Method and device for determining business entity relevance based on knowledge graph

Abstract

The embodiment of the present specification provides a method and device for determining business entity correlation based on a knowledge graph. In the process of determining business entity correlation based on the knowledge graph, the node information is shunted according to the degree of the node, so that the information of the head node is shunted through the common neighbor node as an intermediate node and then converged to the tail node. This technical concept makes the information sent from the head node depend on the degree of the head node, and can make the shunting of the node information comply with the law of conservation of information. In this way, the business processing demand of the asymmetric connection structure can be well met.

Description

Technical Field

[0001] This specification relates to the field of database query optimization technology, and in particular to a method and apparatus for determining the correlation of business entities based on knowledge graphs. Background Technology

[0002] Knowledge graphs can describe various entities or concepts and their relationships in the real world. They can contain a huge semantic network graph, where nodes represent entities or concepts (or entities corresponding to concepts or instances), and edges correspond to entity attributes or relationships between entities. Graph structure features characterize the connections between nodes and are a very important and distinctive type of feature on the graph. Graph structure features are usually determined by heuristic algorithms, offering advantages such as no training required, low cost, interpretability, and suitability for cold starts. They can be applied in graph-based scenarios such as risk control and recommendation systems. Heuristic algorithms are search algorithms based on experience and heuristic knowledge, used to solve complex problems. Heuristic algorithms are proposed in contrast to optimization algorithms and are algorithms constructed based on intuition or experience, providing a feasible solution to the combinatorial optimization problem within acceptable overhead (time and space). Common heuristic algorithms include annealing (SA), genetic algorithms (GA), ant colony optimization (ACO), and artificial neural networks (ANN), among others.

[0003] Graph structural features can be categorized into structural features on individual nodes and structural features on paired nodes. For paired nodes, conventional structural features are usually symmetrical (e.g., undirected graphs), such as CN, AA, Salton, RA, HPI, HDI, Katz, LHN, etc., meaning the structural feature from point A to point B is equal to the structural feature from point B to point A. However, in real-world business scenarios, asymmetric graph structural features are often needed to distinguish the different probabilities of a head node to a tail node versus a tail node to a head node. For example, in a self-driving navigation information push scenario, when various types of buildings or locations (e.g., residential buildings, commercial areas, tourist attractions, etc.) are the destinations, there is a high probability of needing a "parking lot." However, when a parking lot is the destination, the probability of needing a single building location (e.g., residential buildings, commercial areas, shops, restaurants, etc.) in the surrounding environment may be much lower. In other words, the probabilities of the two processes occurring are different. Thus, for the problem of asymmetric needs, symmetric resource allocation (RA) may not adequately meet the relevant business requirements. Summary of the Invention

[0004] This specification describes one or more embodiments of a method, apparatus, and system for determining the correlation of business entities based on knowledge graphs, in order to solve one or more problems mentioned in the background art.

[0005] According to a first aspect, a method for determining the relevance of business entities based on a knowledge graph is provided, used to determine the relevance of a second entity to a first entity, wherein the first entity and the second entity correspond to a first node and a second node, respectively, in the knowledge graph; the method includes: obtaining the number of neighboring nodes of the first node; determining, based on the number of neighboring nodes of the first node, various first coefficients for diverting information corresponding to the first node to each common neighboring node of the first node and the second node; using the various first coefficients and the aggregation rules for converging information from each common neighboring node to the second node, determining a second coefficient for diverting information from the first node to the second node; and determining the relevance of the first entity and the second entity through the second coefficient.

[0006] In one embodiment, all the first coefficients are equal and are the reciprocal of the number of neighboring nodes of the first node.

[0007] In one embodiment, the common neighbor nodes of the first node and the second node include a third node, which is associated with the first node via a first connection type. The first connection type corresponds to a first connection weight, and the first coefficient corresponding to the third node is positively correlated with the first connection weight and negatively correlated with the number of neighbor nodes of the first node.

[0008] In a further embodiment, the first coefficient corresponding to the third node is also negatively correlated with the sum of the connection weights corresponding to each of the neighboring nodes of the first node.

[0009] In one embodiment, determining the second coefficient for information diversion from the first node to the second node using the aggregation rules for converging information from each first coefficient and each common neighbor node to the second node includes: determining, using the aggregation rules, third coefficients for converging information diverted from the first node to each common neighbor node to the second node; determining, based on the product of each first coefficient and the corresponding third coefficient, the proportions for diverting information from the first node to the second node via each common neighbor node; and obtaining the second coefficient based on the sum of the proportions. In another embodiment, determining the second coefficient for information diversion from the first node to the second node using the aggregation rules for converging information from each first coefficient and each common neighbor node to the second node includes: determining, using the aggregation rules, third coefficients for converging information diverted from the first node to the second node; and determining the second coefficient based on the product of the sum of the third coefficients and the first coefficient.

[0010] In a further embodiment, the common neighbor nodes of the first node and the second node include a fourth node, and the convergence rule includes: the third coefficient corresponding to the fourth node is negatively correlated with the degree of the fourth node.

[0011] In a further embodiment, the fourth node is associated with the second node via a second connection type corresponding to a second connection weight, and the aggregation rule further includes: a third coefficient corresponding to the fourth node is positively correlated with the second connection weight.

[0012] In one embodiment, determining the second coefficient for information diversion from the first node to the second node using the first coefficients includes: determining the reciprocal of the degree of each common neighbor node as the third coefficient for the corresponding common neighbor node to aggregate the information diverted from the first node to the second node; and determining the second coefficient based on the sum of the third coefficients and the product of the first coefficients.

[0013] In one embodiment, determining the association between the first entity and the second entity using the second coefficient includes: comparing the second coefficient with a predetermined threshold; if the second coefficient is greater than the predetermined threshold, determining that the second entity is an entity associated with the first entity; if the second coefficient is less than the predetermined threshold, determining that the second entity is an entity not associated with the first entity.

[0014] In one embodiment, the method is used in an information push business scenario, where the first entity and the second entity are both candidate information to be pushed, and each common neighbor node is another business entity that is simultaneously associated with the first entity and the second entity as candidate information. The other business entities include at least one of browsing users, clicking users, publishers, keywords, and information categories.

[0015] In a further embodiment, the candidate information includes at least one of documents, product links, images, and videos.

[0016] According to the second aspect, an apparatus for determining the association of business entities based on a knowledge graph is provided, used to determine the association between a second entity and a first entity, wherein the first entity and the second entity correspond to a first node and a second node in the knowledge graph, respectively; the apparatus includes:

[0017] The acquisition unit is configured to acquire the number of neighboring nodes of the first node;

[0018] The traffic splitting configuration unit is configured to determine the information corresponding to the first node based on the number of neighboring nodes of the first node, and to split each first coefficient to each common neighboring node of the first node and the second node respectively.

[0019] The aggregation configuration unit is configured to use aggregation rules to aggregate information from the first node to the second node using each first coefficient and each common neighbor node, and to determine the second coefficients of the second node to which the information from the first node is distributed;

[0020] The determination unit is configured to determine the association between the first entity and the second entity using the second coefficient.

[0021] According to a third aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of the first aspect.

[0022] According to a fourth aspect, a computing device is provided, including a memory and a processor, characterized in that the memory stores executable code, and when the processor executes the executable code, it implements the method of the first aspect.

[0023] The methods and apparatus provided in the embodiments of this specification, in the process of determining the relevance of business entities based on knowledge graphs, distribute node information according to the degree of the node (i.e., the number of neighboring nodes), so that the information of the head node is distributed through common neighboring nodes as intermediate nodes before converging to the tail node. This technical concept ensures that the information emitted from the head node depends on the degree of the head node, and that the distribution of node information conforms to information conservation. Thus, it can well meet the business processing requirements of asymmetric connection structures, while also exhibiting asymmetric results in symmetric connection structures, showing better performance in scenarios that may involve asymmetric knowledge graph connection structures. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 It is a schematic diagram of a specific knowledge graph architecture;

[0026] Figure 2 This is a flowchart illustrating the process of determining the correlation of business entities based on a knowledge graph according to one embodiment of this specification;

[0027] Figure 3 This diagram illustrates a specific example of a two-way graph for determining the relationships between business entities.

[0028] Figure 4 This specification shows a schematic block diagram of an apparatus for determining the correlation of business entities based on a knowledge graph, according to one embodiment of the present specification. Detailed Implementation

[0029] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0030] Figure 1 A schematic diagram of a specific knowledge graph architecture is shown.

[0031] like Figure 1 As shown, a knowledge graph can be used to describe the relationships between multiple business entities under at least one connection type. Business entities in a knowledge graph can be entities from various business sectors. An entity can be anything that objectively exists and can be distinguished from others; it can be a real object or an abstract concept. Examples include a network platform, a company, a network user, a document, a word, etc.

[0032] Relationships are the connections between entities and can be described by relationship types. Taking network users as an example, relationship types could include at least one of the following: shared terminal application (APP), financial platform transfer, saved contacts in the address book, etc. As another example, relationship types for documents and words as entities could include containment and being contained. Various entities under multiple relationship types can form a vast relationship network, i.e., a knowledge graph.

[0033] In real-world business scenarios, it is often necessary to determine the relationships between business entities. For example, in a payment platform's payment scenario, it is necessary to determine whether two receiving accounts belong to the same merchant or the same user; in an information push scenario, it is necessary to determine whether two pieces of information to be pushed (such as products, documents, short videos, images, etc.) are related; in a flight route design scenario, it is necessary to determine which city is the best transit point for a flight route between two cities; and so on.

[0034] In these business scenarios, the relationships between some business entities are symmetrical, such as determining whether two payment accounts belong to the same merchant or the same user. However, some scenarios may be asymmetrical, such as in the context of short video information push, finding a less popular video (with views or clicks below a second threshold, such as 100) from popular videos that the user likes (views or clicks exceeding a first threshold, such as 100,000) and pushing it to the user, and finding a popular video from less popular videos that the user likes and pushing it to the user.

[0035] In asymmetric association scenarios, for any pair of head nodes A and tail nodes B, the probabilities of the path from head node A to tail node B and from tail node B to head node A are different. Conventional techniques mostly consider only the case where the graph structure features are symmetrical, meaning the structural features from A to B are the same as those from B to A. Examples include the structural features in CN, AA, Salton, RA, HPI, HDI, Katz, and LHN graphs.

[0036] In view of this, this specification provides a technical concept that considers the distribution and aggregation of information between nodes according to the number of neighboring nodes (the degree of the nodes), thereby describing the impact of information on each other during the transmission process between nodes and mapping it to the correlation between entities. This method can effectively take into account the asymmetry of business entities in the graph structure features and improve the business processing effect based on knowledge graphs.

[0037] The following is combined with Figure 2 The technical concept of this specification is described in detail with reference to a specific embodiment shown.

[0038] Figure 2This specification illustrates a process for determining the relevance of business entities based on a knowledge graph, according to one embodiment. The execution entity of this process can be a computer, device, or server with a certain computing power. The knowledge graph can be stored by the execution entity or by other devices connected to the execution entity; this is not limited here. Figure 1 Similar relational network structures can be stored as multiple (a, r, b) triples, where a and b represent nodes, and r represents the association. In an undirected graph, a triple (a, r, b) describes the connection between nodes a and b. In a directed graph, a triple (a, r, b) describes the connection from node a to node b, and a triple (b, r', a) describes the connection from node b to node a.

[0039] Figure 2 The illustrated process is applicable to determining the relationships between business entities in various business scenarios. Here, we assume we are determining the relationship between the second entity and the first entity, or the probability that the second node can be linked from the first node. In a knowledge graph, the first node is the node corresponding to the first entity, and the second node is the node corresponding to the second entity. The first and second entities correspond to different business entities depending on the business scenario. For example, in an information push scenario, business entities can be pushable information such as documents, product links, images, and videos. These business entities can be associated with other business entities related to the pushable information, such as browsing users, clicking users, publishers, and keywords.

[0040] For ease of description, assume the second node is a second-order neighbor of the first node, and denote the first node as u and the second node as v. That is, consider the paths connecting the common neighbors of the first node u and the second node v. Figure 2 As shown, the process for determining the correlation of business entities based on knowledge graphs may include: Step 202, obtaining the number of neighboring nodes of the first node; Step 204, determining the first coefficients for information corresponding to the first node and distributing it to each common neighboring node of the first node and the second node based on the number of neighboring nodes of the first node; Step 206, determining the second coefficients for information from the first node to the second node by using the first coefficients and the aggregation rules for the aggregation of information from each common neighboring node to the second node; Step 208, determining the correlation between the first entity and the second entity through the second coefficients.

[0041] First, in step 202, the number of neighboring nodes of the first node is obtained.

[0042] It is understandable that the neighboring nodes of the first node can be other nodes that have a direct relationship with the first node. The entity type of the neighboring nodes of the first node can be the same as or different from that of the entity corresponding to the first node. For example, if the first node is video information to be pushed, its neighboring nodes can correspond to at least one entity among video category (such as comedy, science popularization, review and debunking, beauty, clothing, etc.), publisher, viewer, etc.

[0043] The number of neighboring nodes of the first node can also be considered the first quantity, which can be determined statistically from the knowledge graph data. When the knowledge graph is stored as triples, the first quantity can be determined by counting the number of triples containing the first node. Specifically, in a symmetric structure, all triples containing the first node can be counted; in an asymmetric structure, only triples where the first node is the head node can be counted. When the knowledge graph is stored as a relational network, the first quantity can be determined by counting the number of edges connecting the first node. Similarly, in a symmetric structure, all edges connecting the first node can be counted; in an asymmetric structure, only edges originating from the first node can be counted.

[0044] It's worth noting that the first quantity can also be called the degree of the first node. In an asymmetric structure, the first quantity is the out-degree of the first node, that is, the number of branches from which information flows out. Let the first node be denoted as u, and the first quantity can be represented by, for example, d. u express.

[0045] Next, in step 204, the information corresponding to the first node is determined based on the number of neighboring nodes of the first node, and the first coefficients are distributed to each common neighboring node of the first node and the second node respectively.

[0046] According to the technical concept of this specification, the information carried by the first node can be distributed to its neighboring nodes. The amount of information distributed to a single neighboring node is described by a first coefficient, which is the product of the first coefficient and the amount of information carried by the first node. The information carried by the first node may include attribute information corresponding to the first entity, such as descriptive information and feature information of a short video. The information carried by the first node may also include association information obtained from its associated nodes, such as category information obtained from nodes corresponding to its category. Generally, the more neighboring nodes there are, the less information is distributed from the first node to a single neighboring node, and the smaller the first coefficient corresponding to a single neighboring node. A single first coefficient reflects the importance of the corresponding neighboring node to the first node.

[0047] In one optional implementation, the information carried by the first node can be equally distributed among its neighboring nodes. For example, the information distributed to each neighboring node is negatively correlated with the first quantity, such that the first coefficient corresponding to each neighboring node is the reciprocal of the first quantity, denoted as 1 / d. u .

[0048] According to another optional implementation, when the information carried by the first node is distributed to its neighboring nodes, the weights corresponding to the connection types between the first node and each neighboring node (such as connection weights) are considered. The information distributed to a single neighboring node is positively correlated with the corresponding weight, that is, the first coefficient corresponding to a single neighboring node is positively correlated with the corresponding weight. For example, suppose that the common neighboring nodes of the first and second nodes include a third node, and the third node is associated with the first node via a first connection type, the first connection type corresponding to a first connection weight, then the first coefficient corresponding to the third node can be positively correlated with the first connection weight and negatively correlated with the first quantity.

[0049] The weight corresponding to a single connection type can be a predetermined value, a value determined according to predetermined rules (such as the cumulative rules for the number of views, favorites, likes, etc. of a user's neighboring nodes when the first node corresponds to a short video), or a value learned using a graph model (a machine learning model used to process knowledge graphs). There are no restrictions here. For example, for a short video business entity, when associated with a publisher business entity, if the connection type between the corresponding nodes is "publish," the connection weight is, for example, 0.3; while when associated with a viewer business entity, if the connection type between the corresponding nodes is "view," the connection weight is, for example, 0.1.

[0050] Thus, in one embodiment, the first coefficient corresponding to a single neighbor node z can be the ratio of the corresponding weight to the first quantity, such as denoted as w. z / d u In another embodiment, the first coefficient corresponding to a single neighbor node can be the product of the ratio of the corresponding weight to the sum of the importance of all neighbor nodes and the reciprocal of the first number, as denoted as: w z / [(∑ j w j )d u In other embodiments, the first coefficient corresponding to a single neighbor node can also be determined in other ways, which will not be elaborated here.

[0051] It is understood that the first node and the second node can have at least one common neighbor node. A single common neighbor node is both a first-order neighbor node of the first node and a first-order neighbor node of the second node. The information of the common neighbor node can be obtained by reading knowledge graph data. For example, in one embodiment, the neighbor nodes of the first node (e.g., the first neighbor node c) can be obtained from the triple containing the first node a (e.g., (a, r, c)). Then, from the triple containing the neighbor nodes of the first node (e.g., (c, r', b)), it is checked whether the second node b is included. If it is included, the corresponding neighbor node (e.g., c) is determined to be a common neighbor node of the first node and the second node. It is worth noting that in the asymmetric structure, it is also necessary to consider that the triple containing the first node a is a triple with the first node as the head node, and the triple containing the neighbor nodes of the first node is a triple with the corresponding neighbor node as the head node.

[0052] Common neighbor nodes are included in the neighbor nodes of the first node. Therefore, the first coefficient of each common neighbor node can be determined by the method for determining the first coefficient of the neighbor nodes described above.

[0053] Then, through step 206, the second coefficient of the information from the first node to the second node is determined by using the aggregation rules of each first coefficient and each common neighbor node to converge to the second node.

[0054] It is understandable that after the information from the first node is distributed to its various neighboring nodes, the information on the common neighboring nodes may converge to the second node. In other words, not all the information distributed to a single common neighboring node necessarily converges to the second node. The degree to which the information from the first node converges to the second node through the common neighboring nodes can be represented by a second coefficient. Under the technical concept of this specification, the second coefficient can be determined by the degree to which information converges to the second node through each of the respective common neighboring nodes.

[0055] Suppose that the degree to which a single common neighbor node converges information to the second node is described by a corresponding third coefficient. Similar to the first coefficient, this third coefficient reflects the importance of the second node to its respective common neighbor node. Generally, the more neighbor nodes a single common neighbor node has, the less information it carries is diverted to the second node, and the smaller the corresponding third coefficient. Here, for a single common neighbor node, in a symmetric graph structure (such as an undirected graph), its number of neighbor nodes is its degree; in an asymmetric graph structure (such as a directed graph), its number of neighbor nodes is its out-degree, which will not be elaborated further here.

[0056] Let z represent any common neighbor node between the first and second nodes (e.g., the fourth node). Referring to the method for determining the first coefficient, the information carried by node z can be equally distributed among its neighbors, or it can be distributed according to the connection type weights. Taking equal distribution as an example, assume the number of neighbors of node z is d. z The information carried by node z is diverted to the third coefficient of the second node v, for example, 1 / d. z When considering connection type weights, such as the fourth node being associated with the second node via a second connection type, the second connection type corresponds to a second connection weight w. v Then, the third coefficient corresponding to the fourth node can be positively correlated with the second connection weight and negatively correlated with the degree of the fourth node. For example, the third coefficient for the information carried by node z being diverted to the second node v can be w. v / d z or w v / [(∑ k w k )d u Let k be a neighboring node of node z. In an optional embodiment, the information carried by node z can also be distributed to the second node according to other information transmission rules, for example, a third coefficient of 1 / log(d) z ).

[0057] Thus, the amount of information diverted from the first node to the second node via a single common neighbor node z can be positively correlated with both the corresponding first coefficient and the corresponding third coefficient. For example, the product of the first and third coefficients corresponding to node z can be used to describe the proportion of information diverted from the first node to the second node via node z.

[0058] Furthermore, by summing the proportions of information distributed from the first node to the second node via each common neighbor node, we can obtain a second coefficient for information distribution from the first node to the second node. As an example, in the ARA (Asymmetric Resource Allocation) algorithm, when the information of a single node is evenly distributed among its neighbor nodes, the second coefficient S is: S ARA [u, v] = 1 / d u ∑ z∈N(u)∩N(v) 1 / d z Or S ARA [u, v] = ∑ z∈N(u)∩N(v) 1 / d u ×1 / d z Where N represents neighboring nodes, then z represents the common neighboring nodes of nodes u and v. When the information of a single node is distributed to its neighbors according to weights, the second coefficient S can be: S ARA[u, v] = ∑ z∈N(u)∩N(v) w z / d u ×w v / d z In many other examples, the second coefficient can be determined in even more ways, such as S ArA [u, v] = 1 / d u ∑ z∈N(u)∩N(v) 1 / log(d z (and so on, which will not be listed here.)

[0059] Furthermore, in step 208, the correlation between the first entity and the second entity is determined by the second coefficient.

[0060] It is understandable that the second coefficient describes the importance of the second node relative to the first node, and the importance of the second entity relative to the first entity; this importance also reflects the correlation between the second entity and the first entity. Therefore, the correlation between the first entity and the second entity can be determined through the second coefficient.

[0061] In one embodiment, a second coefficient can be compared with a predetermined threshold. If the second coefficient is greater than the predetermined threshold, it can be determined that the second entity and the first entity are related; if the second coefficient is less than the predetermined threshold, it can be determined that the second entity and the first entity are not related. Furthermore, cases where the second coefficient equals the predetermined threshold can be included in either the result of determining relatedness or the result of determining non-relatedness, depending on business requirements.

[0062] Those skilled in the art will readily understand that the purpose of specific correlation determination results can be tailored to actual business needs. For example, in an information push scenario, both the first entity and the second entity are information to be pushed (such as product information, short videos, images, news, or popular science documents, etc.), and common neighbor nodes can be information attributes, categories, or publishers, clickers, viewers, etc. If the candidate information (the first entity) is pushed to the current user and accepted by the user (e.g., by clicking), the correlation between the second entity and the first entity can be used to decide whether to push other candidate information (the second entity) to the current user. For instance, if the determination result indicates that the second entity is correlated with the first entity, candidate information (the second entity) can be pushed to the current user, and so on.

[0063] It is worth noting that in asymmetric structures, the determination processes and conclusions regarding whether the second entity is related to the first entity, and whether the first entity is related to the second entity, are independent and cannot be applied interchangeably. Even in undirected or bidirectional graphs, due to the different degrees of the first and second nodes, the determination results may differ when the roles of the head and tail nodes are reversed.

[0064] To clarify the specific application of the technical solutions provided in this specification Figure 3 This diagram illustrates a specific example of determining the relevance of business entities within a concrete knowledge graph structure.

[0065] like Figure 3 As shown, in an information push service scenario, the correlation between business entities corresponding to nodes u and v is detected via a knowledge graph consisting of a bidirectional graph (and similarly for an undirected graph). u d v Let d be the degree (number of first-order neighbors) of node u and node v, respectively. u =4, d v =2. Nodes u and v have two common neighbor nodes, denoted as z1 and z2, with degrees d and d respectively. z1 =5, d z2 =3. For ease of description, this example assumes that when information is distributed among nodes, it is evenly distributed among each neighboring node.

[0066] Given the relationship between node v and node u, the information flows from node u to node v as follows:

[0067] Under path u-z1-v, the first coefficient corresponding to the information splitting from node u by node z1 is: 1 / d u =1 / 4, the third coefficient corresponding to the information splitting from node z1 to node v is 1 / d z1 =1 / 5;

[0068] Under path u-z2-v, the first coefficient corresponding to the information splitting from node u at node z2 is: 1 / d u =1 / 4, the third coefficient corresponding to the information splitting from node z2 to node v is 1 / d z2 =1 / 3;

[0069] Then, the second coefficient for the convergence of information from node u to node v via two neighboring nodes z1 and z2 is: S ARA [u, v] = (1 / d) u )×(1 / d z1 )+(1 / d u )×(1 / d z2 )=(1 / 4)×(1 / 5)+(1 / 4)×(1 / 3)=2 / 15.

[0070] On the other hand, when determining the association between node u and node v, the information flows from node v to node u as follows:

[0071] Under path v-z1-u, the first coefficient corresponding to the information splitting from node v to node z1 is: 1 / d v =1 / 2, the third coefficient corresponding to the information splitting from node z1 to node u is 1 / d z1 =1 / 5;

[0072] Under path v-z2-u, the first coefficient corresponding to the information splitting from node v to node z2 is: 1 / d v =1 / 4, the third coefficient corresponding to the information splitting from node z2 to node u is 1 / d z2 =1 / 3;

[0073] Then the second coefficient for the convergence of information from node v to node u via two neighboring nodes z1 and z2 is: S ARA [v, u] = (1 / d) v )×(1 / d z1 )+(1 / d v )×(1 / d z2 )=(1 / 2)×(1 / 5)+(1 / 2)×(1 / 3)=4 / 15.

[0074] As can be seen from the examples above, based on Figure 3 The bidirectional graph shown illustrates that the information splitting process between nodes u and v is not reversed or equivalent. In other words, even in a symmetrical structure, the information splitting process exhibits asymmetry. Assuming a predetermined threshold of 0.2, we can determine that nodes v and u are not related, while nodes u and v are related. Taking a short video recommendation scenario as an example, nodes u and v correspond to short video A and short video B respectively, and nodes z1 and z2 can correspond to users who simultaneously clicked on short video A and short video B, or users who simultaneously followed the publishers of short video A and short video B, etc. If short video A (corresponding to node u) is recommended to a user, since short video B (corresponding to node v) is not related, we can decide not to recommend short video B to that user. Conversely, if short video B is recommended to a user, since short video A (corresponding to node u) is related, we can decide to further recommend short video A to that user.

[0075] Understandable, Figure 3 The given examples are for illustrative purposes only and the values ​​are relatively small. In real-world cases, the difference in the degree of nodes may be much greater. For instance, short video A may be a popular video with a degree in the millions, while short video B may be a newly released video or a video posted by an account with a small following, with a degree in the thousands or tens. In this case, the difference in the second coefficient between the two directions is huge and has a stronger distinguishing effect.

[0076] in addition, Figure 3 The examples shown can also be applied to other business scenarios. For example, in the payment scenario of a payment platform, in the business scenario of determining whether two receiving accounts belong to the same merchant or the same user, nodes u and v correspond to receiving account A and receiving account B respectively, and nodes z1 and z2 can correspond to payment accounts that have made payments to both receiving account A and receiving account B, or the geographical location of the associated store, etc.; in the flight route design scenario, in the scenario of determining which city is better for connecting flights between two cities, nodes u and v correspond to the departure city and the destination city respectively, and nodes z1 and z2 can correspond to candidate connecting cities, etc.; and so on.

[0077] Reviewing the above process, in determining the relevance of business entities based on knowledge graphs, node information is distributed according to the degree of the nodes. This means that information from the head node is distributed through its common neighbors as intermediate nodes before converging at the tail node. This technical concept ensures that the information emanating from the head node depends on its degree and that the information distribution between nodes conforms to information conservation. This effectively meets the needs of asymmetric connection structures while also exhibiting asymmetric results in symmetric connection structures, showing better performance in scenarios involving asymmetric knowledge graph connection structures. For example, in information push scenarios, it can achieve more personalized push effects, reflected in an increase in the diversity of pushed information, ultimately impacting user click-through rates.

[0078] For example, during the experimentation with the technical solution provided in this specification, it was deployed in a search-based keyword push scenario. Considering the different probabilities of the trigger item and target item occurring in different directions during I2I (item to item) calculation, and introducing ARA asymmetric graph structure features into the downstream model, significant improvements were observed in metrics such as average number of categories exposed per user, average number of search terms exposed per user, average number of categories exposed per user, and average number of categories clicked per user. Experiments show that the performance of the relevant information push metrics can be improved by approximately five percent.

[0079] According to another embodiment, this specification also provides an apparatus for determining the correlation of business entities based on a knowledge graph. This apparatus can be installed in a computer, device, server, etc., with a certain computing power.

[0080] Figure 4 A specific embodiment of a device 400 for determining the relevance of business entities based on a knowledge graph is shown. This device can be used to determine the relevance of a second entity to a first entity, wherein the first entity and the second entity correspond to a first node and a second node in the knowledge graph, respectively. Figure 4As shown, the device 400 includes:

[0081] Unit 401 is configured to obtain the number of neighboring nodes of the first node;

[0082] The traffic splitting configuration unit 402 is configured to determine the information corresponding to the first node based on the number of neighboring nodes of the first node, and to split each first coefficient to each common neighboring node of the first node and the second node respectively.

[0083] The aggregation configuration unit 403 is configured to use the aggregation rules of each first coefficient and each common neighbor node to aggregate to the second node, and to determine the second coefficient of the first node to distribute the information of the second node to the second node.

[0084] The determination unit 404 is configured to determine the relationship between the first entity and the second entity using a second coefficient.

[0085] In an optional embodiment, all first coefficients are equal and are the reciprocal of the number of neighboring nodes of the first node.

[0086] In another optional embodiment, the common neighbor nodes of the first node and the second node include a third node. The third node is associated with the first node via a first connection type, the first connection type corresponding to a first connection weight, and the first coefficient corresponding to the third node is positively correlated with the first connection weight and negatively correlated with the number of neighbor nodes of the first node.

[0087] For example, the single first coefficient corresponding to a single common neighbor node is also negatively correlated with the sum of the connection weights corresponding to each of the first node's various neighbor nodes.

[0088] According to one possible design, the convergence configuration unit 403 is further configured as follows:

[0089] The aggregation rules are used to determine how to aggregate the information that each common neighbor node receives from the first node to the respective third coefficients of the second node;

[0090] Based on the product of each first coefficient and the corresponding third coefficient, determine the proportions by which information from the first node is distributed to the second node via each common neighbor node.

[0091] The second coefficient is obtained by summing the various proportions.

[0092] When information from each node is distributed to its neighboring nodes in an even manner, the aggregation configuration unit 403 is further configured as follows:

[0093] The aggregation rule is used to determine the third coefficients of each common neighbor node that are diverted from the first node and then aggregated to the second node.

[0094] The second coefficient is determined by multiplying the sum of the third coefficients with the first coefficient.

[0095] In one embodiment, assuming any common neighbor node of the first node and the second node is denoted as the fourth node, the aggregation rule may include: the third coefficient corresponding to the fourth node is negatively correlated with the degree of the fourth node.

[0096] Furthermore, assuming that the fourth node is associated with the second node via the second connection type, the second connection type corresponds to the second connection weight, and the third coefficient corresponding to the fourth node is also positively correlated with the second connection weight.

[0097] In one embodiment, the determination unit 404 is further configured as follows:

[0098] The second coefficient is compared with a predetermined threshold; the judgment result is determined based on the comparison result.

[0099] Wherein: if the second coefficient is greater than a predetermined threshold, the second entity is determined to be an entity that is related to the first entity; if the second coefficient is less than the predetermined threshold, the second entity is determined to be an entity that is not related to the first entity.

[0100] In the scenario where device 400 is used for information push services, both the first entity and the second entity can be candidate information to be pushed. Each common neighbor node is another business entity that is simultaneously associated with the candidate information of both the first and second entities. The other business entities include at least one of the following: browsing users, clicking users, publishers, keywords, and information categories. The candidate information may include, but is not limited to, at least one of the following: documents, product links, images, and videos.

[0101] It is worth noting that, Figure 4 The device 400 shown is Figure 2 The method embodiments shown correspond to this, therefore, Figure 2 The relevant descriptions of the methods in [the document] can also be applied to [other applications]. Figure 4 The device 400 shown will not be described in detail here.

[0102] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform a combination Figure 2 The methods described above.

[0103] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to achieve a combination Figure 2 The methods described above.

[0104] Those skilled in the art will recognize that the functions described in the embodiments of this specification in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0105] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the technical concept in this specification. It should be understood that the above description is only a specific embodiment of the technical concept in this specification and is not intended to limit the scope of protection of the technical concept in this specification. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments in this specification should be included within the scope of protection of the technical concept in this specification.

Claims

1. A method for determining the relevance of business entities based on a knowledge graph, applicable to information push business scenarios, used to determine the relevance of a second entity to a first entity, wherein the first entity and the second entity are both candidate information to be pushed, and respectively correspond to a first node and a second node in the knowledge graph; the candidate information includes at least one of documents, product links, images, and videos; the method includes: Get the number of neighboring nodes of the first node; Based on the number of neighboring nodes of the first node, the information corresponding to the first node is determined, and the first coefficients are distributed to each common neighboring node of the first node and the second node respectively. Using the aggregation rules that allow each common neighbor node to converge to the second node, the third coefficients for the convergence of information diverted from the first node to the second node by each common neighbor node are determined. Based on each first coefficient and each third coefficient, the second coefficients for the diversion of information from the first node to the second node are determined. Each common neighbor node is an other business entity that is simultaneously associated with candidate information of the first entity and the second entity. The other business entities include at least one of browsing users, clicking users, publishers, keywords, and information categories. The correlation between the first entity and the second entity is determined by the second coefficient.

2. The method as described in claim 1, wherein, All the first coefficients are equal, and each is the reciprocal of the number of neighboring nodes of the first node.

3. The method as described in claim 1, wherein, The common neighbor nodes of the first node and the second node include a third node. The third node is associated with the first node via a first connection type, the first connection type corresponding to a first connection weight, and the first coefficient corresponding to the third node is positively correlated with the first connection weight and negatively correlated with the number of neighbor nodes of the first node.

4. The method of claim 3, wherein, The first coefficient corresponding to the third node is also negatively correlated with the sum of the connection weights corresponding to each of the neighboring nodes of the first node.

5. The method of claim 1, wherein, Determining the second coefficient based on the first coefficient and each of the third coefficients includes: Based on the product of each first coefficient and the corresponding third coefficient, determine the proportions by which the information of the first node is distributed to the second node via each common neighbor node. The second coefficient is obtained by summing the various proportions.

6. The method of claim 1, wherein, The determination of the second coefficient based on the first coefficient and each of the third coefficients includes: The second coefficient is determined by multiplying the sum of the third coefficients with the first coefficient.

7. The method of claim 1, wherein, The common neighbor nodes of the first node and the second node include the fourth node, and the aggregation rule includes: the third coefficient corresponding to the fourth node is negatively correlated with the degree of the fourth node.

8. The method of claim 7, wherein, The fourth node is associated with the second node via a second connection type, the second connection type corresponding to a second connection weight, and the aggregation rule further includes: the third coefficient corresponding to the fourth node is positively correlated with the second connection weight.

9. The method of claim 1, wherein, The determination of the association between the first entity and the second entity using the second coefficient includes: The second coefficient is compared with a predetermined threshold. If the second coefficient is greater than the predetermined threshold, the second entity is determined to be an entity associated with the first entity; If the second coefficient is less than the predetermined threshold, the second entity is determined to be an entity that is not associated with the first entity.

10. An apparatus for determining the relevance of business entities based on a knowledge graph, applicable to information push business scenarios, used to determine the relevance of a second entity to a first entity, wherein the first entity and the second entity are both candidate information to be pushed, and respectively correspond to a first node and a second node in the knowledge graph; the candidate information includes at least one of documents, product links, images, and videos; the apparatus includes: The acquisition unit is configured to acquire the number of neighboring nodes of the first node; The traffic splitting configuration unit is configured to determine the information corresponding to the first node based on the number of neighboring nodes of the first node, and to split each first coefficient to each common neighboring node of the first node and the second node respectively. The aggregation configuration unit is configured to use aggregation rules to aggregate information from each common neighbor node to the second node, determine each third coefficient for the aggregation of information from each common neighbor node to the second node, and determine the second coefficient for the information from the first node to the second node based on each first coefficient and each third coefficient. Each common neighbor node is other business entities that are simultaneously associated with candidate information of the first entity and the second entity. The other business entities include at least one of browsing users, clicking users, publishers, keywords, and information categories. The determination unit is configured to determine the association between the first entity and the second entity using the second coefficient.

11. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-9.

12. A computing device, comprising a memory and a processor, characterized in that, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-9.

Images