An online adaptive information recommendation method and system based on graph search enhancement

Abstract

The application belongs to the field of information recommendation, and provides an online adaptive information recommendation method and system based on graph retrieval enhancement. In response to a recommendation request of a target user, a pre-trained graph embedding model is accessed to find an initial embedding vector corresponding to the target user. A real-time graph database is accessed to obtain users in the graph database that have a relationship with the target user, obtaining a candidate pool. According to a preset screening strategy, k users are screened from the candidate pool as seed nodes. For each seed node, the corresponding embedding vector is found from the graph embedding model, and independent queries are performed on each embedding vector. Approximate nearest neighbor search is performed in parallel, and after merging and deduplication of the obtained search results, a final candidate pool is formed. The candidate users are sorted according to the distance between the embedding vectors of the candidate users and the initial embedding vector. The application can efficiently adapt to the dynamic changes of graph data and significantly improve the recommendation effect.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of information recommendation, and particularly relates to an online adaptive information recommendation method and system based on graph search enhancement. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] The social network recommendation method generally represents users and items as low-dimensional vectors, and optimizes the recommendation effect through user similarity, network structure and content information. The current social network recommendation method has the following problems:

[0004] High retraining cost: The graph data such as social network is highly dynamic (for example, new friend relationships are constantly established between users). In order to ensure the model effect, the entire graph embedding model needs to be retrained frequently, which consumes huge computing resources and is time-consuming.

[0005] Inefficient long-tail recall: The user relationship network in the real world follows a power-law distribution, that is, a small number of central users have a large number of connections, while most ordinary users (i.e. long-tail users) have sparse connections. When performing approximate nearest neighbor search, a single user embedding vector often cannot effectively recall the long-tail users related to it, resulting in insufficient coverage and diversity of recommendations. SUMMARY

[0006] In order to solve the above problems, the present application proposes an online adaptive information recommendation method and system based on graph search enhancement. The present application dynamically expands the query by constructing a query seed set in real time during the reasoning phase, without adjusting the model or frequent retraining, so as to efficiently adapt to the dynamic changes of graph data and significantly improve the recommendation effect.

[0007] According to some embodiments, the present application adopts the following technical solution:

[0008] An online adaptive information recommendation method based on graph search enhancement, comprising the following steps:

[0009] In response to a recommendation request of a target user, accessing a pre-trained graph embedding model, and according to a unique identifier of the target user in the recommendation request, finding a corresponding initial embedding vector from the graph embedding model;

[0010] Accessing a real-time graph database to obtain users in the graph database who have direct and mutual confirmed connection relationships with the target user, and obtaining a candidate pool;

[0011] According to a preset screening strategy, screening k users from the candidate pool as seed nodes;

[0012] Traverse the determined k seed nodes, for each seed node, according to the unique identifier of its user, find the corresponding embedding vector from the graph embedding model, and aggregate the embedding vectors of all seed nodes to form an enhanced query set;

[0013] Each embedding vector in the enhanced query set is independently queried, and parallel approximate nearest neighbor search is performed. After merging and deduplicating the obtained search results, a final candidate pool is formed;

[0014] Sort each candidate user in the final candidate pool according to the distance between its embedding vector and the initial embedding vector to obtain a sorted recommendation list.

[0015] As an optional implementation, the pre-trained graph embedding model uses a double-layer graph attention network model, and is trained using the obtained historical graph snapshot, which contains the relationships and attributes of all users within a certain time window. During training, the parameters of the graph embedding model are optimized through contrastive learning.

[0016] As an optional implementation, the initial embedding vector is used to represent the global features of the user at the last training of the graph embedding model, and contains the comprehensive information of the user.

[0017] As an optional implementation, according to a preset filtering strategy, k users are selected from the candidate pool as seed nodes, and the process includes: the filtering strategy is a random baseline strategy, and k users are randomly and uniformly sampled from the candidate pool as seed nodes, and k is an integer greater than 0.

[0018] As an optional implementation, according to a preset filtering strategy, k users are selected from the candidate pool as seed nodes, and the process includes: the filtering strategy is a relationship recency strategy, which evaluates the recency of candidate users through a linear weighting model, determines the time window in which each candidate user has a direct and mutually confirmed connection relationship with the target user, assigns weights according to the time window, the weight value is higher when the time window is closer, and determines the top k users according to the maximum weight score as seed nodes.

[0019] As an optional implementation, according to a preset filtering strategy, k users are selected from the candidate pool as seed nodes, and the process includes: the filtering strategy is an interaction closeness and social circle diversity strategy, which first calculates an interaction intensity score for each candidate user in the candidate pool through a linear combination of interaction features, the interaction features include count type features and active day type features, and according to the interaction intensity score, the top d candidate users with the highest scores are selected, d>k.

[0020] The embedding vectors of the d candidate users are obtained, hierarchical clustering is performed, the d candidate users are divided into multiple clusters, a center point of each cluster is calculated, and the center points are ranked according to cosine similarity with the initial embedding vector of the target user, and users of the k center points with the highest similarity are selected as seed nodes.

[0021] As an optional implementation, the process of selecting k users from the candidate pool as seed nodes according to a preset screening strategy includes that the screening strategy is an integration strategy.

[0022] The recency of the candidate users is evaluated through a linear weighting model, a time window in which the candidate users have a direct and mutually confirmed connection relationship with the target user is determined, a weight is assigned according to the time window, the weight value is higher when the time window is closer, and the top k1 candidate users are determined according to the maximum weight score, where k1 is less than k.

[0023] An interaction intensity score is calculated for each candidate user in the candidate pool through a linear combination of interaction features, the interaction features include count type features and active day type features, and the top k2 candidate users with the highest scores are selected according to the interaction intensity scores, where k2 is less than k, and k1+k2=k.

[0024] The two parts of candidate users are combined to form a final seed node set.

[0025] As an optional implementation, the process of aggregating the embedding vectors of all seed nodes to form an enhanced query set includes combining the embedding vectors of all seed nodes together to form a set of k embedding vectors, i.e., an enhanced query set.

[0026] As an optional implementation, in the process of performing independent queries on each embedding vector in the enhanced query set and performing parallel approximate nearest neighbor search, the same recall budget is used for each parallel query, parallel processing is performed using distributed computing resources, and the delay of the entire multi-path retrieval process is kept at the same order of magnitude as the delay of the single-path retrieval process.

[0027] An online adaptive information recommendation system based on graph retrieval enhancement includes:

[0028] An initial embedding vector finding module is configured to, in response to a recommendation request of a target user, access a pre-trained graph embedding model, and find a corresponding initial embedding vector from the graph embedding model according to a unique identifier of the target user in the recommendation request.

[0029] A candidate pool determination module is configured to access a real-time graph database, obtain users in the graph database that have a direct and mutually confirmed connection relationship with the target user, and obtain a candidate pool.

[0030] The seed node screening module is configured to screen k users from the candidate pool as seed nodes according to a preset screening strategy;

[0031] The enhanced query set module is configured to traverse the determined k seed nodes, for each seed node, find the corresponding embedding vector from the graph embedding model according to the unique identifier of the user thereof, and aggregate the embedding vectors of all seed nodes to form an enhanced query set;

[0032] The parallel search module is configured to independently query each embedding vector in the enhanced query set, and perform parallel approximate nearest neighbor search, and after merging and deduplicating the obtained search results, form a final candidate pool;

[0033] The sorting module is configured to sort each candidate user in the final candidate pool according to the distance between the embedding vector thereof and the initial embedding vector, to obtain a sorted recommendation list.

[0034] Compared with the prior art, the beneficial effects of the present application are:

[0035] The present application dynamically expands the query by constructing the query seed set in real time in the reasoning phase, without the need to adjust the model or frequently retrain, so as to efficiently adapt to the dynamic changes of the graph data, significantly improve the recommendation effect, especially in the community scene such as friend recommendation, and finally generate a recommendation list containing both the candidate person highly similar to the user himself and the potential friend with higher diversity from different social circles of the user, thereby achieving a significant effect better than the traditional method in recall rate and business indicators.

[0036] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0037] The drawings accompanying the specification of the present application form a part of the present application, and the schematic embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application.

[0038] Figure 1 is a flow diagram of an online adaptive information recommendation method based on graph retrieval enhancement of an embodiment. DETAILED DESCRIPTION

[0039] The present application will be further described below in combination with the drawings and embodiments.

[0040] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0041] It is also important to note that the use of the term 'example' herein is intended to be non-limiting, and is used to illustrate an example of the present exemplary embodiments. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0042] The embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0043] Embodiment One

[0044] An online adaptive information recommendation method based on graph search enhancement, as shown in Figure 1 includes the following steps:

[0045] In response to a recommendation request of a target user, a pre-trained graph embedding model is accessed, and according to a unique identifier of the target user in the recommendation request, an initial embedding vector corresponding to the target user is found in the graph embedding model;

[0046] A real-time graph database is accessed to obtain users in the graph database who have a direct and mutual confirmed connection relationship with the target user, and a candidate pool is obtained;

[0047] According to a preset filtering strategy, k users are filtered out from the candidate pool as seed nodes;

[0048] The determined k seed nodes are traversed, for each seed node, an embedding vector corresponding to the seed node is found from the graph embedding model according to a unique identifier of the user of the seed node, and embedding vectors of all seed nodes are aggregated to form an enhanced query set;

[0049] Each embedding vector in the enhanced query set is independently queried, and parallel approximate nearest neighbor search is performed, and after merging and deduplication of the obtained search results, a final candidate pool is formed;

[0050] Each candidate user in the final candidate pool is sorted according to the distance between the embedding vector of the candidate user and the initial embedding vector, and a sorted recommendation list is obtained.

[0051] The following is a detailed description of each step.

[0052] Firstly, the triggering condition, when receiving a real-time recommendation request for a certain target user s (i.e. the inquirer), the flow of the method is triggered. The request is usually initiated by the application front end, aiming to generate a potential friend list for the target user s.

[0053] Firstly, interact with a pre-trained and independently deployed graph embedding model. The model is the basis of the entire recommendation system, and its characteristics are as follows:

[0054] The structure of the graph embedding model, in this embodiment, a double-layer graph attention network (Graph Attention Networks, GAT) is adopted.

[0055] Training data: the model is trained offline on historical and large-scale graph snapshot data. The snapshot captures the relationships (as edges) and attributes (as node features) of all users within a certain time window in the past.

[0056] Learning goal: through methods such as contrastive learning, optimize so that in the real social network, users with close relationships and similar interests are also close to each other in the generated low-dimensional vector space.

[0057] Through the embedding vector query service, according to the unique identifier (User ID) of the target user s, the corresponding initial embedding vector E (s) is found from the model storage in real time.

[0058] The initial embedding vector E (s) has the following key properties:

[0059] Static: it reflects the user's global features at the last training of the model. Any changes in the graph since the last training (for example, the target user s has added a friend) will not be reflected in this initial embedding vector.

[0060] Global: the vector is learned in the entire user network and contains comprehensive information about the user, but may not be able to finely express the features of the multiple, specific local social circles to which the user belongs.

[0061] Therefore, the initial embedding vector E (s) constitutes the baseline coordinate of this recommendation task. It can be used for traditional, non-adaptive approximate nearest neighbor search (Approximate Nearest Neighbor, ANN) for searching, but in this invention, its more important role is as the starting point and anchor point for subsequent online community enhancement operations. The core purpose of the subsequent steps is to make up for the static nature of the vector.

[0062] Next, the online community is built to enhance the seed set.

[0063] In this embodiment, based on the current latest graph data, an enhanced query seed set that can represent multiple social circles of the target user s is constructed in real time and adaptively for the target user s at the inference stage of the request (Test-Time).

[0064] Access a real-time graph database that stores the latest full user relationship graph containing billions of nodes and edges.

[0065] Through an efficient graph traversal query, the system finds and obtains all 1-hop bi-directional friends of the target user s, i.e., these friends are users who have direct and mutual confirmed connection relationships with the target user s.

[0066] A community member candidate pool Fu is obtained, which contains all the friends of the target user s at present. This step ensures that the subsequent seed selection is based on the latest social relationship as of the request initiation time, which is the basis for online adaptation.

[0067] According to the preset and flexibly combinable "community importance" strategy, k most representative friends are intelligently selected from the community member candidate pool Fu as community seed nodes. These strategies aim to capture the dynamics, closeness and diversity of the community from different dimensions, and specific strategies can include:

[0068] Strategy V1: Random baseline strategy

[0069] Strategy principle: As a baseline, verify the gain brought by the multi-vector query itself.

[0070] Implementation: Randomly and uniformly sample k friends from the community member candidate pool Fu as seeds. Through online A / B testing, it is shown that even this simplest strategy can bring significant business indicator improvement compared to the single-vector query, proving the effectiveness of the multi-modal representation.

[0071] Strategy V2: Relationship recency strategy

[0072] Strategy principle: Recent friend relationships are more likely to reflect the user's current and future social interests.

[0073] Implementation: This strategy evaluates the recency of candidate friends through a linear weighting model. Specifically, the system checks when the friend relationship was established and assigns different weights. For example, set decreasing weights for friends who have established relationships in the past 1 day, 7 days, 14 days, and 28 days, then calculate the weight score, and select the top k friends with the highest score as seeds.

[0074] Strategy V3: Interaction Intensity and Social Circle Diversity Strategy

[0075] Strategy Principle: This strategy combines two levels of consideration: first, the closest friends are the core of the community; second, these core friends may belong to different social circles and need to ensure coverage.

[0076] Implementation: It includes the following two stages:

[0077] First stage (interaction intensity screening): First, calculate an interaction intensity score for each friend in the candidate pool by linear combination of interaction features, including count-based features (such as chat sending number, snapshot sending number, and story browsing number, etc.) and active day-based features. According to the score of the interaction intensity score, select the top d friends (d>k) with the highest score as high-quality seed candidates.

[0078] The calculation process can use existing models or other algorithms.

[0079] Second stage (clustering dimensionality reduction for diversity): To avoid selecting seeds from the same highly active social circle, the system obtains the embedding vectors of the d high-quality candidates, performs hierarchical clustering (Hierarchical Clustering) on them, and divides them into c clusters (c is usually between 5 and 8). Then, calculate the center point of each cluster and sort them according to the cosine similarity between the center point and the target user s vector, and finally select the most similar k center points as seeds. This step ensures that the final seeds can represent different social circles of the user.

[0080] Strategy V4: Ensemble Strategy

[0081] Principle: Different strategies provide complementary signals, and combining them can achieve better and more robust performance.

[0082] Implementation: Allocate the number of seeds and integrate the results of multiple strategies. Online A / B testing, which integrates the relationship recency strategy (V2) and the interaction intensity strategy (part of V3), is a good example. The specific implementation can be: select k1 seeds with the highest recency and k2 seeds with the most interaction intensity (k1+k2=k), and together they form the final seed set.

[0083] After selecting k representative community seed nodes through the above seed selection mechanism, these discrete node identifiers are converted into a structured multi-modal query set that can be executed by an approximate nearest neighbor search algorithm.

[0084] Specifically, it includes traversing the determined k seed nodes. For each seed node si (where i from 1 to k), the embedding vector query service is invoked again.

[0085] Input: User ID of each seed node.

[0086] Output: Pre-trained embedding vector E(si) corresponding to each seed node si.

[0087] Aggregate the embedding vectors of all the k seed nodes found. Generate a set of k vectors, i.e., the "Community-Enhanced Query Set", denoted as {E(s1), E(s2),..., E(sk)}.

[0088] The generated query set has the following key features:

[0089] Multi-modal: This query set is no longer a single vector representing the user, but a set of vectors. According to the design of the seed selection strategy (especially strategy V4), these vectors may be distributed in different regions in the embedding space, respectively representing the characteristics of different social circles to which the user belongs. This fundamentally solves the problem that a single vector cannot capture the user's multiple identities and interests.

[0090] Dynamic and real-time: Since the seed nodes are selected online based on the latest graph data, the content of this query set is dynamically changing, reflecting the latest state of the user's community in real time. For example, if the user has just joined a new interest group and added a few friends, these new friends are likely to be selected as seeds, thereby immediately affecting the recommendation results and guiding the user to explore more deeply in this new social circle.

[0091] Optional Decoupling from the Searcher: The initial vector E(s) of the target user s can not be included in the query set. Experiments show that using only the vectors of community members (seeds) for querying is equally effective, and may even be better. The underlying assumption is that this approach can more evenly explore various social circles, avoiding excessive bias towards the circle most similar to the user themselves, thereby improving the diversity of recommendations.

[0092] Finally, this constructed, multi-modal query set will be used for the next step of retrieval as input to perform actual user recall.

[0093] After the community-enhanced query set containing k seed vectors is constructed, this step will perform efficient candidate user recall and aggregate the results of multiple recall channels into a unified and high-quality recommendation list.

[0094] Perform multi-path parallel approximate nearest neighbor search:

[0095] Input: The community-enhanced query set containing k embedding vectors {E(s1), E(s2),..., E(sk)} generated in Step 3.

[0096] Perform approximate nearest neighbor search (ANN) service using the retrieval system. This service pre-constructs the embedding vectors of all users into an efficient index. To support high concurrency and reduce costs, the index library of the retrieval system can be partitioned according to dimensions such as geographic location.

[0097] Execution process: The k seed vectors in the query set are submitted to the ANN service as k independent queries in parallel.

[0098] Budget allocation: If the total recall target is N candidates, allocate N / k recall budget for each query. For example, if k=5 and the target recall is 1000 people, each seed vector E(si) will be used to retrieve the 200 closest candidate users in the vector space.

[0099] Parallelization: These k queries can be initiated simultaneously, making full use of distributed computing resources to ensure that the delay of the entire multi-path retrieval process remains at the same order of magnitude as single-path retrieval.

[0100] Collect all the candidate user lists returned by the k parallel retrievals. For example, the list of 200 people returned from seed s1, R1, the list returned from seed s2, R2,..., and up to Rk. Then, perform a union operation on these lists and remove duplicate user IDs to form an aggregated candidate pool. The size of this candidate pool is usually less than N (because different seeds may recall the same candidates), but its diversity is much higher than that of single-vector queries.

[0101] Final ranking: For each candidate user in the aggregated candidate pool, the system will score according to a unified ranking standard to determine the final presentation order to the user. Specifically:

[0102] Ranking based on distance from the query: Calculate the distance (such as L2 distance or cosine distance) between the embedding vector E(c) of each user c in the candidate pool and the embedding vector E(s) of the original target user s.

[0103] Ascending order: sort the candidates according to the calculated distance from small to large. The closer the distance, the more similar the candidate user is to the target user s itself, and the higher the ranking.

[0104] In some embodiments, the method further comprises the steps of post-processing and returning:

[0105] Truncation: according to business requirements, select the top M users (for example, 100) from the sorted list as the final recommendation result.

[0106] Filtering: apply business rules, such as filtering out users who have sent friend requests, users who have been blocked, etc.

[0107] Return result: return the post-processed and final recommendation list to the application front end and display it to the user.

[0108] Embodiment two

[0109] An online adaptive information recommendation system based on graph search enhancement, comprising:

[0110] An initial embedding vector finding module configured to, in response to a recommendation request of a target user, access a pre-trained graph embedding model, and find a corresponding initial embedding vector from the graph embedding model according to a unique identifier of the target user in the recommendation request;

[0111] A candidate pool determination module configured to access a real-time graph database, and obtain users in the graph database that have a direct and mutual confirmed connection relationship with the target user, to obtain a candidate pool;

[0112] A seed node screening module configured to screen k users from the candidate pool as seed nodes according to a preset screening strategy;

[0113] An enhanced query set module configured to traverse the determined k seed nodes, for each seed node, find a corresponding embedding vector from the graph embedding model according to the unique identifier of the user, and aggregate the embedding vectors of all seed nodes to form an enhanced query set;

[0114] A parallel search module configured to independently query each embedding vector in the enhanced query set, and perform parallel approximate nearest neighbor search, and after merging and deduplicating the obtained search results, form a final candidate pool;

[0115] An ordering module configured to sort each candidate user in the final candidate pool according to the distance between its embedding vector and the initial embedding vector, to obtain a sorted recommendation list.

[0116] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and the like) having computer-usable program code embodied in the medium. CD - ROM

[0117] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 Figure 1

[0118] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 Figure 1

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 Figure 1

[0120] The above description is only preferred embodiments of the present application, and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made by those skilled in the art without departing from the spirit and principle of the present application shall fall within the scope of protection of the present application.​​​​​​​

Claims

1. An online adaptive information recommendation method based on graph search enhancement, characterized in that, The method comprises the following steps: In response to a recommendation request of a target user, a pre-trained graph embedding model is accessed, and an initial embedding vector corresponding to the target user is found from the graph embedding model according to a unique identifier of the target user in the recommendation request; A real-time graph database is accessed to obtain users in the graph database that have a direct and mutual confirmation connection relationship with the target user, thereby obtaining a candidate pool; According to a preset filtering strategy, k users are filtered from the candidate pool as seed nodes; The process of filtering k users from the candidate pool as seed nodes according to the preset filtering strategy comprises that the filtering strategy is an interactive closeness and social circle diversity strategy, an interactive intensity score of each candidate user in the candidate pool is calculated through a linear combination of interactive features, the interactive features include count type features and active day type features, and d candidate users with the highest scores are filtered according to the interactive intensity scores, d > k; The embedding vectors of the d candidate users are obtained, hierarchical clustering is performed, the d candidate users are divided into multiple clusters, the center points of the clusters are calculated, and the center points are sorted according to the cosine similarity with the initial embedding vector of the target user, and k users with the most similar center points are selected as seed nodes; The process of filtering k users from the candidate pool as seed nodes according to the preset filtering strategy comprises that the filtering strategy is an integrated strategy: The recency of the candidate users is evaluated through a linear weighting model, a time window in which each candidate user has a direct and mutual confirmation connection relationship with the target user is determined, weights are assigned according to the time window, the weight value is higher when the time window is closer, and the top k1 candidate users are determined according to the maximum weight score, k1 < k; An interactive intensity score of each candidate user in the candidate pool is calculated through a linear combination of interactive features, the interactive features include count type features and active day type features, and k2 candidate users with the highest scores are filtered according to the interactive intensity scores, k2 < k, and k1 + k2 = k; The two parts of candidate users are combined to form a final seed node set; The k seed nodes are determined, and for each seed node, the embedding vector corresponding to the user is found from the graph embedding model according to the unique identifier of the user, and the embedding vectors of all seed nodes are aggregated to form an enhanced query set; Each embedding vector in the enhanced query set is independently queried, and parallel approximate nearest neighbor search is performed, and after merging and deduplicating the obtained search results, a final candidate pool is formed; Each candidate user in the final candidate pool is sorted according to the distance between the embedding vector and the initial embedding vector, and a sorted recommendation list is obtained.

2. The method of claim 1, wherein the online adaptive information recommendation method based on graph search enhancement is characterized in that, The pre-trained graph embedding model adopts a double-layer graph attention network model, and is trained by using obtained historical graph snapshots, the historical graph snapshots contain the relationships and attributes of all users in a historical time window, and the parameters of the graph embedding model are optimized through contrastive learning during training.

3. The method of claim 1, wherein the enhanced online adaptive information recommendation based on graph search is characterized by, The initial embedding vector is used to represent the global features of the user at the last training of the graph embedding model, and contains comprehensive information of the user.

4. The method of claim 1, wherein the online adaptive information recommendation method based on graph search enhancement is characterized by, The process of aggregating the embedding vectors of all seed nodes to form the enhanced query set includes combining the embedding vectors of all seed nodes together to form a set of k embedding vectors, i.e., the enhanced query set.

5. The method of claim 1, wherein the enhanced online adaptive information recommendation based on graph search is characterized by, In the process of performing independent queries on each embedding vector in the enhanced query set and performing parallel approximate nearest neighbor search, the same recall budget is respectively set for each parallel query, and distributed computing resources are used for parallel processing to ensure that the delay of the entire multi-path retrieval process and the delay of the single-path retrieval process remain in the same order of magnitude.

6. An online adaptive information recommendation system based on graph search enhancement, characterized in that, The method comprises the following steps: The initial embedding vector searching module is configured to, in response to a recommendation request of a target user, access a pre-trained graph embedding model, and search for a corresponding initial embedding vector from the graph embedding model according to a unique identifier of the target user in the recommendation request; The candidate pool determining module is configured to access a real-time graph database to obtain users in the graph database that have a direct and mutual confirmation connection relationship with the target user, and obtain a candidate pool; The seed node screening module is configured to screen k users from the candidate pool as seed nodes according to a preset screening strategy. The process of screening k users from the candidate pool as seed nodes according to the preset screening strategy includes that the screening strategy is an interactive closeness and social circle diversity strategy, an interactive intensity score of each candidate user in the candidate pool is calculated through a linear combination of interactive features, the interactive features include count type features and active day type features, d candidate users with the highest scores are screened out according to the interactive intensity scores, and d > k. The embedding vectors of the d candidate users are obtained, hierarchical clustering is performed, the d candidate users are divided into multiple clusters, the center points of the clusters are calculated, and the center points are sorted according to cosine similarity with the initial embedding vector of the target user, and k users with the most similar center points are selected as seed nodes. The process of screening k users from the candidate pool as seed nodes according to the preset screening strategy includes that the screening strategy is an integrated strategy: The recency of the candidate users is evaluated through a linear weighting model, a time window in which each candidate user has a direct and mutual confirmation connection relationship with the target user is determined, weights are assigned according to the time window, the weight value is higher when the time window is closer, and the top k1 candidate users are determined according to the maximum weight score, k1 < k. An interactive intensity score of each candidate user in the candidate pool is calculated through a linear combination of interactive features, the interactive features include count type features and active day type features, and k2 candidate users with the highest scores are screened out according to the interactive intensity scores, k2 < k, and k1 + k2 = k. The two parts of candidate users are combined to form a final seed node set. The enhanced query set module is configured to traverse the determined k seed nodes, for each seed node, search for a corresponding embedding vector from the graph embedding model according to a unique identifier of the user of the seed node, and aggregate the embedding vectors of all seed nodes to form an enhanced query set. The parallel search module is configured to independently query each embedding vector in the enhanced query set respectively, perform parallel approximate nearest neighbor search, and form a final candidate pool after merging and deduplicating each search result obtained. The sorting module is configured to sort each candidate user in the final candidate pool according to the distance between the embedding vector of the candidate user and the initial embedding vector, and obtain a sorted recommendation list.

Images