Method, device, computer device and storage medium for recommending objects

Abstract

The application relates to the technical field of artificial intelligence, and provides a method and device for recommending an object, computer equipment and a storage medium. The method comprises the following steps: acquiring scene mode features extracted from scene mode data matched with a target user identifier; the scene mode data is used for describing a scene mode corresponding to the target user identifier; for each candidate recommended object, based on user features corresponding to the target user identifier, scene mode features and object features corresponding to the candidate recommended object, a multi-layer perception mapping process is performed to obtain a predicted recommendation result of each candidate recommended object in the scene mode; and according to the predicted recommendation result, a target recommended object is selected from the candidate recommended objects, and the target recommended object is recommended based on the target user identifier. By using the above method, accurate recommendation of an object matched with a scene mode can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a method and device for recommending objects, computer equipment and storage medium. BACKGROUND

[0002] With the development of computer technology, there are more and more human-computer interaction scenarios. For example, objects used by users in daily life, such as small programs running in application programs, public numbers, accessed web pages, etc., through object interaction, greatly facilitate users in daily business, travel, shopping, information verification and other aspects of business handling and daily life. Among them, the display mode of the interactive object in the interaction interface is generally the fixed sorting of the system default or the user self-setting.

[0003] However, the traditional object recommendation method generally recommends to the user the object that the user has used recently. Based on this object recommendation method, in some scenarios, the user still needs to use the object that the user needs through relatively complex operations such as on-site code scanning, and this object recommendation method has the problem of inaccurate recommendation of objects. SUMMARY

[0004] Therefore, it is necessary to provide a method, device, computer equipment and storage medium for recommending objects, which can accurately recommend objects.

[0005] A method for recommending objects, the method comprising:

[0006] obtaining a scene mode feature extracted from scene mode data matched with a target user identifier; the scene mode data is used to describe a scene mode corresponding to the target user identifier;

[0007] For each candidate recommendation object, based on the user description feature corresponding to the target user identifier, the scene mode feature, and the object feature corresponding to the candidate recommendation object, a multi-layer perception mapping process is performed to obtain a predicted recommendation result of each candidate recommendation object in the scene mode;

[0008] According to the predicted recommendation result, a target recommendation object is selected from the candidate recommendation object, and the target recommendation object is recommended based on the target user identifier.

[0009] In one embodiment, the candidate recommendation object includes an object associated with the current status data and an object corresponding to an object identifier recorded in the object use history data.

[0010] In one embodiment, the method further comprises:

[0011] An object usage record, the object usage record including state data corresponding to each object being used by a user and an object identifier of each object;

[0012] Based on co-occurrence data in the object usage data, an association relationship between state data and object identifiers that meet association requirements is established based on the co-occurrence data and the state data and the object identifiers;

[0013] Based on the association relationship, an object corresponding to an object identifier associated with the current state data is determined.

[0014] In one of the embodiments, the filtering, from the candidate recommended objects, a target recommended object according to the predicted recommendation result, and recommending the target recommended object based on the target user identifier, includes:

[0015] Based on the number of recommended objects and the predicted recommendation result, a target recommended object meeting the number of recommended objects is filtered from the candidate recommended objects, the number of recommended objects being the same as the number of displayable recommended objects of an object display page of a terminal corresponding to the target user identifier;

[0016] The target recommended object is pushed to the terminal corresponding to the target user identifier, so that the terminal displays the target recommended object on the object display page.

[0017] A device for recommending an object, the device including:

[0018] A scene mode feature acquisition module is configured to acquire a scene mode feature extracted from scene mode data matched with a target user identifier, the scene mode data being used to describe a scene mode corresponding to the target user identifier;

[0019] A predicted recommendation result analysis module is configured to, for each candidate recommended object, perform a multi-layer perception mapping process based on a user description feature corresponding to the target user identifier, the scene mode feature, and an object feature corresponding to the candidate recommended object, to obtain a predicted recommendation result of each candidate recommended object in the scene mode;

[0020] A target recommended object filtering module is configured to filter, from the candidate recommended objects, a target recommended object according to the predicted recommendation result, and recommend the target recommended object based on the target user identifier.

[0021] In one of the embodiments, the scene mode data includes at least one of current state data corresponding to the target user identifier and object usage history data corresponding to the target user identifier;

[0022] The scene mode feature includes at least one of a current situation feature extracted from the current situation data and a historical use feature extracted from the object use history data.

[0023] In one of the embodiments, the current situation data includes a geographical attribute and a state time corresponding to the geographical attribute; the device for recommending an object further includes a geographical position information obtaining module and a state time and geographical attribute determining module; the geographical position information obtaining module is configured to obtain geographical position information collected by a terminal corresponding to the target user identifier when responding to a user interaction operation; the state time and geographical attribute determining module is configured to obtain a corresponding state time based on the geographical position information and determine a geographical attribute of a region corresponding to the geographical position information.

[0024] In one of the embodiments, the target recommended object includes a target recommended sub-application configured to run in a running environment of a parent application;

[0025] The geographical position information obtaining module is further configured to obtain geographical position information collected by the terminal when the parent application running on the terminal corresponding to the target user identifier enters a sub-application access page;

[0026] The target recommended object filtering module is further configured to push recommendation information of the target recommended sub-application to the terminal corresponding to the target user identifier, so that the terminal performs at least one of preloading the target recommended sub-application and displaying the target recommended sub-application on the sub-application access page.

[0027] In one of the embodiments, the current situation feature includes a current situation feature vector; the device for recommending an object further includes a current situation feature vector obtaining module configured to determine an element position and an element value of the current situation data in a feature matrix corresponding to a scene type of the current situation data, to obtain a current situation feature matrix, based on the scene type, and to perform feature mapping processing on the current situation feature matrix to obtain a current situation feature vector.

[0028] In one of the embodiments, the historical use feature includes a time sequence feature vector; the device for recommending an object further includes a time sequence feature vector obtaining module;

[0029] The time sequence feature vector obtaining module is configured to perform time sequence feature processing on an object use history data sequence obtained by sorting the object use history data according to time, to obtain a time sequence feature vector corresponding to the object use history data, based on the time corresponding to the object use history data.

[0030] In one of the embodiments, the device for recommending an object further includes a training sample obtaining module and a training module;

[0031] The training sample obtaining module is configured to obtain a training sample, the training sample comprising user description features, object features and scene mode data corresponding to the target user identifier;

[0032] The training module is configured to train based on the training sample, initial feature extraction parameters and initial mapping parameters until a prediction result obtained based on trained feature extraction parameters and trained mapping parameters meets a training end condition; the trained feature extraction parameters are used to extract the scene mode features from the scene mode data;

[0033] The prediction recommendation result analysis module is further configured to perform multi-layer perception mapping processing based on the trained mapping parameters.

[0034] In one of the embodiments, the scene mode data in the training sample comprises object use history data and history state data;

[0035] The feature extraction parameters comprise history use feature extraction parameters used to extract history use features from the object use history data and state feature extraction parameters used to extract state features from state data, the state data comprising history state data and current state data.

[0036] In one of the embodiments, the training sample comprises positive samples and negative samples; the training sample obtaining module further comprises a matching relationship determining module, a positive sample constructing module and a negative sample constructing module;

[0037] The matching relationship determining module is configured to determine a matching relationship between an object and a history state based on a history state corresponding to each piece of data in the object use history data;

[0038] The positive sample constructing module is configured to construct a positive sample based on a matching object and a matching history state with a matching relationship, the history state data in the positive sample being history state data corresponding to the matching history state, and the object features in the positive sample being object features corresponding to the matching object;

[0039] The negative sample constructing module is configured to construct a negative sample based on a non-matching object and a non-matching history state without a matching relationship, the history state data in the negative sample being history state data corresponding to the non-matching history state, and the object features in the negative sample being object features corresponding to the non-matching object.

[0040] In one of the embodiments, the candidate recommendation objects comprise objects associated with the current state data and objects corresponding to object identifiers recorded in the object use history data.

[0041] In one of the embodiments, the device for recommending an object further comprises an object usage record obtaining module, an association relationship establishing module and an object determining module;

[0042] The object usage record obtaining module is configured to obtain object usage records, the object usage records comprising corresponding state data when each object is used by a user and object identifiers of the objects;

[0043] The association relationship establishing module is configured to establish an association relationship between state data and object identifiers in co-occurrence data in the object usage data based on the co-occurrence data and the object identifiers;

[0044] The object determining module is configured to determine an object corresponding to an object identifier associated with the state data based on the association relationship.

[0045] In one of the embodiments, the device for recommending an object further comprises a corpus data obtaining module, a word vector embedding processing module and a user association feature determining module;

[0046] The corpus data obtaining module is configured to obtain corpus data, the corpus data comprising sequentially connected identification pairs, the identification pairs comprising user identifiers and object identifiers, and adjacent user identifiers and object identifiers in the corpus data indicating that a user corresponding to the adjacent user identifier has a usage behavior on an object corresponding to the adjacent object identifier;

[0047] The word vector embedding processing module is configured to perform word vector embedding processing on the corpus data to obtain a user embedding vector corresponding to each user identifier in the corpus data and an object embedding vector corresponding to each object identifier in the corpus data;

[0048] The user association feature determining module is configured to take a user embedding vector corresponding to the target user identifier as a user description feature corresponding to the target user identifier, and take an object embedding vector corresponding to the candidate recommended object as an object feature corresponding to the candidate recommended object.

[0049] In one of the embodiments, the corpus data obtaining module comprises a relationship network obtaining module and a random walk module;

[0050] The relationship network obtaining module is configured to obtain a relationship network constructed based on user identifiers and object identifiers, nodes in the relationship network representing the user identifiers and the object identifiers, and edges between nodes in the relationship network representing usage behaviors of users corresponding to the user identifiers on objects corresponding to the object identifiers;

[0051] The random walk module is configured to start from a node corresponding to each user identifier, perform random walk along edges of the node in the relationship network, and obtain the corpus data based on a walk track.

[0052] In one of the embodiments, the relationship network acquisition module comprises a behavior matrix acquisition module and a relationship network construction module.

[0053] The behavior matrix acquisition module is configured to acquire a behavior matrix with user identifiers and object identifiers representing rows and columns respectively, determine a target element with a target value in the behavior matrix, and the target value represents that a user corresponding to the user identifier has a usage behavior on an object corresponding to the object identifier.

[0054] The relationship network construction module is configured to take the user identifiers and the object identifiers represented by the rows and columns of the behavior matrix as nodes, determine node edges based on the row and column positions of the target element in the behavior matrix, and obtain a relationship network.

[0055] In one of the embodiments, the target recommendation object screening module comprises a quantity determination module and a pushing module.

[0056] The quantity determination module is configured to screen target recommendation objects from the candidate recommendation objects based on a quantity of recommendation objects and the predicted recommendation result, and the quantity of recommendation objects is the same as a quantity of displayable recommendation objects of an object display page of a terminal corresponding to the target user identifier.

[0057] The pushing module is configured to push the target recommendation objects to a terminal corresponding to the target user identifier, so that the terminal displays the target recommendation objects on the object display page.

[0058] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0059] Acquire scene mode features extracted from scene mode data matched with a target user identifier; the scene mode data is used to describe a scene mode corresponding to the target user identifier.

[0060] For each candidate recommendation object, perform multi-layer perception mapping processing based on user description features corresponding to the target user identifier, scene mode features, and object features corresponding to the candidate recommendation object, to obtain a predicted recommendation result of each candidate recommendation object in the scene mode.

[0061] According to the predicted recommendation result, screen target recommendation objects from the candidate recommendation objects, and recommend the target recommendation objects based on the target user identifier.

[0062] A computer readable storage medium, having stored thereon a computer program, the computer program being executed by a processor to implement the following steps:

[0063] Obtaining scene mode features extracted from scene mode data matching the target user identifier; the scene mode data is used to describe a scene mode corresponding to the target user identifier;

[0064] For each candidate recommended object, based on the user description features corresponding to the target user identifier, the scene mode features, and the object features corresponding to the candidate recommended object, a multi-layer perception mapping process is performed to obtain a predicted recommendation result of each candidate recommended object in the scene mode;

[0065] According to the predicted recommendation result, a target recommended object is selected from the candidate recommended objects, and the target recommended object is recommended based on the target user identifier.

[0066] The above-mentioned recommended object method, device, computer equipment and storage medium, by using scene mode data for describing the scene mode corresponding to the target user identifier, obtaining scene mode features extracted from the scene mode data, to accurately obtain scene mode features matching the target user identifier, for each candidate recommended object, based on the user description features corresponding to the target user identifier, the scene mode features, and the object features corresponding to the candidate recommended object, a multi-layer perception mapping process can be performed by combining multi-dimensional features, by analyzing each candidate recommended object separately, the predicted recommendation result of each candidate recommended object in the scene mode is accurately obtained, according to the predicted recommendation result, a target recommended object is selected from the candidate recommended objects, and the target recommended object is recommended based on the target user identifier, which can realize accurate recommendation of objects matching the scene mode. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 Application environment diagram for the recommended object method in one embodiment;

[0068] Figure 2 Flowchart of the recommended object method in one embodiment;

[0069] Figure 3 Structural diagram of the multi-layer perception machine in one embodiment;

[0070] Figure 4 Page diagram of the sub-application access page in one embodiment;

[0071] Figure 5 Page diagram of the sub-application access page in another embodiment;

[0072] Figure 6 Fig. 16 is a schematic diagram of a relationship network constructed based on user identification and object identification in one embodiment;

[0073] Figure 7 Fig. 17 is a schematic diagram of random walk based on the relationship network in one embodiment;

[0074] Figure 8 Fig. 18 is a schematic diagram of two different behavior matrices in one embodiment;

[0075] Figure 9 Fig. 19 is a schematic diagram of a method of recommending objects in another embodiment;

[0076] Figure 10 Fig. 20 is a schematic diagram of a data processing flow of predicting for each candidate object in one embodiment;

[0077] Figure 11 Fig. 21 is a schematic diagram of a small program for displaying recommendations in one embodiment;

[0078] Figure 12 Fig. 22 is a structural block diagram of an apparatus for recommending objects in one embodiment;

[0079] Figure 13 Fig. 23 is an internal structural diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0080] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application.

[0081] The scheme provided by the embodiments of the present application can involve technologies such as artificial intelligence (AI) and machine learning (ML). Artificial intelligence is the theory, technology and application system for using a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use the knowledge to obtain optimal results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making. Machine learning is a multi-field interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. Machine learning is a branch of artificial intelligence that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure and continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental approach to enabling computers to have intelligence, and its applications are widespread in various fields of artificial intelligence. Based on technologies such as artificial intelligence and machine learning, scene mode features extracted from scene mode data matching a target user identifier can be obtained, the scene mode data is used to describe a scene mode corresponding to the target user identifier, for each candidate recommended object, based on user description features corresponding to the target user identifier, scene mode features and object features corresponding to the candidate recommended object, multi-layer perception mapping processing is performed to obtain a predicted recommendation result of each candidate recommended object in the scene mode, a target recommended object is selected from the candidate recommended objects according to the predicted recommendation result, and the target recommended object is recommended based on the target user identifier, so as to realize accurate recommendation of objects.

[0082] The method for recommending an object provided by the present application can be applied to, for example, Figure 1The application environment shown. Among them, the application environment provides a computer device, for example, terminal 120 or terminal 130 and server 110. The information push method can be applied to server 110 or terminal 120 or terminal 130, and can also be applied to a system including a terminal and a server, and is realized through the interaction of the terminal and the server. For example, the server 110 obtains the scene mode features extracted from the scene mode data matched with the target user identifier, and the scene mode data is used to describe the scene mode corresponding to the target user identifier. The server 110 performs a multi-layer perception mapping process on each candidate recommended object based on the user description features corresponding to the target user identifier, the scene mode features, and the object features corresponding to the candidate recommended object, obtains the predicted recommended result of each candidate recommended object in the scene mode, and the server 110 filters out the target recommended object from the candidate recommended object according to the predicted recommended result, and pushes the recommended information of the target recommended object to the terminal corresponding to the target user identifier. The terminal corresponding to the target user identifier can be any one of terminal 120 or terminal 130. For example, the terminal 120 receives the recommended information of the target recommended object and automatically displays the target recommended object, or displays the target recommended object based on user operation, and the target recommended object can be an applet for running in an application program. Among them, the terminal can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices, for example, the terminal 120 is a smart phone and the terminal 130 is a portable wearable device. The server 110 can be realized by an independent server or a server cluster composed of multiple servers. The terminal 120 or the terminal 130 and the server 110 can be directly or indirectly connected through wired or wireless communication mode, for example, through network connection.

[0083] In one embodiment, as shown in Figure 2 , a method for recommending an object is provided, which is applied to a computer device in Figure 1 for example, including the following steps:

[0084] Step 202, obtaining the scene mode features extracted from the scene mode data matched with the target user identifier.

[0085] The scene mode data is used to describe a scene mode corresponding to the target user identifier, and the scene mode data includes a detailed description of the scene mode. In an embodiment, the scene mode is used to represent characteristics of a user, and the scene mode can correspond to user information of one data dimension or at least two data dimensions. The specific dimensions and the number of dimensions can be selected and set according to data dimensions to be considered when the object is recommended. For example, the scene mode can be used to represent a current state of the user, the scene mode can be used to represent a use habit of the user using the object, the scene mode can be used to represent a specific interaction operation of the user, and in addition, the scene mode can be used to represent information of each of the above dimensions.

[0086] The target user identifier refers to an identifier of a user to which the object recommendation is performed. The computer device can identify the target user identifier through an interaction operation of the user, or determine the target user identifier according to a configured trigger condition and user information. For example, when the user performs an interaction operation, the computer device determines the target user identifier by responding to the interaction operation of the user. For another example, when a terminal satisfies a configured trigger condition, the target user identifier is determined according to user information of a user corresponding to the terminal.

[0087] The scene mode data can be obtained in real time or from pre-stored data. For example, changing information such as time and place can be obtained in real time. For another example, data information that can be recorded can be pre-stored and obtained from pre-stored data. The scene mode data has a matching relationship with the user identifier, and the scene mode data matched with the target user identifier can be obtained through the target user identifier and the matching relationship.

[0088] The scene mode feature is a characteristic representation of the scene mode data, and the computer device extracts the corresponding scene mode feature based on the scene mode data. The characteristic representation refers to a data expression mode that can be used to represent multi-dimensional features of data through characteristic processing of intuitive and readable data. For example, vectorization representation of data, analysis result representation of data, or chart representation of data.

[0089] Specifically, the computer device obtains the scene mode feature extracted from the scene mode data matched with the target user identifier, including: obtaining the scene mode data matched with the target user identifier, and extracting the scene mode feature from the scene mode data.

[0090] In step 204, for each candidate recommended object, a multi-layer perception mapping process is performed based on the user description feature corresponding to the target user identifier, the scene mode feature, and the object feature corresponding to the candidate recommended object, to obtain a predicted recommendation result of each candidate recommended object in the scene mode.

[0091] The candidate recommendation object refers to a selection range of a target recommendation object recommended to a user. The computer respectively performs prediction for each candidate recommendation object. The prediction process includes: based on the user description feature corresponding to the target user identification, the scene mode feature, and the object feature corresponding to the candidate recommendation object, performing multi-layer perception mapping processing.

[0092] The user description feature refers to a characteristic representation of user-related data. The user-related data can include data for describing the characteristics of the user himself, and can also include data for describing the use characteristics of the user to each object, etc. Taking the user description feature as one of the multi-layer mapping perception features can accurately represent the data characteristics of the user data dimension, and facilitate obtaining accurate prediction results through multi-layer perception mapping processing.

[0093] The object feature corresponding to the candidate recommendation object is a characteristic representation of object-related data. The object-related data can include data for describing the characteristics of the object itself, and can also include data for describing each use object of the object, etc. Taking the object feature as one of the multi-layer mapping perception features can accurately represent the data characteristics of the object data dimension, and facilitate obtaining accurate prediction results through multi-layer perception mapping processing.

[0094] In an embodiment, the user description feature and the object feature can be obtained by analyzing the association relationship of all users and all objects. The all users refer to a set of users who have object use records, and the all objects refer to a set of objects used by the users. The association relationship analysis of the all users and the all objects refers to a process of analyzing each user in the all users and each object in the all objects based on the use behavior of any user to any object.

[0095] The multi-layer perception mapping processing refers to a process of mapping data in multiple dimensions to a single data. Specifically, the multi-layer perception mapping processing can be realized by a multi-layer perception machine (also known as an artificial neural network model). For example, the multi-layer perception machine can be a three-layer perception machine. Figure 3As shown, the multi-layer perception includes, in addition to the input and output layers, a plurality of hidden layers in between, and the layers of the multi-layer perception are fully connected. Specifically, the bottom layer of the multi-layer perception is the input layer, the middle layer is the hidden layer, and the last layer is the output layer. The neurons of the hidden layer are fully connected with the input layer. Assuming that the input layer is represented by a vector X, the output of the hidden layer is f(W1X+b1), W1 is the connection coefficient between the input layer and the hidden layer, b1 is the bias of the input layer to the hidden layer, and the function f can be a commonly used sigmoid function or tanh function. The hidden layer to the output layer is a multi-class logistic regression, i.e., a softmax regression, and the output of the output layer is: softmax(W2X1+b2), X1 represents the output f(W1X+b1) of the hidden layer, W2 is the connection coefficient between the hidden layer and the output layer, and b2 is the bias of the hidden layer to the output layer.

[0096] The mapping processing by the multi-layer perception is essentially to take the user description feature, the object feature, and the scene mode feature as the input data of the multi-layer perception, perform mapping processing based on the parameters of each layer in the multi-layer perception, and obtain the output result as the predicted recommendation result of the candidate recommended object in the corresponding scene mode.

[0097] The parameters of each layer in the multi-layer perception include the connection weights and the biases between the layers, including W1, b1, W2, b2, etc. The determination of the parameters of each layer in the multi-layer perception can be obtained through model training. The specific training process includes: first, randomly initializing all the parameters of the multi-layer perception, then iteratively training, constantly calculating the gradient and updating the parameters, until the training end condition is met. The training end condition can be at least one of the conditions that the accuracy meets the accuracy requirement, the error meets the error requirement, or the number of iterations meets the number of times requirement.

[0098] In step 206, the target recommended object is selected from the candidate recommended objects according to the predicted recommendation result, and the target recommended object is recommended to the target user identifier based on the target user identifier.

[0099] The target recommended object refers to the object recommended to the user corresponding to the target user identifier for use by the user. The computer device selects the target recommended object that meets the recommendation requirement from the candidate recommended objects according to the predicted recommendation result of each candidate recommended object in the scene mode.

[0100] The prediction recommendation result can be a recommendation probability or a recommendation level determined based on the recommendation probability. Taking the recommendation probability as an example, the target recommendation object meeting the recommendation requirement can be a candidate recommendation object with a recommendation probability greater than a preset probability, or can be a candidate recommendation object meeting a recommendation quantity selected from a sorting result after sorting according to the recommendation probability. For example, five candidate recommendation objects have recommendation probabilities of 0.5, 0.6, 0.7, 0.75, and 0.9, respectively. If the recommendation requirement is a recommendation probability greater than 0.8, the target recommendation object is the candidate recommendation object with a recommendation probability of 0.9. If the recommendation requirement is the top two candidate recommendation objects with a greater recommendation probability, the target recommendation objects are the candidate recommendation objects with recommendation probabilities of 0.85 and 0.9.

[0101] The target recommendation object can be recommended based on the target user identifier. The recommendation information corresponding to the target recommendation object can be pushed to a terminal corresponding to the target user identifier, so that the terminal performs further processing, such as displaying the target recommendation object, or preloading the target recommendation object, and the like. In this way, the user experience of using the object can be improved by directly displaying the recommended target recommendation object or improving the loading speed of the object.

[0102] The above method for recommending an object can obtain a scene mode feature extracted from scene mode data for describing a scene mode corresponding to a target user identifier, so as to accurately obtain a scene mode feature matched with the target user identifier. For each candidate recommendation object, a multi-dimensional feature can be combined for multi-layer perception mapping processing based on a user description feature corresponding to the target user identifier, the scene mode feature, and an object feature corresponding to the candidate recommendation object. By separately analyzing each candidate recommendation object, an accurate prediction recommendation result of each candidate recommendation object in the scene mode can be obtained. The target recommendation object can be selected from the candidate recommendation objects according to the prediction recommendation result, and the target recommendation object can be recommended based on the target user identifier. In this way, the accurate selection and recommendation of an object matched with the scene mode can be realized.

[0103] In one embodiment, the scene mode data includes at least one of present situation data corresponding to the target user identifier and object use history data corresponding to the target user identifier. The scene mode feature includes at least one of a present situation feature extracted from the present situation data and a history use feature extracted from the object use history data.

[0104] The present situation data refers to data describing the environment or scene in which the user is currently located. By feature processing the present situation data corresponding to the target user identifier, the present situation feature corresponding to the target user identifier can be obtained to realize the feature representation of the present situation data. In an embodiment, the present situation data can be at least one of POI (point of interest) information and time information of the user, wherein the POI information can be a specific geographic location or a regional attribute, and the time information can be a day of the week, a time of day, etc.

[0105] The object use history data refers to the record data of the use or use habit of the user to the object. The use of the object by the user refers to the behavior of the user clicking or other operations on the object to start the running of the object. For example, the object use history data can include the record of the use of each object by the user in a set time range, wherein the set time range can be one month, half a month or a week with the current time as the last time. The record of the use of each object by the user is, for example, that the user used the payment sub-application in the shopping mall at 14:30 on December 12, 2020, and the user used the ordering sub-application in the office building at 11:30 on December 15, 2020. The sub-application refers to a program running in the running environment of the parent application. For example, a small program running in the application program of WeChat, Alipay, etc.

[0106] By feature processing the object use history data corresponding to the target user identifier, the history use feature corresponding to the object use history data can be obtained, which realizes the feature representation of the object use history data.

[0107] In this embodiment, the computer device can obtain the corresponding scene mode feature by obtaining the present situation data corresponding to the target user identifier, or the object use history data corresponding to the target user identifier, or the present situation data and the object use history data corresponding to the target user identifier, to realize the accurate feature description of the scene mode and thus realize the accurate object recommendation.

[0108] In an embodiment, the present situation data includes a regional attribute and a state time corresponding to the regional attribute; the method further includes: obtaining the geographic location information collected by the terminal corresponding to the target user identifier in response to the user interaction operation; obtaining the corresponding state time based on the geographic location information, and determining the regional attribute of the region corresponding to the geographic location information.

[0109] A region refers to a result of dividing a geographical area, and a region attribute is used to describe common features of the divided geographical area. For example, a region can be divided based on a function that can be implemented, such as a commercial district, a hospital, a subway station, an airport, or a train station, or can be divided based on an administrative region, such as Beijing, Shanghai, or Guangzhou. The region attribute can be a single specified dimension feature, such as a function attribute, or can be a multi-dimensional attribute, such as a Guangzhou subway station, and further can be accurate to a specific unit that implements a function, such as the first hospital in city A or the central mall in city B.

[0110] A state time corresponding to the region attribute refers to a time obtained in a process of obtaining the region attribute and used to represent a state in which a user is currently located. In an embodiment, if the computer device is a terminal, the state time can be a system time obtained when the terminal collects geographical position information, if the computer device is a server, the state time can be time information carried when the terminal transmits the obtained geographical position information to the server, or can be a system time obtained when the server receives the geographical position information, or can be a time corresponding to when the server obtains the region attribute based on the geographical position information.

[0111] The geographical position information is used to represent a position in which the user is currently located, and the geographical position information is a kind of information that changes with the movement of the user. The user interaction operation refers to an interaction operation between the user and the terminal. The user interaction operation is a behavior that is actively triggered by the user. The terminal collects the geographical position information when responding to the user interaction operation, so that the time when the terminal collects the geographical position information can be accurately defined, and the terminal can collect the geographical position information when needed, thereby avoiding unnecessary waste of data processing resources of the terminal.

[0112] Taking the computer device as a server as an example, the server obtains geographical position information collected by a terminal corresponding to a target user identifier when the terminal responds to a user interaction operation, obtains a state time corresponding to the geographical position information based on the geographical position information, and determines a region attribute of a region corresponding to the geographical position information.

[0113] Taking the computer device as a terminal as an example, the terminal collects geographical position information when responding to a user interaction operation, obtains a state time corresponding to the geographical position information based on the geographical position information, and determines a region attribute of a region corresponding to the geographical position information.

[0114] In an embodiment, the target recommendation object includes a target recommendation sub-application used to run in a running environment of a parent application; and the geographical position information collected by the terminal corresponding to the target user identifier when the terminal responds to the user interaction operation includes: when the parent application run by the terminal corresponding to the target user identifier enters a sub-application access page, the geographical position information collected by the terminal.

[0115] Recommending target recommendation objects based on target user identifiers includes: pushing recommendation information of target recommendation sub-applications to the terminal corresponding to the target user identifier, so that the terminal performs at least one of the following processes: preloading the target recommendation sub-application and displaying the target recommendation sub-application on the sub-application access page.

[0116] Here, a sub-application refers to a program that runs within the runtime environment of the parent application and can be used without downloading or installing it. The parent application is a program that runs on the terminal by downloading and installing an application package. A sub-application access page refers to a page accessed directly within the parent application through actions within the parent application, such as scrolling down, searching for sub-applications, or triggering icons. This page may include displays of the target recommended sub-applications. Figure 4 As shown. In practical applications, such as Figure 5 As shown, to facilitate user selection, the sub-application access page can display not only recommended sub-applications, but also recently used sub-applications and user-defined sub-applications (My Sub-applications).

[0117] Accessing a child application's page from a running parent application can be achieved by the user performing a pre-defined interactive action. In one specific application, the running parent application accesses the child application's page in response to the user's page pull-down action. When the terminal detects the pull-down action in the parent application's page, it collects the terminal's current geolocation information. In another specific application, the main page of the running parent program may include the child application's page; entering the parent program's main page constitutes accessing the child application's page.

[0118] Preloading refers to the process of loading some key content before all other data is loaded, in order to reduce waiting time. This is to avoid situations where the page remains blank for an extended period due to excessively large amounts of content.

[0119] In one specific embodiment, the method for recommending sub-applications includes: when the parent application running on the terminal corresponding to the target user identifier enters the sub-application access page, the computer device obtains scene mode data that matches the target user identifier and describes the scene mode corresponding to the target user identifier; extracts scene mode features from the scene mode data; for each candidate recommended sub-application, performs multi-layer perception mapping processing based on the user description features corresponding to the target user identifier, the scene mode features, and the sub-application features corresponding to the target candidate recommended sub-application; obtains the predicted recommendation result of each candidate recommended sub-application in the scene mode; and pushes the recommendation information of the target recommended sub-application to the terminal corresponding to the target user identifier, so that the terminal performs at least one of the following processes: preloading the target recommended sub-application and displaying the target recommended sub-application on the sub-application access page.

[0120] By the above method of recommending an object, the sub-application running in the running environment of the mother application can be accurately recommended. Specifically, taking the sub-application running in the running environment of the mother application such as WeChat or Alipay as a small program as an example, compared with the traditional small program display methods, such as the first method: the product manager arranges the small programs according to the fixed order (for example, the development order of the small program); the second method: the user places the small program frequently used by the user in the frequently used menu; and the third method: the user places the recently used small program on the top of the user menu. The above processing methods have certain limitations, for example, in the first method, if the small programs are sorted according to the fixed order, the user will have difficulty in finding the required small program. The second method allows the user to adjust the order of the small program according to the user's own usage habits, but this method is limited to frequently used small programs, and cannot meet the sudden or cyclic small program demand due to business trips, travel, takeout, shopping or short videos. For the third method, since the user may use the small program periodically, locally, or specially, saving only the recently used small program cannot fundamentally improve the user experience, resulting in that in actual use, the user still needs to perform a relatively complex operation or on-site code scanning to open the required small program on the terminal, such as subway code scanning, airport code scanning for epidemic prevention, and the like, which brings many inconveniences in use.

[0121] The above method of recommending an object limits the acquisition time of the scene mode data by acquiring the scene mode data for describing the scene mode corresponding to the target user identifier when the mother application running on the terminal corresponding to the target user identifier enters the sub-application access page, avoids waste of data processing resources of the terminal, and the computer device acquires the scene mode features extracted from the scene mode data to accurately obtain the scene mode features matched with the target user identifier. For each candidate recommended sub-application, based on the user description features corresponding to the target user identifier, the scene mode features, and the sub-application features corresponding to the candidate recommended sub-application, multi-dimensional features can be combined for multi-perception mapping processing. By analyzing each candidate recommended sub-application separately, the prediction recommendation result of each candidate recommended sub-application in the scene mode is accurately obtained. According to the prediction recommendation result, the target recommended sub-application is selected from the candidate recommended sub-application, and the target recommended sub-application is recommended based on the target user identifier, which can accurately select and recommend the sub-application matched with the scene mode.

[0122] In one embodiment, the status features include a status feature vector; the status feature vector is one of the featureized representations of the status data. The method of recommending an object further includes a process of obtaining the status feature vector, and the specific process includes:

[0123] Based on the scene type corresponding to the status data, the element position and element value of the status data in the feature matrix corresponding to the scene type are determined to obtain a status feature matrix; and the status feature matrix is subjected to feature mapping processing to obtain a status feature vector.

[0124] The scene type refers to a data category to which the status data belongs, for example, a data category corresponding to a week, a data category corresponding to a time point in a day, and a data category corresponding to a regional attribute. Each status data has a corresponding scene type, and each scene type is configured with a corresponding feature matrix. Therefore, each status data has a corresponding element in the feature matrix. Specifically, the element value is taken, and the row and column positions of the corresponding element in the feature matrix are determined. For example, the feature matrix corresponding to the week is a 7-row 1-column matrix. If the status data is Wednesday, the first element value of the third row is taken, and the second element value of the remaining rows is taken, where the first element value and the second element value are different, for example, the first element value is 1 and the second element value is 0. For another example, the feature matrix corresponding to the time point can be a matrix configured by dividing a time period into one hour or half an hour, for example, a 24-row 1-column time matrix or a 48-row 1-column matrix, etc. The element position in the feature matrix corresponding to the time point can be determined according to the time range corresponding to the time, and the element value can be taken in the same way as the element value in the feature matrix corresponding to the week, which is not described again.

[0125] The feature mapping processing refers to a data processing process in which the data in the feature matrix is represented by a vector. Specifically, one feature matrix can correspond to one vector, that is, when the number of feature matrices is greater than or equal to two, the number of status feature vectors can be greater than or equal to one, or all feature matrices can correspond to one vector, that is, when the number of feature matrices is greater than or equal to two, the number of status feature vectors can be one.

[0126] Further, the feature mapping processing can be implemented by using parameters for feature mapping obtained through training, wherein the parameters for feature mapping can be obtained based on sample data carrying labeled data.

[0127] In this embodiment, by mapping the status data through the feature matrix, the data under different status can be uniformly represented, and the status feature vector accurately describing the status data can be obtained conveniently and quickly.

[0128] In one embodiment, the historical use feature includes a time sequence feature vector; the time sequence feature vector is one of the featureized representations of the historical data of the object. The method of recommending the object further includes a process of obtaining the time sequence feature vector, and the specific process includes:

[0129] The object use history data sequence obtained by sorting according to time is subjected to time sequence feature processing based on the time corresponding to the object use history data, to obtain a time sequence feature vector corresponding to the object use history data.

[0130] The object use history data includes a plurality of object use records of the user corresponding to the target user identifier within a certain time range, each object use record has a corresponding object use time, for example, object A is used at 11:30 on December 15, 2020. The plurality of object use records are sorted based on the time sequence of the corresponding object use time, to obtain the object use history data sequence. The object use history data can also be recorded directly based on the time sequence of the object use time, in which case the object use history data sequence can be obtained directly from the recorded data.

[0131] The time sequence feature processing refers to a feature processing process of the object use history data sequence, which can be implemented by an LSTM (Long Short-Term Memory) model or an RNN (Recurrent Neural Networks) model. The LSTM model and other models for time sequence feature processing can be trained to obtain model parameters for time sequence feature processing, and the model for time sequence feature processing can be trained based on sample data carrying labeled data.

[0132] By sorting the object use history data by time to obtain the object use history data sequence and then the time sequence feature vector, the cycle information of the user using the object and the user's recent use of the object can be reflected, which facilitates accurate expression of the user's use habits of using the object, thereby achieving convenient and fast obtaining of features corresponding to the object use history data.

[0133] In one embodiment, the method of recommending an object further includes training of a multi-layer perception model, i.e., an artificial neural network model. The training process includes:

[0134] The training sample includes user description features, object features, and scene mode data corresponding to the target user identifier. The training is performed based on the initial feature extraction parameters and the initial mapping parameters until the prediction result obtained based on the trained feature extraction parameters and the trained mapping parameters meets the training end condition. The trained feature extraction parameters are used to extract scene mode features from the scene mode data. The mapping processing of the multi-layer perception includes mapping processing of the multi-layer perception based on the trained mapping parameters.

[0135] The training sample is data carrying a predicted recommendation result, used for supervised model training to obtain required parameters in actual application, including feature extraction parameters for extracting scene mode features from scene mode data, and mapping parameters for mapping processing of multilayer perception.

[0136] Further, the training sample is a sample corresponding to data corresponding to the target user identifier. The model and parameters trained thereby are ensured to correspond to the target user identifier. That is, the model can be trained in the same way for each user to obtain feature extraction parameters and mapping parameters corresponding to the user.

[0137] The training sample includes user description features, object features, and scene mode data corresponding to the target user identifier. The feature extraction of the scene mode data is based on the feature extraction parameters. In the process of training the multilayer perceptron for mapping processing of multilayer perception based on the mapping parameters, the feature extraction parameters are also needed for training to obtain scene mode features. Then, the multilayer perceptron is trained based on the user description features, object features, and scene mode features to obtain the feature extraction parameters and the mapping parameters.

[0138] The training end condition of the model can be that the accuracy meets the accuracy requirement, the error meets the error requirement, or the number of iterations meets the number requirement, which can be set according to actual needs.

[0139] In this embodiment, the model training based on the feature extraction parameters and the mapping parameters can effectively and accurately extract scene mode features based on the feature extraction parameters in subsequent application, and can achieve more accurate mapping processing of multilayer perception. In addition, since the sample data includes user description features, object features, and scene mode data corresponding to the target user identifier, the obtained feature extraction parameters and mapping parameters can better map and fuse the data related to the user corresponding to the target user identifier, thereby obtaining more accurate predicted recommendation results.

[0140] In one embodiment, the scene mode data in the training sample includes object use history data and history state data; the feature extraction parameters include history use feature extraction parameters for extracting history use features from the object use history data, and state feature extraction parameters for extracting state features from the state data, the state data including the history state data and present state data.

[0141] In the case that the scene mode data includes both the object use history data and the history state data, the feature extraction parameters also include two types, corresponding to the object use history data and the history state data respectively. Specifically, the feature extraction parameters include the history use feature extraction parameters and the state feature extraction parameters, the history use feature extraction parameters are used to extract the history use features from the object use history data, and the state feature extraction parameters are used to extract the corresponding state features from the history state data or the current state data. Among them, in the training stage of the model, the scene mode data includes the history state data, and the state features corresponding to the history state data are the history state features, in the application stage of the model, the scene mode data includes the current state data, and the state features corresponding to the current state data are the current state features.

[0142] In the embodiment, through different feature extraction parameters corresponding to different scene mode data, the corresponding extraction of the features corresponding to various scene mode data can be realized, so as to obtain more accurate scene mode features.

[0143] In an embodiment, the training samples include positive samples and negative samples. The positive sample is a sample corresponding to a correct result, and the negative sample is a sample corresponding to an incorrect result. From the label carried by the sample data, the positive sample can be a sample with a probability of 1 for predicting a recommended result, and the negative sample can be a sample with a probability of 0 for predicting a recommended result. Obtaining the training samples includes:

[0144] Based on the history state corresponding to each piece of data in the object use history data, the matching relationship between the object and the history state is determined; based on the matching object and the matching history state with the matching relationship, the positive sample is constructed, the history state data in the positive sample is the history state data corresponding to the matching history state, and the object feature in the positive sample is the object feature corresponding to the matching object; based on the unmatched object and the unmatched history state without the matching relationship, the negative sample is constructed, the history state data in the negative sample is the history state data corresponding to the unmatched history state, and the object feature of the negative sample is the object feature corresponding to the unmatched object.

[0145] In the case that the scene mode data includes both the object use history data and the history state data, the training of the model is for the same user, and the user description features and the object use history data can be relatively stable, so fixed data can be selected, but the history state data and the object data of each use of the object can be different, and in the prediction process, the most important thing is the matching between the object and the state, therefore, the positive and negative samples need to be constructed based on whether the object and the history state have the matching relationship.

[0146] The matching relationship between the object and the historical state can be obtained based on a historical state corresponding to each piece of data in the object use history data. Each piece of data in the object use history data includes corresponding state data at the time of use, and the state data includes time information and regional attributes.

[0147] Based on the matching object and the matching historical state having the matching relationship, a positive sample is constructed by using historical state data corresponding to the matching historical state, object features corresponding to the matching object, and the determined user description features and the object use history data, wherein the probability of the predicted recommendation result corresponding to the positive sample is 1.

[0148] Based on the non-matching object and the non-matching historical state not having the matching relationship, a negative sample is constructed by using historical state data corresponding to the non-matching historical state, object features corresponding to the non-matching object, and the determined user description features and the object use history data, wherein the probability of the predicted recommendation result corresponding to the negative sample is 0.

[0149] In this embodiment, the matching relationship between the object and the state can be accurately expressed by constructing the positive and negative samples, which facilitates taking the matching relationship as an important reference in the model training process, so as to obtain accurate and reliable parameters after the model training is completed.

[0150] In one embodiment, the candidate recommendation object includes an object associated with the current state data and an object corresponding to an object identifier recorded in the object use history data.

[0151] The object associated with the current state data refers to an object having a strong correlation with the current state. For example, most users in a hospital use a small program or a public number corresponding to the hospital, wherein the regional attribute of the current geographic location is the hospital, which can be used as a kind of current state data, and the small program or the public number corresponding to the hospital is an object strongly associated with the hospital. For another example, most users in a subway station use a small program for scanning the code to enter the station. The regional attribute of the current geographic location is the subway station, which can be used as a kind of current state data, and the small program for scanning the code to enter the station is an object strongly associated with the subway station.

[0152] The object corresponding to the object identifier recorded in the object use history data refers to an object used by the user. By taking the object associated with the current state data and the object used by the user as the candidate recommendation object, the candidate recommendation object can be obtained by preliminary screening from a large number of objects, which can not only realize preliminary screening of the object, reduce the range of objects to be predicted, and reduce the data processing amount, but also can obtain the object that the user may need based on the current state and the historical use habit of the user, thereby facilitating accurate recommendation of the object.

[0153] In one embodiment, the method further includes a determination process of the object corresponding to the object identifier associated with the current state data. Specifically, the determination process includes:

[0154] The object usage record includes state data corresponding to each object used by the user and object identification of each object. An association relationship between the state data and the object identification that meets the association requirement is established based on co-occurrence data of the state data and the object identification in the object usage data. The object corresponding to the object identification associated with the current state data is determined based on the association relationship.

[0155] The object usage record includes each usage record of each object by each user. The object usage record includes state data corresponding to each object used by the user and object identification of each object.

[0156] Co-occurrence refers to the simultaneous occurrence of multiple data. Since each object usage record includes state data and object identification, the co-occurrence data can be obtained by counting the number of co-occurrences of each group of state data and object identification in the object usage data. The co-occurrence data can be the ratio of the number of co-occurrences to the total number of object usage data, or a calculation result obtained by other setting algorithms. The association requirement is a filtering condition for the co-occurrence data. Taking the ratio of the co-occurrence data as an example, the association requirement can be that the ratio is greater than a set threshold. Based on the association requirement, the state data and the object identification that meet the association requirement are filtered out, and an association relationship between the state data and the object identification is established.

[0157] In determining the object corresponding to the object identification associated with the current state data, only the state data corresponding to the current state data needs to be found. Based on the association relationship between the state data and the object identification, the object identification associated with the current state data can be obtained, so as to determine the object corresponding to the object identification associated with the current state data.

[0158] In this embodiment, the association relationship between the state data and the object identification is determined based on the co-occurrence data of the state data and the object identification, so as to realize the association between the state data and the object, and facilitate accurate determination of the object corresponding to the current state data.

[0159] In one embodiment, the method for recommending an object further includes a data processing process of obtaining user description features and object features. Specifically, it includes:

[0160] Acquire corpus data, which includes sequentially connected identifier pairs. Each identifier pair includes a user identifier and an object identifier. Adjacent user identifiers and object identifiers in the corpus data indicate that the user corresponding to the adjacent user identifier has a usage behavior of the object corresponding to the adjacent object identifier. Perform word vector embedding processing on the corpus data to obtain the user embedding vector corresponding to each user identifier and the object embedding vector corresponding to each object identifier in the corpus data. Use the user embedding vector corresponding to the target user identifier as the user description feature corresponding to the target user identifier, and use the object embedding vector corresponding to the candidate recommendation object as the object feature corresponding to the candidate recommendation object.

[0161] Corpus data refers to data used for word vector embedding processing. Corpus data can be used to describe the relationship between each vector and each object in the entire user and object set.

[0162] The corpus data consists of sequentially concatenated identifier pairs, each pair containing a user identifier and an object identifier. For example, a user identifier can be represented as U. i To represent, an object identifier can be represented using A i To represent it, for example, taking one piece of corpus data as U1-A1-U5-A6-U8-A 10 -U 12 For example, to facilitate the description of the connection relationship between user identifiers and object identifiers, we will use sequentially connected identifier pairs for illustration. Specifically, for corpus data, by splitting every two adjacent user identifiers and object identifiers, we can obtain multiple identifier pairs, such as U1-A1, U5-A6, U8-A1, etc. 10 Viewed as identifier pairs, 0-U1, A1-U5, A6-U8, A 10 -U 12 Treat it as a pair of labels.

[0163] Furthermore, adjacent user identifiers and object identifiers in the corpus data indicate that the user corresponding to the adjacent user identifier has a usage behavior on the object corresponding to the adjacent object identifier, for example, U1-A1-U5-A6-U8-A 10 -U 12 In the example, user 1 has a usage behavior on object 1, user 5 has a usage behavior on object 1 and object 6, user 8 has a usage behavior on object 6 and object 10, and user 12 has a usage behavior on object 10.

[0164] The word vector embedding refers to representing each word in the corpus data as a vector, that is, converting the corpus data into a numerical tensor. For example, [-0.91, 2, 1.8, -0.82, 0.65,...], which can be understood as embedding the word into a mathematical space. The word embedding processing refers to the vector formed based on the network weight obtained by training through the neural network. The network weight obtained by training can be implemented based on the word2vec (word to vector) algorithm or the doc2vec algorithm.

[0165] The computer device performs word vector embedding processing on the corpus data. Since the corpus data covers each vector in the full amount of users and each object in the full amount of objects, the user embedding vector corresponding to each user identifier in the corpus data and the object embedding vector corresponding to each object identifier in the corpus data can be obtained, so as to realize the characteristic representation of each user and each object.

[0166] In the embodiment, by performing word embedding processing on the corpus data based on the user identifier and the object identifier, the relationship between the user and the object can be fused and embedded and represented by a vector to obtain the user embedding vector and the object embedding vector, and the accurate representation of the user description feature and the object feature is realized.

[0167] In one embodiment, the corpus data is obtained, including: obtaining a relationship network constructed based on the user identifier and the object identifier, the nodes in the relationship network representing the user identifier and the object identifier, and the node edges in the relationship network representing the use behavior of the user corresponding to the user identifier on the object corresponding to the object identifier; taking each user identifier corresponding node as a starting point, randomly walking in the relationship network according to the node edge, and obtaining the corpus data based on the walking track.

[0168] Further, the relationship network constructed based on the user identifier and the object identifier is as follows Figure 6As shown, the relationship network includes nodes and node edges, where the nodes represent the user identifiers U and the object identifiers A. For example, if the total number of users is 1000 and the total number of objects is 2000, the relationship network includes 3000 nodes, that is, each user and each object corresponds to a node. The node edges represent the use behaviors of the user identifiers corresponding to users on the object identifiers corresponding to objects, in other words, the node corresponding to a user identifier can only have a node edge with the node corresponding to an object identifier, and the node edge exists only when the user identifier corresponding user uses the object identifier corresponding object. There is no direct node edge between any two nodes corresponding to user identifiers or any two nodes corresponding to object identifiers. Based on the user identifiers, the object identifiers, and the use behaviors of the user identifiers corresponding to users on the object identifiers corresponding to objects, the constructed relationship network can clearly reflect the relationship between users and objects.

[0169] Random walk refers to the process of walking from one node corresponding to a node edge to another node corresponding to the node edge. The starting point of the random walk can be a node corresponding to an object identifier or a node corresponding to a user identifier. Taking the node corresponding to each user identifier as the starting point as an example, starting from each user identifier corresponding node in the relationship network, a plurality of walking tracks are obtained by random walk, and each track includes a predetermined number of nodes. Figure 7 As shown, the walking track with the user identifier U1 as the starting point. Specifically, for each node corresponding to a user identifier, 80 walking tracks are obtained by random walk, each track includes 100 nodes (50 of which are nodes corresponding to user identifiers and 50 of which are nodes corresponding to object identifiers), and the user identifier and the object identifier corresponding to each walking track constitute a piece of corpus data. If there are 1000 users in the relationship network, the number of corpus data is 1000 x 80 = 8000.

[0170] In this embodiment, since the number of objects used by the user is much smaller than the number of users corresponding to the objects, taking each node corresponding to a user identifier as the starting point can more accurately represent the relationship between each user and the objects. The use behaviors of the user identifiers corresponding to users on the object identifiers corresponding to objects are represented by node edges, and the corpus data obtained by random walk according to the node edges can more clearly and accurately express the relationship between users and objects, facilitating obtaining accurate user embedding vectors and object embedding vectors.

[0171] In one embodiment, obtaining a relationship network based on user identifiers and object identifiers includes: obtaining a behavior matrix in which user identifiers and object identifiers represent rows and columns respectively; determining target elements in the behavior matrix that take the value of a target value, where the target value indicates that the user corresponding to the user identifier has a usage behavior on the object corresponding to the object identifier; and using the user identifiers and object identifiers represented by the rows and columns of the behavior matrix as nodes, and determining the node connections based on the row and column positions of the target elements in the behavior matrix to obtain the relationship network.

[0172] The behavior matrix, in this context, represents user behavior when interacting with objects. The rows and columns of the behavior matrix are determined based on user identifiers and object identifiers. For example, assuming N user identifiers and M object identifiers, ... Figure 8 As shown, the number of rows and columns of the row matrix A can both be N+M, that is, the dimension of the row matrix A is (N+M, N+M). For example, the number of rows of the row matrix A is N and the number of columns is M, that is, the dimension of the row matrix A is (N, M). Or the number of rows of the row matrix A is M and the number of columns is N, that is, the dimension of the row matrix A is (M, N).

[0173] The values ​​of the elements in the behavior matrix include target values ​​representing the user's behavior towards the object corresponding to the user identifier. For example, assuming user 1 has a behavior towards object 3, if the dimension of behavior matrix A is (N+M, N+M), then the value of the element in the 1st row and N+3rd column and the value of the element in the N+3rd row and 1st column are both target values. If the dimension of behavior matrix A is (N, M), then the value of the element in the 1st row and 3rd column is the target value.

[0174] Furthermore, the elements in the behavior matrix can take the values ​​of a first element and a second element, where the target value can be the first element. In a specific application, such as... Figure 8 As shown, the first element can be 1, and the second element can be 0. Using a 0-1 matrix to represent user behavior towards objects simplifies the data and reduces the storage space it occupies.

[0175] In this embodiment, using a behavior matrix to represent user behavior towards objects simplifies complex relationships. By using target elements in the behavior matrix that take the target value, and using the user identifier and object identifier represented by the rows and columns of the behavior matrix as nodes, and determining the node connections based on the row and column positions of the target elements in the behavior matrix, the relationship network can be obtained quickly and directly.

[0176] In one embodiment, according to the predicted recommendation results, a target recommendation object is selected from the candidate recommendation objects, and the target recommendation object is recommended based on the target user identifier, including:

[0177] The target recommendation objects are filtered from the candidate recommendation objects according to the number of the recommendation objects and a predicted recommendation result, and the number of the recommendation objects is the same as the number of displayable recommendation objects of an object display page of a terminal corresponding to the target user identifier; and the target recommendation objects are pushed to the terminal corresponding to the target user identifier, so that the terminal displays the target recommendation objects on the object display page.

[0178] The number of the recommendation objects refers to the number of the target recommendation objects that need to be filtered from the candidate recommendation objects. The number of the recommendation objects is the same as the number of displayable recommendation objects of an object display page of a terminal corresponding to the target user identifier. For example, if the number of displayable recommendation objects of the object display page of the terminal is 4, the number of the target recommendation objects filtered is also 4. The number of displayable recommendation objects of the object display page of the terminal can be determined according to a page configuration parameter, for example, according to the size of a displayed icon and the number of displayable icons per row.

[0179] By displaying the target recommendation objects that meet the number of the recommendation objects on the object display page of the terminal, multi-object recommendation can be achieved, which facilitates user selection and improves user experience.

[0180] In one embodiment, as shown in Figure 9 a method for recommending objects is provided, including the following steps:

[0181] In step 902, a behavior matrix represented by a user identifier and an object identifier is obtained, and a target element with a target value in the behavior matrix is determined, and the target value indicates that a user corresponding to the user identifier has a use behavior on an object corresponding to the object identifier.

[0182] In step 904, the user identifier and the object identifier represented by the row and the column of the behavior matrix are taken as nodes, and a node edge is determined based on the row and the column position of the target element in the behavior matrix, and a relationship network is obtained.

[0183] In step 906, a random walk is performed in the relationship network from each node corresponding to a user identifier as a starting point according to the node edge, and corpus data is obtained based on a walk track.

[0184] In step 908, word vector embedding processing is performed on the corpus data, and a user embedding vector corresponding to each user identifier in the corpus data and an object embedding vector corresponding to each object identifier in the corpus data are obtained.

[0185] In step 910, a user embedding vector corresponding to a target user identifier is taken as a user description feature corresponding to the target user identifier, and an object embedding vector corresponding to a candidate recommendation object is taken as an object feature corresponding to the candidate recommendation object.

[0186] Step 912, obtaining object use history data corresponding to the target user identifier, determining a matching relationship between the object and a historical state based on a historical state corresponding to each piece of data in the object use history data.

[0187] Step 914, based on the matching object and the matching historical state with the matching relationship, constructing a positive sample based on the user description feature, the object use history data, the historical state data corresponding to the matching historical state, and the object feature corresponding to the matching object.

[0188] Step 916, based on the non-matching object and the non-matching historical state without the matching relationship, constructing a negative sample based on the user description feature, the object use history data, the historical state data corresponding to the non-matching historical state, and the object feature corresponding to the non-matching object.

[0189] Step 918, training based on the initial parameters according to the positive sample and the negative sample until the prediction result based on the trained parameters meets the training end condition, and the parameters include the historical use feature extraction parameter, the state feature extraction parameter, and the mapping parameter.

[0190] Step 920, when the parent application running on the terminal corresponding to the target user identifier enters the sub-application access page, obtaining the geographical position information collected by the terminal.

[0191] Step 922, obtaining the corresponding state time based on the geographical position information, and determining the regional attribute of the region corresponding to the geographical position information to obtain the present situation data including the state time and the regional attribute, and obtaining the object use history data corresponding to the target user identifier.

[0192] Step 924, determining the element position and element value of the present situation data in the feature matrix corresponding to the scene type based on the scene type corresponding to the present situation data, to obtain the present situation feature matrix;

[0193] Step 926, performing feature mapping processing on the present situation feature matrix based on the state feature extraction parameter to obtain the present situation feature vector.

[0194] Step 928, obtaining the object use history data sequence sorted according to time based on the time corresponding to the object use history data.

[0195] Step 930, performing time sequence feature processing on the object use history data sequence according to the historical use feature extraction parameter to obtain the time sequence feature vector corresponding to the object use history data.

[0196] Step 932, obtaining the state data corresponding to each object used by the user and the object identifier of each object in the object use record.

[0197] Step 934, based on the state data and the co-occurrence data of the object identifier in the object use data, an association relationship between the state data and the object identifier is established, which meets the association requirement of the co-occurrence data.

[0198] Step 936, based on the object associated with the current state data, and the object corresponding to the object identifier recorded in the object use history data determined based on the association relationship, a candidate recommended object is determined.

[0199] Step 938, for each candidate recommended object, a multilayer perception mapping process is performed based on the mapping parameters according to the user description feature corresponding to the target user identifier, the time sequence feature vector, the current state feature vector, and the object feature corresponding to the candidate recommended object, to obtain a predicted recommendation result of each candidate recommended object in the scene mode.

[0200] Step 940, based on the number of recommended objects and the predicted recommendation result, the target recommended object meeting the number is selected from the candidate recommended object, and the number of recommended objects is the same as the number of displayable recommended objects of the object display page of the terminal corresponding to the target user identifier.

[0201] Step 942, the recommendation information of the target recommended sub-application is pushed to the terminal corresponding to the target user identifier, so that the terminal performs at least one of the following processes: preloading the target recommended sub-application and displaying the target recommended sub-application in the sub-application access page.

[0202] The application also provides an application scenario applying the recommended object method described above. Specifically, the recommended object method is applied in the application scenario as follows:

[0203] Figure 10 The figure shows the prediction principle of each candidate recommended applet. Each candidate applet and user is taken as a combination to obtain the user embedding vector and applet embedding vector, the current state feature corresponding to the current state discrete information, and the time sequence feature corresponding to the user historical use applet calculated by LSTM, and then the four corresponding features are input into a multilayer perceptron (MLP) to obtain a recommendation result. The following will introduce the data sources of the four corresponding features respectively.

[0204] First, the user embedding vector and the applet embedding vector.

[0205] By analyzing the use behavior of the applet of all network users in a recent period of time (such as the last month), the user applet relationship network is obtained, and then a network embedding method is used to assign an embedding variable to each user and applet. Specifically, by defining the behavior matrix A of the user and the applet, the dimension of A is (N+M, N+M), where the number of users is N and the number of applets is M. If the i-th user and the j-th applet have interaction behavior, then A i,j = 1 and A j,i = 1 A j，i = 1, otherwise A i,j = 0 and A j,i = 0 A i，j = 0. Where A is defined according to the actual click behavior of the user and the applet in the recent period of time. After obtaining the user applet relationship network, the vector representation of each user and applet is obtained by the network embedding method. Common network embedding methods include DeepWalk and Node2Vec node embedding methods. Taking the DeepWalk node embedding method as an example, based on the user applet relationship network, starting from each user node in the relationship network, according to the node edges in the relationship network, randomly walking multiple (such as 80) trajectories, each trajectory including 100 nodes. Then, all the trajectories walked are input into the word2vec word vector embedding algorithm as corpus data. Through the word2vec algorithm, the feature vector of each user and each applet is obtained. Wherein, the feature vectors of nodes close to each other in the relationship network are similar.

[0206] Then the acquisition of the current situation features and the time sequence features. Specifically includes:

[0207] The current situation features refer to expressing the current situation of the user in a feature form. Typical examples include the POI information of the user's current block, the current time information, the user's cycle information, and the user's recent use information. Among them, the information can be divided into two categories according to its form:

[0208] One is the current situation discrete information, such as POI information, time features, etc. This kind of information is discrete, such as the day of the week, the time of day, and the current regional attribute information (such as shopping malls, hospitals, subway stations, airports, and train stations). These information are discrete information, which are mapped to a certain row of the feature mapping matrix through a feature mapping module to obtain the corresponding current situation vector. The feature mapping module will be trained in the training module to obtain the state feature extraction parameters, so as to map based on the state feature extraction parameters to obtain better description of POI information, time information, etc.

[0209] Another type is the time sequence embedding information: such as periodic applet usage information and recent applet usage information, the information of sequence attribute needs to be constructed and modeled by a sequence model. Typically, the user's historical 30-day applet usage information is input into an LSTM time sequence neural network to obtain time sequence features. The LSTM time sequence neural network will be trained in the training module to obtain historical feature extraction parameters, and then the state feature extraction parameters are processed to obtain time sequence features.

[0210] It should be noted that the training module will train the feature mapping module, the LSTM time sequence neural network and the multi-layer perceptron at the same time during training. The training process can be supervised training, including positive and negative samples. The positive sample can be the actual historical data of the user, and the negative sample can be a sample based on the user not using a certain applet in a certain scene, so as to obtain better parameters.

[0211] Taking the method applied to WeChat as an example, in Figure 11 the left graph, when the user scrolls down the applet interface, the user's current data is obtained, and based on the data processing process shown in Figure 10 , the recommendation result of each candidate applet is obtained. Then, the applets in the candidate set are sorted and filtered to obtain the recommended applets, and the recommended applets are displayed on the user terminal, as shown in the right graph of Figure 11 .

[0212] It should be understood that although each step in the flowcharts of Figure 2 and Figure 9 is displayed in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified in this document, the execution of these steps has no strict order restriction, and these steps can be executed in other orders. Moreover, Figure 2 and Figure 9 at least some of the steps can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with other steps or steps or stages in other steps.

[0213] In one embodiment, as shown in Figure 12 , a device 1200 for recommending an object is provided, which can be a part of a computer device in the form of a software module or a hardware module, or a combination of the two. The device specifically includes: a scene mode feature acquisition module 1202, a predicted recommendation result analysis module 1204, and a target recommendation object filtering module 1206, wherein:

[0214] The scene mode feature acquisition module 1202 is configured to acquire scene mode features extracted from scene mode data matched with the target user identifier, wherein the scene mode data is used to describe a scene mode corresponding to the target user identifier.

[0215] The predicted recommendation result analysis module 1204 is configured to, for each candidate recommendation object, perform a multi-layer perception mapping process based on the user description features corresponding to the target user identifier, the scene mode features, and the object features corresponding to the candidate recommendation object, to obtain a predicted recommendation result of each candidate recommendation object in the scene mode.

[0216] The target recommendation object screening module 1206 is configured to screen a target recommendation object from the candidate recommendation objects according to the predicted recommendation result, and recommend the target recommendation object based on the target user identifier.

[0217] In an embodiment, the scene mode data includes at least one of present situation data corresponding to the target user identifier and object use history data corresponding to the target user identifier, and the scene mode features include at least one of present situation features extracted from the present situation data and historical use features extracted from the object use history data.

[0218] In an embodiment, the present situation data includes a geographical attribute and a state time corresponding to the geographical attribute, and the device for recommending an object further includes a geographical position information acquisition module and a state time and geographical attribute determination module, wherein the geographical position information acquisition module is configured to acquire geographical position information collected by a terminal corresponding to the target user identifier when responding to a user interaction operation, and the state time and geographical attribute determination module is configured to obtain a corresponding state time based on the geographical position information and determine a geographical attribute of a region corresponding to the geographical position information.

[0219] In an embodiment, the target recommendation object includes a target recommendation sub-application used to run in a running environment of a parent application, and the geographical position information acquisition module is further configured to acquire geographical position information collected by the terminal when the parent application run by the terminal corresponding to the target user identifier enters a sub-application access page, and the target recommendation object screening module is further configured to push recommendation information of the target recommendation sub-application to the terminal corresponding to the target user identifier, so that the terminal performs at least one of preloading the target recommendation sub-application and displaying the target recommendation sub-application on the sub-application access page.

[0220] In an embodiment, the present situation features include a present situation feature vector, and the device for recommending an object further includes a present situation feature vector acquisition module configured to determine an element position and an element value of the present situation data in a feature matrix corresponding to a scene type of the present situation data based on the scene type of the present situation data, to obtain a present situation feature matrix, and to perform feature mapping processing on the present situation feature matrix to obtain the present situation feature vector.

[0221] In an embodiment, the historical use features include a time sequence feature vector; the device for recommending an object further includes a time sequence feature vector obtaining module configured to perform time sequence feature processing on the sequence of the object use history data sorted according to time based on the time corresponding to the object use history data, to obtain a time sequence feature vector corresponding to the object use history data.

[0222] In an embodiment, the device for recommending an object further includes a training sample obtaining module and a training module.

[0223] The training sample obtaining module is configured to obtain a training sample, the training sample including a user description feature corresponding to a target user identifier, an object feature, and scene mode data; the training module is configured to perform training based on an initial feature extraction parameter and an initial mapping parameter according to the training sample, until a prediction result obtained based on a trained feature extraction parameter and a trained mapping parameter meets a training end condition; the trained feature extraction parameter is configured to extract a scene mode feature from the scene mode data; and the prediction result analysis module is further configured to perform a multi-layer perception mapping processing based on the trained mapping parameter.

[0224] In an embodiment, the scene mode data in the training sample includes object use history data and historical state data; the feature extraction parameter includes a historical use feature extraction parameter for extracting historical use features from the object use history data, and a state feature extraction parameter for extracting state features from state data, the state data including the historical state data and present state data.

[0225] In an embodiment, the training sample includes positive samples and negative samples; the training sample obtaining module further includes a matching relationship determining module, a positive sample constructing module, and a negative sample constructing module.

[0226] The matching relationship determining module is configured to determine a matching relationship between an object and a historical state based on a historical state corresponding to each piece of data in the object use history data; the positive sample constructing module is configured to construct a positive sample based on a matching object and a matching historical state that have the matching relationship, the historical state data in the positive sample being historical state data corresponding to the matching historical state, and the object feature in the positive sample being an object feature corresponding to the matching object; and the negative sample constructing module is configured to construct a negative sample based on a non-matching object and a non-matching historical state that do not have the matching relationship, the historical state data in the negative sample being historical state data corresponding to the non-matching historical state, and the object feature of the negative sample being an object feature corresponding to the non-matching object.

[0227] In an embodiment, the candidate recommended objects include an object associated with the present state data and an object corresponding to an object identifier recorded in the object use history data.

[0228] In one embodiment, the apparatus for recommending an object further comprises an object usage record obtaining module, an association relationship establishing module, and an object determining module.

[0229] The object usage record obtaining module is configured to obtain object usage records, the object usage records comprising corresponding state data when each object is used by a user and object identifiers of the objects; the association relationship establishing module is configured to establish an association relationship between state data and object identifiers based on co-occurrence data in the object usage data, the co-occurrence data meeting association requirements; and the object determining module is configured to determine an object corresponding to an object identifier associated with the state data based on the association relationship.

[0230] In one embodiment, the apparatus for recommending an object further comprises a corpus data obtaining module, a word vector embedding processing module, and a user association feature determining module.

[0231] The corpus data obtaining module is configured to obtain corpus data, the corpus data comprising sequentially connected identifier pairs, the identifier pairs comprising user identifiers and object identifiers, adjacent user identifiers and object identifiers in the corpus data indicating that a user corresponding to the adjacent user identifier has a usage behavior on an object corresponding to the adjacent object identifier; the word vector embedding processing module is configured to perform word vector embedding processing on the corpus data to obtain a user embedding vector corresponding to each user identifier in the corpus data and an object embedding vector corresponding to each object identifier in the corpus data; and the user association feature determining module is configured to take a user embedding vector corresponding to a target user identifier as a user description feature corresponding to the target user identifier, and take an object embedding vector corresponding to a candidate recommended object as an object feature corresponding to the candidate recommended object.

[0232] In one embodiment, the corpus data obtaining module comprises a relationship network obtaining module and a random walk module. The relationship network obtaining module is configured to obtain a relationship network constructed based on user identifiers and object identifiers, nodes in the relationship network representing the user identifiers and the object identifiers, and node edges in the relationship network representing usage behaviors of users corresponding to the user identifiers on objects corresponding to the object identifiers; and the random walk module is configured to perform random walk in the relationship network from a node corresponding to each user identifier as a starting point, and obtain corpus data based on a walk track.

[0233] In one embodiment, the relationship network obtaining module comprises a behavior matrix obtaining module and a relationship network constructing module.

[0234] The behavior matrix obtaining module is configured to obtain a behavior matrix represented by user identifiers and object identifiers, determine a target element in the behavior matrix with a target value, and the target value indicates that the user corresponding to the user identifier has a use behavior on the object corresponding to the object identifier.

[0235] In one embodiment, the target recommendation object screening module includes a quantity determining module and a pushing module. The quantity determining module is configured to screen target recommendation objects from the candidate recommendation objects according to the quantity of the recommendation objects and the predicted recommendation result, and the quantity of the recommendation objects is the same as the number of the displayable recommendation objects of the object display page of the terminal corresponding to the target user identifier. The pushing module is configured to push the target recommendation objects to the terminal corresponding to the target user identifier, so that the terminal displays the target recommendation objects on the object display page.

[0236] The specific limitations of the device for recommendation objects can refer to the limitations of the method for recommendation objects in the foregoing, and will not be described here. Each module in the device for recommendation objects described above can be realized by software, hardware, and combinations thereof, in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0237] In one embodiment, a computer device is provided, which can be a server, and an internal structure diagram thereof can be as shown in Figure 13 The computer device includes a processor, a memory, and a network interface connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store user description features, object features, historical object use data, and model parameters. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a method for recommendation objects.

[0238] Those skilled in the art can understand that Figure 13 The structure shown in the foregoing is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the diagram, or combine certain components, or have a different arrangement of components.

[0239] In an embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the above-mentioned method embodiments when executing the computer program.

[0240] In an embodiment, a computer readable storage medium is provided, storing a computer program, the computer program being executed by a processor to implement the steps in the above-mentioned method embodiments.

[0241] In an embodiment, a computer program product or computer program is provided, including computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the steps in the above-mentioned method embodiments.

[0242] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0243] Any combination of the technical features in the above embodiments can be made. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0244] The above embodiments only express several implementation ways of the present application, and the description is specific and detailed, but it should not be understood as a limitation to the patent scope of the application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method of recommending an object, characterized by, The method comprises: obtaining a scene mode feature extracted from scene mode data matched with a target user identifier; the scene mode data is used for describing a scene mode corresponding to the target user identifier; the scene mode data comprises present situation data corresponding to the target user identifier and object use history data corresponding to the target user identifier; the scene mode feature comprises a present situation feature vector extracted from the present situation data and a history use feature extracted from the object use history data; the present situation feature vector is obtained by performing feature mapping processing on a present situation feature matrix, the present situation feature matrix comprising a feature matrix corresponding to a week, a feature matrix corresponding to a time point in a day and a feature matrix corresponding to a regional attribute; for each candidate recommended object, performing multi-perception mapping processing based on a user description feature corresponding to the target user identifier, the scene mode feature and an object feature corresponding to the candidate recommended object, to obtain a predicted recommendation result of each candidate recommended object in the scene mode; the candidate recommended object comprises an object associated with the present situation data and an object corresponding to an object identifier recorded in the object use history data; the object comprises a sub-application used for running in a running environment of a mother application; according to the predicted recommendation result, screening a target recommended sub-application from the candidate recommended object, and recommending the target recommended sub-application based on the target user identifier, so that a terminal corresponding to the target user identifier preloads the target recommended sub-application.

2. The method of claim 1, wherein, The present situation data comprises a regional attribute and a state time corresponding to the regional attribute; the method further comprises: obtaining geographical position information collected by a terminal corresponding to the target user identifier when responding to a user interactive operation; obtaining a corresponding state time based on the geographical position information, and determining a regional attribute of a region corresponding to the geographical position information.

3. The method of claim 2, wherein, The method further comprises: when a mother application running on the terminal corresponding to the target user identifier enters a sub-application access page, obtaining geographical position information collected by the terminal; The method further comprises: pushing recommendation information of the target recommended sub-application to the terminal corresponding to the target user identifier, so that the terminal preloads the target recommended sub-application.

4. The method of claim 1, wherein, The method further comprises: based on a scene type corresponding to the present situation data, determining an element position and an element value of the present situation data in a feature matrix corresponding to the scene type, to obtain a present situation feature matrix; performing feature mapping processing on the present situation feature matrix to obtain a present situation feature vector.

5. The method of claim 1, wherein, The method further comprises: based on a time corresponding to the object use history data, performing time sequence feature processing on an object use history data sequence obtained by sorting the object use history data according to the time, to obtain a time sequence feature vector corresponding to the object use history data.

6. The method of claim 1, wherein, The method further comprises: obtaining a training sample, the training sample comprising user description features corresponding to the target user identifier, object features, and scene pattern data; training based on initial feature extraction parameters and initial mapping parameters according to the training sample until a prediction result obtained based on trained feature extraction parameters and trained mapping parameters meets a training end condition; the trained feature extraction parameters being used to extract the scene pattern features from the scene pattern data; the multi-layer perception-based mapping processing comprises: performing multi-layer perception-based mapping processing based on the trained mapping parameters.

7. The method of claim 6, wherein, The scene pattern data in the training sample comprises object use history data and historical state data. The feature extraction parameters comprise historical use feature extraction parameters for extracting historical use features from the object use history data, and state feature extraction parameters for extracting state features from state data, the state data comprising historical state data and current state data.

8. The method of claim 7, wherein, The training sample comprises positive samples and negative samples. The obtaining of the training sample comprises: determining a matching relationship between an object and a historical state based on a historical state corresponding to each piece of data in the object use history data; constructing a positive sample based on a matching object and a matching historical state that have the matching relationship, the historical state data in the positive sample being historical state data corresponding to the matching historical state, and the object features in the positive sample being object features corresponding to the matching object; constructing a negative sample based on a non-matching object and a non-matching historical state that do not have the matching relationship, the historical state data in the negative sample being historical state data corresponding to the non-matching historical state, and the object features in the negative sample being object features corresponding to the non-matching object.

9. The method of claim 1, wherein, The method further comprises: obtaining corpus data, the corpus data comprising sequentially connected identifier pairs, each identifier pair comprising a user identifier and an object identifier, adjacent user identifiers and object identifiers in the corpus data indicating that a user corresponding to the adjacent user identifier has a use behavior on an object corresponding to the adjacent object identifier; performing word vector embedding processing on the corpus data to obtain a user embedding vector corresponding to each user identifier in the corpus data and an object embedding vector corresponding to each object identifier in the corpus data; taking a user embedding vector corresponding to the target user identifier as a user description feature corresponding to the target user identifier, and taking an object embedding vector corresponding to the candidate recommended object as an object feature corresponding to the candidate recommended object.

10. The method of claim 9, wherein, The obtaining of the corpus data comprises: obtaining a relationship network constructed based on user identifiers and object identifiers, nodes in the relationship network representing user identifiers and object identifiers, and edges between nodes in the relationship network representing use behaviors of users corresponding to the user identifiers on objects corresponding to the object identifiers; starting from a node corresponding to each user identifier, performing random walk in the relationship network according to the edges between the nodes, and obtaining the corpus data based on a walk track.

11. The method of claim 10, wherein, The obtaining of the relationship network constructed based on user identifiers and object identifiers comprises: An action matrix represented by user identifiers and object identifiers is obtained, target elements in the action matrix with a target value are determined, and the target value represents that a user corresponding to the user identifier has a use behavior on an object corresponding to the object identifier; A relationship network is obtained by taking the user identifiers and the object identifiers represented by the rows and columns of the action matrix as nodes and determining node edges based on the positions of the target elements in the action matrix.

12. An apparatus for recommending an object, the apparatus comprising: The apparatus comprises: A scene mode feature acquisition module is configured to obtain scene mode features extracted from scene mode data matched with a target user identifier; the scene mode data is used to describe a scene mode corresponding to the target user identifier; the scene mode data comprises present status data corresponding to the target user identifier and object use history data corresponding to the target user identifier; the scene mode features comprise a present status feature vector extracted from the present status data and a historical use feature extracted from the object use history data; the present status feature vector is obtained by performing feature mapping processing on a present status feature matrix; the present status feature matrix comprises a feature matrix corresponding to a week, a feature matrix corresponding to a time point in a day, and a feature matrix corresponding to a regional attribute; A prediction recommendation result analysis module is configured to, for each candidate recommendation object, perform multi-layer perception mapping processing based on a user description feature corresponding to the target user identifier, the scene mode features, and an object feature corresponding to the candidate recommendation object to obtain a prediction recommendation result of each candidate recommendation object in the scene mode; the candidate recommendation object comprises an object associated with the present status data and an object corresponding to an object identifier recorded in the object use history data; the object comprises a sub-application used to run in a running environment of a parent application; A target recommendation sub-application screening module is configured to screen a target recommendation sub-application from the candidate recommendation object according to the prediction recommendation result, and recommend the target recommendation sub-application based on the target user identifier, so that a terminal corresponding to the target user identifier preloads the target recommendation sub-application.

13. The apparatus of claim 12, wherein, The present status data comprises a regional attribute and a state time corresponding to the regional attribute; the apparatus further comprises: A geographic location information acquisition module is configured to obtain geographic location information collected by a terminal corresponding to the target user identifier when responding to a user interaction operation; A state time and regional attribute determination module is configured to obtain a corresponding state time based on the geographic location information, and determine a regional attribute of a region corresponding to the geographic location information.

14. The apparatus of claim 13, wherein, The geographic location information acquisition module is further configured to obtain geographic location information collected by a terminal when a parent application running on the terminal enters a sub-application access page; The target recommendation sub-application screening module is further configured to push recommendation information of the target recommendation sub-application to a terminal corresponding to the target user identifier, so that the terminal preloads the target recommendation sub-application.

15. The apparatus of claim 12, wherein, The apparatus further comprises: The present situation feature matrix determination module is configured to determine an element position and an element value of the present situation data in a feature matrix corresponding to a scene type corresponding to the present situation data, to obtain a present situation feature matrix. The present situation feature vector obtaining module is configured to perform feature mapping processing on the present situation feature matrix, to obtain a present situation feature vector.

16. The apparatus of claim 12, wherein, The device further comprises: The time sequence feature vector obtaining module is configured to perform time sequence feature processing on the object use history data sequence obtained by sorting according to the time, to obtain a time sequence feature vector corresponding to the object use history data based on the time corresponding to the object use history data.

17. The apparatus of claim 12, wherein, The device further comprises: The training sample obtaining module is configured to obtain a training sample, the training sample comprising a user description feature, an object feature, and scene mode data corresponding to the target user identifier; The training module is configured to perform training based on an initial feature extraction parameter and an initial mapping parameter according to the training sample, until a prediction result obtained based on a trained feature extraction parameter and a trained mapping parameter meets a training end condition; the trained feature extraction parameter is used to extract the scene mode feature from the scene mode data; The prediction recommendation result analysis module is further configured to perform multi-layer perception mapping processing based on the trained mapping parameter.

18. The apparatus of claim 17, wherein, The scene mode data in the training sample comprises object use history data and historical state data; The feature extraction parameter comprises a historical use feature extraction parameter used to extract a historical use feature from the object use history data, and a state feature extraction parameter used to extract a state feature from the state data, the state data comprising the historical state data and the present situation data.

19. The apparatus of claim 18, wherein, The training sample comprises positive samples and negative samples; The training sample obtaining module further comprises: The matching relationship determination module is configured to determine a matching relationship between an object and a historical state based on a historical state corresponding to each piece of data in the object use history data; The positive sample construction module is configured to construct a positive sample based on a matching object and a matching historical state having the matching relationship, the historical state data in the positive sample being historical state data corresponding to the matching historical state, and the object feature in the positive sample being an object feature corresponding to the matching object; The negative sample construction module is configured to construct a negative sample based on a non-matching object and a non-matching historical state not having the matching relationship, the historical state data in the negative sample being historical state data corresponding to the non-matching historical state, and the object feature of the negative sample being an object feature corresponding to the non-matching object.

20. The apparatus of claim 12, wherein, The device further comprises: The corpus data obtaining module is configured to obtain corpus data, the corpus data comprising sequentially connected identifier pairs, the identifier pair comprising a user identifier and an object identifier, and adjacent user identifiers and object identifiers in the corpus data indicating that a user corresponding to the adjacent user identifier has a use behavior on an object corresponding to the adjacent object identifier. The word vector embedding processing module is configured to perform word vector embedding processing on the corpus data to obtain a user embedding vector corresponding to each user identifier in the corpus data and an object embedding vector corresponding to each object identifier in the corpus data. The user association feature determination module is configured to take the user embedding vector corresponding to the target user identifier as a user description feature corresponding to the target user identifier, and take the object embedding vector corresponding to the candidate recommended object as an object feature corresponding to the candidate recommended object.

21. The apparatus of claim 20, wherein, The corpus data obtaining module comprises: The relationship network obtaining module is configured to obtain a relationship network constructed based on user identifiers and object identifiers, wherein nodes in the relationship network represent user identifiers and object identifiers, and edges between nodes in the relationship network represent usage behaviors of users corresponding to the user identifiers on objects corresponding to the object identifiers. The random walk module is configured to perform random walk in the relationship network from a node corresponding to each user identifier as a starting point according to the edges between the nodes, and obtain the corpus data based on a walk track.

22. The apparatus of claim 21, wherein, The relationship network obtaining module comprises: The behavior matrix obtaining module is configured to obtain a behavior matrix with user identifiers and object identifiers representing rows and columns respectively, determine a target element in the behavior matrix with a target value, and the target value represents that a user corresponding to the user identifier has a usage behavior on an object corresponding to the object identifier. The relationship network constructing module is configured to take user identifiers and object identifiers represented by rows and columns of the behavior matrix as nodes, determine edges between the nodes based on positions of the target element in the rows and columns of the behavior matrix, and obtain the relationship network. 23.A computer device, comprising a memory and a processor, wherein the memory stores a computer program. The processor executes the computer program to implement the method in any one of claims 1 to 11.

24. A computer readable storage medium storing a computer program, wherein the computer program comprises instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-23. The computer program is executed by the processor to implement the method in any one of claims 1 to 11.

25. A computer program product comprising computer instructions, characterized in that, The computer instructions are executed by the processor to implement the method in any one of claims 1 to 11.

Images