相似度,计算新用户对候选数据集的第二评分信息。
在本申请实施例中,第二评分信息包括新用户对每个候选数据的第二评分值。可选的,计算第二评分信息的方式具体可以为:
Figure PCTCN2022090754-appb-000006
其中,Like(u,cj)为新用户对第j个候选数据cj的第二评分值,sim(u,wi)为新用户与第i位相似用户的相似度,score(wi,cj)为第i位相似用户对第j个候选数据cj的第一评分值。
步骤S590:根据第二评分信息,对候选数据集进行排序处理,得到推荐数据集。
在本申请实施例中,具体的,可以按照第二评分值从大到小的次序,对候选数据集中多个候选数据进行降序处理,得到推荐数据集。或者,也可以在对候选数据集中多个候选数据进行降序处理之后,将降序处理的结果分为p组推荐数据(p为正整数,且p≥2),再对每组推荐数据进行乱序处理,得到乱序后的p组推荐数据,以形成推荐数据集。
可见,实施上述步骤S560至步骤S590,通过筛选新用户的相似用户,并结合相似用户的数据评分情况预测新用户对不同候选数据的兴趣偏好,实现了对候选数据集的个性化排序,从而优化新用户的搜索满意度。
可以理解的是,本申请中涉及的数据库可以采用Elastic Search、MongoDB或Redis等,不做具体限定。
可见,实施上述方法实施例,通过筛选新用户的相似用户,并结合相似用户的数据评分情况预测新用户对不同候选数据的兴趣偏好,实现了对候选数据集的协同过滤,能够缓解用户侧冷启动问题,提升对新用户进行数据推荐的准确性,进而改善了新用户对数据搜索功能的使用体验。
本申请实施例还提供一种基于图神经网络的数据推荐装置。请参阅图6,图6是本申请实施例所应用的一种基于图神经网络的数据推荐装置的结构框图。如图6所示,该基于图神经网络的数据推荐装置600包括:
获取模块610,用于获取新用户的目标搜索信息。
第一生成模块620,用于根据目标搜索信息,生成候选数据集。
确定模块630,用于确定与新用户相关联的关联用户。
获取模块610,还用于获取关联用户的第一搜索数据。
第二生成模块640,用于根据新用户、关联用户和第一搜索数据,生成新用户的社交关系图。
预测模块650,用于将社交关系图输入到预设图神经网络模型中进行特征预测,得到新 用户的目标特征信息。
排序模块660,用于根据目标特征信息,对候选数据集进行排序处理,得到推荐数据集。
需要说明的是,本实施例的具体实现过程可参见上述方法实施例的具体实现过程,亦不再赘述。
本申请实施例还提供了一种电子设备,该电子设备包括存储器和处理器,存储器上存储有程序,程序被处理器执行时实现一种基于图神经网络的数据推荐方法,该方法包括:获取新用户的目标搜索信息,并根据目标搜索信息,生成候选数据集;确定与新用户相关联的关联用户;获取关联用户的第一搜索数据;根据新用户、关联用户和第一搜索数据,生成新用户的社交关系图;将社交关系图输入到预设图神经网络模型中进行特征预测,得到新用户的目标特征信息;根据目标特征信息,对候选数据集进行排序处理,得到推荐数据集。
本申请实施例还提供了一种存储介质,用于计算机可读存储,存储介质存储有一个或者多个程序,一个或者多个程序可被一个或者多个处理器执行,以实现一种基于图神经网络的数据推荐方法,该方法包括:获取新用户的目标搜索信息,并根据目标搜索信息,生成候选数据集;确定与新用户相关联的关联用户;获取关联用户的第一搜索数据;根据新用户、关联用户和第一搜索数据,生成新用户的社交关系图;将社交关系图输入到预设图神经网络模型中进行特征预测,得到新用户的目标特征信息;根据目标特征信息,对候选数据集进行排序处理,得到推荐数据集。该存储介质可以是非易失性,也可以是易失性,不做具体限定。
本领域普通技术人员可以理解,上文中所公开方法中的全部或某些步骤、系统、设备中的功能模块/单元可以被实施为软件、固件、硬件及其适当的组合。
在硬件实施方式中,在以上描述中提及的功能模块/单元之间的划分不一定对应于物理组件的划分;例如,一个物理组件可以具有多个功能,或者一个功能或步骤可以由若干物理组件合作执行。某些物理组件或所有物理组件可以被实施为由处理器,如中央处理器、数字信号处理器或微处理器执行的软件,或者被实施为硬件,或者被实施为集成电路,如专用集成电路。这样的软件可以分布在计算机可读介质上,计算机可读介质可以包括计算机存储介质(或非暂时性介质)和通信介质(或暂时性介质)。如本领域普通技术人员公知的,术语计算机存储介质包括在用于存储信息(诸如计算机可读指令、数据结构、程序模块或其他数据)的任何方法或技术中实施的易失性和非易失性、可移除和不可移除介质。计算机存储介质包括但不限于RAM、ROM、EEPROM、闪存或其他存储器技术、CD-ROM、数字多功能盘(DVD)或其他光盘存储、磁盒、磁带、磁盘存储或其他磁存储装置、或者可以用于存储期望的信息并且可以被计算机访问的任何其他的介质。此外,本领域普通技术人员公知的是,通信介质通常包含计算机可读指令、数据结构、程序模块或者诸如载波或其他传输机制之类的调制数据信号中的其他数据,并且可包括任何信息递送介质。
以上参照附图说明了本申请的优选实施例,并非因此局限本申请的权利范围。本领域技术人员不脱离本申请的范围和实质内所作的任何修改、等同替换和改进,均应在本申请的权利范围之内。

Claims (20)

  1. 一种基于图神经网络的数据推荐方法,其中,所述方法包括:
    获取新用户的目标搜索信息,并根据所述目标搜索信息,生成候选数据集;
    确定与所述新用户相关联的关联用户;
    获取所述关联用户的第一搜索数据;
    根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;
    将所述社交关系图输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;
    根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
  2. 根据权利要求1所述的方法,其中,所述预设图神经网络模型的训练步骤包括:
    获取训练样本以及所述训练样本的标注数据,所述训练样本包括第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图,所述第一用户和所述第二用户为任意两位用户,所述标注数据用于表示所述第一用户与所述第二用户是否相似;
    将所述第一社交关系样本图和所述第二社交关系样本图输入预设图神经网络模型;
    通过所述预设图神经网络模型生成所述第一社交关系样本图的第一特征信息以及所述第二社交关系样本图的第二特征信息,并对所述第一特征信息和所述第二特征信息进行融合,得到融合结果;
    利用所述标注数据对所述融合结果进行验证,得到验证结果,若所述验证结果未达到预设准确率,则重新执行所述获取训练样本以及所述训练样本的标注数据的步骤,若所述验证结果达到所述预设准确率,训练结束。
  3. 根据权利要求2所述的方法,其中,所述获取训练样本以及所述训练样本的标注数据,包括:
    取第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图作为训练样本;
    对所述第一社交关系样本图和所述第二社交关系样本图进行节点搜索;
    若在所述第二社交关系样本图中搜索到所述第一用户对应的第一节点,和/或,在所述第一社交关系样本图中搜索到所述第二用户对应的第二节点,将所述训练样本的标注数据确定为所述第一用户与所述第二用户相似;
    若未在所述第一社交关系样本图搜索到所述第二节点,并且未在所述第二社交关系样本图中搜索到所述第一节点,将所述训练样本的标注数据确定为所述第一用户与所述第二用户不相似。
  4. 根据权利要求1至3任一项所述的方法,其中,所述根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集,包括:
    根据所述目标特征信息,获取所述新用户的多位相似用户以及所述相似用户与所述新用户的相似度;
    获取多位所述相似用户对所述候选数据集的第一评分信息;
    根据所述第一评分信息,并结合每位所述相似用户与所述新用户的相似度,计算所述新用户对所述候选数据集的第二评分信息;
    根据所述第二评分信息,对所述候选数据集进行排序处理,得到推荐数据集。
  5. 根据权利要求4所述的方法,其中,所述根据所述目标特征信息,获取所述新用户的多位相似用户以及所述相似用户与所述新用户的相似度,包括:
    获取多位已知用户的第三特征信息;
    根据所述目标特征信息以及每位所述已知用户的第三特征信息,计算所述新用户与所述已知用户之间的相似度,并取相似度在预设阈值范围内的已知用户作为相似用户。
  6. 根据权利要求5所述的方法,其中,所述获取多位已知用户的第三特征信息,包括:
    获取多位已知用户的历史搜索信息以及所述已知用户对所述历史搜索信息的操作数据,所述历史搜索信息包括多个第二搜索数据;
    根据所述历史搜索信息和所述操作数据,生成评分矩阵;
    从所述评分矩阵中获取每位已知用户对应的行向量,所述行向量包括多个第一评分元素,多个所述第一评分元素与多个所述第二搜索数据一一对应;
    根据所述已知用户对应的行向量,确定所述已知用户的第三特征信息。
  7. 根据权利要求6所述的方法,其中,所述候选数据集包括多个候选数据;所述根据所述历史搜索信息和所述操作数据,生成评分矩阵之后,所述方法还包括:
    从所述评分矩阵中获取每个第二搜索数据对应的列向量,所述列向量包括多个第二评分元素,多个所述第二评分元素与多位所述已知用户一一对应;
    根据所述第二搜索数据对应的列向量,确定所述第二搜索数据的第四特征信息;
    所述获取多位所述相似用户对所述候选数据集的第一评分信息,包括:
    从多个所述第二搜索数据的第四特征信息中,确定每个所述候选数据的第四特征信息;
    根据所述相似用户的第三特征信息以及所述候选数据的第四特征信息,计算所述相似用户对所述候选数据的第一评分值。
  8. 一种基于图神经网络的数据推荐装置,其中,所述装置包括:
    获取模块,用于获取新用户的目标搜索信息;
    第一生成模块,用于根据所述目标搜索信息,生成候选数据集;
    确定模块,用于确定与所述新用户相关联的关联用户;
    所述获取模块,还用于获取所述关联用户的第一搜索数据;
    第二生成模块,用于根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;
    预测模块,用于将所述社交关系图输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;
    排序模块,用于根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
  9. 一种电子设备,其中,所述电子设备包括存储器、处理器、存储在所述存储器上并可 在所述处理器上运行的程序以及用于实现所述处理器和所述存储器之间的连接通信的数据总线,所述程序被所述处理器执行时实现一种基于图神经网络的数据推荐方法,所述方法包括:
    获取新用户的目标搜索信息,并根据所述目标搜索信息,生成候选数据集;
    确定与所述新用户相关联的关联用户;
    获取所述关联用户的第一搜索数据;
    根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;
    将所述社交关系图输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;
    根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
  10. 根据权利要求9所述的电子设备,其中,所述预设图神经网络模型的训练步骤包括:
    获取训练样本以及所述训练样本的标注数据,所述训练样本包括第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图,所述第一用户和所述第二用户为任意两位用户,所述标注数据用于表示所述第一用户与所述第二用户是否相似;
    将所述第一社交关系样本图和所述第二社交关系样本图输入预设图神经网络模型;
    通过所述预设图神经网络模型生成所述第一社交关系样本图的第一特征信息以及所述第二社交关系样本图的第二特征信息,并对所述第一特征信息和所述第二特征信息进行融合,得到融合结果;
    利用所述标注数据对所述融合结果进行验证,得到验证结果,若所述验证结果未达到预设准确率,则重新执行所述获取训练样本以及所述训练样本的标注数据的步骤,若所述验证结果达到所述预设准确率,训练结束。
  11. 根据权利要求10所述的电子设备,其中,所述获取训练样本以及所述训练样本的标注数据,包括:
    取第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图作为训练样本;
    对所述第一社交关系样本图和所述第二社交关系样本图进行节点搜索;
    若在所述第二社交关系样本图中搜索到所述第一用户对应的第一节点,和/或,在所述第一社交关系样本图中搜索到所述第二用户对应的第二节点,将所述训练样本的标注数据确定为所述第一用户与所述第二用户相似;
    若未在所述第一社交关系样本图搜索到所述第二节点,并且未在所述第二社交关系样本图中搜索到所述第一节点,将所述训练样本的标注数据确定为所述第一用户与所述第二用户不相似。
  12. 根据权利要求9至11任一项所述的电子设备,其中,所述根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集,包括:
    根据所述目标特征信息,获取所述新用户的多位相似用户以及所述相似用户与所述新用户的相似度;
    获取多位所述相似用户对所述候选数据集的第一评分信息;
    根据所述第一评分信息,并结合每位所述相似用户与所述新用户的相似度,计算所述新 用户对所述候选数据集的第二评分信息;
    根据所述第二评分信息,对所述候选数据集进行排序处理,得到推荐数据集。
  13. 根据权利要求12所述的电子设备,其中,所述根据所述目标特征信息,获取所述新用户的多位相似用户以及所述相似用户与所述新用户的相似度,包括:
    获取多位已知用户的第三特征信息;
    根据所述目标特征信息以及每位所述已知用户的第三特征信息,计算所述新用户与所述已知用户之间的相似度,并取相似度在预设阈值范围内的已知用户作为相似用户。
  14. 根据权利要求13所述的电子设备,其中,所述获取多位已知用户的第三特征信息,包括:
    获取多位已知用户的历史搜索信息以及所述已知用户对所述历史搜索信息的操作数据,所述历史搜索信息包括多个第二搜索数据;
    根据所述历史搜索信息和所述操作数据,生成评分矩阵;
    从所述评分矩阵中获取每位已知用户对应的行向量,所述行向量包括多个第一评分元素,多个所述第一评分元素与多个所述第二搜索数据一一对应;
    根据所述已知用户对应的行向量,确定所述已知用户的第三特征信息。
  15. 一种存储介质,用于计算机可读存储,其中,所述存储介质存储有一个或者多个程序,所述一个或者多个程序可被一个或者多个处理器执行,以实现一种基于图神经网络的数据推荐方法,所述方法包括:
    获取新用户的目标搜索信息,并根据所述目标搜索信息,生成候选数据集;
    确定与所述新用户相关联的关联用户;
    获取所述关联用户的第一搜索数据;
    根据所述新用户、所述关联用户和所述第一搜索数据,生成所述新用户的社交关系图;
    将所述社交关系图输入到预设图神经网络模型中进行特征预测,得到所述新用户的目标特征信息;
    根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集。
  16. 根据权利要求15所述的存储介质,其中,所述预设图神经网络模型的训练步骤包括:
    获取训练样本以及所述训练样本的标注数据,所述训练样本包括第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图,所述第一用户和所述第二用户为任意两位用户,所述标注数据用于表示所述第一用户与所述第二用户是否相似;
    将所述第一社交关系样本图和所述第二社交关系样本图输入预设图神经网络模型;
    通过所述预设图神经网络模型生成所述第一社交关系样本图的第一特征信息以及所述第二社交关系样本图的第二特征信息,并对所述第一特征信息和所述第二特征信息进行融合,得到融合结果;
    利用所述标注数据对所述融合结果进行验证,得到验证结果,若所述验证结果未达到预设准确率,则重新执行所述获取训练样本以及所述训练样本的标注数据的步骤,若所述验证结果达到所述预设准确率,训练结束。
  17. 根据权利要求16所述的存储介质,其中,所述获取训练样本以及所述训练样本的标注数据,包括:
    取第一用户的第一社交关系样本图以及第二用户的第二社交关系样本图作为训练样本;
    对所述第一社交关系样本图和所述第二社交关系样本图进行节点搜索;
    若在所述第二社交关系样本图中搜索到所述第一用户对应的第一节点,和/或,在所述第一社交关系样本图中搜索到所述第二用户对应的第二节点,将所述训练样本的标注数据确定为所述第一用户与所述第二用户相似;
    若未在所述第一社交关系样本图搜索到所述第二节点,并且未在所述第二社交关系样本图中搜索到所述第一节点,将所述训练样本的标注数据确定为所述第一用户与所述第二用户不相似。
  18. 根据权利要求15至17任一项所述的存储介质,其中,所述根据所述目标特征信息,对所述候选数据集进行排序处理,得到推荐数据集,包括:
    根据所述目标特征信息,获取所述新用户的多位相似用户以及所述相似用户与所述新用户的相似度;
    获取多位所述相似用户对所述候选数据集的第一评分信息;
    根据所述第一评分信息,并结合每位所述相似用户与所述新用户的相似度,计算所述新用户对所述候选数据集的第二评分信息;
    根据所述第二评分信息,对所述候选数据集进行排序处理,得到推荐数据集。
  19. 根据权利要求18所述的存储介质,其中,所述根据所述目标特征信息,获取所述新用户的多位相似用户以及所述相似用户与所述新用户的相似度,包括:
    获取多位已知用户的第三特征信息;
    根据所述目标特征信息以及每位所述已知用户的第三特征信息,计算所述新用户与所述已知用户之间的相似度,并取相似度在预设阈值范围内的已知用户作为相似用户。
  20. 根据权利要求19所述的存储介质,其中,所述获取多位已知用户的第三特征信息,包括:
    获取多位已知用户的历史搜索信息以及所述已知用户对所述历史搜索信息的操作数据,所述历史搜索信息包括多个第二搜索数据;
    根据所述历史搜索信息和所述操作数据,生成评分矩阵;
    从所述评分矩阵中获取每位已知用户对应的行向量,所述行向量包括多个第一评分元素,多个所述第一评分元素与多个所述第二搜索数据一一对应;
    根据所述已知用户对应的行向量,确定所述已知用户的第三特征信息。
PCT/CN2022/090754 2022-02-16 2022-04-29 基于图神经网络的数据推荐方法及装置、电子设备 WO2023155306A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210143649.3 2022-02-16
CN202210143649.3A CN114491294A (zh) 2022-02-16 2022-02-16 基于图神经网络的数据推荐方法及装置、电子设备

Publications (1)

Publication Number Publication Date
WO2023155306A1 true WO2023155306A1 (zh) 2023-08-24

Family

ID=81482125

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/090754 WO2023155306A1 (zh) 2022-02-16 2022-04-29 基于图神经网络的数据推荐方法及装置、电子设备

Country Status (2)

Country Link
CN (1) CN114491294A (zh)
WO (1) WO2023155306A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112115367A (zh) * 2020-09-28 2020-12-22 北京百度网讯科技有限公司 基于融合关系网络的信息推荐方法、装置、设备和介质
CN112966165A (zh) * 2021-02-03 2021-06-15 北京大学 基于图神经网络的交互式社区搜索方法及装置
CN113282831A (zh) * 2021-06-10 2021-08-20 北京爱奇艺科技有限公司 一种搜索信息的推荐方法、装置、电子设备及存储介质
CN113392289A (zh) * 2021-06-08 2021-09-14 北京三快在线科技有限公司 搜索推荐方法、装置、电子设备
WO2021233199A1 (zh) * 2020-05-19 2021-11-25 华为技术有限公司 搜索推荐模型的训练方法、搜索结果排序的方法及装置
WO2022024663A1 (ja) * 2020-07-30 2022-02-03 オムロン株式会社 モデル生成装置、経路探索装置、モデル生成方法、及びモデル生成プログラム

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021233199A1 (zh) * 2020-05-19 2021-11-25 华为技术有限公司 搜索推荐模型的训练方法、搜索结果排序的方法及装置
WO2022024663A1 (ja) * 2020-07-30 2022-02-03 オムロン株式会社 モデル生成装置、経路探索装置、モデル生成方法、及びモデル生成プログラム
CN112115367A (zh) * 2020-09-28 2020-12-22 北京百度网讯科技有限公司 基于融合关系网络的信息推荐方法、装置、设备和介质
CN112966165A (zh) * 2021-02-03 2021-06-15 北京大学 基于图神经网络的交互式社区搜索方法及装置
CN113392289A (zh) * 2021-06-08 2021-09-14 北京三快在线科技有限公司 搜索推荐方法、装置、电子设备
CN113282831A (zh) * 2021-06-10 2021-08-20 北京爱奇艺科技有限公司 一种搜索信息的推荐方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN114491294A (zh) 2022-05-13

Similar Documents

Publication Publication Date Title
US11580104B2 (en) Method, apparatus, device, and storage medium for intention recommendation
King et al. Computer‐assisted keyword and document set discovery from unstructured text
CN111753060B (zh) 信息检索方法、装置、设备及计算机可读存储介质
CN111708873B (zh) 智能问答方法、装置、计算机设备和存储介质
US9449271B2 (en) Classifying resources using a deep network
US10515424B2 (en) Machine learned query generation on inverted indices
US10354170B2 (en) Method and apparatus of establishing image search relevance prediction model, and image search method and apparatus
CN105630856B (zh) 在线用户简档的自动聚合
US10146862B2 (en) Context-based metadata generation and automatic annotation of electronic media in a computer network
CN104750798B (zh) 一种应用程序的推荐方法和装置
WO2019223552A1 (zh) 文章推荐方法、装置、计算机设备及存储介质
Chen et al. Mobile app tagging
CN111797214A (zh) 基于faq数据库的问题筛选方法、装置、计算机设备及介质
TW201447797A (zh) 內容個人化之多相排序方法和系統
CN112052387B (zh) 一种内容推荐方法、装置和计算机可读存储介质
US11249993B2 (en) Answer facts from structured content
Wei et al. Online education recommendation model based on user behavior data analysis
CN110162535B (zh) 用于执行个性化的搜索方法、装置、设备以及存储介质
WO2023155306A1 (zh) 基于图神经网络的数据推荐方法及装置、电子设备
CN109885758A (zh) 一种基于二部图的新型随机游走的推荐方法
CN113761125A (zh) 动态摘要确定方法和装置、计算设备以及计算机存储介质
CN116610853A (zh) 搜索推荐方法、搜索推荐系统、计算机设备及存储介质
CN110222156B (zh) 发现实体的方法和装置、电子设备、计算机可读介质
CN113722484A (zh) 基于深度学习的谣言检测方法、装置、设备及存储介质
Li et al. Research on hot news discovery model based on user interest and topic discovery

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: 22926619

Country of ref document: EP

Kind code of ref document: A1