WO2023065640A1

WO2023065640A1 - Model parameter adjustment method and apparatus, electronic device and storage medium

Info

Publication number: WO2023065640A1
Application number: PCT/CN2022/090461
Authority: WO
Inventors: 李雷来; 王健宗; 瞿晓阳
Original assignee: 平安科技（深圳）有限公司
Priority date: 2021-10-22
Filing date: 2022-04-29
Publication date: 2023-04-27
Also published as: CN113779423A

Abstract

The present application relates to intelligent decision making, and discloses a model parameter adjustment method and apparatus, an electronic device and a storage medium. The method comprises: acquiring first social information, second social information, and third social information, the first social information and the second social information belonging to the same category, and the first social information and the third social information belonging to different categories; separately encoding the first social information, the second social information, and the third social information, to obtain a first feature vector, a second feature vector, and a third feature vector; determining a distance between the first feature vector and the second feature vector, to obtain a first distance; determining a distance between the first feature vector and the third feature vector, to obtain a second distance; determining a loss function on the basis of a difference between the first distance and the second distance; and adjusting a model parameter of a recommendation model according to the loss function, so as to train the recommendation model. By means of implementing embodiments of the present application, generalization ability of a recommendation model is improved.

Description

A model parameter adjustment method, device, electronic equipment and storage medium

priority statement

This application claims the priority of the Chinese patent application with the application number 202111231066.8 submitted to the China Patent Office on October 22, 2021, and the title of the invention is "A Model Parameter Adjustment Method, Device, Electronic Equipment, and Storage Medium", the entire content of which Incorporated in this application by reference.

technical field

The present application relates to the field of artificial intelligence, and in particular to a model parameter adjustment method, device, electronic equipment and storage medium.

Background technique

In today's society, social behavior has almost completely revolved around the Internet, such as blog posts, social status, hot comments and other social information flows are often generated from a large number of social events. Generally speaking, this social information flow has the characteristics of timing, large quantity, fast update, and high complexity. However, at the current stage, social information is often used for training the recommendation model, and the trained recommendation model is used to recommend social information for users. Because only the social information is mechanically input into the recommendation model during training, there may be a problem that the social information recommendation is not accurate enough when using the trained recommendation model to recommend social information for users. In other words, the generalization ability of this recommendation model is poor. Therefore, the inventor realizes that how to improve the generalization ability of the recommendation model has become an urgent technical problem to be solved at the current stage.

Contents of the invention

Embodiments of the present application provide a model parameter adjustment method, device, electronic device, and storage medium, which improve the generalization capability of the recommendation model.

The first aspect of the present application provides a method for adjusting model parameters, including:

Acquiring first social information, second social information and third social information, where the first social information and the second social information belong to the same category, and the first social information and the third social information belong to different categories;

Encoding the first social information, the second social information and the third social information respectively to obtain a first feature vector, a second feature vector and a third feature vector;

determining a distance between the first eigenvector and the second eigenvector to obtain a first distance;

determining a distance between the first eigenvector and the third eigenvector to obtain a second distance;

determining a loss function based on the difference between the first distance and the second distance;

Adjusting model parameters of the recommendation model according to the loss function to train the recommendation model.

The second aspect of the present application provides a model parameter adjustment device, the device includes an acquisition module, an encoding module, a first determination module, a second determination module, a third determination module and a training module,

The acquisition module is configured to acquire first social information, second social information and third social information, the first social information and the second social information belong to the same category, the first social information and the second social information 3. Social information belongs to different categories;

The encoding module is configured to encode the first social information, the second social information and the third social information respectively to obtain a first feature vector, a second feature vector and a third feature vector;

The first determination module is configured to determine a distance between the first feature vector and the second feature vector to obtain a first distance;

The second determination module is configured to determine a distance between the first feature vector and the third feature vector to obtain a second distance;

The third determination module is configured to determine a loss function according to the difference between the first distance and the second distance;

The training module is configured to adjust model parameters of the recommendation model according to the loss function, so as to train the recommendation model.

The third aspect of the present application provides an electronic device for model parameter adjustment, which includes a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory, and are generated to be executed by the processor to perform the following steps:

A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, and the stored computer program is executed by the processor to implement the following steps:

It can be seen that in the above technical solution, by determining the first distance between the feature vectors corresponding to the social information belonging to the same category, and determining the second distance between the feature vectors corresponding to the social information belonging to different categories, and then can be based on The difference between the first distance and the second distance determines the loss function. Because the loss function is determined according to the first distance between the feature vectors corresponding to the social information belonging to the same category and the second distance between the feature vectors corresponding to the social information belonging to different categories, so when using this loss function to adjust the recommendation model The model parameters can make the representation of similar data more abundant, which in turn enhances the feature extraction ability of the recommendation model and improves the generalization ability of the recommendation model.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present application. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

in:

Fig. 1 is a schematic flow chart of a model parameter adjustment method provided by the embodiment of the present application;

FIG. 2 is a schematic diagram of a heterogeneous social graph provided by an embodiment of the present application;

FIG. 3 is a schematic diagram of a homogeneous social graph obtained based on the heterogeneous social graph shown in FIG. 2;

Fig. 4 is a schematic flowchart of another model parameter adjustment method provided by the embodiment of the present application;

FIG. 5 is a schematic diagram of a model parameter adjustment device provided in an embodiment of the present application;

FIG. 6 is a schematic structural diagram of an electronic device in a hardware operating environment involved in an embodiment of the present application.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the application with reference to the drawings in the embodiments of the application. Apparently, the described embodiments are only some of the embodiments of the application, not all of them. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the scope of protection of this application.

Each will be described in detail below.

The terms "first" and "second" in the specification and claims of the present application and the above drawings are used to distinguish different objects, rather than to describe a specific order. Furthermore, the terms "include" and "have", as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes unlisted steps or units, or optionally further includes For other steps or units inherent in these processes, methods, products or apparatuses.

The following describes the embodiments of the present application with reference to the accompanying drawings. It can be understood that in the present application, the execution subject may be an electronic device or a cloud server, which is not limited here. Among them, electronic devices may include various handheld devices with wireless communication functions, vehicle-mounted devices, wearable devices, computing devices or other processing devices connected to wireless modems, as well as various forms of user equipment (User Equipment, UE), mobile Taiwan (Mobile Station, MS), terminal equipment (terminal device) and so on.

Referring to FIG. 1 , FIG. 1 is a schematic flowchart of a method for adjusting model parameters provided by an embodiment of the present application. As shown in Figure 1, the method includes:

101. Acquire first social information, second social information, and third social information, where the first social information and the second social information belong to the same category, and the first social information and the third social information belong to different category.

Wherein, the first social information includes the first word text, the first tag information, the first named entity, the first user identifier, the first time information and the identifier of the first social information. The first word text can include one or more word texts, and one or more word texts are words other than common words and rare words. Common words can be modal particles, stop words, etc., and rare words can be included in open source The words in the data set or presets are not limited here. The first tag information is used to identify a topic or category to which the first social information belongs. The first user identifier is an identifier of a user who publishes the first social information. The first time information is the time when the first social information is published.

Wherein, the second social information includes the second word text, the second label information, the second named entity, the second user identifier, the second time information and the identifier of the second social information. The second word text may include two or more word texts. The second tag information is used to identify a topic or category to which the second social information belongs. The second user identifier is the identifier of the user who publishes the second social information. The second time information is the time when the second social information is published.

Wherein, the third social information includes third word text, third tag information, third named entity, third user identifier, third time information, and third social information identifier. The third word text may include three or more word texts. The third tag information is used to identify a topic or category to which the third social information belongs. The third user identifier is the identifier of the user who publishes the third social information. The third time information is the time when the third social information is published.

It should be noted that in this application, social information may be, for example, text information such as tweets and comments and/or image information, which is not limited here.

In addition, the first word text is obtained by natural language processing according to the first social information, the second word text is obtained by natural language processing according to the second social information, and the third word text is obtained by natural language processing according to the third social information. Do limit.

102. Encode the first social information, the second social information, and the third social information respectively to obtain a first feature vector, a second feature vector, and a third feature vector.

Optionally, before step 102, the method further includes: acquiring a heterogeneous social graph, where the heterogeneous social graph includes a plurality of heterogeneous nodes and at least two heterogeneous nodes among the plurality of heterogeneous nodes A heterogeneous node in the heterogeneous social graph includes the following items: word text, label information, user identification, time information and identification of social information, and the label information is used to identify the category to which the social information belongs ; Generate an isomorphic social graph according to the heterogeneous social graph, the isomorphic social graph includes a plurality of isomorphic nodes and connection edges between at least two of the plurality of isomorphic nodes, the An isomorphic node in the isomorphic social graph is an identifier of social information, and the plurality of isomorphic nodes include the identifier of the first social information, the identifier of the second social information, and the identifier of the third social information ; According to the isomorphic social graph, determine the first weight and the second weight, the first weight is determined according to the connection edge between the identification of the first social information and the identification of the second social information, the The second weight is determined according to the connection edge between the identifier of the first social information and the identifier of the third social information; if the first weight is higher than the first threshold, determine the first social information and the third social information The second social information belongs to the same category; if the second weight is lower than a second threshold, it is determined that the first social information and the third social information belong to different categories.

Wherein, the connection edge between at least two heterogeneous nodes may be a connection edge between at least two heterogeneous nodes of the same type, and/or a connection edge between at least two heterogeneous nodes of different types.

For example, refer to FIG. 2 , which is a schematic diagram of a heterogeneous social graph provided by an embodiment of the present application. As shown in Figure 2, a heterogeneous node (identification of social information) can have connection edges with other nodes, for example, a heterogeneous node (identification of social information) can be connected with word text, label information, user identification, time There are connection edges between nodes such as information, and there may be a connection edge between a heterogeneous node (identification of social information) and an identification of another social information. It can be understood that a heterogeneous node (identification of social information) may have connection edges with nodes such as word text, label information, user identification, and time information, that is, connection edges between at least two heterogeneous nodes of different types ; There may also be a connection edge between a heterogeneous node (identification of social information) and another identification of social information, that is, a connection edge between at least two heterogeneous nodes of the same type.

Wherein, the connection edge between at least two isomorphic nodes may be a connection edge between at least two isomorphic nodes of the same type.

For example, refer to FIG. 3 , which is a schematic diagram of a homogeneous social graph obtained based on the heterogeneous social graph shown in FIG. 2 . As shown in Figure 3, there is a connecting edge between every two social information identifiers (every two isomorphic nodes) among the three social information identifiers (three isomorphic nodes), that is, at least two homogeneous nodes of the same type connect edges between nodes.

Wherein, the first threshold may be the same as or different from the second threshold, which is not limited here. For example, the first threshold is higher than the second threshold.

Optionally, generating a homogeneous social graph according to the heterogeneous social graph includes: mapping the heterogeneous social graph into a homogeneous social graph based on a heterogeneous information network (heterogeneous information networks, HIN) mapping rule.

Wherein, the HIN mapping rule includes one or more of the following: if the word text connected with the identifier of the social information D in the heterogeneous social graph is connected with the word text connected with the identifier of the social information E in the heterogeneous social graph If the similarity is greater than or equal to the third threshold, connect the identification of social information D and the identification of social information E in the homogeneous social graph; if the tag information connected to the identification of social information D in the heterogeneous social graph is connected to If the tag information connected to the logo of social information E in the heterogeneous social graph is the same, then the logo of social information D is connected to the logo of social information E in the homogeneous social graph; if the social information D in the heterogeneous social graph The user ID connected with the ID of is the same as the user ID connected with the ID of social information E in the heterogeneous social graph, then the ID of social information D is connected with the ID of social information E in the homogeneous social graph; if The time information connected with the logo of social information D in the heterogeneous social graph is the same as the time information connected with the logo of social information E in the heterogeneous social graph, then the logo of social information D and The identification of social information E is connected; if the difference between the time information connected with the identification of social information D in the heterogeneous social graph and the time information connected with the identification of social information E in the heterogeneous social graph is less than or equal to the first Four thresholds, connect the identification of social information D and the identification of social information E in the isomorphic social graph.

Wherein, the third threshold is different from the fourth threshold, for example, the third threshold is greater than the fourth threshold, or the third threshold is smaller than the fourth threshold.

It can be seen that in the above technical solution, the isomorphic social graph is obtained based on the heterogeneous social graph, so that the obtained isomorphic social graph is more in line with the actual situation. At the same time, by determining the first weight and the second weight according to the isomorphic social graph, whether different social information belongs to the same category can be determined according to the two weights, thereby improving the accuracy of category determination.

Optionally, before the acquisition of the heterogeneous social graph, the method further includes: acquiring multiple pieces of social information within a preset time; extracting word text and tags contained in each piece of social information among the multiple pieces of social information identification of information, user identification, time information and social information; generating the heterogeneous social graph according to the word text, tag information, user identification, time information and identification of social information contained in each piece of social information.

Wherein, the preset time can be set by an administrator, or configured in a configuration file, which is not limited here.

Wherein, multiple pieces of social information can be included in the same social information block, the number of the social information block is within the preset number range, the preset coding range can be 0 to t+w, and t is greater than or equal to 0 and less than w Integer, w is the length of the time window for maintaining the recommendation model, which can be set by the administrator or configured in the configuration file, and there is no limit here. It should be understood that in this application, the social information contained in the social information block within the preset encoding range is not outdated.

It can be understood that in this application, different social information blocks correspond to different numbers, and the size of the numbers is used to indicate the time sequence in which the social information blocks occur. In addition, the occurrence times of different social information in the same social information block may be different or the same, that is, the time information contained in different social information in the same social information block may be different or the same.

It can be seen that in the above technical solution, the generation of heterogeneous social graph is realized.

Optionally, the determining the first weight and the second weight according to the isomorphic social graph includes: if the connection edge between at least two homogeneous nodes in the isomorphic social graph is based on the heterogeneous social The word text associated with the identification of different social information in the figure is determined, and the word text associated with the identification of the first social information is determined according to the similarity between the word text associated with the identification of the second social information. The first weight; the second weight is determined according to the similarity between the word text associated with the identifier of the first social information and the word text associated with the identifier of the third social information.

Wherein, determining the first weight according to the similarity between the word text associated with the identifier of the first social information and the word text associated with the identifier of the second social information may include: according to the first The first weight is determined by a cosine similarity between the word text associated with the identifier of the social information and the word text associated with the identifier of the second social information.

Wherein, determining the second weight according to the similarity between the word text associated with the identifier of the first social information and the word text associated with the identifier of the third social information includes: according to the first social information The second weight is determined by a cosine similarity between the word text associated with the identifier of the information and the word text associated with the identifier of the third social information.

It can be seen that in the above technical solution, the weight is determined by the similarity between words and texts associated with different social information identifiers, so that it can be more accurately determined whether they belong to the same category according to the weight.

Optionally, the determining the first weight and the second weight according to the isomorphic social graph includes: if the connection edge between at least two homogeneous nodes in the isomorphic social graph is based on the heterogeneous social The label information associated with the identification of different social information in the figure is determined, and the label information associated with the identification of the first social information is determined according to the similarity between the label information associated with the identification of the second social information. The first weight; the second weight is determined according to the similarity between the tag information associated with the identifier of the first social information and the tag information associated with the identifier of the third social information.

Wherein, determining the second weight according to the similarity between the tag information associated with the identifier of the first social information and the tag information associated with the identifier of the third social information includes: according to the first social information The second weight is determined by a cosine similarity between tag information associated with the identifier of the information and tag information associated with the identifier of the third social information.

It can be seen that in the above technical solution, the weight is determined by the similarity between the tag information associated with different social information identifiers, so that it can be more accurately determined whether they belong to the same category according to the weight.

Optionally, the determining the first weight and the second weight according to the isomorphic social graph includes: if the connection edge between at least two homogeneous nodes in the isomorphic social graph is based on the heterogeneous social The time information associated with the identification of different social information in the figure is determined, then according to the difference between the time information associated with the identification of the first social information and the time information associated with the identification of the second social information, determine The first weight: determining the second weight according to the difference between the time information associated with the identifier of the first social information and the time information associated with the identifier of the third social information.

It can be seen that in the above technical solution, the weight is determined by the similarity between the time information associated with different social information identifiers, so that it can be more accurately determined whether they belong to the same category according to the weight.

103. Determine a distance between the first feature vector and the second feature vector to obtain a first distance.

104. Determine a distance between the first eigenvector and the third eigenvector to obtain a second distance.

105. Determine a loss function according to the difference between the first distance and the second distance.

106. Adjust model parameters of the recommendation model according to the loss function, so as to train the recommendation model.

Optionally, the model parameters of layer l when the input of the recommended model is m _i

The model parameters of the first l- 1 layer when the input of the recommendation model is m _j

related; wherein, l is an integer greater than or equal to 2; m _i is the first social information, m _j is the second social information; or, m _i is the first social information, m _j is the third social information.

It can be seen that in the above technical solution, since the input of the recommended model is m _i , the model parameters of the l-th layer

The model parameters of the first l-1 layer when the input of the recommendation model is m _j

Therefore, the model parameters of different layers in the recommendation model can be associated with each other, and the information contained in the model parameters can be enriched, thereby improving the generalization ability of the recommendation model.

optional,

satisfy the following formula:

Among them, heads means that the model parameters of the first l-1 layer are connected in series towards the head direction, N(m _j ) is the adjacency matrix of m _j ,

It is used to extract the model parameters of the first l-1 layer when the input of the recommendation model is m _j ,

It is used to aggregate the model parameters of the first l-1 layers extracted when the input of the recommendation model is m _j .

Optionally, the loss function ζ _t satisfies the following formula:

Among them, m _i is the first social information, m _i+ is the second social information, m _i- is the third social information,

is the first distance,

is the second distance, a is the regularization parameter, and T is the set formed by the combination of every three pieces of social information. In the combination, social information A and social information B belong to the same type, and in the combination, social information A and social information C is of different types.

Referring to FIG. 4 , FIG. 4 is a schematic flowchart of another method for adjusting model parameters provided by the embodiment of the present application. As shown in Figure 4, the method includes:

401. Acquire first social information, second social information, and third social information, where the first social information and the second social information belong to the same category, and the first social information and the third social information belong to different categories. category.

Wherein, step 401 is the same as step 101 in FIG. 1 , and will not be repeated here.

402. Obtain a heterogeneous social graph.

Wherein, for step 402, reference may be made to the relevant description of step 102 in FIG. 1 , and details are not repeated here.

403. Generate a homogeneous social graph according to the heterogeneous social graph.

Wherein, for step 403, reference may be made to the relevant description of step 102 in FIG. 1 , and details are not repeated here.

404. Determine a first weight and a second weight according to the isomorphic social graph.

Wherein, for step 404, reference may be made to the related description of step 102 in FIG. 1 , and details are not repeated here.

405. If the first weight is higher than a first threshold, determine that the first social information and the second social information belong to the same category.

Wherein, for step 405, reference may be made to the relevant description of step 102 in FIG. 1 , and details are not repeated here.

406. If the second weight is lower than a second threshold, determine that the first social information and the third social information belong to different categories.

Wherein, for step 406, reference may be made to the relevant description of step 102 in FIG. 1 , and details are not repeated here.

407. Encode the first social information, the second social information, and the third social information respectively to obtain a first feature vector, a second feature vector, and a third feature vector.

Wherein, step 407 is the same as step 102 in FIG. 1 , and will not be repeated here.

408. Determine a distance between the first feature vector and the second feature vector to obtain a first distance.

Wherein, step 408 is the same as step 103 in FIG. 1 , and will not be repeated here.

409. Determine a distance between the first feature vector and the third feature vector to obtain a second distance.

Wherein, step 409 is the same as step 104 in FIG. 1 , and will not be repeated here.

410. Determine a loss function according to the difference between the first distance and the second distance.

Wherein, step 410 is the same as step 105 in FIG. 1 , and will not be repeated here.

411. Adjust model parameters of the recommendation model according to the loss function, so as to train the recommendation model.

Wherein, step 411 is the same as step 106 in FIG. 1 , and will not be repeated here.

It can be seen that in the above technical solution, by determining the first distance between the feature vectors corresponding to the social information belonging to the same category, and determining the second distance between the feature vectors corresponding to the social information belonging to different categories, and then can be based on The difference between the first distance and the second distance determines the loss function. Because the loss function is determined according to the first distance between the feature vectors corresponding to the social information belonging to the same category and the second distance between the feature vectors corresponding to the social information belonging to different categories, so when using this loss function to adjust the recommendation model The model parameters can make the representation of similar data more abundant, which in turn enhances the feature extraction ability of the recommendation model and improves the generalization ability of the recommendation model. At the same time, by obtaining the isomorphic social graph based on the heterogeneous social graph, the obtained isomorphic social graph is more in line with the actual situation. At the same time, by determining the first weight and the second weight according to the isomorphic social graph, whether different social information belongs to the same category can be determined according to the two weights, thereby improving the accuracy of category determination.

Referring to FIG. 5 , FIG. 5 is a schematic diagram of a model parameter adjustment device provided in an embodiment of the present application. Wherein, as shown in FIG. 5 , a model parameter adjustment device 500 provided in the embodiment of the present application includes an acquisition module 501, an encoding module 502, a first determination module 503, a second determination module 504, a third determination module 505 and a training module 506,

The obtaining module 501 is configured to obtain first social information, second social information and third social information, the first social information and the second social information belong to the same category, the first social information and the The third social information belongs to different categories; the encoding module 502 is configured to encode the first social information, the second social information and the third social information respectively to obtain the first feature vector, the second feature vector and the third eigenvector; the first determining module 503 is used to determine the distance between the first eigenvector and the second eigenvector to obtain the first distance; the second determining module 504 uses To determine the distance between the first eigenvector and the third eigenvector to obtain a second distance; the third determination module 505 is configured to obtain a second distance according to the distance between the first distance and the second distance The difference is to determine a loss function; the training module 506 is configured to adjust model parameters of the recommendation model according to the loss function, so as to train the recommendation model.

Optionally, the model parameter adjustment apparatus 500 further includes a generation module 507 and an acquisition module 501, further configured to acquire a heterogeneous social graph, where the heterogeneous social graph includes multiple heterogeneous nodes and at least A connection edge between two heterogeneous nodes, a heterogeneous node in the heterogeneous social graph includes the following items: word text, label information, user identification, time information and social information identification, and the label information is used To identify the category to which the social information belongs; the generating module 507 is configured to generate an isomorphic social graph according to the heterogeneous social graph, the isomorphic social graph includes a plurality of isomorphic nodes and at least one of the plurality of isomorphic nodes A connection edge between two isomorphic nodes, one isomorphic node in the isomorphic social graph is an identifier of social information, and the plurality of isomorphic nodes include the identifier of the first social information, the second The identification of social information and the identification of the third social information; the first determination module 503 is further configured to determine a first weight and a second weight according to the isomorphic social graph, and the first weight is based on the first A connection edge between the identifier of the social information and the identifier of the second social information is determined, and the second weight is determined according to a connection edge between the identifier of the first social information and the identifier of the third social information; The first determining module 503 is further configured to determine that the first social information and the second social information belong to the same category if the first weight is higher than the first threshold; the first determining module 503 is further configured to determine if If the second weight is lower than a second threshold, it is determined that the first social information and the third social information belong to different categories.

Optionally, the model parameter adjustment device 500 also includes an extraction module 508, the acquisition module 501 is also used to acquire multiple pieces of social information within a preset time; the extraction module 508 is also used to extract each of the multiple pieces of social information Identification of word text, tag information, user ID, time information and social information contained in social information; generating module 507 is also used to generate word text, tag information, user ID, time information and social Information identification, generating the heterogeneous social graph.

Optionally, when determining the first weight and the second weight according to the isomorphic social graph, the first determining module 503 is configured to: if the connection edge between at least two isomorphic nodes in the isomorphic social graph Determined according to the word text associated with different social information identifiers in the heterogeneous social graph, then according to the difference between the word text associated with the first social information identifier and the word text associated with the second social information identifier The first weight is determined according to the similarity between them; the second weight is determined according to the similarity between the word text associated with the identifier of the first social information and the word text associated with the identifier of the third social information .

Relevant; wherein, l is an integer greater than or equal to 2; m _i is the first social information, m _j is the second social information; or, m _i is the first social information, m _j is the third social information.

optional,

satisfy the following formula:

Optionally, the loss function ζ _t satisfies the following formula:

is the first distance,

Referring to FIG. 6 , FIG. 6 is a schematic structural diagram of an electronic device of a hardware operating environment involved in an embodiment of the present application.

An embodiment of the present application provides an electronic device for model parameter adjustment, including a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured Executed by the processor to execute instructions comprising the steps in any one of the model parameter adjustment methods. Wherein, as shown in FIG. 6, the electronic equipment of the hardware operating environment involved in the embodiment of the present application may include:

Processor 601, such as a CPU.

The storage 602, optionally, the storage may be a high-speed RAM storage, or a stable storage, such as a disk storage.

The communication interface 603 is configured to realize connection and communication between the processor 601 and the memory 602 .

Those skilled in the art can understand that the structure of the electronic device shown in FIG. 6 is not limited thereto, and may include more or less components than shown in the figure, or combine some components, or arrange different components.

As shown in FIG. 6 , the memory 602 may include an operating system, a network communication module, and one or more programs. An operating system is a program that manages and controls the hardware and software resources of a server and supports the operation of one or more programs. The network communication module is used to realize the communication between various components inside the memory 602, and communicate with other hardware and software inside the electronic device.

In the electronic device shown in FIG. 6 , the processor 601 is used to execute one or more programs in the memory 602 to implement the following steps:

For the specific implementation of the electronic device involved in the present application, reference may be made to the various embodiments of the above-mentioned model parameter adjustment method, which will not be repeated here.

The present application also provides a computer-readable storage medium, the computer-readable storage medium is used to store a computer program, and the stored computer program is executed by the processor to implement the following steps:

For the specific implementation of the computer-readable storage medium involved in the present application, reference may be made to the various embodiments of the above-mentioned model parameter adjustment method, and details are not repeated here. The computer-readable storage medium may be non-volatile or volatile.

It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the described action sequence. limitations, as certain steps may be performed in other orders or simultaneously depending on the application. Secondly, those skilled in the art should also know that the embodiments described in the specification belong to preferred embodiments, and the actions and modules involved are not necessarily required by this application.

As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still understand the foregoing The technical solutions described in each embodiment are modified, or some of the technical features are replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the various embodiments of the application.

Claims

A method for adjusting model parameters, including:

Acquiring first social information, second social information and third social information, where the first social information and the second social information belong to the same category, and the first social information and the third social information belong to different categories;

Encoding the first social information, the second social information and the third social information respectively to obtain a first feature vector, a second feature vector and a third feature vector;

determining a distance between the first eigenvector and the second eigenvector to obtain a first distance;

determining a distance between the first eigenvector and the third eigenvector to obtain a second distance;

determining a loss function based on the difference between the first distance and the second distance;

Adjusting model parameters of the recommendation model according to the loss function to train the recommendation model.
The method according to claim 1, wherein, in the encoding of the first social information, the second social information and the third social information respectively, the first feature vector, the second feature vector and the Before the third eigenvector, the method also includes:

Obtain a heterogeneous social graph, the heterogeneous social graph includes a plurality of heterogeneous nodes and connection edges between at least two heterogeneous nodes in the plurality of heterogeneous nodes, one heterogeneous social graph in the heterogeneous social graph The node includes the following items: word text, label information, user identification, time information and identification of social information, and the label information is used to identify the category to which the social information belongs;

Generate an isomorphic social graph according to the heterogeneous social graph, the isomorphic social graph includes a plurality of isomorphic nodes and connection edges between at least two isomorphic nodes among the plurality of isomorphic nodes, the homogeneous social graph includes An isomorphic node in the structural social graph is an identifier of social information, and the plurality of isomorphic nodes include an identifier of the first social information, an identifier of the second social information, and an identifier of the third social information;

According to the isomorphic social graph, determine a first weight and a second weight, the first weight is determined according to a connection edge between the identifier of the first social information and the identifier of the second social information, and the first weight is determined according to the connection edge between the identifier of the first social information and the identifier of the second social information The second weight is determined according to the connecting edge between the identifier of the first social information and the identifier of the third social information;

If the first weight is higher than a first threshold, then determining that the first social information and the second social information belong to the same category;

If the second weight is lower than a second threshold, it is determined that the first social information and the third social information belong to different categories.
The method according to claim 2, wherein, before said acquiring the heterogeneous social graph, said method further comprises:

Obtain multiple pieces of social information within a preset time;

Extracting the word text, label information, user identification, time information and social information identification contained in each social information in the plurality of social information;

The heterogeneous social graph is generated according to the word text, label information, user identification, time information and identification of social information contained in each piece of social information.
The method according to claim 2, wherein said determining the first weight and the second weight according to the isomorphic social graph comprises:

If the connection edge between at least two isomorphic nodes in the homogeneous social graph is determined according to the word text associated with the identifiers of different social information in the heterogeneous social graph, then according to the identifier of the first social information The similarity between the associated word text and the word text associated with the identifier of the second social information determines the first weight;

The second weight is determined according to the similarity between the word text associated with the identifier of the first social information and the word text associated with the identifier of the third social information.
The method according to claim 4, wherein the input of the recommended model is the model parameters of the l-th layer when mi
The model parameters of the first l-1 layer when the input of the recommendation model is m j
Relevant; wherein, l is an integer greater than or equal to 2; m i is the first social information, m j is the second social information; or, m i is the first social information, m j is the third social information.
The method according to claim 5, wherein,
satisfy the following formula:

Among them, heads means that the model parameters of the first l-1 layer are connected in series towards the head direction, N(m j ) is the adjacency matrix of m j ,
It is used to extract the model parameters of the first l-1 layer when the input of the recommendation model is m j ,
It is used to aggregate the model parameters of the first l-1 layers extracted when the input of the recommendation model is m j .
The method according to claim 4, wherein the loss function ζ t satisfies the following formula:

Among them, m i is the first social information, m i+ is the second social information, m i- is the third social information,
is the first distance,
is the second distance, a is the regularization parameter, and T is the set formed by the combination of every three pieces of social information. In the combination, social information A and social information B belong to the same type, and in the combination, social information A and social information C is of different types.
A model parameter adjustment device, wherein the device includes an acquisition module, an encoding module, a first determination module, a second determination module, a third determination module and a training module,

The acquisition module is configured to acquire first social information, second social information and third social information, the first social information and the second social information belong to the same category, the first social information and the second social information 3. Social information belongs to different categories;

The encoding module is configured to encode the first social information, the second social information and the third social information respectively to obtain a first feature vector, a second feature vector and a third feature vector;

The first determination module is configured to determine a distance between the first feature vector and the second feature vector to obtain a first distance;

The second determination module is configured to determine a distance between the first feature vector and the third feature vector to obtain a second distance;

The third determination module is configured to determine a loss function according to the difference between the first distance and the second distance;

The training module is configured to adjust model parameters of the recommendation model according to the loss function, so as to train the recommendation model.
An electronic device for model parameter adjustment, including a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and generated by the processing The controller executes to execute the instructions of the following steps:

Acquiring first social information, second social information and third social information, where the first social information and the second social information belong to the same category, and the first social information and the third social information belong to different categories;

Encoding the first social information, the second social information and the third social information respectively to obtain a first feature vector, a second feature vector and a third feature vector;

determining a distance between the first eigenvector and the second eigenvector to obtain a first distance;

determining a distance between the first eigenvector and the third eigenvector to obtain a second distance;

determining a loss function based on the difference between the first distance and the second distance;

Adjusting model parameters of the recommendation model according to the loss function to train the recommendation model.
The electronic device according to claim 9, wherein, in the encoding of the first social information, the second social information and the third social information respectively, a first feature vector and a second feature vector are obtained and before the third eigenvector, the steps also include:

Obtain a heterogeneous social graph, the heterogeneous social graph includes a plurality of heterogeneous nodes and connection edges between at least two heterogeneous nodes in the plurality of heterogeneous nodes, one heterogeneous social graph in the heterogeneous social graph The node includes the following items: word text, label information, user identification, time information and identification of social information, and the label information is used to identify the category to which the social information belongs;

Generate an isomorphic social graph according to the heterogeneous social graph, the isomorphic social graph includes a plurality of isomorphic nodes and connection edges between at least two isomorphic nodes among the plurality of isomorphic nodes, the homogeneous social graph includes An isomorphic node in the structural social graph is an identifier of social information, and the plurality of isomorphic nodes include an identifier of the first social information, an identifier of the second social information, and an identifier of the third social information;

According to the isomorphic social graph, determine a first weight and a second weight, the first weight is determined according to a connection edge between the identifier of the first social information and the identifier of the second social information, and the first weight is determined according to the connection edge between the identifier of the first social information and the identifier of the second social information The second weight is determined according to the connecting edge between the identifier of the first social information and the identifier of the third social information;

If the first weight is higher than a first threshold, then determining that the first social information and the second social information belong to the same category;

If the second weight is lower than a second threshold, it is determined that the first social information and the third social information belong to different categories.
The electronic device according to claim 10, wherein, before said acquiring the heterogeneous social graph, said step further comprises:

Obtain multiple pieces of social information within a preset time;

Extracting the word text, label information, user identification, time information and social information identification contained in each social information in the plurality of social information;

The heterogeneous social graph is generated according to the word text, label information, user identification, time information and identification of social information contained in each piece of social information.
The electronic device according to claim 10, wherein said determining the first weight and the second weight according to the isomorphic social graph comprises:

If the connection edge between at least two isomorphic nodes in the homogeneous social graph is determined according to the word text associated with the identifiers of different social information in the heterogeneous social graph, then according to the identifier of the first social information The similarity between the associated word text and the word text associated with the identifier of the second social information determines the first weight;

The second weight is determined according to the similarity between the word text associated with the identifier of the first social information and the word text associated with the identifier of the third social information.
The electronic device according to claim 12, wherein the input of the recommended model is the model parameter of the l-th layer when mi
The model parameters of the first l-1 layer when the input of the recommendation model is m j
Relevant; wherein, l is an integer greater than or equal to 2; m i is the first social information, m j is the second social information; or, m i is the first social information, m j is the third social information.
The electronic device according to claim 13, wherein,
satisfy the following formula:

Among them, heads means that the model parameters of the first l-1 layer are connected in series towards the head direction, N(m j ) is the adjacency matrix of m j ,
It is used to extract the model parameters of the first l-1 layer when the input of the recommendation model is m j ,
It is used to aggregate the model parameters of the first l-1 layers extracted when the input of the recommendation model is m j .
A computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, and the stored computer program is executed by the processor to implement the following steps:

Acquiring first social information, second social information and third social information, where the first social information and the second social information belong to the same category, and the first social information and the third social information belong to different categories;

Encoding the first social information, the second social information and the third social information respectively to obtain a first feature vector, a second feature vector and a third feature vector;

determining a distance between the first eigenvector and the second eigenvector to obtain a first distance;

determining a distance between the first eigenvector and the third eigenvector to obtain a second distance;

determining a loss function based on the difference between the first distance and the second distance;

Adjusting model parameters of the recommendation model according to the loss function to train the recommendation model.
The computer-readable storage medium according to claim 15, wherein, after encoding the first social information, the second social information and the third social information respectively, the first feature vector, the second Before the second eigenvector and the third eigenvector, the steps also include:

Obtain a heterogeneous social graph, the heterogeneous social graph includes a plurality of heterogeneous nodes and connection edges between at least two heterogeneous nodes in the plurality of heterogeneous nodes, one heterogeneous social graph in the heterogeneous social graph The node includes the following items: word text, label information, user identification, time information and identification of social information, and the label information is used to identify the category to which the social information belongs;

Generate an isomorphic social graph according to the heterogeneous social graph, the isomorphic social graph includes a plurality of isomorphic nodes and connection edges between at least two isomorphic nodes among the plurality of isomorphic nodes, the homogeneous social graph includes An isomorphic node in the structural social graph is an identifier of social information, and the plurality of isomorphic nodes include an identifier of the first social information, an identifier of the second social information, and an identifier of the third social information;

According to the isomorphic social graph, determine a first weight and a second weight, the first weight is determined according to a connection edge between the identifier of the first social information and the identifier of the second social information, and the first weight is determined according to the connection edge between the identifier of the first social information and the identifier of the second social information The second weight is determined according to the connecting edge between the identifier of the first social information and the identifier of the third social information;

If the first weight is higher than a first threshold, then determining that the first social information and the second social information belong to the same category;

If the second weight is lower than a second threshold, it is determined that the first social information and the third social information belong to different categories.
The computer-readable storage medium according to claim 16, wherein, before said obtaining a heterogeneous social graph, said steps further comprise:

Obtain multiple pieces of social information within a preset time;

Extracting the word text, label information, user identification, time information and social information identification contained in each social information in the plurality of social information;

The heterogeneous social graph is generated according to the word text, label information, user identification, time information and identification of social information contained in each piece of social information.
The computer-readable storage medium according to claim 16, wherein said determining the first weight and the second weight according to the isomorphic social graph comprises:

If the connection edge between at least two isomorphic nodes in the homogeneous social graph is determined according to the word text associated with the identifiers of different social information in the heterogeneous social graph, then according to the identifier of the first social information The similarity between the associated word text and the word text associated with the identifier of the second social information determines the first weight;

The second weight is determined according to the similarity between the word text associated with the identifier of the first social information and the word text associated with the identifier of the third social information.
The computer-readable storage medium according to claim 18, wherein the input of the recommended model is the model parameter of the lth layer when mi
The model parameters of the first l-1 layer when the input of the recommendation model is m j
Relevant; wherein, l is an integer greater than or equal to 2; m i is the first social information, m j is the second social information; or, m i is the first social information, m j is the third social information.
The computer readable storage medium of claim 19, wherein:
satisfy the following formula:

Among them, heads means that the model parameters of the first l-1 layer are connected in series towards the head direction, N(m j ) is the adjacency matrix of m j ,
It is used to extract the model parameters of the first l-1 layer when the input of the recommendation model is m j ,
It is used to aggregate the model parameters of the first l-1 layers extracted when the input of the recommendation model is m j .