Embedded representation management method and device

Abstract

An embedded representation management method and apparatus, the method comprising: in response to a first version number input by a user, loading a first embedded representation corresponding to the first version number from a disk into a memory (S410); training the first embedded representation according to preset training data to obtain a second embedded representation (S420); determining a second version number of the second embedded representation according to the first version number and the scene of the training data (S430); storing the second embedded representation and the second version number in a disk (S440); multi-version management of the embedded representation can be realized, so that the management cost of the embedded representation is reduced, and the management efficiency is improved.

Description

Embedded representation management method and device Technical Field

[0001] The present application relates to the field of information retrieval, and in particular to an embedded representation management method and device. Background Art

[0002] With the massive expansion of internet information, embeddings have gained widespread application in information retrieval. From text-based web news to illustrated products, to short videos carrying even richer information, all can be encoded into an embedding, ultimately playing a vital role in the ranking and recall processes of information retrieval. Therefore, embeddings are a crucial representation of internet information.

[0003] In related technologies, embedding representations are typically managed in a file-based format. However, with the increasing number of pre-training methods, scenarios, and data sizes, existing embedding representation management methods suffer from high management costs and poor efficiency. Therefore, a new embedding representation management method is urgently needed.

[0004] Summary of the Invention

[0005] In view of this, an embedded representation management method and device are proposed.

[0006] In a first aspect, an embodiment of the present application provides an embedded representation management method, the method comprising: in response to a first version number input by a user, loading a first embedded representation corresponding to the first version number from a disk into a memory; training the first embedded representation according to preset training data to obtain a second embedded representation; determining a second version number of the second embedded representation according to the first version number and the scenario of the training data; and storing the second embedded representation and the second version number to the disk.

[0007] An embodiment of the present application can, in response to a first version number input by a user, load a first embedded representation corresponding to the first version number from a disk into a memory, and then train the first embedded representation according to preset training data to obtain a second embedded representation, and then determine a second version number of the second embedded representation based on the first version number and the scenario of the training data, and store the second embedded representation and the second version number on the disk, thereby enabling multi-version management of the embedded representation, thereby reducing the management cost of the embedded representation and improving management efficiency.

[0008] According to the first aspect, in a first possible implementation of the embedding representation management method, the training of the first embedding representation to obtain the second embedding representation includes: in the process of training the first embedding representation, obtaining a first intermediate version of the first embedding representation according to a preset first time interval, the first time interval being at the level of days; when the latest obtained first intermediate version meets a preset first evaluation condition, storing the latest obtained first intermediate version to the disk; and when a preset training end condition is met, ending the training to obtain the second embedding representation.

[0009] In this embodiment, during the training of the first embedding representation, the first intermediate version of the first embedding representation can be obtained according to a preset first time interval (day level), and when the newly obtained first intermediate version meets the preset first evaluation condition, the newly obtained first intermediate version is stored to the disk; then, when the preset training end condition is met, the training is terminated to obtain the second embedding representation. In this way, during the training of the first embedding representation, the first intermediate version of the first embedding representation can be stored to the disk at the daily level, so that when an anomaly is found during the training process, it can be rolled back based on the stored daily intermediate version (full intermediate version), thereby reducing the operation and maintenance costs of the embedding representation development.

[0010] According to a first possible implementation manner of the first aspect, in a second possible implementation manner of the embedded representation management method, the training of the first embedded representation to obtain the second embedded representation includes: in the process of training the first embedded representation, obtaining a second intermediate version of the first embedded representation according to a preset second time interval, the second time interval being at the hour level or the minute level; determining a first difference between the latest obtained second intermediate version and the previous second intermediate version; and backing up the latest obtained second intermediate version to the memory when the first difference meets a preset second evaluation condition.

[0011] In this embodiment, during the training of the first embedded representation, a second intermediate version of the first embedded representation can be obtained according to a preset second time interval (hourly or minutely), and a first difference between the latest obtained second intermediate version and the previous second intermediate version can be determined; if the first difference satisfies a preset second evaluation condition, the latest obtained second intermediate version is backed up to the memory. In this way, during the training of the first embedded representation, an hourly / minutely backup of the second intermediate version of the first embedded representation can be achieved, so that when an anomaly is found during the training process, a version rollback can be performed based on the hourly / minutely backup version (incremental intermediate version), thereby reducing the operation and maintenance costs of the embedded representation development.

[0012] According to the second possible implementation manner of the first aspect, in a third possible implementation manner of the embedding representation management method, the training of the first embedding representation to obtain the second embedding representation includes: when the latest acquired first intermediate version does not meet the first evaluation condition, or when the first difference does not meet the second evaluation condition, according to a preset version fallback rule, selecting a fallback version from the second intermediate version in the memory or the first intermediate version on the disk; and continuing training based on the fallback version.

[0013] In this embodiment, if the newly acquired first intermediate version does not meet the first evaluation condition, or if the first difference does not meet the second evaluation condition, a fallback version can be selected from the second intermediate version in memory or the first intermediate version on disk according to a preset version fallback rule, and training can continue based on the fallback version. In this way, the training process of the first embedded representation can be dynamically monitored, and when an anomaly is found, the version can be automatically rolled back. This allows for automatic adjustments to anomalies or online disturbances in the embedded representation training process without manual intervention, thereby significantly reducing the operational and maintenance costs of embedded representation development.

[0014] According to a second possible implementation manner of the first aspect, in a fourth possible implementation manner of the embedding representation management method, the method further includes: loading a first neighbor graph corresponding to the first embedding representation from the disk into the memory; training the first embedding representation to obtain a second embedding representation, including: during the training of the first embedding representation, dynamically updating the first neighbor graph according to the second time interval to obtain a neighbor graph corresponding to each second intermediate version; after obtaining the second embedding representation, determining a second neighbor graph corresponding to the second embedding representation.

[0015] In this embodiment, during the training of the first embedding representation, the first neighbor graph can be dynamically updated according to the second time interval to obtain the neighbor graph corresponding to each second intermediate version, and after obtaining the second embedding representation, the second neighbor graph corresponding to the second embedding representation is determined, so that the first neighbor graph can be dynamically updated during the training of the first embedding representation, so as to perform version difference comparison based on the neighbor graph corresponding to the second intermediate version.

[0016] According to a fourth possible implementation manner of the first aspect, in a fifth possible implementation manner of the embedded representation management method, determining the first difference between the latest acquired second intermediate version and the previous second intermediate version includes: acquiring a third neighbor graph and a fourth neighbor graph, the third neighbor graph referring to the neighbor graph corresponding to the latest acquired second intermediate version, and the fourth neighbor graph referring to the neighbor graph corresponding to the previous second intermediate version; determining the changed nodes in the third neighbor graph based on the fourth neighbor graph; determining the neighbor change information and node change information of each changed node, the neighbor change information including at least one of the number of neighbor changes, the ratio of neighbor changes and the local neighbor similarity score, and the node change information including at least one of the node offset direction and the node offset distance; determining the first difference between the latest acquired second intermediate version and the previous second intermediate version based on the neighbor change information and node change information of each changed node.

[0017] In this embodiment, when determining the first difference, the third neighbor graph and the fourth neighbor graph can be obtained first, and then the changed nodes in the third neighbor graph can be determined based on the fourth neighbor graph, as well as the neighbor change information and node change information of each changed node can be determined. Then, based on the neighbor change information and node change information of each changed node, the first difference between the latest obtained second intermediate version and the previous second intermediate version can be determined, so that the first difference between the latest obtained second intermediate version and the previous second intermediate version can be determined quickly and accurately to improve processing efficiency.

[0018] According to the first aspect or any one of the first possible implementation manner of the first aspect to the fifth possible implementation manner of the first aspect, in a sixth possible implementation manner of the embedded representation management method, storing the second embedded representation and the second version number to the disk includes: storing the second embedded representation to the disk in a multi-level differential storage manner.

[0019] In this embodiment, the second embedded representation can be stored in a disk in a multi-level differential storage manner, thereby greatly saving storage resources and effectively reducing storage consumption.

[0020] According to the sixth possible implementation manner of the first aspect, in the seventh possible implementation manner of the embedded representation management method, storing the second embedded representation on the disk in a multi-level differential storage manner includes: determining a fifth embedded representation from the second embedded representation, the fifth embedded representation referring to an embedded representation in the second embedded representation that has changed relative to the first embedded representation; storing the fifth embedded representation on the disk; and establishing a storage mapping table corresponding to the second embedded representation based on a first address and a second address, the first address referring to an address on the disk of an embedded representation in the second embedded representation that has not changed relative to the first embedded representation, and the second address referring to an address on the disk of the fifth embedded representation.

[0021] In this embodiment, when the second embedded representation is stored in the disk in a multi-level differential storage manner, a fifth embedded representation can be first determined from the second embedded representation, where the fifth embedded representation refers to an embedded representation in the second embedded representation that has changed relative to the first embedded representation; the fifth embedded representation is then stored in the disk; thereafter, a storage mapping table corresponding to the second embedded representation can be established based on the first address and the second address, where the first address refers to the address in the disk of the embedded representation in the second embedded representation that has not changed relative to the first embedded representation, and the second address refers to the address in the disk of the fifth embedded representation, thereby enabling multi-level differential storage of the second embedded representation to be achieved, effectively saving disk storage resources.

[0022] According to the first aspect or any one of the first possible implementation manner of the first aspect to the seventh possible implementation manner of the first aspect, in an eighth possible implementation manner of the embedded representation management method, the method further includes: in response to a user's comparison request for the third embedded representation and the fourth embedded representation, performing dimensionality reduction processing on the third embedded representation and the fourth embedded representation, respectively, to obtain a first dimensionality reduction vector and a second dimensionality reduction vector; determining a second difference between the third embedded representation and the fourth embedded representation; and displaying the first dimensionality reduction vector, the second dimensionality reduction vector, and the second difference.

[0023] In this embodiment, in response to a user's comparison request for the third embedded representation and the fourth embedded representation, dimensionality reduction processing can be performed on the third embedded representation and the fourth embedded representation, respectively, to obtain a first reduced dimensionality vector and a second reduced dimensionality vector; at the same time, the second difference between the third embedded representation and the fourth embedded representation is determined, and then the first reduced dimensionality vector, the second reduced dimensionality vector and the second difference are displayed, thereby enabling visualization of the differences between different versions of the embedded representations.

[0024] In second aspect, an embodiment of the present application provides an embedded representation management device, which includes: a first loading module, which loads a first embedded representation corresponding to a first version number from a disk into a memory in response to a first version number input by a user; a training module, which is used to train the first embedded representation according to preset training data to obtain a second embedded representation; a version number determination module, which is used to determine the second version number of the second embedded representation according to the first version number and the scenario of the training data; and a storage module, which is used to store the second embedded representation and the second version number to the disk.

[0025] An embodiment of the present application can, in response to a first version number input by a user, load a first embedded representation corresponding to the first version number from a disk into a memory, and then train the first embedded representation according to preset training data to obtain a second embedded representation, and then determine a second version number of the second embedded representation based on the first version number and the scenario of the training data, and store the second embedded representation and the second version number on the disk, thereby enabling multi-version management of the embedded representation, thereby reducing the management cost of the embedded representation and improving management efficiency.

[0026] According to the second aspect, in a first possible implementation of the embedded representation management device, the training module includes: a first acquisition submodule, used to obtain a first intermediate version of the first embedded representation according to a preset first time interval during training of the first embedded representation, and the first time interval is at the day level; a first storage submodule, used to store the latest acquired first intermediate version to the disk if the latest acquired first intermediate version meets a preset first evaluation condition; and a training end submodule, used to end the training and obtain a second embedded representation if the preset training end condition is met.

[0027] In this embodiment, during the training of the first embedding representation, the first intermediate version of the first embedding representation can be obtained according to a preset first time interval (day level), and when the newly obtained first intermediate version meets the preset first evaluation condition, the newly obtained first intermediate version is stored to the disk; then, when the preset training end condition is met, the training is terminated to obtain the second embedding representation. In this way, during the training of the first embedding representation, the first intermediate version of the first embedding representation can be stored to the disk at the daily level, so that when an anomaly is found during the training process, it can be rolled back based on the stored daily intermediate version (full intermediate version), thereby reducing the operation and maintenance costs of the embedding representation development.

[0028] According to a first possible implementation manner of the second aspect, in a second possible implementation manner of the embedded representation management device, the training module includes: a second acquisition submodule, used to obtain a second intermediate version of the first embedded representation according to a preset second time interval during training of the first embedded representation, and the second time interval is in hours or minutes; a difference determination submodule, used to determine a first difference between the latest acquired second intermediate version and the previous second intermediate version; and a second storage submodule, used to back up the latest acquired second intermediate version to the memory when the first difference meets a preset second evaluation condition.

[0029] In this embodiment, during the training of the first embedded representation, a second intermediate version of the first embedded representation can be obtained according to a preset second time interval (hourly or minutely), and a first difference between the latest obtained second intermediate version and the previous second intermediate version can be determined; if the first difference satisfies a preset second evaluation condition, the latest obtained second intermediate version is backed up to the memory. In this way, during the training of the first embedded representation, an hourly / minutely backup of the second intermediate version of the first embedded representation can be achieved, so that when an anomaly is found during the training process, a version rollback can be performed based on the hourly / minutely backup version (incremental intermediate version), thereby reducing the operation and maintenance costs of the embedded representation development.

[0030] According to a second possible implementation manner of the second aspect, in a third possible implementation manner of the embedded representation management device, the training module includes: a fallback sub-module, which is used to select a fallback version from the second intermediate version in the memory or the first intermediate version on the disk according to a preset version fallback rule when the latest acquired first intermediate version does not meet the first evaluation condition, or when the first difference does not meet the second evaluation condition; and a training sub-module, which continues training based on the fallback version.

[0031] In this embodiment, if the newly acquired first intermediate version does not meet the first evaluation condition, or if the first difference does not meet the second evaluation condition, a fallback version can be selected from the second intermediate version in memory or the first intermediate version on disk according to a preset version fallback rule, and training can continue based on the fallback version. In this way, the training process of the first embedded representation can be dynamically monitored, and when an anomaly is found, the version can be automatically rolled back. This allows for automatic adjustments to anomalies or online disturbances in the embedded representation training process without manual intervention, thereby significantly reducing the operational and maintenance costs of embedded representation development.

[0032] According to a second possible implementation manner of the second aspect, in a fourth possible implementation manner of the embedded representation management device, the device further includes: a second loading module, used to load the first neighbor graph corresponding to the first embedded representation from the disk into the memory; the training module includes: a dynamic update submodule, used to dynamically update the first neighbor graph according to the second time interval during the training of the first embedded representation to obtain a neighbor graph corresponding to each second intermediate version; and a neighbor graph determination submodule, used to determine the second neighbor graph corresponding to the second embedded representation after obtaining the second embedded representation.

[0033] In this embodiment, during the training of the first embedding representation, the first neighbor graph can be dynamically updated according to the second time interval to obtain the neighbor graph corresponding to each second intermediate version, and after obtaining the second embedding representation, the second neighbor graph corresponding to the second embedding representation is determined, so that the first neighbor graph can be dynamically updated during the training of the first embedding representation, so as to perform version difference comparison based on the neighbor graph corresponding to the second intermediate version.

[0034] According to a fourth possible implementation manner of the second aspect, in a fifth possible implementation manner of the embedded representation management device, the difference determination submodule is used to: obtain a third neighbor graph and a fourth neighbor graph, the third neighbor graph refers to the neighbor graph corresponding to the latest obtained second intermediate version, and the fourth neighbor graph refers to the neighbor graph corresponding to the previous second intermediate version; determine the changed nodes in the third neighbor graph based on the fourth neighbor graph; determine the neighbor change information and node change information of each changed node, the neighbor change information includes at least one of the number of neighbor changes, the ratio of neighbor changes and the local neighbor similarity score, and the node change information includes at least one of the node offset direction and the node offset distance; determine the first difference between the latest obtained second intermediate version and the previous second intermediate version based on the neighbor change information and node change information of each changed node.

[0035] In this embodiment, when determining the first difference, the third neighbor graph and the fourth neighbor graph can be obtained first, and then the changed nodes in the third neighbor graph can be determined based on the fourth neighbor graph, as well as the neighbor change information and node change information of each changed node can be determined. Then, based on the neighbor change information and node change information of each changed node, the first difference between the latest obtained second intermediate version and the previous second intermediate version can be determined, so that the first difference between the latest obtained second intermediate version and the previous second intermediate version can be determined quickly and accurately to improve processing efficiency.

[0036] According to the second aspect or any one of the first possible implementation manner of the second aspect to the fifth possible implementation manner of the second aspect, in a sixth possible implementation manner of the embedded representation management device, the storage module includes: a differential storage submodule, used to store the second embedded representation to the disk in a multi-level differential storage manner.

[0037] In this embodiment, the second embedded representation can be stored in a disk in a multi-level differential storage manner, thereby greatly saving storage resources and effectively reducing storage consumption.

[0038] According to the sixth possible implementation manner of the second aspect, in the seventh possible implementation manner of the embedded representation management device, the differential storage submodule is used to: determine a fifth embedded representation from the second embedded representation, the fifth embedded representation referring to an embedded representation in the second embedded representation that has changed relative to the first embedded representation; store the fifth embedded representation on the disk; and establish a storage mapping table corresponding to the second embedded representation based on a first address and a second address, the first address referring to an address on the disk of an embedded representation in the second embedded representation that has not changed relative to the first embedded representation, and the second address referring to an address on the disk of the fifth embedded representation.

[0039] In this embodiment, when the second embedded representation is stored in the disk in a multi-level differential storage manner, a fifth embedded representation can be first determined from the second embedded representation, where the fifth embedded representation refers to an embedded representation in the second embedded representation that has changed relative to the first embedded representation; the fifth embedded representation is then stored in the disk; thereafter, a storage mapping table corresponding to the second embedded representation can be established based on the first address and the second address, where the first address refers to the address in the disk of the embedded representation in the second embedded representation that has not changed relative to the first embedded representation, and the second address refers to the address in the disk of the fifth embedded representation, thereby enabling multi-level differential storage of the second embedded representation to be achieved, effectively saving disk storage resources.

[0040] According to the second aspect or any one of the first possible implementation manner of the second aspect to the seventh possible implementation manner of the second aspect, in an eighth possible implementation manner of the embedded representation management device, the device further includes: a dimensionality reduction module, which, in response to a user's comparison request for the third embedded representation and the fourth embedded representation, performs dimensionality reduction processing on the third embedded representation and the fourth embedded representation, respectively, to obtain a first dimensionality reduction vector and a second dimensionality reduction vector; a difference determination module, which is used to determine the second difference between the third embedded representation and the fourth embedded representation; and a display module, which is used to display the first dimensionality reduction vector, the second dimensionality reduction vector and the second difference.

[0041] In this embodiment, in response to a user's comparison request for the third embedded representation and the fourth embedded representation, dimensionality reduction processing can be performed on the third embedded representation and the fourth embedded representation, respectively, to obtain a first reduced dimensionality vector and a second reduced dimensionality vector; at the same time, the second difference between the third embedded representation and the fourth embedded representation is determined, and then the first reduced dimensionality vector, the second reduced dimensionality vector and the second difference are displayed, thereby enabling visualization of the differences between different versions of the embedded representations.

[0042] In a third aspect, an embodiment of the present application provides an embedded representation management device, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the above-mentioned first aspect or one or more of the multiple possible implementation methods of the first aspect when executing the instructions.

[0043] In a fourth aspect, an embodiment of the present application provides a non-volatile computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, implement the embedded representation management method of the above-mentioned first aspect or one or more of the multiple possible implementation methods of the first aspect.

[0044] In the fifth aspect, an embodiment of the present application provides a computer program product, including a computer-readable code, or a non-volatile computer-readable storage medium carrying a computer-readable code. When the computer-readable code runs in an electronic device, the processor in the electronic device executes the embedded representation management method of the above-mentioned first aspect or one or more of the multiple possible implementations of the first aspect.

[0045] These and other aspects of the present application will become more readily apparent from the following description of the embodiment(s). BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate exemplary embodiments, features, and aspects of the application and, together with the description, serve to explain the principles of the application.

[0047] FIG1 is a schematic diagram showing a system architecture of an embedded presentation management system according to an embodiment of the present application.

[0048] FIG2 is a schematic diagram showing a software architecture of an embedded presentation management system according to an embodiment of the present application.

[0049] FIG3 is a schematic diagram showing a component structure of an embedded presentation management system according to an embodiment of the present application.

[0050] FIG4 shows a flowchart of an embedded representation management method according to an embodiment of the present application.

[0051] FIG5 shows a schematic diagram of a neighbor graph according to an embodiment of the present application.

[0052] FIG6 is a schematic diagram showing the version evolution of an embedded representation according to an embodiment of the present application.

[0053] FIG7 shows a schematic diagram of multi-level differential storage according to an embodiment of the present application.

[0054] FIG8 shows a schematic diagram of visualization of an embedded representation according to an embodiment of the present application.

[0055] FIG9 shows a schematic diagram of visualization of an embedded representation according to an embodiment of the present application.

[0056] FIG10 shows a schematic diagram of visualization of an embedded representation according to an embodiment of the present application.

[0057] FIG11 shows a block diagram of an embedded representation management device according to an embodiment of the present application. DETAILED DESCRIPTION

[0058] Various exemplary embodiments, features, and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise indicated.

[0059] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0060] In addition, numerous specific details are provided in the detailed description below to better illustrate the present application. Those skilled in the art will appreciate that the present application can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art are not described in detail in order to highlight the main purpose of the present application.

[0061] In the field of information retrieval, embedding refers to the vectorized representation of internet information, such as users, web pages, and videos. For example, a web page can be vectorized to generate a meaningful embedding. Each web page corresponds to one embedding, and different web pages have different embeddings. Therefore, embedding is a crucial representation of internet information.

[0062] With the explosive growth of internet data, pre-trained models for embedding representations are becoming more and more common. For example, models such as generative pre-training (GPT) and bidirectional encoder representations from transformers (BERT) are widely used in natural language understanding, while models such as visual BERT and you only look once (YOLO) are widely used in computer vision. Models such as graph convolutional neural networks (GCN) are used for large-scale graph learning. In addition, as retrieval needs become more and more widespread, multi-modal pre-trained models for embedding representations have also emerged, which allow information between different modalities to be associated. As a result, embedding representations are becoming more and more diverse, presenting multi-version and multi-granularity representations.

[0063] For example, for a web page on the Internet, its text content can obtain embeddings related to natural language processing (Graph Convolutional Networks, NLP). At the same time, its image or video content can also be encoded into a visual embedding representation. In addition, it has a graph structure-related embedding representation in the web page link relationship graph (or in the user-web page bipartite graph).

[0064] In the related art, existing embedding representations are typically managed in a file-based format, which can handle version management on the order of ten. However, with the continuous increase in pre-training methods, scenarios, and data scale, existing embedding representation management methods are not suitable for managing multiple versions of embedding representations, resulting in high management costs and poor management efficiency.

[0065] In order to solve the above technical problems, the present application provides an embedded representation management method, which includes: in response to a first version number input by a user, loading a first embedded representation corresponding to the first version number from a disk into a memory; training the first embedded representation according to preset training data to obtain a second embedded representation; determining a second version number of the second embedded representation according to the first version number and the scenario of the training data; and storing the second embedded representation and the second version number to the disk.

[0066] The embedded representation management method of an embodiment of the present application can load a first embedded representation corresponding to the first version number from the disk into the memory in response to a first version number input by the user, and then train the first embedded representation according to preset training data to obtain a second embedded representation, and then determine the second version number of the second embedded representation according to the scenario of the first version number and the training data, and store the second embedded representation and the second version number on the disk, thereby realizing multi-version management of the embedded representation, thereby reducing the management cost of the embedded representation and improving management efficiency.

[0067] The embedded representation management method of the embodiments of the present application can be applied to electronic devices, including a server and a terminal device connected to the server. The server can be a cloud server, a server cluster, etc., and the terminal device can be a personal computer, a laptop, a smartphone, a tablet computer, etc. This application does not limit the specific types of the server and terminal device.

[0068] From the perspective of product implementation, the embedded representation management method of the embodiment of the present application can be implemented as an embedded representation management system (or embedded representation management platform). The embedded representation management system can serve offline model training, online streaming learning, and online reasoning of embedded representations. The embedded representation management system can be used to implement multi-version management of embedded representations, such as version tracking and control of embedded representations, version difference comparison, multi-level differential storage, visualization, and other functions. Those skilled in the art can set the specific functions implemented by the embedded representation management system according to actual conditions, and this application does not limit this.

[0069] Figure 1 shows a schematic diagram of the system architecture of an embedded presentation management system according to an embodiment of the present application. As shown in Figure 1 , the system architecture of the embedded presentation management system comprises three layers, namely a hardware layer 110 , a software system layer 120 , and an application layer 130 .

[0070] The hardware layer 110 includes various hardware required to implement the embedded presentation management system, including terminal devices 111, servers 112, network cards 113, memory / disk 114, and central processing units (CPUs) / graphics processing units (GPUs) 115. In practical applications, the hardware layer 110 may also include other hardware, which is not limited in this application.

[0071] Terminal device 111 refers to an electronic device that can be connected to the embedded representation management system and can perform interface interactive operations, including but not limited to personal computers, laptops, smart phones, tablets, etc. This application does not limit the specific type of terminal device 111. Server 112 refers to a physical machine used to deploy various modules of the embedded representation management system. Server 112 can be a cloud server, server cluster, etc. This application does not limit the specific type of server 112. Network card 113 refers to a physical device used for communication between various hardware of the embedded representation management system. The communication here can be, for example, communication between servers within a server cluster, communication between terminal devices and servers, etc. Memory / disk 114 is used to store multiple versions of embedded representations. The memory query speed is fast and is mainly used to store embedded representation versions that need to be called frequently, such as embedded representation versions used for online reasoning, embedded representation versions being trained, etc. The disk is mainly used to store historical backup embedded representations. CPU / GPU 115 is used to complete relevant calculations of the embedded representation management system.

[0072] In the hardware layer of Figure 1, the network card 113, memory / disk 114, and CPU / GPU 115 are shown as separate hardware examples. In actual applications, the network card 113, memory / disk 114, and CPU / GPU 115 can be located in the server 112, and the terminal device 111 can also include the network card 113. Those skilled in the art can set the installation location and installation method of various types of hardware according to actual conditions, and this application does not impose any restrictions on this.

[0073] The software system layer 120 includes components or modules for implementing the core functions of the embedded representation management system, and may specifically include a version tracking and control module 121 , a multi-level differential storage module 122 , a version difference comparison module 123 and a visualization module 124 .

[0074] Application layer 130 includes components or modules for implementing offline and online processes of the embedded representation management system, and specifically may include an offline processing module 131 and an online processing module 132. In some examples, application layer 130 may also include a dynamic monitoring module (not shown) for dynamically monitoring the offline and online training processes of the embedded representation.

[0075] Figure 2 shows a schematic diagram of the software architecture of the embedded presentation management system according to an embodiment of the present application. As shown in Figure 2 , the software architecture of the embedded presentation management system consists of three layers, namely, a management and computing layer 210 , a service layer 220 , and a persistence layer 230 .

[0076] The management computing layer 210 includes the program code for the core modules of the embedded representation management system. For example, the management computing layer 210 may include the program code for the embedded representation management system's version tracking and control module, multi-level differential storage module, version difference comparison module, visualization module, and dynamic monitoring module. This provides functions such as embedded representation version tracking and control, storage, difference comparison, and dynamic monitoring during training. The management computing layer 210 is connected to the training / inference platform 240 to dynamically monitor the embedded representation training process (including offline training and online streaming training).

[0077] The service layer 220 provides services such as distributed storage, real-time reading and writing, and updating of embedded representations in the embedded representation management system. The service layer 220 can communicate with the management computing layer 210 and the training / inference platform 240 via the TCP / IP protocol. The service layer 220 can adopt a distributed system. In the case where the service layer 220 adopts a distributed system, the service layer 220 may include a master server (master) and multiple slave servers (servers), namely slave server 1, slave server 2, ..., slave server n (n is a positive integer). The master server and each slave server can communicate via the TCP / IP protocol, but there is no connection between the slave servers and no communication. The master server mainly provides global ID management and mapping, as well as unified management of operations such as embedded representation push (storing the embedded representation from memory to disk), expansion, and disaster recovery. The slave server is used to provide high-speed read and write services.

[0078] The persistence layer 230 is used to store multiple historical versions of the embedded representation on disk. As shown in Figure 2, the persistence layer 230 stores multiple versions of the embedded representation as it evolves along three branches (branch 1, branch 2, and branch 3). The persistence layer 230 communicates with the service layer 220 via the TCP / IP protocol. The persistence layer 230 can also adopt a distributed system. In one example, the distributed nodes of the persistence layer 230 share a single machine with the slave servers of the service layer 220, thereby reducing remote communication and improving processing efficiency.

[0079] Figure 3 illustrates a schematic diagram of the component architecture of an embedded representation management system according to an embodiment of the present application. As shown in Figure 3 , to improve the compatibility of the embedded representation management system with the training / inference platform 350, the embedded representation management system of this embodiment of the present application employs a layered design, comprising, from the upper user interaction layer to the lower storage layer, the following: client interaction layer 310, management and computing layer 320, service layer 330, and persistence layer 340.

[0080] The front-end of the client interaction layer 310 can be designed using commonly used front-end frameworks (e.g., Vue, React, Qt, etc.), including but not limited to web pages and applications (APPs), to provide users with a simple and efficient user interface. The client interaction layer 310 can also provide operations and displays for functions such as version difference comparison and dynamic monitoring embedded in the presentation management system.

[0081] The management computing layer 320 is the core computing component of the embedded representation management system. It may include one or more embedded representation management computing hosts, responsible for performing computations related to the embedded representation management system. Furthermore, the management computing layer 320 communicates with the training / inference platform 350 via TCP / IP, enabling collaborative tasks between embedded representation training / inference and embedded representation management. Furthermore, the training / inference platform 350 communicates with the training / inference platform operation interface 360 ​​via TCP / IP to facilitate user interface operations and displays.

[0082] Both the service layer 330 and the persistence layer 340 are used to implement distributed storage of embedded representations. The service layer 330 can provide high-speed read and write operations for embedded representations, supporting multi-version distributed training, online inference services, and more. The service layer 330 can be implemented as a distributed system consisting of a master server and multiple slave servers, where the multiple slave servers are designated as slave 1, slave 2, ..., and slave n (n is a positive integer). The master server and the slave servers can communicate via the TCP / IP protocol, but the slave servers are not connected to each other and do not communicate with each other.

[0083] The persistence layer 340 is used to provide disk storage of embedded representations of multiple historical versions. The persistence layer 340 can also adopt a distributed system. As shown in Figure 3, the persistence layer 340 is implemented as a distributed storage system composed of database 1, database 2, ..., database n. The database can be a key-value database, such as RocksDB, etc. This application does not limit the specific type of database. In one example, the database in the persistence layer 340 can share a machine with the slave server of the service layer 330, that is, database 1 shares a machine with slave server 1, database 2 shares a machine with slave server 2, ..., database n shares a machine with slave server n.

[0084] As can be seen from Figure 3, the embodiment shown in Figure 3 uses a distributed multi-level storage approach to store and manage embedded representations. From the perspective of the training / inference platform 350, the service layer 330 provides a distributed multi-machine memory storage solution, providing the training / inference platform 350 with fast read / write capabilities for embedded representations, supporting efficient offline training and real-time online inference services; the persistence layer 340 stores all embedded representation versions of historical training and uses a distributed multi-machine disk / hard disk storage solution (such as solid state drives (SSDs) and mechanical hard disk drives (HDDs)).

[0085] From a hardware implementation perspective, the embodiment shown in Figure 3 places the persistence layer 340 and service layer 330 in a cluster. Each machine in the cluster includes at least one server service for the service layer and one database service for the persistence layer. The server service provides in-memory storage, while the database service provides disk or hard drive storage. In this design, when pulling up an embedded representation, each machine in the cluster only needs to read the embedded representation from its local hard drive and write it to its local memory. This eliminates the need for communication between server services and database services, thus reducing unnecessary communication overhead.

[0086] Regarding the storage of the embedded representation of the service layer 330, when allocating the embeddings to be stored to each machine, a hash map (hashmap) + linear table (array) mapping method can be used, and the mapping-related processing is uniformly executed by the master. Each version of the embedded representation has a corresponding embedding representation table (Embedding table), which is used to store multiple embeddings and the ID of each embedding. For each embedding table, the master of the service layer maintains a hash map: hashmap <emb_id,pair<server_id,index> >, where emb_id represents the embedding ID, server_id represents the server ID in the service layer, and index represents the index of the embedding with the ID emb_id in the array. Based on the above hash mapping, any embedding ID in the embedding table can be directly mapped to an index on a specific machine (server).

[0087] Each server stores the embedding table in a linear table format. To achieve load balancing, for static embedding tables, an "equal split" approach can be used to evenly distribute the embedding table across different servers. For dynamically scalable embedding tables, a "prioritize allocation to servers with less storage" strategy can be adopted.

[0088] FIG4 shows a flow chart of an embedded representation management method according to an embodiment of the present application. As shown in FIG4 , the embedded representation management method includes:

[0089] Step S410: In response to a first version number input by a user, a first embedded representation corresponding to the first version number is loaded from a disk into a memory.

[0090] When training an embedding representation, the user must first specify the first version number of the first embedding representation to be trained. For example, the user can enter the first version number of the first embedding representation to be trained in the training / inference platform operation interface. In response to the first version number entered by the user, the embedding representation management system can load the first embedding representation corresponding to the first version number from the disk into the memory. The first embedding representation here refers to the full vector of the embedding representation corresponding to the first version number. In other words, the first embedding representation includes multiple dense vectors (i.e., embedding representations) and IDs (identities) corresponding to each dense vector.

[0091] For example, in an embedded representation management system, the service layer corresponds to memory, and the persistence layer corresponds to disk. After the user enters the first version number in the training / inference platform operation interface, the training / inference platform will send the first version number to the management computing layer of the embedded representation management system. After receiving the first version number, the management computing layer will send a pull request (the operation of loading the embedded representation from disk to memory) for the first version number to the service layer. Pulling the first version number here refers to pulling the full vector of the first embedded representation corresponding to the first version number. After receiving the pull request for the first version number sent by the management computing layer, the service layer will send a pull task for the first version number to the persistence layer. After receiving the pull task for the first version number sent by the service layer, the persistence layer immediately reads the first embedded representation corresponding to the first version number concurrently, serializes the read first embedded representation, and sends it to the service layer via the TCP / IP protocol. If the persistence layer and the service layer share a node, each node will write the read first embedded representation directly to local memory.

[0092] The service layer comprises multiple servers. After receiving data from the persistence layer, it splits the data into multiple copies and distributes them evenly across the servers. The specific distribution strategy can adopt common distributed system strategies (taking into account load balancing, scaling, and other factors), with the master implementing a unified strategy. After receiving data from the persistence layer, each server deserializes the data into an ID and a dense vector, then stores it in memory. This allows the first embedding representation corresponding to the first version number to be loaded from disk into memory. This means that the full vector of the first embedding representation is stored in memory.

[0093] In one possible implementation, when loading the first embedded representation corresponding to the first version number from disk into memory, if a first neighbor graph corresponding to the first embedded representation is also stored on disk, the first neighbor graph may also be loaded from disk into memory at the same time. The first neighbor graph is a K-nearest neighbor graph (KNNGraph) constructed based on the first embedded representation, where K is a positive integer.

[0094] In one possible implementation, if the user does not input the first version number, the latest version number of the embedded representation in the embedded representation management system is used by default, ie, the first embedded representation corresponding to the latest version number is loaded from the disk into the memory.

[0095] Step S420: Training the first embedding representation according to preset training data to obtain a second embedding representation.

[0096] After loading the first embedding representation corresponding to the first version number from disk into memory, the first embedding representation can be trained based on preset training data. When training the first embedding representation, the first embedding representation can be used as a model parameter for training. In other words, the first embedding representation can be trained through model training.

[0097] When training the first embedding representation according to the preset training data, the service layer (i.e., memory) is equivalent to providing a dictionary of the first embedding representation. The training / inference platform can obtain the dense vector (i.e., embedding representation) corresponding to the training data from the service layer through the ID of the training data, and use the obtained dense vector as the model input to train the model (offline training or online streaming training), and then send the gradients for the first embedding representation obtained during the training process to the service layer. After receiving the gradients for the first embedding representation, the service layer first aggregates (reduce) the gradients, and then updates the first embedding representation based on the aggregated gradients. When the preset training end conditions are met (for example, the number of training rounds reaches a preset round threshold, the model's loss function converges within a certain interval, etc.), the training is terminated and the second embedding representation is obtained. In one possible implementation, if there is no embedding representation corresponding to the ID of the training data in the first embedding representation, a corresponding embedding representation can be generated for the training data by random initialization.

[0098] The training data can be determined based on context, such as news, web pages, or products. For example, assuming the training data context is products, and the first embedding representation is the embedding representation of user information, then the first embedding representation is trained based on the training data, resulting in a second embedding representation of user information in the product context. In other words, the second embedding representation corresponds to the context of the training data. Different scenarios, different training data, and different second embedding representations obtained through training will also vary.

[0099] In one possible implementation, during the training of the first embedded representation, a first intermediate version of the first embedded representation may be obtained according to a preset first time interval. The first time interval may be set to a daily level, for example, 1 day, 3 days, 5 days, etc. Those skilled in the art may set the specific length of the first time interval based on actual circumstances, and this application does not impose any limitations thereon. During the training of the first embedded representation, multiple first intermediate versions may be obtained.

[0100] After each acquisition of the latest first intermediate version, a determination can be made as to whether the latest first intermediate version meets a preset first evaluation criterion. The first evaluation criterion can be used to evaluate the quality of the first intermediate version, and can be conducted offline or online. Offline evaluation typically involves collecting off-path data based on A / B testing to evaluate the quality of the first intermediate version, while online evaluation typically evaluates the quality of the first intermediate version based on the performance of its online inference service.

[0101] In the case that the newly acquired first intermediate version meets the first evaluation condition, the newly acquired first intermediate version can be stored on the disk to achieve the daily disk storage of the first intermediate version of the first embedded representation. The daily disk storage here means storing the full vector of the first intermediate version of the first embedded representation on the disk. In other words, the daily disk storage of the first intermediate version of the first embedded representation is full storage. When storing the newly acquired first intermediate version on the disk, a multi-level differential storage method can be used. Multi-level differential storage is an incremental storage and indexing method. When storing, only the embedded representations that have changed relative to the previous version in the new version are stored. For the embedded representations that have not changed, they are directly indexed to the embedded representation of the previous version or the index of the embedded representation of the previous version. The specific implementation of multi-level differential storage will be described in detail in the following embodiments.

[0102] In this way, during the training of the first embedding representation, the first intermediate version of the first embedding representation can be stored on disk at the daily level, so that when an abnormality is found during the training process, it can be rolled back based on the stored daily intermediate version (full intermediate version), thereby reducing the operation and maintenance costs of the embedding representation development.

[0103] The above embodiment uses a timed triggering method (i.e., triggering the disk flushing after a first time interval has elapsed) to trigger the daily flushing of the first intermediate version of the first embedded representation. Alternatively, a quantitative triggering method (i.e., triggering the flushing after completing training on a fixed amount of data) can be used to trigger the daily flushing of the first intermediate version of the first embedded representation. Those skilled in the art can configure the triggering method for daily flushing based on actual circumstances, and this application does not impose any restrictions thereon.

[0104] In one possible implementation, during the training of the first embedding representation, as the training progresses, some of the embedding representations in the first embedding representation will change. The management computing layer can traverse each changed node according to the previous version of the neighbor graph KNNGraph, and update its M-order neighbors (M is a positive integer) (i.e., rebuild the M-order neighbors of each changed node) to achieve dynamic update of the neighbor graph KNNGraph.

[0105] In one possible implementation, each time the latest first intermediate version is obtained, the neighbor graph of the previous version can be updated to obtain the neighbor graph corresponding to the latest obtained first intermediate version, and when the latest obtained first intermediate version is stored on the disk, the neighbor graph corresponding to the latest obtained first intermediate version is also stored on the disk.

[0106] In one possible implementation, during the training of the first embedding representation, a second intermediate version of the first embedding representation may be obtained according to a preset second time interval. The second time interval may be set to the hour level or the minute level. For example, assuming the second time interval is at the hour level, the second time interval may be set to 1 hour, 2 hours, etc.; assuming the second time interval is at the minute level, the second time interval may be set to 15 minutes, 30 minutes, etc. Those skilled in the art may set the specific length of the second time interval according to actual circumstances, and this application does not impose any restrictions thereon. During the training of the first embedding representation, multiple second intermediate versions may be obtained.

[0107] After each acquisition of the latest second intermediate version, a first difference between the latest second intermediate version and the previous second intermediate version can be determined. The first difference between the latest second intermediate version and the previous second intermediate version can be determined by comparing their neighbor graphs.

[0108] In one possible implementation, during training of the first embedding representation, the first neighbor graph may be dynamically updated according to the second time interval to obtain a neighbor graph corresponding to each second intermediate version. The neighbor graph corresponding to the first second intermediate version is updated based on the first neighbor graph, and the neighbor graph corresponding to the (i+1)th second intermediate version is updated based on the neighbor graph corresponding to the (i)th second intermediate version (where i is a positive integer), thereby enabling dynamic updating of the first neighbor graph during training.

[0109] Among them, the specific process of updating the neighbor graph corresponding to the i-th second intermediate version to obtain the neighbor graph corresponding to the i+1-th second intermediate version can be illustrated as follows: for the embedded representation of the gradient update in the i+1-th second intermediate version (relative to the i-th second intermediate version), the M-order neighbors (M is a positive integer) of the node corresponding to the gradient-updated embedded representation can be searched in the neighbor graph corresponding to the i-th second intermediate version, and then according to the distance between the gradient-updated embedded representation and its M-order neighbors (for example, cosine distance, euclidean distance, etc.), its neighbor nodes are updated, thereby completing the update of the neighbor graph and obtaining the neighbor graph corresponding to the i+1-th second intermediate version.

[0110] When determining the first difference between the most recently acquired second intermediate version and the previous second intermediate version, the third and fourth neighbor graphs can be first obtained. The third neighbor graph refers to the neighbor graph corresponding to the most recently acquired second intermediate version, and the fourth neighbor graph refers to the neighbor graph corresponding to the previous second intermediate version. Then, based on the fourth neighbor graph, the changed nodes in the third neighbor graph are determined. A changed node in the third neighbor graph refers to a node in the third neighbor graph that has changed due to training relative to the fourth neighbor graph.

[0111] In one example, the embedding representation that participated in the training between the last second intermediate version and the latest second intermediate version can be regarded as the embedded representation that has changed, and then the corresponding node of the changed embedded representation in the third nearest neighbor graph can be determined as the changed node in the third nearest neighbor graph. For example, in the training data of the training between the last second intermediate version and the latest second intermediate version, it is possible to check which embedding representation IDs appear. If the embedding representation ID appears in the training data, it can be considered that the embedding representation corresponding to the ID participated in the training between the last second intermediate version and the latest second intermediate version. Then, the embedding representation corresponding to the ID that appears in the training data can be determined as the embedded representation that has changed, and then the corresponding node of the changed embedded representation in the third nearest neighbor graph can be determined as the changed node in the third nearest neighbor graph.

[0112] In another example, the changed nodes in the third nearest neighbor graph can be determined by comparing the embedded representations corresponding to the nodes in the third nearest neighbor graph with the embedded representations corresponding to the nodes in the fourth nearest neighbor graph. This application does not limit the specific method for determining the changed nodes in the third nearest neighbor graph.

[0113] After determining the changed nodes in the third nearest neighbor graph, the neighbor change information and node change information of each changed node can be determined. Among them, the neighbor change information may include at least one of the number of neighbor changes, the neighbor change ratio and the local neighbor similarity score, and the node change information may include at least one of the node offset direction and the node offset distance. Assuming that the preset number of neighbors of each node is P (P is a positive integer), for any changed node, the M-order new neighbors of the changed node can be determined from the third nearest neighbor graph, and the M-order old neighbors of the changed node can be determined from the fourth nearest neighbor graph, and the number of changes in the M-order new neighbors of the changed node relative to the M-order old neighbors is determined as the number of neighbor changes of the changed node. The neighbor change ratio of the changed node is the ratio of the number of neighbor changes of the changed node to the preset number of neighbors P. The local neighbor similarity score of the changed node can be determined by the following formula (1):

[0114] score=S(KNN(w1),KNN(w2)) (1)

[0115] In formula (1), w1 represents the changed node, w2 represents the corresponding node of the changed node in the fourth neighbor graph, KNN(w1) represents the K neighbors of w1 in the third neighbor graph (i.e., the K nearest neighbors), KNN(w2) represents the K neighbors of w2 in the fourth neighbor graph (i.e., the K nearest neighbors), S(KNN(w1), KNN(w2)) represents the similarity between the K neighbors of w1 and the K neighbors of w2, which can be determined by the Jaccard coefficient; score represents the local neighbor similarity score of the changed node w1, and its value is between 0 and 1.

[0116] The node offset direction and node offset distance of a changed node can be determined based on a vector difference between the embedded representation of the changed node after the change and the embedded representation before the change. Specifically, the length of the vector difference between the embedded representation of the changed node after the change and the embedded representation before the change can be determined as the node offset distance of the changed node, and the direction of the vector difference can be determined as the node offset direction of the changed node.

[0117] After determining the neighbor change information and node change information of each changed node using the above method, the first difference between the most recently acquired second intermediate version and the previous second intermediate version can be determined based on the neighbor change information and node change information of each changed node. In this way, the first difference between the most recently acquired second intermediate version and the previous second intermediate version can be determined quickly and accurately, thereby improving processing efficiency.

[0118] Figure 5 shows a schematic diagram of a neighbor graph according to an embodiment of the present application. As shown in Figure 5, the neighbor graph 510 can be regarded as a neighbor graph corresponding to the previous second intermediate version, and the neighbor graph 520 can be regarded as a neighbor graph corresponding to the most recently acquired second intermediate version. The neighbor graph 510 includes 14 nodes, namely node 1, node 2, ..., node 14. Relative to the neighbor graph 510, only node 4 has changed in the neighbor graph 520, and the other nodes have not changed. In the neighbor graph 510, the first-order neighbors of node 4 are nodes (3, 5, 6), and in the neighbor graph 520, the first-order neighbors of node 4 are nodes (5, 6, 9). It can be seen that the number of neighbor nodes changed for node 4 is 1, and the ratio of neighbor nodes changed is 1 / 3.

[0119] After determining the first difference between the most recently obtained second intermediate version and the previous second intermediate version, it can be determined whether the first difference meets the preset second evaluation condition. The second evaluation condition can be used to evaluate the quality of the second intermediate version. The second evaluation condition may include at least one of the following conditions: the number of neighbor changes of each change node does not exceed the preset neighbor change number threshold, the neighbor change ratio of each change node does not exceed the preset neighbor change ratio threshold, the local neighbor similarity score of each change node is greater than or equal to the preset local neighbor similarity score threshold, the offset distance of each change node does not exceed the preset node offset threshold, the offset direction of each change node does not exceed the preset node offset direction threshold, etc. In actual applications, those skilled in the art can set the specific content of the second evaluation condition according to actual conditions, and this application does not impose any restrictions on this.

[0120] For example, as shown in Figure 5, if the offset distance of node 4 is greater than the preset node offset threshold, the newly acquired second intermediate version can be considered to not meet the second evaluation condition. The semantics of the offset of node 4 may be, for example, that the user corresponding to node 4 suddenly changes from liking "ancient 2D" to liking "extreme sports." This offset of node 4 may be caused by noise or anomalies introduced by the training data or model. By comparing the two second intermediate versions, this anomaly can be quickly discovered, triggering corresponding processing such as automatic version rollback.

[0121] When the first difference satisfies the second evaluation condition, the most recently obtained second intermediate version can be backed up to the memory, thereby realizing hourly / minute-level backup of the second intermediate version of the first embedded representation. The hourly / minute-level backup here is incremental storage, and a multi-level differential storage method can also be used here. Since the difference between the second intermediate versions is calculated frequently and the hourly / minute-level increments are not too large, the hourly / minute-level backups can be stored directly in the server's memory to support high-speed access at the management computing layer. In some examples, the number of hourly / minute-level backup versions can also be set. In this case, the latest few backup versions can be retained in the memory based on the number of versions.

[0122] In this way, during the training of the first embedding representation, the second intermediate version of the first embedding representation can be backed up at the hourly / minutely level, so that when an anomaly is found during the training process, the version can be rolled back based on the hourly / minutely level backup version (incremental intermediate version), thereby reducing the operation and maintenance costs of the embedding representation development.

[0123] The above embodiment uses a timed triggering method (i.e., triggering the backup after reaching the second time interval) to trigger the hourly / minute-level backup of the second intermediate version of the first embedded representation. A quantitative triggering method (i.e., triggering the backup after completing training of a fixed amount of data) can also be used to trigger the hourly / minute-level backup of the second intermediate version of the first embedded representation. Those skilled in the art can set the triggering method for the hourly / minute-level backup according to actual circumstances, and this application does not impose any restrictions on this.

[0124] In one possible implementation, when the latest acquired first intermediate version does not meet the first evaluation condition (for example, when the online effect of the latest acquired first intermediate version is significantly reduced), or when the first difference between the latest acquired second intermediate version and the previous second intermediate version does not meet the second evaluation condition, a fallback version can be selected from the second intermediate version in the memory or the first intermediate version on the disk according to the preset version fallback rule, and training can be continued based on the fallback version.

[0125] Among them, the preset version rollback rules may include node-level rollback rules and full-level rollback rules. Node-level rollback is mainly for version rollback when a single or a small number (the specific number can be set according to actual conditions) of nodes go astray. Accordingly, the node-level rollback rule can be set so that the number of deviated nodes does not exceed the preset deviated node threshold. When performing node-level rollback, a rollback version can be selected from the second intermediate version in the memory, and then training can continue based on the rollback version. The deviated node here refers to a node whose neighbor change information or node change information exceeds the corresponding threshold.

[0126] A full rollback is primarily used to roll back versions when a large number of nodes have deviated. Accordingly, the full rollback rule can be set to occur when the number of deviated nodes exceeds a preset threshold. When performing a full rollback, a rollback version is selected from the first intermediate version stored on disk, which is then pulled from disk into memory. Training then continues based on this rollback version.

[0127] In this way, the training process of the first embedding representation can be dynamically monitored, and when an anomaly is found, the version can be automatically rolled back. This allows automatic adjustments to anomalies or online disturbances in the embedding representation training process without human intervention, thereby significantly reducing the operation and maintenance costs of embedding representation development.

[0128] Step S430: Determine a second version number of the second embedded representation based on the first version number and the scenario of the training data.

[0129] After obtaining the second embedded representation, the metadata of the second embedded representation can be determined based on the first version number and the scenario of the training data. The metadata of the second embedded representation may include data such as the first version number, the scenario of the training data, and the dimension (for example, 128 dimensions) and quantity (for example, 100 million) of the second embedded representation. In one example, the metadata of the second embedded representation may also include other data such as the model used during training, the training method (for example, offline training or online streaming training), and the time when the training data was collected. Those skilled in the art can determine the specific content of the metadata of the second embedded representation based on actual conditions, and this application does not impose any restrictions on this. Then, the second version number of the second embedded representation can be determined based on the metadata of the second embedded representation. The version number of the second embedded representation obtained through training is different depending on the training data collected in different time periods for different scenarios.

[0130] After determining the second version number of the second embedded representation, an association between the second version number and the first version number may be established. In one example, the association between the second version number and the first version number may be added to the metadata of the second embedded representation. In another example, the metadata of the second embedded representation may further include the IDs of dense vectors (i.e., embedded representations) in the second embedded representation that have not changed relative to the first embedded representation, thereby enabling rapid switching of embedded representation versions.

[0131] Figure 6 shows a schematic diagram of the version evolution of the embedded representation according to an embodiment of the present application. As shown in Figure 6, starting from the initial fused embedded representation (i.e., the embedded representation obtained by joint training based on data from web pages, news, and product scenarios), that is, starting from the embedded representation version numbered VR-1.0, through continuous training, multiple versions and branches are evolved for business launch or comparative analysis. As shown in Figure 6, the version numbers of the embedded representations in different scenarios and different time periods are different.

[0132] Step S440: Store the second embedded representation and the second version number to the disk.

[0133] After obtaining the second embedded representation and the second version number, the second embedded representation and the second version number may be stored in a persistence layer, that is, stored in a disk.

[0134] In one possible implementation, the second embedded representation can be stored on disk using multi-level differential storage. Multi-level differential storage is an incremental storage and indexing method that stores only those embedded representations that have changed relative to the previous version. Unchanged embedded representations are directly indexed to the previous version's embedded representation or an index of the previous version's embedded representation.

[0135] Since most of the dense vectors in the first embedding remain unchanged during training, and only a small portion of them change, the second embedding is stored to disk using multi-level differential storage. While the entire second embedding appears to be stored externally, only the incremental portion is stored internally. This approach significantly conserves storage resources and effectively reduces storage consumption.

[0136] In one possible implementation, when storing the second embedded representation to disk in a multi-level differential storage manner, a fifth embedded representation can be first determined from the second embedded representation, where the fifth embedded representation refers to an embedded representation in the second embedded representation that has changed relative to the first embedded representation; the fifth embedded representation is then stored to disk; and thereafter, a storage mapping table corresponding to the second embedded representation is established based on the first address and the second address, where the first address refers to the address on disk of the embedded representation in the second embedded representation that has not changed relative to the first embedded representation, and the second address refers to the address on disk of the fifth embedded representation.

[0137] Figure 7 shows a schematic diagram of multi-level differential storage according to an embodiment of the present application. As shown in Figure 7, the embedded representation with version number V1 includes five trained vectors, namely u1, u2, u3, u4, and u5, whose actual storage addresses on the disk are 0x01, 0x02, 0x03, 0x04, and 0x05, respectively. The storage mapping table corresponding to the embedded representation with version number V1 is Map-V1, which records the actual storage addresses of these five vectors.

[0138] The embedded representation with version number V1 is trained to obtain the embedded representation with version number V2. After training, among the five vectors of version V2, vectors u1, u2, and u3 have changed compared to version V1. The changed u1, u2, and u3 can be incrementally stored on disk, and their actual storage addresses are 0x11, 0x12, and 0x13 respectively. Then, a storage mapping table Map-V2 corresponding to the embedded representation with version number V2 can be established. In Map-V2, the changed u1, u2, and u3 are mapped to the new storage addresses, while the unchanged u4 and u5 still use the same mapping addresses as in Map-V1.

[0139] The embedded representation with version number V2 is trained to obtain the embedded representation with version number V3. After training, of the five vectors of version V3, only u2 has changed compared to version V2. The changed u2 can be incrementally stored on disk, and its actual storage address is 0x21. Then, a storage mapping table Map-V3 corresponding to the embedded representation with version number V3 is established. In Map-V3, the changed u2 is mapped to the new storage address, while the unchanged u1, u3, u4, and u5 still use the same mapping addresses as in Map-V2.

[0140] In this way, when storing the three versions of the embedded representation (a total of 15 vectors), only the storage space of 9 vectors is actually used, saving the storage space of 6 vectors.

[0141] In a possible implementation, before the second embedded representation is stored in the hard disk, a quality assessment may be performed on the second embedded representation. If the second embedded representation passes the quality assessment, the second embedded representation is stored in the disk in the above manner.

[0142] In one possible implementation, after obtaining the second embedded representation, a second neighbor graph corresponding to the second embedded representation can be determined, and an association relationship between the second neighbor graph and the second embedded representation can be established. The second neighbor graph can also be stored to disk at the same time as the second embedded representation is stored to disk.

[0143] In one possible implementation, after training is completed and the second embedding representation is obtained, if the service needs to be put online immediately after training, the second embedding representation in memory may not be deleted after the second embedding representation is stored to the hard disk (i.e., written to disk), so that the online inference service can directly access the same second embedding representation used during training for online inference. If the trained embedding representation is put online, the embedding representation to be used is loaded from disk into memory using a method similar to step S410, and online inference is performed based on the embedding representation in memory.

[0144] In a possible implementation, the embedded representation management method of the present application further provides a visual display of the embedded representation, which is used to display the differences between different versions of the embedded representation. The visual display of the embedded representation is applied to the client interaction layer of the embedded representation management system.

[0145] In response to the user's request to compare the third embedding representation and the fourth embedding representation, the third embedding representation and the fourth embedding representation can be loaded from the disk into the memory respectively, and the third embedding representation and the fourth embedding representation can be subjected to dimensionality reduction processing respectively to obtain a first dimensionality reduction vector and a second dimensionality reduction vector. For example, the 128-dimensional third embedding representation can be subjected to dimensionality reduction processing to obtain a two-dimensional first dimensionality reduction vector, and the 128-dimensional fourth embedding representation can be subjected to dimensionality reduction processing to obtain a two-dimensional second dimensionality reduction vector. Among them, the dimensionality reduction processing can adopt existing related technologies, such as principal components analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), uniform manifold approximation and projection (UMAP), etc. This application does not limit the specific method of dimensionality reduction processing.

[0146] A second difference between the third and fourth embedded representations can also be determined in a manner similar to determining the first difference in the above embodiment. The first reduced dimensionality vector, the second reduced dimensionality vector, and the second difference are then displayed via a visualization interface of the client interaction layer. The relationship between the third and fourth embedded representations is as follows: the fourth embedded representation is obtained by training the third embedded representation.

[0147] Figure 8 shows a schematic diagram of a visualization of an embedding representation according to an embodiment of the present application. As shown in Figure 8, the version number of the third embedding representation is VW-1.0, and the version number of the fourth embedding representation is VN-2.0. The left portion of Figure 8 shows the overall distribution of the first reduced dimensionality vector obtained after the dimensionality reduction processing of the third embedding representation in the vector space. The right portion of Figure 8 shows the overall distribution of the second reduced dimensionality vector obtained after the dimensionality reduction processing of the fourth embedding representation in the vector space. Figure 8 can intuitively show the overall difference between the third embedding representation and the fourth embedding representation in the vector space.

[0148] Figure 9 shows a schematic diagram of an embedded representation visualization according to an embodiment of the present application. As shown in Figure 9, the histogram in the figure is a distribution histogram of the local neighbor similarity scores of the third embedded representation and the fourth embedded representation. In this visualization interface, the number of neighbors, distance function, and dimensionality reduction method related to the calculation of the local neighbor similarity score can also be selected.

[0149] FIG10 is a schematic diagram showing a visualization of an embedding representation according to an embodiment of the present application. As shown in FIG10 , the left portion is a neighbor graph of the third embedding representation, and the right portion is a neighbor graph of the fourth embedding representation. FIG10 also shows the sorting of changed nodes. When a changed node is selected, for example, node 4 is selected, the neighbor change information and node change information of the changed node will be displayed below, specifically including: the number of neighbor changes (first order), the neighbor change ratio (first order), the local neighbor similarity score, the node offset direction, and the node offset distance.

[0150] In addition, the visualization of the embedded representation also provides a way to view the domain space distribution of nodes in the third and fourth embedded representations to observe the learning effect of the embedded representation.

[0151] FIG11 is a block diagram of an embedded representation management device according to an embodiment of the present application. As shown in FIG11 , the surgical embedded representation management device includes:

[0152] A first loading module 1110, in response to a first version number input by a user, loads a first embedded representation corresponding to the first version number from a disk into a memory;

[0153] A training module 1120, configured to train the first embedding representation based on preset training data to obtain a second embedding representation;

[0154] A version number determining module 1130, configured to determine a second version number of the second embedded representation based on the first version number and the scenario of the training data;

[0155] The storage module 1140 is configured to store the second embedded representation and the second version number in the disk.

[0156] In one possible implementation, the training module 1120 includes: a first acquisition submodule, used to obtain a first intermediate version of the first embedding representation according to a preset first time interval during training of the first embedding representation, and the first time interval is at the day level; a first storage submodule, used to store the latest acquired first intermediate version to the disk when the latest acquired first intermediate version meets a preset first evaluation condition; and a training end submodule, used to end the training and obtain a second embedding representation when a preset training end condition is met.

[0157] In one possible implementation, the training module 1120 includes: a second acquisition submodule, used to obtain a second intermediate version of the first embedding representation according to a preset second time interval during the training of the first embedding representation, where the second time interval is in hours or minutes; a difference determination submodule, used to determine a first difference between the latest acquired second intermediate version and the previous second intermediate version; and a second storage submodule, used to back up the latest acquired second intermediate version to the memory when the first difference meets a preset second evaluation condition.

[0158] In one possible implementation, the training module 1120 includes: a fallback sub-module, which is used to select a fallback version from the second intermediate version in the memory or the first intermediate version on the disk according to a preset version fallback rule when the latest acquired first intermediate version does not meet the first evaluation condition, or when the first difference does not meet the second evaluation condition; and a training sub-module, which continues training based on the fallback version.

[0159] In one possible implementation, the device also includes: a second loading module, used to load the first neighbor graph corresponding to the first embedding representation from the disk into the memory; the training module 1120 includes: a dynamic update submodule, used to dynamically update the first neighbor graph according to the second time interval during the training of the first embedding representation to obtain a neighbor graph corresponding to each second intermediate version; a neighbor graph determination submodule, used to determine the second neighbor graph corresponding to the second embedding representation after obtaining the second embedding representation.

[0160] In one possible implementation, the difference determination submodule is used to: obtain a third neighbor graph and a fourth neighbor graph, the third neighbor graph refers to the neighbor graph corresponding to the most recently obtained second intermediate version, and the fourth neighbor graph refers to the neighbor graph corresponding to the previous second intermediate version; determine the changed nodes in the third neighbor graph based on the fourth neighbor graph; determine the neighbor change information and node change information of each changed node, the neighbor change information includes at least one of the number of neighbor changes, the ratio of neighbor changes and the local neighbor similarity score, and the node change information includes at least one of the node offset direction and the node offset distance; determine the first difference between the most recently obtained second intermediate version and the previous second intermediate version based on the neighbor change information and node change information of each changed node.

[0161] In a possible implementation, the storage module includes: a differential storage submodule, configured to store the second embedded representation to the disk in a multi-level differential storage manner.

[0162] In one possible implementation, the differential storage submodule is used to: determine a fifth embedded representation from the second embedded representation, where the fifth embedded representation refers to an embedded representation in the second embedded representation that has changed relative to the first embedded representation; store the fifth embedded representation on the disk; and establish a storage mapping table corresponding to the second embedded representation based on a first address and a second address, where the first address refers to an address on the disk of an embedded representation in the second embedded representation that has not changed relative to the first embedded representation, and the second address refers to an address on the disk of the fifth embedded representation.

[0163] In one possible implementation, the device further includes: a dimensionality reduction module, which, in response to a user's comparison request for the third embedded representation and the fourth embedded representation, performs dimensionality reduction processing on the third embedded representation and the fourth embedded representation, respectively, to obtain a first reduced dimensionality vector and a second reduced dimensionality vector; a difference determination module, which is used to determine a second difference between the third embedded representation and the fourth embedded representation; and a display module, which is used to display the first reduced dimensionality vector, the second reduced dimensionality vector, and the second difference.

[0164] An embodiment of the present application provides an embedded representation management device, comprising: a processor and a memory for storing processor-executable instructions; wherein the processor is configured to implement the above method when executing the instructions.

[0165] An embodiment of the present application provides a non-volatile computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions implement the above method when executed by a processor.

[0166] An embodiment of the present application provides a computer program product, including a computer-readable code, or a non-volatile computer-readable storage medium carrying the computer-readable code. When the computer-readable code runs in a processor of an electronic device, the processor in the electronic device executes the above method.

[0167] A computer-readable storage medium may be a tangible device that can hold and store instructions for use by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination thereof.

[0168] The computer-readable program instructions or codes described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions to be stored in the computer-readable storage medium in each computing / processing device.

[0169] The computer program instructions for performing the operations of the present application may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, by utilizing the state information of computer-readable program instructions to personalize an electronic circuit, such as a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer-readable program instructions, thereby implementing various aspects of the present application.

[0170] Various aspects of the present application are described herein with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer-readable program instructions.

[0171] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device is generated that implements the functions / actions specified in one or more blocks in the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium, where these instructions cause the computer, programmable data processing device, and / or other device to operate in a specific manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing various aspects of the functions / actions specified in one or more blocks in the flowchart and / or block diagram.

[0172] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more blocks in the flowchart and / or block diagram.

[0173] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, systems, methods and computer program products according to multiple embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a part for a module, program segment or instruction, and the part for the module, program segment or instruction comprises one or more executable instructions for realizing the logical function of the specification. In some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous boxes can actually be performed substantially in parallel, and they can sometimes also be performed in the opposite order, depending on the function involved.

[0174] It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented by hardware (such as a circuit or ASIC (Application Specific Integrated Circuit)) that performs the corresponding function or action, or can be implemented by a combination of hardware and software, such as firmware.

[0175] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple situations. A single processor or other unit can implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0176] The embodiments of the present application have been described above. The above description is illustrative and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for managing embedded representations, characterized in that: The method comprises: In response to a first version number input by a user, loading a first embedded representation corresponding to the first version number from a disk into a memory; Training the first embedding representation according to preset training data to obtain a second embedding representation; Determining a second version number of the second embedding representation according to the first version number and the scenario of the training data; The second embedded representation and the second version number are stored in the disk.

2. The method according to claim 1, characterized in that The training of the first embedding representation to obtain a second embedding representation includes: During the training of the first embedding representation, obtaining a first intermediate version of the first embedding representation according to a preset first time interval, where the first time interval is on the order of days; If the newly obtained first intermediate version meets the preset first evaluation condition, storing the newly obtained first intermediate version to the disk; When the preset training end condition is met, the training is ended and the second embedding representation is obtained.

3. The method according to claim 2, characterized in that The training of the first embedding representation to obtain a second embedding representation includes: During the training of the first embedding representation, obtaining a second intermediate version of the first embedding representation according to a preset second time interval, where the second time interval is in hours or minutes; determining a first difference between a most recently obtained second intermediate version and a previous second intermediate version; When the first difference satisfies a preset second evaluation condition, the most recently obtained second intermediate version is backed up to the memory.

4. The method according to claim 3, characterized in that The training of the first embedding representation to obtain a second embedding representation includes: When the newly obtained first intermediate version does not meet the first evaluation condition, or when the first difference does not meet the second evaluation condition, a fallback version is selected from the second intermediate version in the memory or the first intermediate version on the disk according to a preset version fallback rule; Training continues based on the fallback version.

5. The method according to claim 3, characterized in that The method further comprises: Loading a first neighbor graph corresponding to the first embedding representation from the disk into the memory; The training of the first embedding representation to obtain a second embedding representation includes: During the training of the first embedding representation, dynamically updating the first neighbor graph according to the second time interval to obtain a neighbor graph corresponding to each second intermediate version; After obtaining the second embedded representation, a second neighbor graph corresponding to the second embedded representation is determined.

6. The method according to claim 5, characterized in that Determining a first difference between the latest obtained second intermediate version and the previous second intermediate version includes: Obtaining a third neighbor graph and a fourth neighbor graph, wherein the third neighbor graph refers to a neighbor graph corresponding to the most recently obtained second intermediate version, and the fourth neighbor graph refers to a neighbor graph corresponding to the previous second intermediate version; Determining, according to the fourth nearest neighbor graph, the changed nodes in the third nearest neighbor graph; Determining neighbor change information and node change information of each changed node, wherein the neighbor change information includes at least one of a number of neighbor changes, a neighbor change ratio, and a local neighbor similarity score, and the node change information includes at least one of a node offset direction and a node offset distance; A first difference between the latest obtained second intermediate version and the previous second intermediate version is determined according to the neighbor change information and the node change information of each changed node.

7. The method according to any one of claims 1 to 6, characterized in that The storing the second embedded representation and the second version number to the disk includes: The second embedded representation is stored in the disk in a multi-level differential storage manner.

8. The method according to claim 7, characterized in that The step of storing the second embedded representation to the disk in a multi-level differential storage manner includes: Determining a fifth embedded representation from the second embedded representation, wherein the fifth embedded representation refers to an embedded representation in the second embedded representation that is changed relative to the first embedded representation; storing the fifth embedded representation to the disk; A storage mapping table corresponding to the second embedded representation is established based on a first address and a second address, wherein the first address refers to an address of an embedded representation in the second embedded representation that has not changed relative to the first embedded representation in the disk, and the second address refers to an address of the fifth embedded representation in the disk.

9. The method according to any one of claims 1 to 8, characterized in that The method further comprises: In response to a user's request for comparison of the third embedding representation and the fourth embedding representation, performing dimensionality reduction processing on the third embedding representation and the fourth embedding representation to obtain a first reduced dimensionality vector and a second reduced dimensionality vector; determining a second difference between the third embedded representation and the fourth embedded representation; The first reduced dimensionality vector, the second reduced dimensionality vector, and the second difference are displayed.

10. An embedded representation management device, characterized in that: The device comprises: a first loading module, in response to a first version number input by a user, loading a first embedded representation corresponding to the first version number from a disk into a memory; A training module, configured to train the first embedding representation based on preset training data to obtain a second embedding representation; a version number determination module, configured to determine a second version number of the second embedded representation based on the first version number and the scenario of the training data; A storage module is configured to store the second embedded representation and the second version number in the disk.

11. The device according to claim 10, characterized in that The training module includes: A first acquisition submodule is configured to acquire a first intermediate version of the first embedding representation according to a preset first time interval during training of the first embedding representation, where the first time interval is on a day level; A first storage submodule is configured to store the newly obtained first intermediate version to the disk if the newly obtained first intermediate version meets a preset first evaluation condition; The training end submodule is used to end the training and obtain the second embedding representation when the preset training end conditions are met.

12. The device according to claim 11, characterized in that The training module includes: A second acquisition submodule is configured to acquire a second intermediate version of the first embedding representation according to a preset second time interval during training of the first embedding representation, where the second time interval is in hours or minutes; a difference determination submodule, configured to determine a first difference between a newly acquired second intermediate version and a previous second intermediate version; The second storage submodule is configured to back up the most recently obtained second intermediate version to the memory when the first difference satisfies a preset second evaluation condition.

13. The device according to claim 12, characterized in that The training module includes: a fallback submodule, configured to select a fallback version from the second intermediate version in the memory or the first intermediate version on the disk according to a preset version fallback rule when the newly obtained first intermediate version does not meet the first evaluation condition or when the first difference does not meet the second evaluation condition; The training submodule continues training based on the fallback version.

14. The device according to claim 12, characterized in that The device further comprises: a second loading module, configured to load a first neighbor graph corresponding to the first embedding representation from the disk into the memory; The training module includes: a dynamic updating submodule, configured to dynamically update the first neighbor graph according to the second time interval during training of the first embedding representation to obtain a neighbor graph corresponding to each second intermediate version; A neighbor graph determination submodule is used to determine a second neighbor graph corresponding to the second embedding representation after obtaining the second embedding representation.

15. The device according to claim 14, characterized in that The difference determination submodule is used to: Obtaining a third neighbor graph and a fourth neighbor graph, wherein the third neighbor graph refers to a neighbor graph corresponding to the most recently obtained second intermediate version, and the fourth neighbor graph refers to a neighbor graph corresponding to the previous second intermediate version; Determining, according to the fourth nearest neighbor graph, the changed nodes in the third nearest neighbor graph; Determining neighbor change information and node change information of each changed node, wherein the neighbor change information includes at least one of a number of neighbor changes, a neighbor change ratio, and a local neighbor similarity score, and the node change information includes at least one of a node offset direction and a node offset distance; A first difference between the latest obtained second intermediate version and the previous second intermediate version is determined according to the neighbor change information and the node change information of each changed node.

16. The device according to any one of claims 10 to 15, characterized in that The storage module includes: The differential storage submodule is configured to store the second embedded representation in the disk in a multi-level differential storage manner.

17. The device according to claim 16, characterized in that The differential storage submodule is used for: Determining a fifth embedded representation from the second embedded representation, wherein the fifth embedded representation refers to an embedded representation in the second embedded representation that is changed relative to the first embedded representation; storing the fifth embedded representation to the disk; A storage mapping table corresponding to the second embedded representation is established based on a first address and a second address, wherein the first address refers to an address of an embedded representation in the second embedded representation that has not changed relative to the first embedded representation in the disk, and the second address refers to an address of the fifth embedded representation in the disk.

18. The device according to any one of claims 10 to 17, characterized in that The device further comprises: a dimensionality reduction module, in response to a user's request for comparison of the third embedding representation and the fourth embedding representation, performing dimensionality reduction processing on the third embedding representation and the fourth embedding representation, respectively, to obtain a first reduced dimensionality vector and a second reduced dimensionality vector; a difference determination module, configured to determine a second difference between the third embedded representation and the fourth embedded representation; A display module is configured to display the first reduced dimensionality vector, the second reduced dimensionality vector, and the second difference.

19. An embedded representation management device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to implement the method according to any one of claims 1 to 9 when executing the instructions.

20. A non-volatile computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 9 is implemented.

21. A computer program product, characterized in that The invention comprises a computer-readable code or a non-volatile computer-readable storage medium carrying the computer-readable code, wherein when the computer-readable code is executed in an electronic device, a processor in the electronic device executes the method according to any one of claims 1 to 9.

Images