Recommended method, device, electronic equipment and computer readable storage medium

Abstract

Embodiments of the present application disclose a recommendation method and device, electronic equipment and a computer readable storage medium, which are used to shorten the cold start period of new content and improve the accuracy of new content recommendation. The method comprises: obtaining a recommendation request comprising information of a first target content; inputting the recommendation request into a trained recall model to obtain candidate recommended content output by the recall model; if it is determined that the first target content comprises new content, obtaining a new and old content mapping relationship by connecting each attribute of the new content with a target attribute of a second target content in a first association graph, the first association graph being used to describe the association relationship between the content and the attribute and the association relationship between the attributes, the target attribute of the second target content being the same as the attribute of the new content; inputting the new and old content mapping relationship and the candidate recommended content into a trained network model to obtain an output result of the trained network model; and determining the recommended content according to the output result.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a recommended method, apparatus, electronic device, and computer-readable storage medium. Background Technology

[0002] With the continuous development of big data and the Internet, recommendation systems are being used more and more widely in e-commerce, consulting, knowledge learning and short video entertainment.

[0003] Recommendation systems can recommend content that best matches a user's needs or interests based on data such as user and content characteristics. For example, a recommendation system can recommend news items that match a user's preferences based on data such as user preferences and browsing history.

[0004] Currently, recommendation systems often require a cold start for new content to enter the system and be recommended. However, existing recommendation systems suffer from problems such as long cold start periods for new content and low recommendation accuracy. Summary of the Invention

[0005] This application provides a recommendation method, apparatus, electronic device, and computer-readable storage medium, which can solve the problems of long cold start cycles for new content and low recommendation accuracy in existing recommendation methods.

[0006] In a first aspect, embodiments of this application provide a recommended method, the method comprising:

[0007] Obtain the recommendation request, which includes information about the primary target content;

[0008] Input the recommendation request into the pre-trained recall model to obtain the recall results output by the pre-trained recall model. The recall results include candidate recommendation content.

[0009] If it is determined that the first target content includes new content, then the pre-constructed first association graph is obtained, and the mapping relationship between the new and old content is obtained by connecting each attribute of the new content with the target attribute of the second target content in the first association graph. The new content is the content to be cold-started. The first association graph is used to describe the association relationship between content and attributes as well as the association relationship between attributes. The target attribute of the second target content is the same as the attribute of the new content.

[0010] Input the mapping relationship between new and old content and the candidate recommended content into the trained network model to obtain the output of the trained network model. The output is used to describe the probability of user interaction with new and old content.

[0011] Recommended content is determined based on the output results.

[0012] As can be seen from the above, the embodiments of this application obtain the mapping relationship between new and old content by connecting the attributes of new content with the target attributes of the second target content in the first association graph. This allows new content with the same attributes to be connected with old content, realizing the cold start prediction of new content. This enables new content to quickly enter the recommendation system, shortens the cold start cycle of new content, and also enables the network model to perform effective and accurate reasoning, thus improving the recommendation accuracy of new content.

[0013] In some possible implementations of the first aspect, the method further includes, before obtaining the recommendation request:

[0014] Acquire content data and user data;

[0015] Based on the content data, construct the first association graph;

[0016] The recall model is trained using the first association graph and / or the multi-reference behavior fusion sequence to obtain the trained recall model. The multi-reference behavior fusion sequence is constructed based on user data.

[0017] The network model is trained using content data and user data to obtain the trained network model.

[0018] In this implementation, the first association graph mines the potential associations between content and attributes, and the multi-reference behavior fusion sequence mines the potential associations between content. Training the recall model based on the first association graph and / or the multi-reference behavior fusion sequence can make the knowledge graph denser, thereby obtaining a recall model with higher recommendation accuracy, thus improving the recommendation accuracy of candidate content.

[0019] Among some possible implementations of the first aspect, constructing a first association graph based on content data includes:

[0020] Based on the attribute information of each content, obtain the associated attribute information and the same attribute information. The associated attribute information is used to describe the associated attributes between each content, and the same attribute information is used to describe the same attributes between each content. The content data includes the attribute information of the content.

[0021] The content and attributes are connected based on attribute information, the same attributes are connected based on the same attribute information, and related attributes are connected based on related attribute information to obtain the first association graph.

[0022] In this implementation, not only are individual content items connected to their corresponding attributes, but content items with the same attributes are also connected through attributes, and content items with related attributes are also connected through attributes. This allows for the discovery of the inherent relationships between content and attributes, as well as between attributes, thereby improving the accuracy of content recommendation.

[0023] In some possible implementations of the first aspect, the user data includes user action behavior data and content citation data within a preset time period; a multi-citation behavior fusion sequence is constructed based on the user data, including:

[0024] Based on operational behavior data, determine the new content to be added within a preset time period;

[0025] For each new piece of content, based on the content citation data, the template content referenced by the new content is connected with the next new content in the time sequence to obtain a multi-citation behavior fusion sequence. The content citation data is used to describe the content referenced by each piece of content and the template content referenced by each piece of content.

[0026] In this implementation, the template content referenced by the newly added content is connected with the next newly added content in the time sequence. This allows for the integration of user history and content relationships to build a deeper graph, thereby improving the accuracy of content recommendation.

[0027] In some possible implementations of the first aspect, the mapping relationship between the new and old content is obtained by connecting the various attributes of the new content with the target attributes of the second target content in the first association graph, including:

[0028] Construct the attribute matrix for the new content;

[0029] Connect each attribute of the new content with the target attribute of the second target content in the first association graph to obtain the second association graph;

[0030] For each attribute in the attribute matrix, based on the second association graph, the identifier of the second target content with the target attribute is determined, and the attributes in the attribute matrix are replaced with the identifier of the new content to obtain the mapping relationship between the old and new content.

[0031] In this implementation, the attribute information of the new content is supplemented by constructing an attribute matrix of the new content; by obtaining the content identifier with the same attributes as the new content and constructing a mapping matrix of the new and old content, the mapping relationship between the new and old content is obtained, which can help complete the cold start prediction of the new content.

[0032] In some possible implementations of the first aspect, if the first target content does not include new content, the method further includes: inputting candidate recommendation content into the trained network model and obtaining the output result of the trained network model.

[0033] In some possible implementations of the first aspect, the method further includes loading the trained network model via a lazy evaluation mechanism for directed acyclic graph tasks before inputting candidate recommendations into the trained network model.

[0034] In this implementation, the trained network model is loaded and saved through the lazy evaluation mechanism of the directed acyclic graph task, which allows the model to be loaded before the cold start inference begins, thus ensuring the real-time performance of cold start inference.

[0035] Secondly, embodiments of this application provide a recommended apparatus, which includes:

[0036] The request module is used to obtain recommendation requests, which include information about the first target content.

[0037] The recall module is used to input recommendation requests into a pre-trained recall model and obtain the recall results output by the pre-trained recall model. The recall results include candidate recommendation content.

[0038] The module for determining the mapping relationship between new and old content is used to obtain a pre-constructed first association graph if the first target content is determined to include new content. The new content is the content to be cold-started. The first association graph is used to describe the association relationship between content and attributes as well as the association relationship between attributes. The target attributes of the second target content are the same as the attributes of the new content.

[0039] The sorting module is used to input the mapping relationship between new and old content and candidate recommended content into the trained network model, obtain the output results of the trained network model, and use the output results to describe the interaction probability between the user and the new and old content; the recommended content is determined based on the output results.

[0040] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any of the first aspects above.

[0041] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the first aspects above.

[0042] Fifthly, embodiments of this application provide a chip system including a processor coupled to a memory. The processor executes a computer program stored in the memory to implement the method described in any of the first aspects above. The chip system may be a single chip or a chip module composed of multiple chips.

[0043] Sixthly, embodiments of this application provide a computer program product that, when run on an electronic device, causes the electronic device to execute the method described in the first aspect.

[0044] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0045] Figure 1 A schematic diagram of a recommendation process provided in an embodiment of this application;

[0046] Figure 2 A schematic block diagram of a recommendation system provided in an embodiment of this application;

[0047] Figure 3 Another schematic block diagram of the recommendation system provided in the embodiments of this application;

[0048] Figure 4 A flowchart illustrating a recommended method provided in an embodiment of this application;

[0049] Figure 5A This is a schematic diagram illustrating the relationship between the content and attributes provided in the embodiments of this application;

[0050] Figure 5B This is a schematic diagram illustrating the relationship between attributes provided in the embodiments of this application;

[0051] Figure 6 This is a schematic diagram of the association map in the recommended scenario provided in the embodiments of this application;

[0052] Figure 7A This is a schematic diagram of the second association map construction process provided in an embodiment of this application;

[0053] Figure 7B This is a schematic diagram of the second correlation map provided in an embodiment of this application;

[0054] Figure 8A A schematic diagram of the offline recall phase provided in an embodiment of this application;

[0055] Figure 8B Another schematic diagram of the offline recall phase provided in the embodiments of this application;

[0056] Figure 9 Another schematic flowchart of the recommended method provided in the embodiments of this application;

[0057] Figure 10 A schematic diagram of the recommended process provided for embodiments of this application;

[0058] Figure 11 Structural block diagram of the recommended device provided in the embodiments of this application;

[0059] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0060] In the following description, specific details such as particular system architectures and technologies are set forth for illustrative purposes and not for limiting purposes, in order to provide a thorough understanding of the embodiments of this application.

[0061] Typically, the recommendation process can be divided into a recall phase and a ranking phase. For example, see... Figure 1 The illustrated embodiment of this application provides a schematic diagram of a recommendation process. The recommendation model includes, but is not limited to, a recall layer and a ranking layer. The recall layer is used in the recall stage, and the ranking layer is used in the ranking stage. Both the recall and ranking layers typically rely on relevant network models. The recall layer usually uses an embedding model, which constructs a graph from the original input data and, after generating a sequence through point walks based on the graph, embeds features into this sequence to obtain a feature embedding vector. The ranking layer typically uses an artificial intelligence (AI) model, such as a reinforcement learning model or a knowledge graph augmentation model.

[0062] like Figure 1 As shown, the candidate content set contains tens of millions of content data points. These tens of millions of content data points are input into the recall layer. The recall layer model performs recall operations on the tens of millions of content data points based on various feature information (such as user feature information) to quickly narrow down the tens of millions of content data points to thousands, and then inputs the thousands of content data points into the ranking layer. The ranking layer model performs ranking operations on the thousands of content data points based on various feature information to obtain ten-level content data points, and then recommends these ten-level content data points to users.

[0063] It is understandable that the recall phase is characterized by large data volume, high speed, simple model and few features, while the ranking phase is characterized by small data volume, accurate ranking, complex model and many features.

[0064] For new content, recommendation systems often need to use cold start technology to allow new content to enter the recommendation system and be recommended.

[0065] Cold start can refer to the situation where new content, having not participated in training in the past and lacking reference information from other content, is difficult to be recommended by the recommendation system and requires a long period of "warm-up" before it can be recommended.

[0066] During their long-term research, the inventors of this application discovered that after new content enters the database of a recommendation system, the knowledge graph built based on the old content or other modules of the recommendation system cannot be updated in a timely manner. This results in the new content not being able to enter the recommendation system quickly and not being recommended immediately, leading to an excessively long cold start period for new content. In addition, when recommending content based on a knowledge graph or other models that cannot be updated in a timely manner, the recommendation accuracy of new content will also be reduced.

[0067] To address the issues mentioned above, this application proposes a recommendation scheme. This scheme first constructs a first association graph based on existing content data. After obtaining new input content, the attributes of the new content are connected to the corresponding attributes of the existing content in the first association graph to obtain a mapping relationship between the new and existing content. This mapping relationship is then input into a trained network model to obtain the interaction probability between the user and the new and existing content. Recommended content is obtained based on the interaction probability. This shortens the cold start cycle for new content and improves the accuracy of new content recommendations.

[0068] The recommendation scheme provided in this application can be applied to a recommendation system. For example, see... Figure 2 The diagram shown is a schematic block diagram of a recommendation system provided in an embodiment of this application. The recommendation system may include at least one terminal device 100 and a server 200, and the terminal device 100 and the server 200 are communicatively connected.

[0069] Terminal device 100 can be a user terminal such as a mobile phone, tablet, or computer. The type of terminal device 100 may vary depending on the application scenario of the recommendation system. For example, when the recommendation system is used to recommend news or short videos, terminal device 100 can be a mobile phone or tablet; when the recommendation system is used for solution recommendations in the telecommunications operations and maintenance field, terminal device 100 can be a computer used by operations and maintenance personnel. This application embodiment does not limit the specific type of terminal device 100.

[0070] Server 200 can refer to a cloud server, which can communicate and connect with terminal device 100 via the Internet. Server 200 may include multiple server entities with different functions. For example, server 200 may include an application server, a data storage server, and a recommendation server. The data storage server can be used to store massive amounts of content data, and the recommendation server is used to implement recommendation functions. Of course, server 200 is presented as a single server entity, and this single server entity performs multiple different functions.

[0071] Terminal device 100 is used to send a recommendation request to server 200, and after receiving the recommendation content returned by server 200 in response to the recommendation request, presents the recommendation content to the user.

[0072] Server 200 is used to receive recommendation requests from terminal device 100; in response to the recommendation request, it performs recall and sorting operations based on stored content data and user data, obtains recommended content, and returns the recommended content to terminal device 100.

[0073] See, as an example rather than a limitation. Figure 3 The diagram shown represents another schematic block diagram of the recommendation system provided in this application embodiment. This recommendation system may include a recommendation platform 31, an application server 32, a recommendation device 33, and a data storage server 34. The recommendation device 33 is used to implement the recommendation function; it can be a server entity independent of the application server 32, or it can be deployed on the application server 32 as a module of the application server 32, without limitation. In this case, the aforementioned server 200 may include the application server 32, the recommendation device 33, and the data storage server 34.

[0074] Specifically, the recommendation platform 31 presents a front-end interface on the user 35's terminal device. On this front-end interface, the user can perform actions such as editing, clicking, browsing, and referencing content. When the user 35 interacts with the content on the front-end interface, the recommendation platform 31 records the user's action data and transmits this data in batches to the back-end application server 32. This user action data may include, for example, actions such as clicking, browsing, editing, and referencing. After receiving the user action data, the application server 32 processes it accordingly and stores it on the data storage server 34.

[0075] In addition, user 35 can also proactively initiate or trigger a recommendation request through the front-end page of recommendation platform 31. Recommendation platform 31 can respond to user actions by sending the recommendation request to application server 32, which then forwards the request to recommendation device 33. The recommendation request may carry natural language text, which could be search terms entered by the user in the search box or on the front-end interface.

[0076] After application server 32 sends a recommendation request to recommendation device 33, recommendation device 33 responds to the recommendation request by obtaining user operation behavior data through application server 32 and performing structured processing on the user operation behavior data to obtain structured data. Based on the structured data, it performs a content data recall operation to obtain recall results. During the recall operation, recommendation device 33 requests content data from data storage server 34. Data storage server 32 responds to this request by sending the stored content data to recommendation device 33 through application server 32. After obtaining the recall results, recommendation device 33 performs precise sorting of the recall results to obtain ordered recommended content and returns the ordered list of recommended content to application server 32. Application server 32 then recommends the list of recommended content to user 35 through recommendation platform 31.

[0077] Figure 3 The recommendation system can be applied to various recommendation scenarios, and the list of recommended content will differ depending on the scenario. For example, Figure 3 The diagram illustrates scenarios such as solution recommendations, homepage recommendations, and "you might like" recommendations. In the solution recommendation scenario, data storage server 34 stores massive amounts of solution data maintained by operations and maintenance personnel. The recommendation device 33 filters and recommends solutions from this massive amount of data. In the "you might like" recommendation scenario, data storage server 34 stores massive amounts of news data. The recommendation device 33 filters and recommends news articles from this massive amount of news data. In the homepage recommendation scenario, data storage server 34 stores massive amounts of document content data. The recommendation device 33 filters and recommends content from this massive amount of document content data, providing homepage content recommendations for search pages or web pages.

[0078] After exemplarily introducing the recommendation system architecture and application scenarios provided in the embodiments of this application, the recommendation scheme provided in the embodiments of this application will be described in detail below.

[0079] Please see Figure 4 This is a flowchart illustrating a recommended method provided in an embodiment of this application. The method may include the following steps:

[0080] Step S401: Server 200 obtains the recommendation request, which includes information about the first target content.

[0081] The recommendation request can be sent by the terminal device 100. The terminal device 100 can send a recommendation request to the server 200 after detecting a user's triggering action. For example, after a user enters search keywords in the search box of a mobile webpage, the mobile phone sends a recommendation request to the server, which carries the search keywords. Of course, in other embodiments, the server 200 can also be triggered to enter the recommendation process through other means.

[0082] In addition to information about the primary target content, recommendation requests can also include user information, such as user identifiers.

[0083] The information of the first target content can include relevant information about the first target content. For example, in a news recommendation scenario, when a user browses and clicks on relevant news on their mobile phone, the phone responds to the click action by sending a recommendation request to the server, which then responds to the request. In this case, the first target content can refer to the news article clicked by the user, and its information can include one or more of the title, content, and category of the clicked news article. The specific information of the first target content can be set as needed. Of course, the recommendation request can also carry the user's user ID so that the server can look up the user's relevant data based on the user ID. This relevant data may include, but is not limited to, the user's historical operational behavior data and user preference data. Furthermore, the server can make recommendations based on the user's relevant data to make the recommended news more suitable for the user.

[0084] For example, in a solution recommendation scenario within the telecommunications operations and maintenance field, a user fills in relevant solution information on a computer. In response to the user's action, the computer generates a recommendation request and sends it to the server. This recommendation request carries the solution information filled in by the user and the user's information. In this case, the primary target content is the solution filled in by the user, and the information within the primary target content is the relevant solution information.

[0085] In step S402, server 200 inputs the recommendation request into the pre-trained recall model and obtains the recall result output by the pre-trained recall model, which includes candidate recommendation content.

[0086] Recall models are used in the recall phase of the recommendation process, and can be, for example, graph embedding models. These models perform recall operations, filtering candidate recommended content from a content dataset.

[0087] Step S403: Server 200 determines whether the first target content includes new content. If the first target content includes new content, proceed to step S404; if the first target content does not include new content, proceed to step S406.

[0088] The first target content includes one or more items, and there may or may not be new content among these items.

[0089] The information in the first target content refers to the relevant data used to describe the first target content. Based on the information in the first target content, it can be determined whether the first target content includes new content.

[0090] For example, the content specifically refers to a solution, and the information in the first target content includes the solution name. In the solution name-solution ID table, solution names and solution IDs are matched one-to-one; for existing solutions, each solution name has a unique solution ID. Then, the solution name-solution ID table is traversed to determine if a solution ID corresponding to the solution name carried in the recommendation request can be found. If a solution ID exists for the solution name carried in the recommendation request, it indicates that the solution is an existing solution; otherwise, if a solution ID does not exist for the solution name carried in the recommendation request, it indicates that the solution is a new solution.

[0091] Step S404: Server 200 obtains the pre-constructed first association graph and obtains the mapping relationship between the new content and the target attribute of the second target content in the first association graph by connecting each attribute of the new content with the target attribute of the second target content in the first association graph. The new content is the content to be cold-started. The first association graph is used to describe the association relationship between content and attribute as well as the association relationship between attributes. The target attribute of the second target content is the same as the attribute of the new content.

[0092] The first association graph is pre-constructed based on content data. Content data includes, but is not limited to, the attribute information of each piece of content. Each piece of content has one or more attributes. For example, in a solution recommendation scenario, content refers to solutions, and content data refers to solution data, which includes the attribute information of each solution; each solution may have the following attributes: change sub-scenario, technical support center, product line, component, product family, and product, etc. As another example, in a news recommendation scenario, content refers to news, and content data refers to news data. In this case, each news item may have attributes such as news category, publication time, and location.

[0093] The first association graph consists of edges and nodes. Content and attributes are the nodes of the graph, and the associations between content and attributes are the edges of the graph.

[0094] In some embodiments, attribute information of each content is first obtained, and then for each content, attributes belonging to that content are connected to that content based on its attribute information to construct a first association graph. For example, see Figure 5A The diagram illustrates the relationship between content and attributes. It shows two pieces of content, each containing two attributes, and each piece of content connected to its corresponding two attributes. In this first association graph, the edges between content nodes and attribute nodes describe the relationships between content and attributes.

[0095] In other embodiments, the first association graph can describe not only the relationships between content and attributes, but also the relationships between attributes themselves. In this case, the relationships between attributes are also edges of the graph. When constructing the first association graph, not only are content nodes connected to attribute nodes, but also nodes with the same attribute and related attribute nodes are connected.

[0096] Based on the attribute information of each piece of content, we can determine the common attribute information and the related attribute information. Related attribute information describes the related attributes between pieces of content; that is, it shows that a certain attribute of one piece of content is related to a certain attribute of another piece of content. Common attribute information describes the shared attributes between pieces of content; that is, it shows that a certain attribute of one piece of content is the same as a certain attribute of another piece of content. Specifically, the aforementioned related attribute information can be obtained based on the content's attribute information through expert prior information and domain knowledge.

[0097] After obtaining the related attribute information and the same attribute information, related attributes can be connected based on the related attribute information, and the same attributes can be connected based on the same attribute information. Furthermore, for each piece of content, attributes belonging to that content are connected to that content based on the attribute information to construct a first association graph. At this point, the first association graph includes not only edges between content nodes and attribute nodes, but also edges between attribute nodes, where these two attribute nodes can be the same attribute node or related attribute nodes.

[0098] For example, see Figure 5B The diagram shown illustrates the relationships between attributes. Figure 5B The example shows two content items, each containing two attributes. For each content item, based on its attribute information, the attributes belonging to that content are concatenated with that content; furthermore, the common attributes of the two content items are concatenated, specifically referring to... Figure 5B The dashed line between two unshaded triangles indicates that the triangles connected by the dashed line represent two identical attributes; connecting the related attributes of two pieces of content specifically means... Figure 5B The dashed line between the shaded and unshaded triangles indicates that the unshaded and shaded triangles connected by the dashed line represent two related attributes.

[0099] By comparison Figure 5A and Figure 5B It can be seen that, Figure 5A It simply connects the content and attributes, and Figure 5B It not only connects content and attributes, but also connects attributes to each other. In comparison, Figure 5BBy connecting similar attributes and related attributes, we can deeply explore the underlying information between content and attributes, as well as between attributes.

[0100] By connecting similar attributes and related attributes, deeper correlation information is uncovered, solving the problem of poor content recommendation in sparse scenarios where there is no interaction between users and content, thereby improving the accuracy of content recommendation.

[0101] Of course, in some other embodiments, when constructing the first association graph, in addition to connecting content and attributes, it is also possible to connect only the same attributes or related attributes, without having to connect both the same attributes and related attributes simultaneously. This can also improve the accuracy of content recommendation.

[0102] For example, see Figure 6 The diagram illustrates the association between different scenarios in the proposed solutions. Scenario A includes attributes A1 and A2, and Scenario B includes attributes B1 and B2. Assume that the identity document (ID) of Scenario A is DI2022063000004, attribute A1 is the technical support center (specifically, a mobile service point in location C), and attribute A2 is a product attribute. The ID of Scenario B is DI20211108000886, attribute B1 is the technical support center (specifically, a mobile service point in location C), and attribute B2 is a product attribute.

[0103] In this scenario, Solution A is the new solution, and Solution B is the old solution; Solution A references Solution B. In the solution recommendation scenario, the attributes of the solution can include, but are not limited to: changing sub-scenarios, technical support center, product line, components, and products.

[0104] First, based on the attribute information of Scheme A and Scheme B, it is determined that attribute A1 and attribute B1 are the same attribute. Both attribute A1 and attribute B1 are technical support center attributes, and both technical support center attributes are mobile stations in location C.

[0105] Then, based on the prior information and domain knowledge of the experts, and based on the attribute information of Scheme A and Scheme B, it was determined that both attribute A2 and attribute B2 are product attributes. Although the product names are different, they both belong to products related to performance tuning. Therefore, attribute A2 and attribute B2 are determined to be related attributes.

[0106] Finally, based on the attribute information of scheme A, attributes A1 and A2 are both connected to scheme A; based on the attribute information of scheme B, attributes B1 and B2 are both connected to scheme B; in addition, the same attributes and related attributes are also connected, that is, attributes A1 and B1 are connected, and attributes A2 and B2 are connected, to construct an attribute-enhanced association graph, which is the first association graph mentioned above.

[0107] After constructing the first association graph based on the content data, the first association graph can be stored on a data storage server. When the server 200 needs to perform a cold start with new content, it reads the pre-stored first association graph and connects the new content with the old content in the first association graph to obtain the mapping relationship between the new and old content.

[0108] The second target content is content that already exists in the first association graph, and this content has a target attribute. For new content, all content in the first association graph is old content, meaning the second target content is old content.

[0109] The second target content refers to content that has one or more target attributes. The target attributes are the same as those of the first target content, or in other words, the target attributes are attributes of the first target content. That is, when a piece of content in the first association graph has a target attribute, that content is the second target content.

[0110] For example, the first objective is as described above. Figure 6 Solution A has two attributes, A1 and A2. Solution B has two attributes, B1 and B2. Since attribute B1 of Solution B is the same as attribute A1 of Solution A, attribute B1 is the target attribute. Therefore, Solution B has the target attribute B1 and is the second target content.

[0111] In the first association graph, find the target attribute that is the same as the attribute of the first target content, then connect the attribute of the first target content with the corresponding target attribute to construct the second association graph, and obtain the relationship between the old and new content based on the second association graph.

[0112] In some embodiments, an attribute matrix of the new content is first constructed. Specifically, a preset number of attributes are obtained from the attributes of the new content, and these preset number of attributes are combined to form an attribute matrix. The preset number can be set according to actual needs; it can be greater than, less than, or equal to the number of attributes of the new content. When the preset number is greater than the number of attributes of the new content, a corresponding number of attributes can be randomly extracted from the attributes of the new content to complete the attribute matrix. For example, in the field of solution recommendation, the solution with solution ID 60145 is a new solution, which has three attributes: "BTS3900", "Software Upgrade", and "Software Update". Since each solution in the original attribute matrix has four attributes, i.e., the preset number is four, which is greater than the number of attributes of the new solution. In this case, the three attributes of the new solution are first used as three attributes in the attribute matrix, and then one attribute is randomly selected from the three attributes as the fourth attribute of the attribute matrix to obtain an attribute matrix including four attributes. Assuming that "Software Upgrade" is randomly selected as the fourth attribute of the attribute matrix, the attribute matrix at this time can be exemplarily shown in Table 1 below.

[0113] Table 1

[0114]

[0115] When the preset number is less than the number of attributes in the new content, a preset number of attributes can be randomly selected from the attributes of the new content, and the selected attributes can be combined into an attribute matrix.

[0116] It should be noted that for each new piece of content, a certain number of attributes can be randomly selected (with the possibility of repeated selection) to form an attribute matrix. Based on the attribute matrix, the attributes of the new content can be supplemented into the network model and the first association graph.

[0117] After constructing the attribute matrix of the new content, each attribute of the new content is connected with the target attribute of the second target content in the first association graph to obtain the second association graph.

[0118] For example, the attribute matrix of the new content is shown in Table 1 above. For the attribute "BTS3900", the solution with ID 603124 has this attribute; for the attribute "software upgrade", the solution with ID 603137 has this attribute; for the attribute "SoftwareUpdate", the solution with ID 602994 has this attribute; and for the second attribute "software upgrade", the solution with ID 603056 has this attribute.

[0119] A second association graph is constructed by connecting the same attributes of the new and old schemes. Specifically, "BTS3900" in scheme 60145 is connected to "BTS3900" in scheme 603124, "software upgrade" in scheme 60145 is connected to "software upgrade" in scheme 603137, "software update" in scheme 60145 is connected to "software update" in scheme 602994, and "software upgrade" in scheme 60145 is connected to "software upgrade" in scheme 603056. In this way, based on the first association graph, the attribute information of the new scheme is supplemented into the first association graph by connecting the various attributes of scheme 60145 to the attributes of the corresponding old schemes, thus constructing the second association graph.

[0120] For example, see Figure 7A The diagram showing the second association map construction process and Figure 7B The diagram shown is a schematic of the second correlation map. Figure 7A and Figure 7B In the scenario of solution recommendation, the first target content is the new solution, and the second target content is the old solution in the first association graph.

[0121] like Figure 7A As shown, new scheme 1, new scheme 2 and new scheme 3 are all schemes to be cold-started. The first association graph includes old scheme 1 to old scheme 9. Old scheme 1 has attribute 1, old scheme 2 has attribute 1, old scheme 3 has attributes 1 and 2, old scheme 4 has attributes 2 and 4, old scheme 5 has attribute 4, old scheme 6 has attribute 4, and old scheme 7, old scheme 8 and old scheme 9 all have attribute 3.

[0122] New scheme 1 has attribute 1, new scheme 2 has attributes 2 and 4, and new scheme 3 has attributes 3 and 4.

[0123] For each new solution, an attribute matrix is ​​constructed, and based on this matrix, the attributes of the new solution are linked to the corresponding attributes of the old solutions. For example... Figure 7AAs shown, attribute 1 of the new scheme 1 is the same as attribute 1 of the old schemes 1, 2, and 3. Therefore, connect the new scheme 1 with attribute 1 using a solid line, and then connect attribute 1 of the new scheme 1 with attribute 1 of the old schemes 1, 2, and 3 using a dashed line. Similarly, attribute 2 of the new scheme 2 is the same as attribute 2 of the old schemes 3 and 4. Therefore, connect the new scheme 2 with attribute 2 using a solid line, and then connect attribute 2 of the new scheme 2 with attribute 2 of the old schemes 3 and 4 using a dashed line. Attribute 4 of the new scheme 2 is the same as attribute 4 of the old schemes 4, 5, and 6. Therefore, connect the new scheme 2 with attribute 4 using a solid line, and then connect attribute 4 of the new scheme 2 with attribute 4 of the old schemes 3, 4, 5, and 6 using a dashed line.

[0124] The attribute 4 of the new scheme 3 is the same as the attribute 4 of the old schemes 4, 5, and 6. Therefore, after connecting the new scheme 3 and attribute 4 with a solid line, the attribute 4 of the new scheme 3 is then connected with the attribute 4 of the old schemes 3, 5, and 6 with a dashed line. The attribute 3 of the new scheme is the same as the attribute 3 of the old schemes 7, 8, and 9. Therefore, after connecting the new scheme 3 and attribute 3 with a solid line, the attribute 3 of the new scheme 3 is then connected with the attribute 3 of the old schemes 7, 8, and 9 with a dashed line.

[0125] By connecting the new and old solutions to their shared attributes, they can be indirectly linked. After connecting the shared attributes of the new and old solutions, one can obtain, for example... Figure 7B The second correlation map is shown. (See example...) Figure 7B As shown, new scheme 1, new scheme 2 and new scheme 3 in the second association graph are all connected to the old scheme through the same attribute.

[0126] After obtaining the second association graph, for each attribute in the attribute matrix, based on the second association graph, the identifier of the second target content with the target attribute is determined, and the attributes in the attribute matrix are replaced with the identifier of the new content to obtain the mapping relationship between the old and new content.

[0127] The mapping relationship between new and old content can be specifically represented as a new-old content mapping matrix. This is achieved by finding content in the second association graph that shares the same attributes as the new content, and replacing the attributes in the attribute matrix with the identifiers of the second target content.

[0128] For example, suppose the attribute matrix of the new scheme is as shown in Table 1 above, with the specific identifier being the scheme ID. Then, from the second association graph, we know that for the attribute "BTS3900", the scheme with ID 603124 has this attribute, so we replace "BTS3900" in Table 1 with "603124"; for the attribute "software upgrade", the scheme with ID 603137 has this attribute, so we replace the first "software upgrade" in Table 1 with "603137"; similarly, for the attribute "software update", the scheme with ID 602994 has this attribute, so we replace "software update" in Table 1 with "602994"; for the second "software upgrade" attribute, the scheme with ID 603056 has this attribute, so we replace the second "software upgrade" in Table 1 with "603056". The resulting mapping matrix of the old and new content can be exemplarily shown in Table 2 below. Among them, Schemes 603124, 603137, 602994 and 603056 are all old schemes in the first correlation map.

[0129] Table 2

[0130] 60145 603124 603137 602994 603056

[0131] It is worth noting that, by constructing an attribute matrix for new content, the attribute information of the new content can be supplemented; by obtaining content identifiers with the same attributes as the new content and connecting the same attributes of the old and new content, the mapping relationship between the old and new content can be obtained. In this way, the attributes recognized by the old model can be used to describe the new content, and the knowledge graph can be updated in a timely manner according to the attribute information of the new content, thereby assisting in the cold start prediction of new content.

[0132] In step S405, server 200 inputs the mapping relationship between new and old content and candidate recommended content into the trained network model to obtain the output result of the trained network model. The output result is used to describe the probability of user interaction with new and old content.

[0133] The network model mentioned above can be a Knowledge Graph Convolutional Network (KGCN) model, or other types of network models, which are not limited here.

[0134] This network model is a pre-trained and saved model. If there is no new content in the first target content, a cold start is not required; in this case, candidate recommendations can be directly input into the saved network model for inference. If there is new content in the first target content, the mapping relationship between old and new content and candidate recommendations can be input into the network model for inference.

[0135] The network model is trained for a preset number of rounds, and the prediction accuracy of each round is compared. The network model corresponding to the round with the highest prediction accuracy is selected as the completed network model. During each round of model training, candidate recommendation content can be used as input to the network model, and the prediction results output by the network model can be obtained. In practical applications, the saved network model can be trained once every preset period. In the offline network model training phase, the network model can be trained using the candidate recommendation content output from the offline recall phase to obtain the completed network model. Specifically, the offline recall phase can construct a knowledge graph based on content data and user data, and then use the knowledge graph to train the graph embedding model, or use the knowledge graph, multi-reference behavior fusion sequence, and first association graph to train the graph embedding model.

[0136] It should be noted that when the network model is a KGCN model, before training begins, the entire knowledge graph typically needs to be read, and then tens of millions of content and attribute information entries in the knowledge graph need to be divided. Before using Python's GIL unlocking technique, Python needs to process these tens of millions of content and attribute information entries one by one. Even if each entry takes 0.001 seconds, processing ten million entries would still take nearly 167 minutes, resulting in low efficiency throughout the training process. To address this, this embodiment uses Python's GIL unlocking technique to process multiple entries simultaneously. Compared to processing one entry every 0.001 seconds, processing multiple entries every 0.001 seconds significantly improves the efficiency of the entire training process.

[0137] If the first target content includes new content, i.e. content requiring a cold start, then the trained network model is used for cold start inference. The trained network model performs cold start inference for new content to infer the probability of user interaction with both new and old content, and the inferred interaction probability is output as the result.

[0138] After training the network model and obtaining the trained network model, the trained network model can be saved and then loaded when cold start inference of new content is required later.

[0139] In some embodiments, each time a cold start inference begins, the saved network model needs to be reloaded, and the mapping relationship between the old and new content is input into the network model after loading. Loading the network model typically takes more than three seconds, resulting in a time difference between inputting the mapping relationship and outputting the inference result exceeding three seconds. Since cold start inference has high real-time requirements, a time difference greater than three seconds does not meet the real-time requirements for cold start inference with new content.

[0140] In other embodiments, to ensure the real-time performance of cold-start inference for new content, a lazy evaluation mechanism using a Directed Acyclic Graph (DAG) task can be used to load the network model. This DAG lazy evaluation mechanism allows the stored network model to be loaded before cold-start inference begins, and cold-start inference can be performed immediately upon obtaining the mapping relationship between new and old content, without waiting for the network model to load. Thus, the time from inputting the mapping relationship to outputting the inference result can be less than 0.1 seconds, meeting the real-time requirements of cold-start inference. Furthermore, after completing inference, the network model loaded using the DAG lazy evaluation mechanism can re-enter a state waiting for input until the inference service is shut down or restarted.

[0141] It is worth noting that when the network model is a KGCN model, the graph convolutional network (GCN) can be used to mine the association attributes of content on the knowledge graph, thereby effectively capturing the relevance between content and improving the accuracy of new content recommendation. In addition, based on the first association graph, new content with the same attributes is connected with old content, which allows the KGCN model to make effective and accurate inferences about new content based on the mapping relationship between new and old content, thereby improving the accuracy of new content recommendation. Furthermore, using the mapping relationship between new and old content as the input of the model allows the network model to dynamically process new content, avoiding the inability to make inferences in real time due to frequent updates to the network model.

[0142] Step S406: Server 200 inputs the candidate recommendation content into the trained network model and obtains the output result of the trained network model.

[0143] If the initial target content contains no new content (i.e., no content awaiting cold start), then instead of connecting old and new content to obtain a mapping relationship, the candidate recommended content can be directly input into the network model for inference. In this case, the network model's output describes the probability of user interaction with each piece of content.

[0144] In other words, if there is no new content in the input of the network model, the network model can use the original attribute matrix for inference; if there is new content in the input of the network model, the mapping relationship between the old and new content needs to be used as the input of the network model, so that the updated attribute matrix can be input into the network model, and the network model can effectively and accurately infer the new content.

[0145] Step S407: Server 200 determines the recommended content based on the output results.

[0146] The above output result can be the output result of step S405, in which case since the first target content includes new content, the output result includes the interaction probability of the new content and the interaction probability of the old content; or it can be the output result of step S406, in which case the output result does not include the interaction probability of the new content.

[0147] The output can represent the interaction probability of each piece of content. The higher the interaction probability, the higher the recommendation priority. In some embodiments, the content is sorted according to the interaction probability, and recommended content is selected from the sorted content. For example, content with an interaction probability greater than or equal to a preset probability threshold can be selected as recommended content, or the content can be sorted from largest to smallest interaction probability, and the top 10 content can be selected as recommended content.

[0148] It should be noted that, in this embodiment of the application, the attributes of the new content are connected with the target attributes of the second target content in the first association graph to obtain the mapping relationship between the new and old content. This allows new content with the same attributes to be connected with old content, and the attribute information of the new content is added to the network model and the first association graph. This enables cold start prediction of new content, thereby allowing new content to quickly enter the recommendation system, shortening the cold start cycle of new content, and enabling the network model to perform effective and accurate reasoning, thus improving the recommendation accuracy of new content.

[0149] Based on the above embodiments, the server 200 can first obtain content data and user data, then train a network model and a recall model based on the content data and user data, obtain the trained network model and the trained recall model, and save the trained network model and the trained recall model.

[0150] User data and content data can be stored in a data storage server, and server 200 can obtain the required content data and user data from the data storage server.

[0151] User data may include, but is not limited to, user action behavior data within a preset time period. This user action behavior data is obtained by recording user actions related to content within the preset time period. User actions related to content include, but are not limited to, clicking, browsing, editing, and referencing. The preset time period can be set according to actual application needs.

[0152] It is understandable that user data can include not only user behavior data but also other information, such as basic user information like age, gender, and preferences. Of course, user data can also exclude user behavior data and only include user information. The specific information included in user data may vary depending on the application scenario, and this is not a limitation here.

[0153] When training a recall model based on content data and user data, the model can be trained using at least one of a first association graph and a multi-reference behavior fusion sequence, or it can be trained without using at least one of the first association graph and the multi-reference behavior fusion sequence. The multi-reference behavior fusion sequence is constructed based on user data.

[0154] In some embodiments, see Figure 8A The diagram shown illustrates an offline recall stage provided in an embodiment of this application. The server 200 first performs a graph construction operation, then performs a recall operation based on the graph embedding component, outputting candidate recommended content. Here, graph construction refers to building a knowledge graph based on user data and content data; the graph embedding component refers to recall based on the graph embedding component or graph embedding model.

[0155] In essence, recall refers to using graph embedding models to describe content or attributes using feature vectors, and describing the relationships between content items and between content and attributes using the relationships between feature vectors. Based on these relationships, a small number of highly relevant pieces of content are coarsely selected from a large dataset as candidate recommendations. For example, from 100,000 pieces of content, recall can coarsely select 1,000 highly relevant candidate items, which then form the candidate recommendations.

[0156] exist Figure 8A Instead of using at least one of the first association graph and the multi-reference behavior fusion sequence to train the recall model, the constructed knowledge graph is simply input into the graph embedding component to train the graph embedding model.

[0157] However, the knowledge graph constructed from user data and content data is a sparse graph. A sparse knowledge graph cannot uncover the potential relationships between content or between content and attributes. As a result, the accuracy of the recall results obtained solely based on the sparse knowledge graph is low, that is, the content recommendation accuracy is low. Thus, training a recall model based on a sparse knowledge graph will lead to a low content recommendation accuracy of the trained recall model.

[0158] In other embodiments, see Figure 8B The illustration shows another schematic diagram of the offline recall stage provided in this application embodiment. The server 200 can first construct a knowledge graph based on user data and content data; then, based on the knowledge graph, and according to a first association graph and / or a multi-reference behavior fusion sequence, obtain the embedding vector of each content through a graph embedding model. The multi-reference behavior fusion sequence is constructed based on user data; and candidate recommended content is obtained based on the embedding vector of each content. At this time, not only is the constructed knowledge graph used as the input of the graph embedding component, but at least one of the first association graph and the multi-reference behavior fusion sequence is also used as the input of the graph embedding component.

[0159] It is worth noting that both the first association graph and the multi-reference behavior fusion sequence can make the knowledge graph denser. Furthermore, the multi-reference behavior fusion sequence can uncover potential relationships between content items, while the first association graph can uncover potential relationships between content and attributes. Training a recall model based on the knowledge graph, combining at least one of the first association graph and the multi-reference behavior fusion sequence, can result in a higher accuracy in content recommendation.

[0160] User data includes user behavior data and content citation data within a preset time period. The preset time period can be set according to actual needs, such as a week or a month. User behavior data may include, but is not limited to, editing, clicking, browsing, and citation. Content citation data describes the content referenced by each piece of content and the template content referenced by each piece of content. Typically, new content will reference old content, and old content may also reference old content, and old content will reference template content.

[0161] In some embodiments, new content within a preset time period can be determined first based on operational behavior data. New content can refer to content newly added within the preset time period. Then, for each new content, the template content referenced by the new content is connected with the next new content in the time sequence based on content reference data to obtain a multi-reference behavior fusion sequence.

[0162] New content may directly reference template content or indirectly reference template content. Indirect reference means that the new content references existing content, which in turn directly or indirectly references template content.

[0163] It is understandable that, based on the time when the new content was added, there is a chronological relationship between the new content and the new content.

[0164] It is worth noting that connecting the template content referenced by newly added content with the next newly added content in the time sequence can integrate user history and content relationships to build a deeper graph of connections, thereby improving the accuracy of content recommendation.

[0165] For example, in a solution recommendation scenario, a preset time period of one month is set to record user behavior data related to the solution within one month.

[0166] In the context of solution recommendation, solution writing refers to the process of creating a new solution. Instead of starting from scratch, it often involves referencing existing solutions (i.e., solution referencing), which are typically generated by modifying an initial template (i.e., template referencing). In this context, new content refers to the new solution, old content refers to the old solution, and template content refers to the template solution.

[0167] Extract user c003689xx's operational behavior data from the beginning of June to the end of July 2022. Assume that part of the operational behavior data for this month is shown in Table 3 below.

[0168] Table 3

[0169]

[0170]

[0171] As shown in Table 3 above, the user wrote a new scheme DI20220630000004 on June 30th. The new scheme DI20220630000004 referenced the old scheme numbered DI20211108000886. The template of DI20211108000886 is DI202111012001542. On July 6th, the user wrote a new scheme DI20220706000001. The new scheme DI20220706000001 referenced the old scheme numbered IN20210522000086. The template of the old scheme IN20210522000086 is IN202102250000033.

[0172] It is understandable that for the new scheme DI20220630000004, the next new scheme in terms of timing is DI20220706000001, which indirectly references the template scheme DI202111012001542.

[0173] For the new scheme DI20220630000004, connect the new scheme DI20220630000004 with the referenced old scheme DI20211108000886, connect the old scheme DI20211108000886 with the template scheme DI202111012001542, and then connect the template scheme DI202111012001542 with the next new scheme DI20220706000001. In this way, the template scheme of the new scheme can be connected with the next new scheme.

[0174] For example, the new proposal and proposal reference written on June 30, the proposal reference and template reference, as well as the template reference on June 30 and the new proposal written on July 6, are structured into a multi-reference behavior fusion sequence, which can be shown in Table 4 below.

[0175] Table 4

[0176] DI20220630000004 DI20211108000886 DI20211108000886 DI202111012001542 DI202111012001542 DI20220706000001

[0177] It is worth noting that by linking new content with template references of adjacent content, sequential adjacency information can be added, forming a structured table.

[0178] As can be seen from the above, the embodiments of this application can combine at least one of the first association graph and the multi-reference behavior fusion sequence to train the recall model, which can make the content recommendation accuracy of the trained recall model higher.

[0179] As can be seen from the above embodiments, the recommendation scheme provided in this application can include an offline stage and an online inference stage. The online inference stage refers to using the trained model to perform inference to obtain the probability of user interaction with content.

[0180] See Figure 9 The illustrated schematic flowchart shows another recommended method provided in the embodiments of this application, which may include the following steps:

[0181] Step S901: Server 200 obtains content data and user data. The content data includes the attribute information of the content.

[0182] Step S902: Server 200 constructs the first association graph based on the content data.

[0183] Step S903: Server 200 constructs a multi-reference behavior fusion sequence based on user data.

[0184] Step S904: Server 200 uses the first association graph and / or multi-reference behavior fusion sequence to train the recall model and obtain the trained recall model.

[0185] Step S905: Server 200 trains the network model based on content data and user data, obtains the trained network model, and saves the trained network model.

[0186] In this embodiment, steps S901 to S905 are the offline stages. In specific applications, the network model training process can be performed once every preset period.

[0187] Step S906: Server 200 obtains a recommendation request, which includes information about the first target content and user information.

[0188] In step S907, server 200 inputs the recommendation request into the trained recall model and obtains the recall results output by the pre-trained recall model. The recall results include candidate recommendation content.

[0189] Understandably, the recall phase of the recommendation process can reduce a large amount of content data (e.g., tens of millions) to a smaller amount (e.g., thousands).

[0190] In step S908, server 200 determines whether the first target content includes new content. If the first target content includes new content, proceed to step S909; if the first target content does not include new content, proceed to step S911.

[0191] In step S909, server 200 obtains the pre-constructed first association graph and obtains the mapping relationship between the new and old content by connecting the various attributes of the new content with the target attributes of the second target content in the first association graph.

[0192] In step S910, server 200 inputs the mapping relationship between new and old content and candidate recommended content into the trained network model to obtain the output result of the trained network model. The output result is used to describe the probability of user interaction with new and old content.

[0193] Step S911: Server 200 inputs the candidate recommendation content into the trained network model and obtains the output result of the trained network model.

[0194] Step S912: Server 200 determines the recommended content based on the output results.

[0195] In this embodiment, steps S906 to S912 are the online inference stage.

[0196] It is understood that the similarities between this embodiment and the above embodiments can be found in the corresponding content above, and will not be repeated here.

[0197] To better illustrate the recommended solutions provided in the embodiments of this application, the following description uses a recommended scenario as an example.

[0198] See Figure 10 The schematic diagram of the recommendation process provided in the embodiment of this application is shown. The recommendation process may include input, recall stage, ranking stage, construction of multi-reference behavior fusion sequence, construction of attribute enhancement association sequence, construction of new scheme mapping enhancement information, and output.

[0199] The input can include solution data, user data, and natural language text. Natural language text can be words entered by the user in the search box. User data includes historical user behavior data for a preset time period. Solution data can specifically be a solution dataset; for example, in a solution recommendation scenario within the telecommunications operations and maintenance field, this solution dataset would specifically be a tele-graph dataset.

[0200] For details on the tele-graph dataset, please refer to https: / / www.datafountain.cn / datasets / 6847.

[0201] In the process of constructing a multi-reference behavior fusion sequence, a multi-reference behavior fusion sequence can be constructed based on the user's historical operation behavior data.

[0202] In the process of constructing the attribute-enhanced association sequence, the inputs are the attribute information and association attribute information of each solution. This association attribute information can be determined based on domain knowledge and expert prior information, and it describes the association attributes between each solution and the solutions themselves. Based on the solution attribute information and the association attribute information, an attribute-enhanced solution association sequence is constructed, and the output is an attribute-enhanced association graph. Specifically, this attribute-enhanced association graph is the first association graph described above; the specific construction process can be found in the construction process of the first association graph described above.

[0203] After constructing the first association graph and the multi-reference behavior fusion sequence, these can be used as inputs to the graph embedding component in the online recall phase. The input to the graph embedding component can include the knowledge graph obtained through graph construction operations, the first association graph, and the multi-reference behavior fusion sequence. A knowledge graph is constructed based on content data, user data, and natural language text. Then, the knowledge graph, the multi-reference behavior fusion sequence, and the attribute-enhanced association graph (i.e., the first association graph) are used as inputs to the graph embedding component for recall, outputting candidate recommendation solutions. In other words, the recall model is trained offline using the first association graph and the multi-reference behavior fusion sequence to obtain a trained recall model. During the online phase, this trained recall model is used for recall operations.

[0204] After obtaining candidate recommendation schemes, the candidate recommendation schemes are used as input to the KGCN model.

[0205] During the construction of new scheme mapping enhancement information, the system identifies whether a new scheme exists among the externally input schemes based on the scheme data. If a new scheme exists among the externally input schemes, an attribute-enhanced scheme association sequence is constructed based on the information of the new scheme and the first association graph to obtain the new scheme mapping enhancement information (i.e., the new-old scheme mapping matrix). Specifically, an attribute matrix of the new scheme is constructed, and then a second association graph is constructed by connecting the attributes of the new scheme with the same attributes of the old schemes in the first association graph. The new-old scheme mapping matrix is ​​obtained based on the second association graph.

[0206] During the ranking phase, if the external input includes new solutions, the mapping matrix between the old and new solutions, along with candidate recommended solutions, is input into the trained KGCN model to obtain the inference results output by the KGCN model. In this process, ranking can be performed using knowledge graph-enhanced recommendation techniques. Specifically, the trained KGCN model is pre-loaded using a DAG task lazy evaluation technique. Finally, the mapping matrix between the old and new solutions is input into the KGCN model to obtain the interaction probabilities output by the KGCN model, and recommended solutions are selected based on these interaction probabilities.

[0207] If the external input scheme does not include a new scheme, then the candidate recommendation scheme and other information are used as input to the trained KGCN model to obtain the interaction probability output by the KGCN model, and the recommendation scheme is selected based on the interaction probability.

[0208] It is worth noting that existing solutions do not utilize users' historical citation information to construct multi-citation behavior fusion sequences for in-depth mining of user historical citation information. However, this application's embodiment constructs multi-citation behavior fusion sequences to deeply mine user operation behavior data and uses these sequences to make the relevant edges of the knowledge graph denser, thereby improving the accuracy of the solution recommendation.

[0209] Existing solution recommendations do not use information such as solution attributes and related attributes to enhance the association information. However, the embodiments of this application construct an attribute-enhanced association graph based on solution attribute information and related attribute information, and use the attribute-enhanced association graph to make the knowledge graph denser, so as to mine deeper information and improve the accuracy of solution recommendation.

[0210] Furthermore, to address the cold start problem in existing scheme recommendations, an attribute-enhanced association graph is used. By connecting the same attributes of new and old schemes, the mapping relationship between the new and old schemes is obtained. This allows the KGCN model to obtain information such as the mapping relationship between the new and old schemes, thereby improving the efficiency of cold start and the accuracy of new scheme recommendations.

[0211] Corresponding to the recommended method described in the above embodiments, Figure 11 A structural block diagram of a recommended apparatus provided in an embodiment of this application is shown, which can be integrated onto server 200. For ease of explanation, only the parts relevant to the embodiments of this application are shown.

[0212] Reference Figure 11 The recommended device may include:

[0213] Request module 111 is used to obtain a recommendation request, which includes information about the first target content;

[0214] The recall module 112 is used to input the recommendation request into the pre-trained recall model and obtain the recall result output by the pre-trained recall model. The recall result includes candidate recommendation content.

[0215] The new and old mapping relationship determination module 113 is used to obtain a pre-constructed first association graph if it is determined that the first target content includes new content, and obtain the new and old content mapping relationship by connecting each attribute of the new content with the target attribute of the second target content in the first association graph. The new content is the content to be cold-started. The first association graph is used to describe the association relationship between content and attributes and the association relationship between attributes. The target attribute of the second target content is the same as the attribute of the new content.

[0216] The sorting module 114 is used to input the mapping relationship between new and old content and candidate recommended content into the trained network model, obtain the output result of the trained network model, and the output result is used to describe the interaction probability between the user and the new and old content; and determine the recommended content based on the output result.

[0217] In some possible implementations, the recommended apparatus described above further includes:

[0218] The first association graph construction module is used to acquire content data and user data, including content attribute information; and to construct the first association graph based on the content data.

[0219] The recall model training module is used to train the recall model using the first association graph and / or the multi-reference behavior fusion sequence to obtain the trained recall model. The multi-reference behavior fusion sequence is constructed based on user data.

[0220] The network model training module is used to train the network model using content data and user data to obtain the trained network model.

[0221] In some possible implementations, the first association graph construction module is specifically used to: obtain association attribute information and common attribute information based on the attribute information of each content, wherein the association attribute information is used to describe the association attributes between each content, and the common attribute information is used to describe the common attributes between each content, and the content data includes the attribute information of the content; connect the content and attributes based on the attribute information, connect the common attributes based on the common attribute information, and connect the association attributes based on the association attribute information to obtain the first association graph.

[0222] In some possible implementations, user data includes user action data and content reference data within a preset time period; the device also includes a multi-reference behavior fusion sequence construction module, used to determine new content within the preset time period based on the action data; for each new piece of content, based on the content reference data, the template content referenced by the new content is connected with the next new content in the time sequence to obtain a multi-reference behavior fusion sequence, wherein the content reference data is used to describe the content referenced by each piece of content and the template content referenced by each piece of content.

[0223] In some possible implementations, the new and old mapping relationship determination module 113 is specifically used to: construct an attribute matrix of new content; connect each attribute of the new content with the target attribute of the second target content in the first association graph to obtain the second association graph; for each attribute in the attribute matrix, determine the identifier of the second target content with the target attribute according to the second association graph, and replace the attributes of the attribute matrix with the identifier of the new content to obtain the new and old content mapping relationship.

[0224] In some possible implementations, if the first target content does not include new content, the sorting module 114 is also used to: input the candidate recommendation content into the trained network model and obtain the output result of the trained network model.

[0225] In some possible implementations, the device also includes a model loading module, which loads the trained network model via a lazy evaluation mechanism for directed acyclic graph tasks, before the candidate recommendations are input into the trained network model.

[0226] The aforementioned recommended device has the function of implementing the aforementioned recommended method. This function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the aforementioned function, and the modules can be software and / or hardware.

[0227] It should be noted that the information interaction and execution process between the above-mentioned devices are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0228] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 12 As shown, the electronic device 12 of this embodiment includes: at least one processor 120 ( Figure 12 Only one is shown in the diagram), memory 121, and a computer program 122 stored in memory 121 and executable on at least one processor 120. When processor 120 executes computer program 122, it implements the steps in any of the recommended method embodiments described above. For example, implementing... Figure 4 Each step in the process.

[0229] Electronic device 12 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. This electronic device may include, but is not limited to, a processor 120 and a memory 121. Those skilled in the art will understand that... Figure 12 This is merely an example of electronic device 12 and does not constitute a limitation on electronic device 12. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0230] The processor 120 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0231] In some embodiments, memory 121 may be an internal storage unit of the electronic device 12, such as a hard disk or RAM of the electronic device 12. In other embodiments, memory 121 may be an external storage device of the electronic device 12, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 12. Furthermore, memory 121 may include both internal and external storage units of the electronic device 12. Memory 121 is used to store operating systems, applications, bootloaders, data, and other programs, such as program code for computer programs. Memory 121 can also be used to temporarily store data that has been output or will be output.

[0232] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0233] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0234] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0235] In the embodiments provided in this application, it should be understood that the disclosed devices, electronic devices, and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0236] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0237] The electronic device provided in this application embodiment may include a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method as described in any of the above method embodiments.

[0238] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps described in the various method embodiments above.

[0239] This application provides a computer program product that, when run on an electronic device, enables the electronic device to perform the steps described in the various method embodiments above.

[0240] This application also provides a chip system, which includes a processor coupled to a memory. The processor executes a computer program stored in the memory to implement the methods described in the above embodiments. The chip system may be a single chip or a chip module composed of multiple chips.

[0241] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. Furthermore, in the description of this application specification and the appended claims, the terms "first," "second," "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance. References to "one embodiment" or "some embodiments" in this application specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in yet other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0242] Finally, it should be noted that the above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A recommendation method, characterized in that, The method includes: Obtain a recommendation request, wherein the recommendation request includes information about the first target content; The recommendation request is input into a pre-trained recall model to obtain the recall result output by the pre-trained recall model, and the recall result includes candidate recommendation content; If it is determined that the first target content includes new content, then a pre-constructed first association graph is obtained, and the mapping relationship between the new and old content is obtained by connecting each attribute of the new content with the target attribute of the second target content in the first association graph. The new content is the content to be cold-started. The first association graph is used to describe the association relationship between content and attributes and the association relationship between attributes. The target attribute of the second target content is the same as the attribute of the new content. The mapping relationship between the new and old content and the candidate recommended content are input into the trained network model to obtain the output result of the trained network model. The output result is used to describe the probability of user interaction with the new and old content. Recommended content is determined based on the output results.

2. The method according to claim 1, characterized in that, Prior to obtaining the recommendation request, the method further includes: Acquire content data and user data; Based on the content data, construct the first association graph; The recall model is trained using the first association graph and / or the multi-reference behavior fusion sequence to obtain the trained recall model, wherein the multi-reference behavior fusion sequence is constructed based on the user data; The network model is trained using the content data and the user data to obtain the trained network model.

3. The method according to claim 2, characterized in that, Based on the content data, a first association graph is constructed, including: Based on the attribute information of each content, related attribute information and identical attribute information are obtained. The related attribute information is used to describe the related attributes between the content, and the identical attribute information is used to describe the identical attributes between the content. The content data includes the attribute information of the content. The content and the attribute are connected according to the attribute information, the same attribute is connected according to the same attribute information, and the related attribute is connected according to the related attribute information to obtain the first association graph.

4. The method according to claim 2, characterized in that, The user data includes user action data and content citation data within a preset time period; Constructing the multi-reference behavior fusion sequence based on the user data includes: Based on the operational behavior data, determine the new content to be added within the preset time period; For each of the newly added contents, the template content referenced by the newly added content is connected with the next newly added content in the time sequence according to the content reference data to obtain the multi-reference behavior fusion sequence. The content reference data is used to describe the content referenced by each content and the template content referenced by each content.

5. The method according to any one of claims 1 to 4, characterized in that, By connecting the various attributes of the new content with the target attributes of the second target content in the first association graph, a mapping relationship between the new and old content is obtained, including: Construct the attribute matrix of the new content; Connect each attribute of the new content with the target attribute of the second target content in the first association graph to obtain the second association graph; For each attribute in the attribute matrix, based on the second association graph, the identifier of the second target content having the target attribute is determined, and the attributes in the attribute matrix are replaced with the identifier of the new content to obtain the mapping relationship between the old and new content.

6. The method according to claim 1, characterized in that, If the first target content does not include the new content, the method further includes: The candidate recommendation content is input into the trained network model to obtain the output result of the trained network model.

7. The method according to claim 1 or 6, characterized in that, Before inputting the candidate recommendation content into the trained network model, the method further includes: The trained network model is loaded using a lazy evaluation mechanism for directed acyclic graph tasks.

8. A recommended device, characterized in that, The device includes: The request module is used to obtain a recommendation request, which includes information about the first target content; The recall module is used to input the recommendation request into a pre-trained recall model and obtain the recall result output by the pre-trained recall model, wherein the recall result includes candidate recommendation content; The new and old mapping relationship determination module is used to obtain a pre-constructed first association graph if it is determined that the first target content includes new content, and obtain the new and old content mapping relationship by connecting each attribute of the new content with the target attribute of the second target content in the first association graph. The new content is the content to be cold-started. The first association graph is used to describe the association relationship between content and attribute and the association relationship between attributes. The target attribute of the second target content is the same as the attribute of the new content. The sorting module is used to input the mapping relationship between the new and old content and the candidate recommended content into the trained network model, obtain the output result of the trained network model, and the output result is used to describe the interaction probability between the user and the new and old content; and determine the recommended content based on the output result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images