Method and device for determining recommendation information, computer device and storage medium

Abstract

The application relates to a method and device for determining recommended information, a computer device, a storage medium and a computer program product. The application can be applied to cloud technology, artificial intelligence and various scenes. The method comprises the following steps: obtaining a target recommendation behavior sequence based on each historical recommended information behavior of a target object within a preset time period. The information of each historical recommended information behavior in the target recommendation behavior sequence comprises any one of a recommended information identifier and a first recommended information category of the historical recommended information behavior. The target recommended information category is determined based on the target recommendation behavior sequence. Similar recommended information categories similar to the target recommended information category are determined. The recommended information corresponding to the target recommended information category and the similar recommended information categories is determined as to-be-recommended information. The accuracy and generalization of the recommended information are improved.

Description

Technical Field

[0001] This application relates to the fields of computer and communication technology, and in particular to a method, apparatus, computer device, and storage medium for determining recommendation information. Background Technology

[0002] With the development of internet technology, accessing information online has become an integral part of people's lives, including their leisure, entertainment, and work. Businesses often promote their products online by distributing recommendations to increase brand awareness. Currently, the process involves first categorizing the recommendations based on images, materials, and titles, then assigning each recommendation to a corresponding category, and finally determining the specific recommendations to be pushed to users based on their historical behavior with each recommendation. However, if the images in the recommendations are difficult to identify or contain multiple product signals, the assigned category may differ from the actual category, thus reducing the accuracy and generalization of the recommendations. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, apparatus, computer device, and storage medium for determining recommendation information that can improve the accuracy and generalization of recommendation information, in order to address the above-mentioned technical problems.

[0004] Firstly, this application provides a method for determining recommendation information. The method includes:

[0005] Based on the target object's historical recommendation information behavior within a preset time period, obtain the target recommendation behavior sequence, and the information of each historical recommendation information behavior in the target recommendation behavior sequence, including the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories;

[0006] Determine the category of target recommendation information based on the target recommendation behavior sequence;

[0007] Identify similar recommendation categories that are similar to the target recommendation category;

[0008] The recommended information corresponding to the target recommendation information category and the similar recommendation information category is determined as the information to be recommended.

[0009] In one embodiment, a target recommendation behavior sequence is obtained based on the target object's historical recommendation information behavior within a preset time period, including:

[0010] Based on the behavior time of each historical recommendation behavior, the historical recommendation behavior is sorted to obtain the first recommendation behavior sequence. The information of each historical recommendation behavior in the first recommendation behavior sequence includes the recommendation information identifier of the historical recommendation behavior and the first recommendation information category.

[0011] The information of each historical recommendation behavior in the first recommendation behavior sequence is sampled and processed to obtain the target recommendation behavior sequence.

[0012] In one embodiment, a target recommendation behavior sequence is obtained based on the target object's historical recommendation information behavior within a preset time period, including:

[0013] A graph network is constructed based on each historical recommendation information behavior. The nodes of the graph network are each historical recommendation information behavior. The node attributes of the behavior node include: recommendation information identifier and first recommendation information category. An edge between two behavior nodes indicates that the two behavior nodes have the same node attributes.

[0014] Generate target recommendation behavior sequences based on graph networks.

[0015] In one embodiment, generating a target recommendation behavior sequence based on a graph network includes:

[0016] Based on the behavior time of each historical recommendation information behavior, the initial node is determined in the graph network;

[0017] Based on the initial node, a second recommendation behavior sequence is obtained by random walk in the graph network. The information of each historical recommendation behavior in the second recommendation behavior sequence includes the recommendation information identifier of the historical recommendation behavior and the first recommendation information category.

[0018] Information from each historical recommendation behavior in the second recommendation behavior sequence is sampled and processed to obtain the target recommendation behavior sequence.

[0019] In one embodiment, determining the category of target recommendation information based on the target recommendation behavior sequence includes:

[0020] The feature vectors of each historical recommendation information behavior are obtained based on the target recommendation behavior sequence. The feature vectors include a first feature vector corresponding to the recommendation information identifier and a second feature vector corresponding to the first recommendation information category.

[0021] Calculate the first similarity between the first feature vector and the second feature vector of each historical recommendation information behavior;

[0022] The target recommendation information category is determined by the first similarity of the behavior of each historical recommendation information.

[0023] In one embodiment, the feature vectors of each historical recommendation information behavior are obtained based on the target recommendation behavior sequence, including:

[0024] Based on each first historical recommendation information behavior in the target recommendation behavior sequence, predict the recommendation information identifier of each second historical recommendation information behavior in the target recommendation behavior sequence. The first historical recommendation information behavior is the historical recommendation information behavior that includes the recommendation information identifier, and the second historical recommendation behavior is the historical recommendation information behavior that includes the first recommendation information category.

[0025] Based on each second historical recommendation information behavior in the target recommendation behavior sequence, predict the first recommendation information category for each first historical recommendation information behavior in the target recommendation behavior sequence;

[0026] Based on the recommendation information identifiers of each historical recommendation information behavior in the target recommendation behavior sequence, the first feature vector corresponding to each recommendation information identifier is obtained;

[0027] Based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence, the second feature vector corresponding to each first recommendation information category is obtained.

[0028] In one embodiment, the target recommendation information category is determined by the first similarity of the behaviors of each historical recommendation information, including:

[0029] If the first similarity of the historical recommendation information behavior is greater than or equal to the second similarity threshold, then the first recommendation information category of the historical recommendation information behavior is determined as the target recommendation information category.

[0030] In one embodiment, determining the target recommendation information category based on the first similarity of various historical recommendation information behaviors further includes:

[0031] If the first similarity of historical recommendation information behavior is less than the second similarity threshold, then the second similarity between the first feature vector of historical recommendation information behavior and the category feature vector of each recommendation information category is determined.

[0032] If the recommended information category corresponding to the second similarity with the largest value among all second similarities is consistent with the first recommended information category, then the first recommended information category is determined as the target recommended information category.

[0033] In one embodiment, determining the target recommendation information category based on the first similarity of various historical recommendation information behaviors further includes:

[0034] If the category of the recommended information corresponding to the second similarity with the largest value among all the second similarities is inconsistent with the category of the first recommended information, calculate the similarity difference between the first similarity and the largest second similarity.

[0035] If the similarity difference is greater than or equal to the maximum similarity difference, and the maximum second similarity is greater than the first similarity, then the recommendation information category corresponding to the maximum second similarity is determined as the target recommendation information category.

[0036] In one embodiment, determining the target recommendation information category based on the first similarity of various historical recommendation information behaviors further includes:

[0037] If the similarity difference is less than the minimum similarity difference, then the second recommendation information category is obtained and determined as the target recommendation information category.

[0038] In one embodiment, the method further includes:

[0039] Based on the first recommendation information category of each historical recommendation information behavior and the target recommendation information category, the model parameters of the recall ranking model are updated. The recall ranking model is used to identify the recommendation information category of the recommendation information behavior.

[0040] Secondly, this application also provides a device for determining recommendation information. The device includes:

[0041] The acquisition module is used to acquire a target recommendation behavior sequence based on the target object's historical recommendation information behavior within a preset time period, and the information of each historical recommendation information behavior in the target recommendation behavior sequence, including the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories;

[0042] The determination module is used to determine the target recommendation information category based on the target recommendation behavior sequence; determine the similar recommendation information category that is similar to the target recommendation information category; and determine the recommendation information corresponding to the target recommendation information category and the similar recommendation information category as the information to be recommended.

[0043] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0044] Based on the target object's historical recommendation information behavior within a preset time period, obtain the target recommendation behavior sequence, and the information of each historical recommendation information behavior in the target recommendation behavior sequence, including the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories;

[0045] Determine the category of target recommendation information based on the target recommendation behavior sequence;

[0046] Identify similar recommendation categories that are similar to the target recommendation category;

[0047] The recommended information corresponding to the target recommendation information category and the similar recommendation information category is determined as the information to be recommended.

[0048] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0049] Based on the target object's historical recommendation information behavior within a preset time period, obtain the target recommendation behavior sequence, and the information of each historical recommendation information behavior in the target recommendation behavior sequence, including the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories;

[0050] Determine the category of target recommendation information based on the target recommendation behavior sequence;

[0051] Identify similar recommendation categories that are similar to the target recommendation category;

[0052] The recommended information corresponding to the target recommendation information category and the similar recommendation information category is determined as the information to be recommended.

[0053] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0054] Based on the target object's historical recommendation information behavior within a preset time period, obtain the target recommendation behavior sequence, and the information of each historical recommendation information behavior in the target recommendation behavior sequence, including the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories;

[0055] Determine the category of target recommendation information based on the target recommendation behavior sequence;

[0056] Identify similar recommendation categories that are similar to the target recommendation category;

[0057] The recommended information corresponding to the target recommendation information category and the similar recommendation information category is determined as the information to be recommended.

[0058] The aforementioned method, apparatus, computer equipment, storage medium, and computer program product for determining recommendation information acquire a target recommendation behavior sequence based on the target object's historical recommendation behavior within a preset time period. The information for each historical recommendation behavior in the target recommendation behavior sequence includes a recommendation information identifier and any one of a first recommendation information category. Based on the target recommendation behavior sequence, a target recommendation information category is determined. Similar recommendation information categories similar to the target recommendation information category are determined. The recommendation information corresponding to the target recommendation information category and the similar recommendation information categories is identified as the information to be recommended. By determining the target recommendation information category through the target recommendation behavior sequence obtained from historical recommendation behavior, the accuracy of the determined target recommendation information category is ensured because the information for each historical recommendation behavior includes a recommendation information identifier or a first recommendation information category. Furthermore, similar recommendation information categories similar to the target recommendation information category are considered. Therefore, the determined information to be recommended can consider recommendation information from the target recommendation information category determined based on historical recommendation behavior, as well as recommendation information from similar recommendation information categories, thereby improving both the accuracy and generalizability of the recommendation information. Attached Figure Description

[0059] Figure 1 This is a diagram illustrating the application environment of a method for determining recommendation information in one embodiment.

[0060] Figure 2 This is a flowchart illustrating a method for determining recommendation information in one embodiment;

[0061] Figure 3 This is a schematic diagram of the recommendation information push process interface in one embodiment;

[0062] Figure 4 This is a flowchart illustrating the process of obtaining a target recommendation behavior sequence in one embodiment;

[0063] Figure 5 This is a flowchart illustrating the process of obtaining the target recommendation behavior sequence in another embodiment;

[0064] Figure 6 This is a schematic diagram of constructing a graph network based on behaviors related to historical recommendation information in one embodiment;

[0065] Figure 7 This is a schematic diagram of a process for generating target recommendation behavior sequences based on graph networks in one embodiment;

[0066] Figure 8 This is a schematic diagram of a random walk in a graphical network in one embodiment;

[0067] Figure 9This is a flowchart illustrating the process of determining the category of target recommendation information based on the target recommendation behavior sequence in one embodiment;

[0068] Figure 10 This is a flowchart illustrating the process of obtaining feature vectors for each historical recommendation behavior based on a target recommendation behavior sequence in one embodiment.

[0069] Figure 11 This is a flowchart illustrating the process of determining the target recommendation information category based on the first similarity of the behaviors of each historical recommendation information in one embodiment.

[0070] Figure 12 This is a partial flowchart illustrating the process of determining the target recommendation information category in one embodiment;

[0071] Figure 13 This is a partial flowchart illustrating the process of determining the target recommendation information category in another embodiment;

[0072] Figure 14 This is a partial flowchart illustrating the process of determining the target recommendation information category in another embodiment;

[0073] Figure 15 This is a schematic diagram illustrating the process of updating model parameters for a recall ranking model in one embodiment.

[0074] Figure 16 This is a schematic diagram of the overall process of determining recommendation information in one embodiment;

[0075] Figure 17 This is a structural block diagram of a device for determining recommendation information in one embodiment;

[0076] Figure 18 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0078] Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to achieve data computation, storage, processing, and sharing. Cloud technology is a collective term for network technology, information technology, integration technology, management platform technology, and application technology applied based on the cloud computing business model. It can form resource pools, be used on demand, and is flexible and convenient. Cloud computing technology will become a crucial support. Backend services of technical network systems require substantial computing and storage resources, such as video websites, image websites, and many portal websites. With the rapid development and application of the internet industry, every item may have its own identification mark in the future, requiring transmission to backend systems for logical processing. Data of different levels will be processed separately, and various industry data will require robust system support, which can only be achieved through cloud computing.

[0079] The solution provided in this application relates to Artificial Intelligence as a Service (AIaaS) in cloud technology. AIaaS is currently a mainstream service model for artificial intelligence platforms. Specifically, AIaaS platforms break down several common AI services and provide them as independent or packaged services in the cloud. This service model is similar to opening an AI-themed marketplace: all developers can access and use one or more AI services provided by the platform through API interfaces. Some experienced developers can also use the AI ​​framework and AI infrastructure provided by the platform to deploy and maintain their own dedicated cloud AI services. The solution in this application is specifically illustrated through the following embodiments:

[0080] The method for determining recommendation information provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. The data storage system can store the data that server 104 needs to process, such as the target object's historical recommendation information behavior and various recommendation information within a preset time period. The data storage system can be integrated on server 104, or it can be placed in the cloud or on other servers.

[0081] Specifically, before determining recommendation information, a recall and ranking model needs to be trained. Terminal 102 can send a training instruction to server 104 to enable server 104 to train the model, or server 104 can directly start model training; this is not limited here. When it is necessary to determine recommendation information for a target object, server 104 can obtain historical recommendation information behaviors of the target object within a preset time period and identify the recommendation information category of each historical recommendation information behavior using the trained recall and ranking model obtained by server 104. Thus, server 104 obtains a target recommendation behavior sequence based on each historical recommendation information behavior, determines the target recommendation information category based on the target recommendation behavior sequence, and then determines similar recommendation information categories similar to the target recommendation information category. Therefore, server 104 determines the recommendation information corresponding to the target recommendation information category and the similar recommendation information categories as the information to be recommended.

[0082] The terminal 102 can be, but is not limited to, a desktop computer, laptop computer, smartphone, tablet computer, IoT device, and portable wearable device. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0083] Furthermore, the method for determining recommendation information provided in this application embodiment can be applied to scenarios such as ad push, text message push, program download push, and video message push. Taking the scenario of ad push as an example, if the historical recommendation information behavior is the target object's historical ad browsing behavior, it is possible to determine the specific product categories included in the browsed historical ads. Based on the method for determining recommendation information provided in this application embodiment, the target object's historical product categories are determined, as well as similar product categories. The ad information corresponding to the product categories and similar product categories is then determined as the information to be recommended. Finally, the ad information to be pushed to the target user is determined from the determined information to be recommended.

[0084] Secondly, taking a scenario applied to text information push as an example, if the historical recommendation information behavior is the target object's historical text information browsing behavior, then the specific text categories included in the browsed historical text information can be determined. Based on the method for determining recommendation information provided in this application, the target object's browsed historical text categories can be determined, along with similar text categories. The text information corresponding to the text categories and similar text categories is then determined as the information to be recommended. Finally, the text information to be pushed to the target user is determined from the determined information to be recommended. It should be understood that the foregoing examples are only for understanding this solution and should not be construed as limiting this solution.

[0085] The method for determining recommendation information will be described in detail below. In one embodiment, such as... Figure 2 As shown, a method for determining recommendation information is provided, which can be applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0086] Step 202: Based on the target object's historical recommendation information behavior within a preset time period, obtain the target recommendation behavior sequence, including the recommendation information identifier of the historical recommendation information behavior and any item in the first recommendation information category.

[0087] The target audience refers to the individuals to whom the recommended information will be pushed. Historical recommended information behavior refers to the target audience's actions related to the recommended information, such as clicking, browsing, saving, and sharing it; this is not limited here. Based on the aforementioned application scenarios, the recommended information is determined according to the application scenario and may include, but is not limited to, advertisements, text messages, video content, and app download information.

[0088] Furthermore, historical recommendation information behavior has a corresponding recommendation information identifier and a first recommendation information category. The recommendation information identifier is the identifier corresponding to the recommendation information; for example, recommendation information A1 corresponds to recommendation information identifier 1, and recommendation information A2 corresponds to recommendation information identifier 2.

[0089] Secondly, the first recommendation information category describes the specific category of the recommendation information, and the target recommendation behavior sequence is represented by the recommendation information category identifier in the target recommendation behavior sequence. Taking the scenario applied to advertising push as an example, the first recommendation information category is explained as follows: In the field of product advertising push, the standard category system is a unified product classification system established to unify different product classification standards. It is divided into first-level, second-level, and third-level categories, and each product in each product library belongs to only one sub-category of the product system. The recommendation information category mentioned in the embodiments of this application is the product classification based on the standard category system. For example, the specific category of recommendation information can be clothing or electronic products. The specific category of recommendation information can also be women's clothing, men's clothing, and children's clothing included in clothing. Similarly, the specific category of recommendation information can also be dresses, skirts, and trousers included in women's clothing. It should be understood that the aforementioned products can be virtual products or physical products. For example, products can be electronic products or virtual electronic coupons, etc., without limitation here.

[0090] Specifically, server 104 first obtains the historical recommendation information behaviors of the target object within a preset time period, and then identifies the first recommendation information category of each historical recommendation information behavior through the trained recall and ranking model obtained by server 104. That is, the first recommendation information category is identified based on the trained recall and ranking model. The aforementioned preset time period can be 3 days, 7 days, 15 days, etc., and the specific preset time period needs to be flexibly determined according to the actual situation.

[0091] Based on this, server 104 obtains the target recommendation behavior sequence based on each historical recommendation information behavior. At this time, the information of each historical recommendation information behavior in the target recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information category. For example, the target object includes historical recommendation information behavior B1 for recommendation information A1, historical recommendation information behavior B2 for recommendation information A2, and historical recommendation information behavior B3 for recommendation information A2 within a preset time period. Recommendation information A1 corresponds to recommendation information identifier 1 and the first recommendation information category C1, recommendation information A2 corresponds to recommendation information identifier 2 and the first recommendation information category C2, and recommendation information A3 corresponds to recommendation information identifier 3 and the first recommendation information category C3. Thus, the target recommendation behavior sequence {recommendation information identifier 1, first recommendation information category C2, recommendation information identifier 3} can be obtained. That is, the information of each historical recommendation information behavior in the target recommendation behavior sequence can include either the recommendation information identifier of the historical recommendation information behavior or the first recommendation information category of the historical recommendation information behavior.

[0092] Step 204: Determine the target recommendation information category based on the target recommendation behavior sequence.

[0093] The target recommendation information category includes the final category of each historical recommendation information behavior. The target recommendation information category may be a different recommendation information category from the first recommendation information category.

[0094] Specifically, server 104 needs to determine whether the first recommendation information category identified by the trained recall and ranking model can accurately describe the specific recommendation information corresponding to each historical recommendation information behavior. Based on the determination result, the target recommendation information category is determined; that is, if the determination result accurately describes the first recommendation information category, then the first recommendation information category can be determined as the target recommendation information category. Conversely, if the determination result is not accurate, server 104 needs to adjust the recommendation information category based on subsequent steps to obtain a more accurate recommendation information category, and then determine that as the target recommendation information category.

[0095] Step 206: Determine similar recommendation information categories that are similar to the target recommendation information category.

[0096] The similarity recommendation information category is defined as the category whose similarity to the target recommendation information category is greater than a first similarity threshold. The first similarity threshold can be 0.9 or 0.95, etc., and is not limited here.

[0097] Specifically, server 104 calculates the similarity between the target recommendation category and each recommendation category, and identifies recommendation categories with similarity scores greater than a first similarity threshold as similar recommendation categories. Furthermore, the server can sort the recommendation categories with similarity scores greater than the first similarity threshold from highest to lowest based on their similarity scores, and determine a specific number of similar recommendation categories based on specific application requirements. For example, if the target recommendation category is women's dresses, then similar recommendation categories could be women's tops, or women's skirts, etc.

[0098] Step 208: Determine the recommended information corresponding to the target recommended information category and the similar recommended information category as the information to be recommended.

[0099] The information to be recommended includes recommended information corresponding to the target recommended information category, as well as recommended information corresponding to similar recommended information categories.

[0100] Specifically, server 104 retrieves various recommendation information from the data storage system. These recommendation information pieces form a recommendation information database, and each piece of recommendation information has a corresponding recommendation information category. Based on this, server 104 retrieves recommendation information for the target recommendation information category from the various recommendation information pieces. The recommendation information for the target recommendation information category can be: all recommendation information under the target recommendation information category. Alternatively, the recommendation information for the target recommendation information category can also be: all recommendation information under the target recommendation information category sorted from largest to smallest based on criteria such as click-through rate / favorite rate / share count, and the required number of recommendation information determined by application requirements. Similarly, server 104 retrieves recommendation information for similar recommendation information categories from the various recommendation information pieces, and retrieves the recommendation information corresponding to similar recommendation information categories using a similar method described above.

[0101] In practical applications, after server 104 determines the information to be recommended, it pushes either the information to be recommended or a portion of the information to be recommended when the target user visits again. For example, in the scenario of advertising push, the advertising information corresponding to the determined product category and similar product categories is determined as the information to be recommended, and then the final advertising information to be pushed to the target user is determined from the determined information to be recommended. Secondly, in the scenario of text information push, the text information corresponding to the determined text category and similar text categories is determined as the information to be recommended, and then the final text information to be pushed to the target user is determined from the determined information to be recommended.

[0102] For example, taking an application advertising scenario, where the target recommended information category is women's dresses, and similar recommended information categories are women's tops and women's skirts, this will be explained as follows: Figure 3 As shown, when a user visits the product interface again, advertising information will be pushed to the product interface. Specifically, advertising information 301 corresponding to the target recommendation information category of women's skirt will be pushed, and advertising information 302 corresponding to the similar recommendation information category of women's top and advertising information 303 corresponding to the similar recommendation information category of women's skirt will also be pushed.

[0103] In the aforementioned method for determining recommendation information, the target recommendation information category is determined through the sequence of target recommendation behaviors obtained from historical recommendation behavior. Since the information of each historical recommendation behavior includes a recommendation information identifier or a first recommendation information category, the accuracy of the determined target recommendation information category is guaranteed. Furthermore, similar recommendation information categories that are similar to the target recommendation information category are considered. Therefore, the determined recommendation information can consider recommendation information based on the target recommendation information category determined from historical recommendation behavior, as well as recommendation information from similar recommendation information categories, thereby improving both the accuracy and generalizability of the recommendation information.

[0104] In one embodiment, such as Figure 4 As shown, step 202 involves obtaining a target recommendation behavior sequence based on the target object's historical recommendation information behavior within a preset time period, including:

[0105] Step 402: Based on the behavior time of each historical recommendation behavior, sort the historical recommendation behavior to obtain the first recommendation behavior sequence. The information of each historical recommendation behavior in the first recommendation behavior sequence includes the recommendation information identifier of the historical recommendation behavior and the first recommendation information category.

[0106] The "behavior time" describes the time when a target object performs an action related to the recommended information. For example, if a target object clicks on recommended information A1 at time T1, resulting in historical recommended information behavior B1, then the behavior time of historical recommended information behavior B1 is time T1. Similarly, if a target object shares recommended information A2 at time T2, resulting in historical recommended information behavior B2, then the behavior time of historical recommended information behavior B2 is time T2.

[0107] Specifically, when the target user's actions related to the recommended information are recorded to generate historical recommended information behavior, the time of occurrence of each action can also be recorded, and each historical recommended information behavior has a corresponding time. Based on this, server 104 first obtains each historical recommended information behavior of the target user within a preset time period, and can obtain the corresponding time of each historical recommended information behavior. Then, server 104 sorts each historical recommended information behavior of the target user within the preset time period according to the time of each historical recommended information behavior to obtain the first recommended behavior sequence.

[0108] For example, taking a preset time period of 3 days as an example, if the target object performs related operations on recommendation information A1, recommendation information A2, and recommendation information A3 within 3 days, we can obtain the historical recommendation behavior B1 corresponding to recommendation information A1, the historical recommendation behavior B2 corresponding to recommendation information A2, and the historical recommendation behavior B3 corresponding to recommendation information A3. Based on this, if historical recommendation behavior B1 corresponds to recommendation information identifier 1 and the first recommendation information category C1, and the specific behavior time is T1; historical recommendation behavior B2 corresponds to recommendation information identifier 2 and the first recommendation information category C2, and the specific behavior time is T2; and historical recommendation behavior B3 corresponds to recommendation information identifier 3 and the first recommendation information category C3, and the specific behavior time is T3, where T1 is earlier than T2, and T2 is earlier than T3, we can obtain the first recommendation behavior sequence {(recommendation information identifier 1, first recommendation information category C1), (recommendation information identifier 2, first recommendation information category C2), (recommendation information identifier 3, first recommendation information category C3)}.

[0109] At this point, server 104 has obtained the structured first recommendation behavior sequence. Since the information of each historical recommendation behavior in the first recommendation behavior sequence, including the recommendation information identifier and the first recommendation information category, retains all the information of the historical recommendation behavior, it is equivalent to adding samples for the subsequent model. This makes it easy to learn redundant information, thereby reducing the efficiency of model training. Therefore, this embodiment will make further improvements based on this.

[0110] Step 404: Sample and process the information of each historical recommendation behavior in the first recommendation behavior sequence to obtain the target recommendation behavior sequence.

[0111] Specifically, server 104 samples the information of each historical recommendation behavior in the first recommendation behavior sequence according to a sampling ratio, obtaining the historical recommendation behavior corresponding to the sampling ratio, and taking the recommendation information identifier corresponding to each historical recommendation behavior as the information of that historical recommendation behavior in the target recommendation behavior sequence. Based on this, the first recommendation information category of other historical recommendation behaviors is taken as the information of that historical recommendation behavior in the target recommendation behavior sequence. It should be understood that the sampling process is only used to sample the information of each historical recommendation behavior, and does not change the order of the behavior time corresponding to each historical recommendation behavior in the recommendation behavior sequence.

[0112] For example, assuming a sampling ratio of X%, x% of historical recommendation information behaviors are randomly selected, and the recommendation information identifiers of x% of these behaviors are retained. Next, the first recommendation information category of another (1-x%) of historical recommendation information behaviors is retained, thus obtaining the target recommendation behavior sequence. Specifically, this is determined through a hyperparameter search, where X is 80. That is, the target recommendation behavior sequence includes the recommendation information identifiers of 80% of historical recommendation information behaviors and the first recommendation information category of 20% of historical recommendation information behaviors.

[0113] In this embodiment, historical recommendation behaviors are sorted by their timing to ensure the accuracy of the temporal sequence. Furthermore, information from each historical recommendation behavior is sampled and processed to reduce redundancy while preserving historical information, thereby improving the efficiency of subsequent model training and ultimately enhancing the efficiency of recommending information.

[0114] It is understandable that historical recommendation information behavior includes other rich behavioral information in addition to time and type information. In order to avoid wasting behavioral information, the following will introduce how to consider the correlation of behavioral information of various historical recommendation information behaviors from multiple dimensions, so as to obtain the target recommendation behavior sequence.

[0115] In one embodiment, such as Figure 5 As shown, step 202 involves obtaining a target recommendation behavior sequence based on the target object's historical recommendation information behavior within a preset time period, including:

[0116] Step 502: Construct a graph network based on each historical recommendation information behavior. The nodes of the graph network are each historical recommendation information behavior. The node attributes of the behavior node include: recommendation information identifier and first recommendation information category. An edge between two behavior nodes indicates that the two behavior nodes have the same node attributes.

[0117] Specifically, after obtaining each historical recommendation information behavior, the server 104 constructs a graph network based on the recommendation information identifier, first recommendation information category, and other behavioral information of each historical recommendation information behavior. For example, taking the scenario of advertising push again, the historical recommendation information behavior may also include other behavioral information such as advertiser identifier, industry identifier, and application material identifier.

[0118] Therefore, the graph network includes behavior nodes corresponding to each historical recommendation behavior. Each behavior node's attributes include: the recommendation information identifier of the corresponding historical recommendation behavior and the first recommendation information category. It should be understood that each behavior node's attributes may also include: the behavior time of the corresponding historical recommendation behavior and other behavior information. For example, in an advertising push scenario, each behavior node's attributes may also include: advertiser identifier, industry identifier, and application material identifier, etc. In a video push scenario, each behavior node's attributes may also include: video producer identifier, video publisher identifier, and video material identifier, etc. Based on this, when two behavior nodes have the same node attributes, there will be an edge connecting the two behavior nodes.

[0119] For ease of understanding, we will use historical recommendation behavior B1 corresponding to recommendation information A1, historical recommendation behavior B2 corresponding to recommendation information A2, historical recommendation behavior B3 corresponding to recommendation information A3, and historical recommendation behavior B4 corresponding to recommendation information A4 as examples for illustration.

[0120] Specifically, historical recommendation behavior B1 corresponds to recommendation information identifier 1 and the first recommendation information category C1, with the corresponding behavior time T1 and behavior information D1. Historical recommendation behavior B2 corresponds to recommendation information identifier 2 and the first recommendation information category C2, with the corresponding behavior time T2 and behavior information D1. Historical recommendation behavior B3 corresponds to recommendation information identifier 3 and the first recommendation information category C3, with the corresponding behavior time T2 and behavior information D2. Historical recommendation behavior B4 corresponds to recommendation information identifier 4 and the first recommendation information category C3, with the corresponding behavior time T4 and behavior information D1.

[0121] Based on this, such as Figure 6As shown, behavior node 601 corresponds to historical recommendation behavior B1, behavior node 602 corresponds to historical recommendation behavior B2, behavior node 603 corresponds to historical recommendation behavior B3, and behavior node 604 corresponds to historical recommendation behavior B4. As illustrated in the previous example, the behavior information D1 of historical recommendation behaviors B1, B2, and B4 is the same, meaning that behavior nodes 601, 602, and 604 share the same node attribute (behavior information D1). Therefore, behavior nodes 601, 602, and 604 are connected by edges.

[0122] Similarly, the historical recommendation information behavior B2 and historical recommendation information behavior B3 have the same behavior time T2, meaning that behavior node 602 and behavior node 603 have the same node attribute (behavior time T2). Therefore, behavior node 602 and behavior node 603 are connected by an edge. Furthermore, since the first recommendation information category C3 of historical recommendation information behavior B3 and historical recommendation information behavior B4 is the same, meaning that behavior node 603 and behavior node 604 have the same node attribute (first recommendation information category C3), behavior node 603 and behavior node 604 are connected by an edge. Thus, we obtain the following... Figure 6 The graph network shown is shown.

[0123] Step 504: Generate a target recommendation behavior sequence based on the graph network.

[0124] Specifically, server 104 obtains the target recommended behavior sequence by using a random walk method in the graph network based on the node attributes of each behavior node and the relationship between the connecting edges between the nodes.

[0125] In this embodiment, a graph network is constructed to ensure that the correlation of behavioral information of each historical recommendation behavior can be considered from multiple dimensions during the generation of the target recommendation behavior sequence, so as to avoid wasting behavioral information and thus ensure the accuracy of information in the target recommendation behavior sequence.

[0126] The following section details how to obtain the target recommendation behavior sequence in a graph network by using random walks, based on the node attributes of each behavior node and the relationship between the edges connecting the nodes.

[0127] In one embodiment, such as Figure 7 As shown, step 504, generating a target recommendation behavior sequence based on a graph network, includes:

[0128] Step 702: Determine the initial node in the graph network based on the behavior time of each historical recommendation information behavior.

[0129] The initial node is the behavior node corresponding to the initial position in the random walk method.

[0130] Specifically, server 104 determines the behavior node corresponding to the earliest historical recommendation behavior among the behavior times of each historical recommendation behavior as the initial node.

[0131] Step 704: Based on the initial node, perform a random walk in the graph network to obtain the second recommendation behavior sequence, including the information of each historical recommendation behavior in the second recommendation behavior sequence, including the recommendation information identifier of the historical recommendation behavior and the first recommendation information category.

[0132] Specifically, server 104 traverses the graph network starting from the initial node, and walks to behavior nodes that are connected to the initial node by random walk. Then it continues to walk in a similar manner, and records the historical recommendation information of the behavior node at each walk, thereby generating the second recommendation behavior sequence.

[0133] For ease of understanding, based on Figure 6 For further explanation of the graph network shown, please refer to [link / reference]. Figure 8 If action time T1 precedes action time T2, and action time T2 precedes action time T3, then based on the aforementioned action times T1 to T3, action node 601 can be determined as the initial node, and then the graph network can be traversed starting from action node 601.

[0134] Specifically, such as Figure 8 As shown in Figure (A), server 104 travels from behavior node 601 to behavior node 602 with a connected edge, then from behavior node 602 to behavior node 603 with a connected edge, and finally from behavior node 603 to behavior node 604 with a connected edge, thus completing the traversal of the graph network and obtaining the second recommendation behavior sequence {(recommendation information identifier 1, first recommendation information category C1), (recommendation information identifier 2, first recommendation information category C2), (recommendation information identifier 3, first recommendation information category C3), (recommendation information identifier 4, first recommendation information category C3)}.

[0135] Secondly, in another mode of movement, specifically as follows: Figure 8As shown in Figure (B), server 104 can also traverse from behavior node 601 to behavior node 604 with a connected edge, then from behavior node 604 to behavior node 602 with a connected edge, and finally from behavior node 602 to behavior node 603 with a connected edge, thus completing the traversal of the graph network and obtaining the second recommendation behavior sequence {(recommendation information identifier 1, first recommendation information category C1), (recommendation information identifier 4, first recommendation information category C3), (recommendation information identifier 2, first recommendation information category C2), (recommendation information identifier 3, first recommendation information category C3)}. It should be understood that the foregoing examples are for understanding this scheme and should not be construed as limiting this scheme.

[0136] At this point, server 104 has obtained the structured second recommendation behavior sequence. Since the information of each historical recommendation behavior in the second recommendation behavior sequence, including the recommendation information identifier and the first recommendation information category, retains all the information of the historical recommendation behavior, it is equivalent to adding samples for the subsequent model. This makes it easy to learn redundant information, thus reducing the efficiency of model training. Therefore, similar to the aforementioned embodiments, further improvements are needed.

[0137] Step 706: Sample the information of each historical recommendation behavior in the second recommendation behavior sequence to obtain the target recommendation behavior sequence.

[0138] Specifically, server 104 samples the information of each historical recommendation behavior in the second recommendation behavior sequence to obtain the target recommendation behavior sequence. The specific sampling method and example are similar to step 404, and will not be repeated here.

[0139] In this embodiment, the graph network can consider the correlation of behavioral information of each historical recommendation behavior from multiple dimensions to avoid wasting behavioral information and thus ensure the accuracy of information in the target recommendation behavior sequence. Based on this, the behavioral information of historical recommendation behaviors is further retained, and the redundancy of the recommendation behavior sequence is reduced through sampling, thereby improving the efficiency of subsequent model training and thus improving the efficiency of determining recommendation information.

[0140] In one embodiment, such as Figure 9 As shown, step 204, determining the target recommendation information category based on the target recommendation behavior sequence, includes:

[0141] Step 902: Obtain the feature vector of each historical recommendation information behavior based on the target recommendation behavior sequence. The feature vector includes a first feature vector corresponding to the recommendation information identifier and a second feature vector corresponding to the first recommendation information category.

[0142] Specifically, in order to learn the feature vectors corresponding to the recommendation information identifier and the first recommendation information category, the server 104 adopts an unsupervised learning method based on the word2vec model. It predicts the recommendation information identifiers near the target recommendation behavior sequence based on the recommendation information identifier, and predicts the first recommendation information category near the target recommendation behavior sequence based on the first recommendation information category. Then, through the backward gradient update of the neural network, the feature vectors corresponding to the recommendation information identifier and the first recommendation information category are updated to the optimal vector space, thereby obtaining the feature vectors of each historical recommendation behavior. At this time, the feature vectors of each historical recommendation behavior include the first feature vector corresponding to the recommendation information identifier of the historical recommendation behavior and the second feature vector corresponding to the first recommendation information category.

[0143] Step 904: Calculate the first similarity between the first feature vector and the second feature vector of each historical recommendation information behavior.

[0144] The first similarity is used to describe the difference between the recommendation information indicated by the recommendation information identifier of historical recommendation information behavior and the first recommendation information category.

[0145] Specifically, server 104 calculates any one of the following: cosine distance, Euclidean distance, standardized Euclidean distance, or Mahalanobis distance between the first and second feature vectors of each historical recommendation information behavior, to obtain the first similarity between the first and second feature vectors.

[0146] Step 906: Determine the target recommendation information category based on the first similarity of the behavior of each historical recommendation information.

[0147] Specifically, server 104 determines whether the first recommendation information category can accurately describe the historical recommendation information behavior based on the first similarity of each historical recommendation information behavior, and then determines the target recommendation information category of the historical recommendation information behavior based on the judgment result.

[0148] In this embodiment, the feature vectors of each historical recommendation behavior can accurately describe the feature information of the recommendation information identifier and the first recommendation information category. Therefore, the obtained first similarity can more accurately reflect the difference between the recommendation information identifier and the first recommendation information category that reflects the historical recommendation behavior. Based on this, it can be more accurate to judge whether the first recommendation information category can accurately describe the historical recommendation behavior, thereby improving the accuracy of the determined target recommendation information category.

[0149] In one embodiment, such as Figure 10 As shown, step 902 involves obtaining the feature vectors of each historical recommendation behavior based on the target recommendation behavior sequence, including:

[0150] Step 1002: Based on each first historical recommendation information behavior in the target recommendation behavior sequence, predict the recommendation information identifier of each second historical recommendation information behavior in the target recommendation behavior sequence. The first historical recommendation information behavior is the historical recommendation information behavior that includes the recommendation information identifier, and the second historical recommendation information behavior is the historical recommendation information behavior that includes the first recommendation information category.

[0151] Since the information of each historical recommendation behavior in the target recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and any item in the first recommendation information category, when obtaining the feature vector of each historical recommendation information behavior, the feature vector needs to include the features of the behavior information of each dimension of the historical recommendation information behavior. At this time, processing is required to obtain the complete behavior information of each historical recommendation information behavior.

[0152] Specifically, historical recommendation behaviors including recommendation information identifiers are first defined as first historical recommendation behaviors, and historical recommendation behaviors including first recommendation information categories are defined as second historical recommendation behaviors. Therefore, the target recommendation behavior sequence includes information from each first historical recommendation behavior and information from each second historical recommendation behavior, wherein the information from the first historical recommendation behavior includes recommendation information identifiers, and the information from the second historical recommendation behavior includes first recommendation information categories.

[0153] Based on this, server 104 predicts the recommendation information identifier of each second historical recommendation information behavior according to the recommendation information identifier of each first historical recommendation information behavior in the target recommendation behavior sequence.

[0154] Step 1004: Based on each second historical recommendation information behavior in the target recommendation behavior sequence, predict the first recommendation information category for each first historical recommendation information behavior in the target recommendation behavior sequence.

[0155] Specifically, server 104 predicts the first recommendation information category of each first historical recommendation information behavior based on the first recommendation information category of each second historical recommendation information behavior in the target recommendation behavior sequence.

[0156] Step 1006: Based on the recommendation information identifiers of each historical recommendation information behavior in the target recommendation behavior sequence, obtain the first feature vector corresponding to each recommendation information identifier.

[0157] Specifically, server 104 obtains the first feature vector corresponding to each recommendation information identifier based on the recommendation information identifier of each historical recommendation information behavior in the target recommendation behavior sequence. That is, server 104 obtains the first feature vector corresponding to each recommendation information identifier based on the recommendation information identifier of each first historical recommendation information behavior and the predicted recommendation information identifier of each second historical recommendation information behavior.

[0158] Step 1008: Based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence, obtain the second feature vector corresponding to each first recommendation information category.

[0159] Specifically, server 104 obtains a second feature vector corresponding to each first recommendation information category based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence. That is, server 104 obtains a second feature vector corresponding to each first recommendation information category based on the first recommendation information category of each second historical recommendation information behavior and the predicted first recommendation information category of each first historical recommendation information behavior.

[0160] In this embodiment, the recommendation information identifiers near the target recommendation information identifier are predicted by the recommendation information identifier in the target recommendation behavior sequence, and the first recommendation information category near the first recommendation information category is predicted by the first recommendation information category in the target recommendation behavior sequence. This ensures that the obtained feature vector can include the features of various dimensions of historical recommendation information behavior, thereby improving the information integrity and reliability of the feature vector.

[0161] Since it is necessary to determine whether the first recommendation information category can accurately describe the historical recommendation information behavior, the following will describe the methods for determining the target recommendation information category under different judgment results.

[0162] In one embodiment, such as Figure 11 As shown, step 906 determines the target recommendation information category based on the first similarity of the behavior of each historical recommendation information, including:

[0163] Step 1102: If the first similarity of the historical recommendation information behavior is greater than or equal to the second similarity threshold, then the first recommendation information category of the historical recommendation information behavior is determined as the target recommendation information category.

[0164] The second similarity threshold is 0.9, which was determined based on a large amount of experimental data and practical application requirements.

[0165] Specifically, after obtaining the first similarity of historical recommendation information behavior, server 104 first compares the first similarity with a second similarity threshold. If the first similarity is greater than or equal to the second similarity threshold, it indicates that the difference between the first recommendation information category and the recommendation information indicated by the recommendation information identifier is small, meaning that the first recommendation information category can accurately describe the historical recommendation information behavior. Therefore, the first recommendation information category is directly determined as the target recommendation information category. It should be understood that since there are multiple historical recommendation information behaviors in actual applications, the target recommendation information category includes the recommendation information categories ultimately corresponding to each historical recommendation information behavior.

[0166] In this embodiment, by comparing the first similarity with the second similarity threshold, the first recommendation information category that can accurately describe the historical recommendation information behavior is directly determined as the target recommendation information category, thus ensuring the accuracy of the target recommendation information category and avoiding subsequent error correction adjustments.

[0167] The foregoing embodiments describe the situation where the first recommendation information category can accurately describe historical recommendation information behavior; conversely, the following will describe the situation where the first recommendation information category cannot accurately describe historical recommendation information behavior.

[0168] In one embodiment, such as Figure 12 As shown, step 906, determining the target recommendation information category based on the first similarity of each historical recommendation information behavior, further includes:

[0169] Step 1202: If the first similarity of historical recommendation information behavior is less than the second similarity threshold, then determine the second similarity between the first feature vector of historical recommendation information behavior and the category feature vector of each recommendation information category.

[0170] The second similarity is used to describe the difference between the recommendation information indicated by the recommendation information identifier of historical recommendation information behavior and each recommendation information category.

[0171] Specifically, as can be seen from the foregoing embodiments, the data storage system stores various recommendation information, and the various recommendation information forms a recommendation information library. Each recommendation information has a corresponding recommendation information category. Based on the recommendation information categories of each recommendation information in the recommendation information library, the server 104 can further obtain a recommendation information category library, and each recommendation information category in the recommendation information category library has a corresponding category feature vector.

[0172] Based on this, server 104 calculates any one of the following distances between the first feature vector and each recommendation information category: cosine distance, Euclidean distance, standardized Euclidean distance, or Mahalanobis distance.

[0173] Step 1204: If the recommended information category corresponding to the second similarity with the largest value among all second similarities is consistent with the first recommended information category, then the first recommended information category is determined as the target recommended information category.

[0174] Specifically, server 104 determines the recommendation information category corresponding to the second similarity with the largest value among the second similarity values ​​based on the first feature vector and each recommendation information category. That is, the recommendation information indicated by the recommendation information identifier of historical recommendation information behavior has the smallest difference with the recommendation information category corresponding to the largest value of the second similarity value. Therefore, it is determined that the recommendation information indicated by the recommendation information identifier of historical recommendation information behavior belongs to the recommendation information category corresponding to the largest value of the second similarity value. Thus, it is determined whether the recommendation information category is consistent with the first recommendation information category. If they are consistent, it means that the first recommendation information category may not have been misclassified, and the recommendation information may contain some influencing information. Therefore, the first recommendation information category is still determined as the target recommendation information category.

[0175] In this embodiment, when the first similarity indicates that the first recommendation information category cannot accurately describe the historical recommendation information behavior, the differences between the recommendation information indicated by the recommendation information identifier of the historical recommendation information behavior and each recommendation information category are further considered to avoid the result deviation caused by the influence of information and special information, thereby ensuring the accuracy of the target recommendation information category and improving the flexibility of determining the target recommendation information category.

[0176] In one embodiment, such as Figure 13 As shown, step 906, determining the target recommendation information category based on the first similarity of each historical recommendation information behavior, further includes:

[0177] Step 1302: If the category of the recommended information corresponding to the second similarity with the largest value among the second similarities is inconsistent with the first recommended information category, calculate the similarity difference between the first similarity and the largest second similarity.

[0178] Specifically, server 104 determines the recommendation information category corresponding to the second similarity with the largest value among the second similarities based on the first feature vector and each recommendation information category. That is, the recommendation information indicated by the recommendation information identifier of historical recommendation information behavior has the smallest difference with the recommendation information category corresponding to the largest value of the second similarity. Therefore, it is said that the recommendation information indicated by the recommendation information identifier of historical recommendation information behavior belongs to the recommendation information category corresponding to the largest value of the second similarity. Thus, it is determined whether the recommendation information category is consistent with the first recommendation information category. If they are inconsistent, it means that the first recommendation information category has been misclassified, or there may be misclassification in each recommendation information database. In this case, the similarity difference between the first similarity and the largest second similarity is further calculated.

[0179] For example, if the first similarity between the first feature vector and the second feature vector of historical recommendation information behavior is 0.5, and the second similarity with the largest value among all the second similarities is 0.8, then the similarity difference between the first similarity and the largest second similarity is 0.3.

[0180] Step 1304: If the similarity difference is greater than or equal to the maximum similarity difference, and the maximum second similarity is greater than the first similarity, then the recommendation information category corresponding to the maximum second similarity is determined as the target recommendation information category.

[0181] The maximum similarity difference is 0.3, which is determined based on a large amount of experimental data and practical application requirements.

[0182] Specifically, server 104 determines whether the similarity difference between the first similarity and the largest second similarity is greater than or equal to the largest similarity difference. If so, it indicates that the classification deviation between the recommended information category corresponding to the largest second similarity and the first recommended information category is very large. In this case, it is necessary to compare the size between the largest second similarity and the first similarity. If the largest second similarity is greater than the first similarity, it indicates that the first recommended information category is misclassified, while the recommended information category corresponding to the largest second similarity can more accurately describe the historical recommended information behavior. Therefore, the recommended information category corresponding to the largest second similarity is determined as the target recommended information category. Secondly, if the first similarity is greater than the largest second similarity, the first recommended information category is still determined as the target recommended information category.

[0183] In this embodiment, the target recommendation information category is determined by multiple judgments, avoiding category identification bias caused by special cases. While ensuring the accuracy of the target recommendation information category, the flexibility of determining the target recommendation information category is further improved.

[0184] In one embodiment, such as Figure 14 As shown, step 906, determining the target recommendation information category based on the first similarity of each historical recommendation information behavior, further includes:

[0185] Step 1402: If the similarity difference is less than the minimum similarity difference, then obtain the input second recommendation information category and determine the second recommendation information category as the target recommendation information category.

[0186] The minimum similarity difference can be 0.1 or 0.05, and this minimum similarity difference is determined based on a large amount of experimental data and practical application requirements.

[0187] Specifically, server 104 determines whether the similarity difference between the first similarity and the largest second similarity is less than the minimum similarity difference. If so, it means that the recommendation information category corresponding to the largest second similarity is similar to the recommendation information category described by the first recommendation information category. In this case, manual intervention is required, that is, by having annotators identify and label the categories of historical recommendation information behaviors and input the determined second recommendation information category. Thus, server 104 obtains the input second recommendation information category and determines the second recommendation information category as the target recommendation information category.

[0188] For example, the second similarity score with the highest value is 0.6, and the corresponding recommendation information category is women's trousers. The first similarity score is 0.56, and the first recommendation information category is women's skirts. In this case, the first recommendation information category is very similar to the recommendation information category corresponding to the highest second similarity score. At this time, the labeler identifies and labels the category of historical recommendation information behavior, and the second recommendation information category is women's skirts and trousers. Thus, the second recommendation information category (women's skirts and trousers) is determined as the target recommendation information category.

[0189] In this embodiment, the category of the target recommendation information is determined by multiple judgments to avoid category identification bias caused by special cases. Furthermore, when the similarity difference is small, it is determined by manual intervention to further ensure the accuracy of the target recommendation information category.

[0190] In one embodiment, such as Figure 15 As shown, other methods for determining recommendation information include:

[0191] Step 1502: Based on the first recommendation information category of each historical recommendation information behavior and the target recommendation information category, update the model parameters of the recall ranking model. The recall ranking model is used to identify the recommendation information category of the recommendation information behavior.

[0192] Specifically, based on the adjustment of the first recommendation information category in the aforementioned embodiments, after determining the target recommendation information category, the model parameters of the recall ranking model are updated based on the corrected target recommendation information category and the first recommendation information category, so as to ensure that the recall ranking model can learn more accurate category features.

[0193] In this embodiment, by updating the model parameters of the recall ranking model, the recall ranking model can learn more accurate category features, thereby making the process of the recall ranking model to identify the category of recommendation information behavior more accurate.

[0194] The methods described above for obtaining the target recommendation behavior sequence, determining the target recommendation information category, and similar recommendation information categories will be explained below, using the aforementioned embodiments as an example to illustrate the overall process of determining recommendation information, such as... Figure 16 As shown, a method for determining recommendation information is provided, which can be applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0195] Step 1602: Construct a graph network based on the behavior of each historical recommendation information.

[0196] Specifically, after obtaining each historical recommendation information behavior, the server 104 constructs a graph network based on the recommendation information identifier, first recommendation information category, and other behavioral information of each historical recommendation information behavior. For example, taking the scenario of advertising push again, the historical recommendation information behavior may also include advertiser identifier, industry identifier, and application material identifier, etc.

[0197] Therefore, the graph network includes behavior nodes corresponding to each historical recommendation behavior. Thus, the node attributes of each behavior node include: the recommendation information identifier of the corresponding historical recommendation behavior and the first recommendation information category. It should be understood that the node attributes of each behavior node may also include: the behavior time of the corresponding historical recommendation behavior and other behavior information.

[0198] Step 1604: Determine the initial node in the graph network based on the behavior time of each historical recommendation information behavior.

[0199] The initial node is the behavior node corresponding to the initial position in the random walk method.

[0200] Specifically, server 104 determines the behavior node corresponding to the earliest historical recommendation behavior among the behavior times of each historical recommendation behavior as the initial node.

[0201] Step 1606: Based on the initial node, perform a random walk in the graph network to obtain the second recommended behavior sequence.

[0202] Specifically, server 104 traverses the graph network starting from the initial node, and walks to behavior nodes that are connected to the initial node by random walk. Then it continues to walk in a similar manner, and records the historical recommendation information of the behavior node at each walk, thereby generating the second recommendation behavior sequence.

[0203] Step 1608: Sample and process the information of each historical recommendation behavior in the second recommendation behavior sequence to obtain the target recommendation behavior sequence.

[0204] Specifically, server 104 samples the information of each historical recommendation behavior in the second recommendation behavior sequence to obtain the target recommendation behavior sequence. The specific sampling method and example are similar to step 404, and will not be repeated here.

[0205] Step 1610: Based on each first historical recommendation information behavior in the target recommendation behavior sequence, predict the recommendation information identifier of each second historical recommendation information behavior in the target recommendation behavior sequence.

[0206] Specifically, historical recommendation behaviors including recommendation information identifiers are first defined as first historical recommendation behaviors, and historical recommendation behaviors including first recommendation information categories are defined as second historical recommendation behaviors. Therefore, the target recommendation behavior sequence includes information from each first historical recommendation behavior and information from each second historical recommendation behavior, wherein the information from the first historical recommendation behavior includes recommendation information identifiers, and the information from the second historical recommendation behavior includes first recommendation information categories.

[0207] Based on this, server 104 predicts the recommendation information identifier of each second historical recommendation information behavior according to the recommendation information identifier of each first historical recommendation information behavior in the target recommendation behavior sequence.

[0208] Step 1612: Based on each second historical recommendation information behavior in the target recommendation behavior sequence, predict the first recommendation information category of each first historical recommendation information behavior in the target recommendation behavior sequence.

[0209] Specifically, server 104 predicts the first recommendation information category of each first historical recommendation information behavior based on the first recommendation information category of each second historical recommendation information behavior in the target recommendation behavior sequence.

[0210] Step 1614: Based on the recommendation information identifiers of each historical recommendation information behavior in the target recommendation behavior sequence, obtain the first feature vector corresponding to each recommendation information identifier.

[0211] Specifically, server 104 obtains the first feature vector corresponding to each recommendation information identifier based on the recommendation information identifier of each historical recommendation information behavior in the target recommendation behavior sequence. That is, server 104 obtains the first feature vector corresponding to each recommendation information identifier based on the recommendation information identifier of each first historical recommendation information behavior and the predicted recommendation information identifier of each second historical recommendation information behavior.

[0212] Step 1616: Based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence, obtain the second feature vector corresponding to each first recommendation information category.

[0213] Specifically, server 104 obtains a second feature vector corresponding to each first recommendation information category based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence. That is, server 104 obtains a second feature vector corresponding to each first recommendation information category based on the first recommendation information category of each second historical recommendation information behavior and the predicted first recommendation information category of each first historical recommendation information behavior.

[0214] Step 1618: Calculate the first similarity between the first feature vector and the second feature vector of each historical recommendation information behavior.

[0215] The first similarity is used to describe the difference between the recommendation information indicated by the recommendation information identifier of historical recommendation information behavior and the first recommendation information category.

[0216] Specifically, server 104 calculates any one of the following: cosine distance, Euclidean distance, standardized Euclidean distance, or Mahalanobis distance between the first and second feature vectors of each historical recommendation information behavior, to obtain the first similarity between the first and second feature vectors.

[0217] Step 1620: Determine the target recommendation information category based on the first similarity of the behavior of each historical recommendation information.

[0218] Specifically, server 104 determines whether the first recommendation information category can accurately describe the historical recommendation information behavior based on the first similarity of each historical recommendation information behavior, and then determines the target recommendation information category of the historical recommendation information behavior based on the judgment result.

[0219] The method for determining the category of target recommendation information under different judgment results is similar to the aforementioned embodiments, and will not be repeated here.

[0220] Step 1622: Determine similar recommendation information categories that are similar to the target recommendation information category.

[0221] The similarity recommendation information category is defined as the category whose similarity to the target recommendation information category is greater than a first similarity threshold. The first similarity threshold can be 0.9 or 0.95, etc., and is not limited here.

[0222] Specifically, server 104 calculates the similarity between the target recommendation information category and each recommendation information category, and determines the recommendation information category with a similarity greater than a first similarity threshold as a similar recommendation information category.

[0223] Step 1624: Determine the recommended information corresponding to the target recommendation information category and the similar recommendation information category as the information to be recommended.

[0224] The information to be recommended includes recommended information corresponding to the target recommended information category, as well as recommended information corresponding to similar recommended information categories.

[0225] Specifically, server 104 retrieves each recommendation information from the data storage system, and the recommendation information forms a recommendation information database, with each recommendation information having a corresponding recommendation information category.

[0226] Step 1626: Update the model parameters of the recall ranking model based on the first recommendation information category of each historical recommendation information behavior and the determined target recommendation information category.

[0227] Specifically, based on the adjustment of the first recommendation information category in the aforementioned embodiments, after determining the target recommendation information category, the model parameters of the recall ranking model are updated based on the corrected target recommendation information category and the first recommendation information category, so as to ensure that the recall ranking model can learn more accurate category features.

[0228] It should be understood that the specific implementation methods of steps 1602 to 1626 are similar to those of the aforementioned embodiments, and will not be repeated here.

[0229] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0230] Based on the same inventive concept, this application also provides a device for determining recommendation information to implement the method for determining recommendation information described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the device for determining recommendation information provided below can be found in the limitations of the method for determining recommendation information above, and will not be repeated here.

[0231] In one embodiment, such as Figure 17 As shown, a device for determining recommendation information is provided, including: an acquisition module 1702 and a determination module 1704, wherein:

[0232] The acquisition module 1702 is used to acquire a target recommendation behavior sequence based on the target object's historical recommendation information behavior within a preset time period, and the information of each historical recommendation information behavior in the target recommendation behavior sequence, including the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories;

[0233] The determination module 1704 is used to determine the target recommendation information category based on the target recommendation behavior sequence; and to determine the similar recommendation information category that is similar to the target recommendation information category; and to determine the recommendation information corresponding to the target recommendation information category and the similar recommendation information category as the information to be recommended.

[0234] In one embodiment, the acquisition module 1702 is specifically used to sort the historical recommendation information behaviors based on the behavior time of each historical recommendation information behavior to obtain a first recommendation behavior sequence. The information of each historical recommendation information behavior in the first recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and the first recommendation information category. The module also samples and processes the information of each historical recommendation information behavior in the first recommendation behavior sequence to obtain a target recommendation behavior sequence.

[0235] In one embodiment, the acquisition module 1702 is specifically used to construct a graph network based on each historical recommendation information behavior. The nodes of the graph network are each historical recommendation information behavior. The node attributes of the behavior node include: recommendation information identifier and first recommendation information category. An edge between two behavior nodes indicates that the two behavior nodes have the same node attribute. The module also generates a target recommendation behavior sequence based on the graph network.

[0236] In one embodiment, the acquisition module 1702 is specifically used to determine an initial node in a graph network based on the behavior time of each historical recommendation behavior; and based on the initial node, to obtain a second recommendation behavior sequence by randomly walking in the graph network, the information of each historical recommendation behavior in the second recommendation behavior sequence including the recommendation information identifier of the historical recommendation behavior and the first recommendation information category; and to sample and process the information of each historical recommendation behavior in the second recommendation behavior sequence to obtain a target recommendation behavior sequence.

[0237] In one embodiment, the determining module 1704 is specifically used to obtain the feature vectors of each historical recommendation information behavior according to the target recommendation behavior sequence. The feature vectors include a first feature vector corresponding to the recommendation information identifier and a second feature vector corresponding to the first recommendation information category. The module also calculates the first similarity between the first feature vector and the second feature vector of each historical recommendation information behavior and determines the target recommendation information category based on the first similarity of each historical recommendation information behavior.

[0238] In one embodiment, the determining module 1704 is specifically configured to predict the recommendation information identifier of each second historical recommendation information behavior in the target recommendation behavior sequence based on each first historical recommendation information behavior in the target recommendation behavior sequence, wherein the first historical recommendation information behavior is a historical recommendation information behavior including the recommendation information identifier, and the second historical recommendation information behavior is a historical recommendation information behavior including the first recommendation information category; and predict the first recommendation information category of each first historical recommendation information behavior in the target recommendation behavior sequence based on each second historical recommendation information behavior in the target recommendation behavior sequence; and obtain the first feature vector corresponding to each recommendation information identifier based on the recommendation information identifier of each historical recommendation information behavior in the target recommendation behavior sequence; and obtain the second feature vector corresponding to each first recommendation information category based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence.

[0239] In one embodiment, the determining module 1704 is specifically used to determine the first recommendation information category of the historical recommendation information behavior as the target recommendation information category if the first similarity of the historical recommendation information behavior is greater than or equal to the second similarity threshold.

[0240] In one embodiment, the determining module 1704 is further configured to: if the first similarity of the historical recommendation information behavior is less than the second similarity threshold, determine the second similarity between the first feature vector of the historical recommendation information behavior and the category feature vector of each recommendation information category; if the recommendation information category corresponding to the second similarity with the largest value among the second similarities is consistent with the first recommendation information category, then determine the first recommendation information category as the target recommendation information category.

[0241] In one embodiment, the determining module 1704 is further configured to: if the recommended information category corresponding to the second similarity with the largest value among the second similarities is inconsistent with the first recommended information category, calculate the similarity difference between the first similarity and the largest second similarity; if the similarity difference is greater than or equal to the largest similarity difference, and the largest second similarity is greater than the first similarity, then determine the recommended information category corresponding to the largest second similarity as the target recommended information category.

[0242] In one embodiment, the determining module 1704 is further configured to obtain the input second recommendation information category if the similarity difference is less than the minimum similarity difference, and determine the second recommendation information category as the target recommendation information category.

[0243] In one embodiment, the device for determining recommendation information further includes a parameter updating module;

[0244] The parameter update module is used to update the model parameters of the recall ranking model based on the first recommendation information category of each historical recommendation information behavior and the target recommendation information category. The recall ranking model is used to identify the recommendation information category of the recommendation information behavior.

[0245] Each module in the aforementioned device for determining recommendation information can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0246] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 18 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores historical recommendation information behavior and recommendation information. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a method for determining recommendation information.

[0247] Those skilled in the art will understand that Figure 18 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0248] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0249] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0250] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0251] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0252] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0253] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0254] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for determining recommendation information, characterized in that, The method includes: Based on the target object's historical recommendation information behavior within a preset time period, a target recommendation behavior sequence is obtained. The information of each of the historical recommendation information behaviors in the target recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories. Based on each first historical recommendation information behavior in the target recommendation behavior sequence, predict the recommendation information identifier of each second historical recommendation information behavior in the target recommendation behavior sequence, wherein the first historical recommendation information behavior is the historical recommendation information behavior including the recommendation information identifier, and the second historical recommendation behavior is the historical recommendation information behavior including the first recommendation information category; Based on each of the second historical recommendation information behaviors in the target recommendation behavior sequence, predict the first recommendation information category for each of the first historical recommendation information behaviors in the target recommendation behavior sequence; Based on the recommendation information identifier of each historical recommendation information behavior in the target recommendation behavior sequence, a first feature vector corresponding to each recommendation information identifier is obtained; Based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence, a second feature vector corresponding to each first recommendation information category is obtained; The first similarity between the first feature vector and the second feature vector of each of the historical recommendation information behaviors is calculated. The target recommendation information category is determined by the first similarity of the behaviors of each of the historical recommendation information; Identify similar recommendation information categories that are similar to the target recommendation information category; The recommended information corresponding to the target recommendation information category and the similar recommendation information category is determined as the information to be recommended.

2. The method according to claim 1, characterized in that, The step of obtaining the target recommendation behavior sequence based on the target object's historical recommendation information behavior within a preset time period includes: Based on the behavior time of each of the historical recommendation information behaviors, the historical recommendation information behaviors are sorted to obtain a first recommendation behavior sequence. The information of each of the historical recommendation information behaviors in the first recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and the first recommendation information category. The information of each of the historical recommendation behaviors in the first recommendation behavior sequence is sampled and processed to obtain the target recommendation behavior sequence.

3. The method according to claim 1, characterized in that, The step of obtaining the target recommendation behavior sequence based on the target object's historical recommendation information behavior within a preset time period includes: A graph network is constructed based on the historical recommendation information behaviors. The behavior nodes of the graph network are the historical recommendation information behaviors. The node attributes of the behavior nodes include: the recommendation information identifier and the first recommendation information category. An edge between two behavior nodes indicates that the two behavior nodes have the same node attributes. The target recommendation behavior sequence is generated based on the graph network.

4. The method according to claim 3, characterized in that, The generation of the target recommendation behavior sequence based on the graph network includes: Based on the behavior time of each of the historical recommendation information behaviors, the initial node is determined in the graph network; Based on the initial node, a second recommendation behavior sequence is obtained by random walk in the graph network. The information of each historical recommendation behavior in the second recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and the first recommendation information category. The information of each of the historical recommendation behaviors in the second recommendation behavior sequence is sampled and processed to obtain the target recommendation behavior sequence.

5. The method according to claim 1, characterized in that, Determining the target recommendation information category based on the first similarity of each of the historical recommendation information behaviors includes: If the first similarity of the historical recommendation information behavior is greater than or equal to the second similarity threshold, then the first recommendation information category of the historical recommendation information behavior is determined as the target recommendation information category.

6. The method according to claim 1, characterized in that, The step of determining the target recommendation information category based on the first similarity of the behaviors of each of the historical recommendation information further includes: If the first similarity of the historical recommendation information behavior is less than the second similarity threshold, then the second similarity between the first feature vector of the historical recommendation information behavior and the category feature vector of each recommendation information category is determined. If the recommended information category corresponding to the second similarity with the largest value among all the second similarities is consistent with the first recommended information category, then the first recommended information category is determined as the target recommended information category.

7. The method according to claim 6, characterized in that, The step of determining the target recommendation information category based on the first similarity of the behaviors of each of the historical recommendation information further includes: If the category of the recommended information corresponding to the second similarity with the largest value among all the second similarities is inconsistent with the first recommended information category, calculate the similarity difference between the first similarity and the largest second similarity; If the similarity difference is greater than or equal to the maximum similarity difference, and the maximum second similarity is greater than the first similarity, then the recommendation information category corresponding to the maximum second similarity is determined as the target recommendation information category.

8. The method according to claim 7, characterized in that, The step of determining the target recommendation information category based on the first similarity of the behaviors of each of the historical recommendation information further includes: If the similarity difference is less than the minimum similarity difference, then the input second recommendation information category is obtained, and the second recommendation information category is determined as the target recommendation information category.

9. The method according to any one of claims 5 to 8, characterized in that, The method further includes: Based on the first recommendation information category of each historical recommendation information behavior and the target recommendation information category, the model parameters of the recall ranking model are updated. The recall ranking model is used to identify the recommendation information category of the recommendation information behavior.

10. A device for determining recommendation information, characterized in that, The device includes: The acquisition module is used to acquire a target recommendation behavior sequence based on the target object's historical recommendation information behaviors within a preset time period. The information of each historical recommendation information behavior in the target recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and any one of the first recommendation information categories. The determining module is configured to: predict the recommendation information identifier of each second historical recommendation information behavior in the target recommendation behavior sequence based on each first historical recommendation information behavior in the target recommendation behavior sequence, wherein the first historical recommendation information behavior is a historical recommendation information behavior including the recommendation information identifier, and the second historical recommendation information behavior is a historical recommendation information behavior including the first recommendation information category; predict the first recommendation information category of each first historical recommendation information behavior in the target recommendation behavior sequence based on each second historical recommendation information behavior in the target recommendation behavior sequence; obtain a first feature vector corresponding to each recommendation information identifier based on the recommendation information identifier of each historical recommendation information behavior in the target recommendation behavior sequence; obtain a second feature vector corresponding to each first recommendation information category based on the first recommendation information category of each historical recommendation information behavior in the target recommendation behavior sequence; calculate a first similarity between the first feature vector and the second feature vector of each historical recommendation information behavior; determine the target recommendation information category based on the first similarity of each historical recommendation information behavior; determine similar recommendation information categories similar to the target recommendation information category; and determine the recommendation information corresponding to the target recommendation information category and the similar recommendation information categories as the information to be recommended.

11. The apparatus according to claim 10, characterized in that, The acquisition module is specifically used to sort the historical recommendation information behaviors based on the behavior time of each historical recommendation information behavior to obtain a first recommendation behavior sequence. The information of each historical recommendation information behavior in the first recommendation behavior sequence includes the recommendation information identifier and the first recommendation information category of the historical recommendation information behavior. The module also samples and processes the information of each historical recommendation information behavior in the first recommendation behavior sequence to obtain the target recommendation behavior sequence.

12. The apparatus according to claim 10, characterized in that, The acquisition module is specifically used to construct a graph network based on each of the historical recommendation information behaviors. The behavior nodes of the graph network are each of the historical recommendation information behaviors. The node attributes of the behavior nodes include: the recommendation information identifier and the first recommendation information category. An edge between two behavior nodes indicates that the two behavior nodes have the same node attribute. The target recommendation behavior sequence is generated based on the graph network.

13. The apparatus according to claim 12, characterized in that, The acquisition module is specifically used to determine an initial node in the graph network based on the behavior time of each of the historical recommendation information behaviors; based on the initial node, to randomly walk in the graph network to obtain a second recommendation behavior sequence, wherein the information of each of the historical recommendation information behaviors in the second recommendation behavior sequence includes the recommendation information identifier of the historical recommendation information behavior and the first recommendation information category; and to sample and process the information of each of the historical recommendation information behaviors in the second recommendation behavior sequence to obtain the target recommendation behavior sequence.

14. The apparatus according to claim 10, characterized in that, The determining module is specifically used to determine the first recommendation information category of the historical recommendation information behavior as the target recommendation information category if the first similarity of the historical recommendation information behavior is greater than or equal to the second similarity threshold.

15. The apparatus according to claim 10, characterized in that, The determining module is further configured to: if the first similarity of the historical recommendation information behavior is less than the second similarity threshold, determine the second similarity between the first feature vector of the historical recommendation information behavior and the category feature vector of each recommendation information category; if the recommendation information category corresponding to the second similarity with the largest value among the second similarities is consistent with the first recommendation information category, then determine the first recommendation information category as the target recommendation information category.

16. The apparatus according to claim 15, characterized in that, The determining module is further configured to calculate the similarity difference between the first similarity and the largest second similarity if the category of the recommended information corresponding to the second similarity with the largest value among the second similarities is inconsistent with the first recommended information category; If the similarity difference is greater than or equal to the maximum similarity difference, and the maximum second similarity is greater than the first similarity, then the recommendation information category corresponding to the maximum second similarity is determined as the target recommendation information category.

17. The apparatus according to claim 16, characterized in that, The determining module is further configured to, if the similarity difference is less than the minimum similarity difference, obtain the input second recommendation information category and determine the second recommendation information category as the target recommendation information category.

18. The apparatus according to any one of claims 14 to 17, characterized in that, The device for determining recommendation information further includes a parameter update module; The parameter update module is used to update the model parameters of the recall ranking model based on the first recommendation information category of each historical recommendation information behavior and the target recommendation information category. The recall ranking model is used to identify the recommendation information category of the recommendation information behavior.

19. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

20. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

21. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images