Business recommendation method, device, electronic device and computer-readable storage medium

Abstract

The present invention discloses a business recommendation method, device, electronic device and computer-readable storage medium, which belong to the field of knowledge graph technology. The business recommendation method includes: extracting the first context subgraph that is updated in the business knowledge graph, and extracting the first triplet that is updated in the business knowledge graph; learning the first context subgraph to obtain the vector representation of the first context subgraph; inputting the vector representation of the first context subgraph and the first triplet into the translation model for learning to obtain the updated knowledge representation of the business knowledge graph; and using the updated knowledge representation to make business recommendations. According to the scheme in the present application, when making business recommendations, there is no need to learn the knowledge representation of the entire updated business knowledge graph, and only the updated context subgraph information and triplet information need to be considered, thereby achieving efficient business recommendations.

Description

Technical Field

[0001] The present invention belongs to the field of knowledge graph technology, and specifically relates to a business recommendation method, device, electronic device and computer-readable storage medium. Background Art

[0002] Currently, business recommendations can be made through existing knowledge representation technology by learning vector representations of entities and relationships in business knowledge graphs. However, existing knowledge representation technology mostly performs knowledge representation learning on static knowledge graphs. Therefore, when the business knowledge graph is updated, it is necessary to perform knowledge representation learning on the entire updated business knowledge graph before making business recommendations, which makes it impossible to make business recommendations efficiently. Summary of the invention

[0003] The purpose of the embodiments of the present invention is to provide a business recommendation method, device, electronic device and computer-readable storage medium to solve the current problem that business recommendations cannot be made efficiently when the business knowledge graph is updated.

[0004] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0005] In a first aspect, an embodiment of the present invention provides a service recommendation method, including:

[0006] Extracting a first context subgraph updated in the business knowledge graph, and extracting a first triplet updated in the business knowledge graph;

[0007] Learning the first context subgraph to obtain a vector representation of the first context subgraph;

[0008] Inputting the vector representation of the first context subgraph and the first triple into a translation model for learning to obtain an updated knowledge representation of the business knowledge graph;

[0009] Business recommendations are made using the updated knowledge representation.

[0010] Optionally, the loss function of the translation model is:

[0011]

[0012] Among them, (h, r, t) represents a triple; H(h) = E(h) + g(h), H(h) represents the knowledge representation of entity h, E(h) represents the vector representation of entity h, and g(h) represents the vector representation of the context subgraph of entity h; H(r) = E(r) + g(r), H(r) represents the knowledge representation of relation r, E(r) represents the vector representation of entity r, and g(r) represents the vector representation of the context subgraph of relation r; H(t) = E(t) + g(t), H(t) represents the knowledge representation of entity t, E(t) represents the vector representation of entity t, and g(h) represents the vector representation of the context subgraph of entity t.

[0013] Optionally, extracting the first context subgraph that is updated in the business knowledge graph includes:

[0014] Determine a first entity that has changed in the business knowledge graph;

[0015] An entity associated with the first entity within a preset hop is selected from the business knowledge graph, and the first entity and the selected entity are combined to obtain an updated context subgraph of the first entity.

[0016] Optionally, extracting the first context subgraph that is updated in the business knowledge graph includes:

[0017] Determine a first relationship that has changed in the business knowledge graph;

[0018] determining at least one entity pair associated with the first relationship;

[0019] For each of the entity pairs, a relationship path within a preset length associated with the first relationship is selected from the business knowledge graph, and the first relationship and the selected relationship path are combined to obtain a context subgraph where the first relationship is updated.

[0020] Optionally, the learning the first context subgraph to obtain a vector representation of the first context subgraph includes:

[0021] Inputting the first context subgraph into an attention-enhanced graph convolutional neural network for learning to obtain a vector representation and a weight of each node in the first context subgraph;

[0022] Based on the vector representation and weight of each node, a vector representation of the first context subgraph is calculated.

[0023] Optionally, the method further includes:

[0024] Extracting a second context subgraph of entities and relationships in the business knowledge graph, and extracting a second triple in the business knowledge graph;

[0025] Learning the second context subgraph according to the attention-enhanced graph convolutional neural network to obtain a vector representation of the second context subgraph;

[0026] Inputting the vector representation of the second context subgraph and the second triple into a translation model for learning, thereby obtaining knowledge representations of entities and relationships in the business knowledge graph;

[0027] Business recommendations are made using the knowledge representation of the entities and relationships.

[0028] In a second aspect, an embodiment of the present invention provides a service recommendation device, including:

[0029] An extraction module, used to extract a first context subgraph updated in the business knowledge graph, and to extract a first triplet updated in the business knowledge graph;

[0030] A first learning module, used to learn the first context subgraph to obtain a vector representation of the first context subgraph;

[0031] A second learning module is used to input the vector representation of the first context subgraph and the first triple into a translation model for learning, so as to obtain an updated knowledge representation of the business knowledge graph;

[0032] The recommendation module is used to make business recommendations using the updated knowledge representation.

[0033] Optionally, the loss function of the translation model is:

[0034]

[0035] Among them, (h, r, t) represents a triple; H(h) = E(h) + g(h), H(h) represents the knowledge representation of entity h, E(h) represents the vector representation of entity h, and g(h) represents the vector representation of the context subgraph of entity h; H(r) = E(r) + g(r), H(r) represents the knowledge representation of relation r, E(r) represents the vector representation of entity r, and g(r) represents the vector representation of the context subgraph of relation r; H(t) = E(t) + g(t), H(t) represents the knowledge representation of entity t, E(t) represents the vector representation of entity t, and g(h) represents the vector representation of the context subgraph of entity t.

[0036] Optionally, the extraction module includes:

[0037] A first determining unit, used to determine a first entity that has changed in the business knowledge graph;

[0038] A first selection unit, configured to select an entity associated with the first entity within a preset hop from the business knowledge graph;

[0039] The first combining unit is used to combine the first entity and the selected entity to obtain an updated context subgraph of the first entity.

[0040] Optionally, the extraction module includes:

[0041] A second determining unit, used to determine a first relationship that has changed in the business knowledge graph;

[0042] A third determining unit, configured to determine at least one entity pair associated with the first relationship;

[0043] A second selection unit is used to select, for each of the entity pairs, a relationship path within a preset length associated with the first relationship from the business knowledge graph;

[0044] The second combining unit is used to combine the first relationship and the selected relationship path to obtain an updated context subgraph of the first relationship.

[0045] Optionally, the first learning module includes:

[0046] A learning unit, configured to input the first context subgraph into an attention-enhanced graph convolutional neural network for learning, and obtain a vector representation and a weight of each node in the first context subgraph;

[0047] A computing unit is used to compute the vector representation of the first context subgraph based on the vector representation and weight of each node.

[0048] Optionally, the extraction module is further used to: extract a second context subgraph of entities and relationships in the business knowledge graph, and extract a second triple in the business knowledge graph;

[0049] The first learning module is further used to: learn the second context subgraph according to the attention-enhanced graph convolutional neural network to obtain a vector representation of the second context subgraph;

[0050] The second learning module is further used to: input the vector representation of the second context subgraph and the second triple into the translation model for learning, so as to obtain the knowledge representation of entities and relationships in the business knowledge graph;

[0051] The recommendation module is also used to make business recommendations using the knowledge representation of the entities and relationships.

[0052] In a third aspect, an embodiment of the present invention provides an electronic device, which includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the business recommendation method described above.

[0053] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the business recommendation method as described above is implemented.

[0054] In an embodiment of the present invention, when a business knowledge graph is updated, the electronic device can obtain the updated knowledge representation of the business knowledge graph based on the updated context subgraph information and triple information in the business knowledge graph, and use the updated knowledge representation to make business recommendations. Therefore, when making business recommendations, it is not necessary to learn the knowledge representation of the entire updated business knowledge graph, and only the updated context subgraph information and triple information can be considered, thereby achieving efficient business recommendations; further, by adding context information such as subgraphs, the learned knowledge representation can be made more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of a service recommendation method provided by an embodiment of the present invention;

[0056] Figure 2 is a schematic diagram of an attention-enhanced graph convolutional neural network in an embodiment of the present invention;

[0057] Figure 3A It is a partial structural diagram of a business knowledge graph in an example of the present invention;

[0058] Figure 3B is a schematic diagram of a context subgraph of entity e3 in an example of the present invention;

[0059] Figure 3C is a schematic diagram of a context subgraph of entity e1 in an example of the present invention;

[0060] Figure 4A It is a partial structural diagram of another business knowledge graph in an example of the present invention;

[0061] Figure 4B is a schematic diagram of a context subgraph of relation r1 in an example of the present invention;

[0062] Figure 5 It is a schematic diagram of a part of the structure of the mobile service knowledge graph in the example of the present invention;

[0063] Figure 6is a schematic diagram of the context subgraph of the entity "Wang Xiaoming" in the example of the present invention;

[0064] Figure 7 is a schematic diagram of the translation model training in the example of the present invention;

[0065] Figure 8 is a schematic diagram of a portion of the structure of the updated mobile business knowledge graph in an example of the present invention;

[0066] Fig. 9A is a schematic diagram of the context subgraph of the updated entity “Wang Xiaoming” in the example of the present invention;

[0067] Fig. 9B is a schematic diagram of a context subgraph of an updated entity “sales” in an example of the present invention;

[0068] Fig.10 It is a schematic diagram of the overall process of the service recommendation process in the example of the present invention;

[0069] Fig.11 is a structural diagram of a business recommendation device provided by an embodiment of the present invention;

[0070] Fig.12 It is a structural schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] The terms "first", "second", etc. in the specification and claims of the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0073] In order to facilitate understanding of the embodiments of the present invention, the following contents are first described.

[0074] Knowledge graph: It is essentially a semantic network that describes various entities and concepts in the real world, as well as the relationships between them. The nodes of the knowledge graph represent entities or concepts, and the edges represent the various semantic relationships between entities / concepts. In the knowledge graph, a relationship can be used to describe the association between two entities.

[0075] Knowledge representation: Use machine learning technology to map entities and relationships in the knowledge graph into dense, low-dimensional real-valued vectors, and then perform subsequent calculations and reasoning. The basic principle of knowledge representation is: for a triple (e1, r, e2), the vector of entity e1 plus the vector of relationship r is close to the vector of entity e2, that is, E(e1)+E(r)≈E(e2). Knowledge representation technology can be widely used in semantic similarity calculation, knowledge graph completion, relationship extraction, automatic question answering, recommendation systems, intelligent interactive systems and other fields.

[0076] Graph Convolutional Networks (GCN): A graph is a data format that can be used to represent social networks, communication networks, etc. The nodes in the graph represent individuals in the network, and the connected edges represent the connection relationship between individuals. Graph convolutional neural networks are an effective technology for learning this kind of graph structure. The input is the entire graph. In the convolution layer 1, a convolution operation can be performed on the neighbors of each entity, and the node is updated with the result of the convolution; then it passes through an activation function such as a linear rectifier function (also known as a rectified linear unit, ReLU), and then another convolution layer (Convolution Layer 2) and an activation function; repeat the above process until the number of layers reaches the expected depth. Finally, a vector is learned for each entity in the graph and output.

[0077] Knowledge representation learning is the representation learning of entities and relationships in knowledge graphs (knowledge bases). By projecting entities or relationships into low-dimensional vector spaces, it is possible to represent the semantic information of entities and relationships, and efficiently calculate entities, relationships, and the complex semantic relationships between them, which is of great significance to the construction, reasoning, and application of knowledge bases. However, existing knowledge representation technologies have the following problems:

[0078] (1) Existing knowledge representation technologies use triples in knowledge graphs as input and obtain vectorized representations of entities and relationships in triples through continuous iteration. However, this method does not fully utilize the contextual information in the knowledge graph, resulting in low accuracy of existing knowledge representation; (2) Existing knowledge representation technologies all perform knowledge representation learning on static knowledge graphs. If the knowledge graph is updated, the entire knowledge graph needs to be trained, which requires large amounts of computation and is time-consuming, and cannot achieve real-time updates of knowledge representation. In real applications, knowledge graphs are constantly changing as external information is added or deleted or updated. Therefore, existing knowledge graph vectorization representation technologies cannot meet the ever-changing scenarios in real applications.

[0079] In order to solve the above problems, namely how to make full use of the context information in the knowledge graph to improve the accuracy of knowledge representation and how to achieve continuous updating of knowledge representation with less computational effort, this application proposes a dynamic update mechanism for knowledge representation based on context subgraph.

[0080] The service recommendation method provided by the embodiment of the present invention is described in detail below through specific embodiments and application scenarios in conjunction with the accompanying drawings.

[0081] See also Figure 1 , Figure 1 is a flow chart of a service recommendation method provided by an embodiment of the present invention, the method is applied to an electronic device, such as Figure 1 As shown, the method comprises the following steps:

[0082] Step 101: extract the first context subgraph updated in the business knowledge graph, and extract the first triplet updated in the business knowledge graph.

[0083] In this embodiment, the business knowledge graph may be selected as but not limited to a communication business knowledge graph, a banking business knowledge graph, an advertising business knowledge graph, etc., and may be determined based on the actual application scenario.

[0084] For the above-mentioned first context subgraph, it can be a context subgraph of entities that have changed in the business knowledge graph (also called context information), or it can be a context subgraph of relationships that have changed in the business knowledge graph (also called context information).

[0085] In one implementation, the electronic device may monitor the update changes of the business knowledge graph in real time so as to extract the updated context subgraph information and triple information in the business knowledge graph.

[0086] Step 102: learning the first context subgraph to obtain a vector representation of the first context subgraph.

[0087] In this embodiment, when learning the first context subgraph, learning can be performed based on a graph convolutional neural network, or learning can be performed in another manner, without limitation.

[0088] Optionally, the learning process in the above step 102 may include: inputting the first context subgraph into an attention-enhanced graph convolutional neural network for learning to obtain a vector representation and weight of each node in the first context subgraph; and calculating the vector representation of the first context subgraph based on the vector representation and weight of each node. It should be noted that if the first context subgraph is a context subgraph of a changed entity, the nodes therein correspond to the entity; and if the first context subgraph is a context subgraph of a changed relationship, the nodes therein correspond to the relationship.

[0089] In one embodiment, the attention-enhanced graph convolutional neural network can be Figure 2 For example, if the context subgraph T is input into the attention-enhanced graph convolutional neural network, as Figure 2 As described above, the vector of each node vi in ​​the context subgraph T can be first obtained based on graph convolutional neural network learning, and then the importance of each node vi, i.e., weight a, can be learned based on the attention mechanism. i , and use Formula 1 or Formula 2 to calculate the vector representation of the context subgraph T (i.e., the subgraph embedding vector). If the context subgraph T is an entity context subgraph, Formula 1 can be used: Get the corresponding vector representation; or, if the context subgraph T is a relational context subgraph, Formula 2 can be used: Get the corresponding vector representation; n is the number of nodes (i.e., entities or relationships) in the context subgraph T.

[0090] Step 103: Input the vector representation of the first context subgraph and the first triple into a translation model for learning to obtain an updated knowledge representation of the business knowledge graph.

[0091] In this embodiment, the translation model may be a Translating Embedding (TransE) model. The translation model not only considers the triple information of the knowledge graph, but also considers the context information of the entity or relationship in the knowledge graph, that is, the context subgraph information.

[0092] Optionally, the basic principle of the translation model may be H(h)+H(r)≈H(t), and the loss function of the translation model may be:

[0093]

[0094] Among them, (h, r, t) represents a triple; H(h) = E(h) + g(h), H(h) represents the knowledge representation of entity h, E(h) represents the vector representation of entity h, and g(h) represents the vector representation of the context subgraph of entity h; H(r) = E(r) + g(r), H(r) represents the knowledge representation of relation r, E(r) represents the vector representation of entity r, and g(r) represents the vector representation of the context subgraph of relation r; H(t) = E(t) + g(t), H(t) represents the knowledge representation of entity t, E(t) represents the vector representation of entity t, and g(h) represents the vector representation of the context subgraph of entity t. In this way, through continuous iterative updates by stochastic gradient descent, the entity knowledge representation and relation knowledge representation of the business knowledge graph can be finally learned.

[0095] Step 104: Use the updated knowledge representation to make business recommendations.

[0096] The business recommendation method of the embodiment of the present invention can obtain the updated knowledge representation of the business knowledge graph based on the updated context subgraph information and triple information in the business knowledge graph when the business knowledge graph is updated, and use the updated knowledge representation to make business recommendations. Therefore, when making business recommendations, it is not necessary to learn the knowledge representation of the entire updated business knowledge graph, and only the updated context subgraph information and triple information can be considered, thereby achieving efficient business recommendations; further, by adding context information such as subgraphs, the learned knowledge representation can be made more accurate.

[0097] In an embodiment of the present invention, for an entity in a knowledge graph, entities associated with it within a preset hop (such as one hop, or two hops, etc.) can be selected to form a context subgraph.

[0098] Optionally, the above process of extracting the first context subgraph that has been updated in the business knowledge graph may include: determining the first entity that has changed in the business knowledge graph; selecting an entity that is associated with the first entity within a preset jump from the business knowledge graph, and combining the first entity and the selected entity to obtain the updated context subgraph of the first entity.

[0099] for example Figure 3A As shown in , if the first entity that changes is e3, and the entities and relationships associated within one hop are selected, then the entities associated with the entity e3 are e1 and e4, that is, the context subgraph of the entity e3 is as follows Figure 3B Or, if the first entity that changes is e1, and the entities and relationships associated within one hop are selected, the entities associated with the entity e1 are e2, e3, and e5, that is, the context subgraph of the entity e1 is as follows: Figure 3C shown.

[0100] In an embodiment of the present invention, for a relationship in a knowledge graph, a relationship path within a preset length associated with the entity pair can be selected to form a context subgraph. The preset length can be determined based on actual needs, such as less than or equal to 2.

[0101] Optionally, the above process of extracting the first context subgraph that has been updated in the business knowledge graph may include: determining a first relationship that has changed in the business knowledge graph; determining at least one entity pair associated with the first relationship; for each of the entity pairs, selecting a relationship path within a preset length associated with the first relationship from the business knowledge graph, and combining the first relationship and the selected relationship path to obtain the updated context subgraph of the first relationship.

[0102] for example Figure 4A As shown in , if the first relationship that changes is r1, and the entity pairs associated with this relationship r1 are (e5, e2), (e5, e3), and (e4, e1), a relationship path with a length less than or equal to 2 is selected. Since for the entity pair (e5, e2), in addition to the relationship r1 that can reach e2 from e5, the relationship pair (r5, r2) can also reach e2 from e5. Similarly, for the entity pair (e5, e3), in addition to the relationship r1 that can reach e3 from e5, the relationship pair (r1, r4) can also reach e3 from e5. For the entity pair (e4, e1), only the relationship r1 can reach e1 from e4. Therefore, the relationship paths of the r1 relationship are (r5, r2) and (r1, r4), respectively. The context subgraph of the r1 relationship can be as follows Figure 4B shown.

[0103] It is understandable that the above content introduces the knowledge representation update process. In addition, in this embodiment, the static and unupdated business knowledge graph can also be learned and used to make business recommendations.

[0104] Optionally, the service recommendation method in this embodiment may further include:

[0105] Extracting a second context subgraph of entities and relationships in the business knowledge graph, and extracting a second triple in the business knowledge graph; wherein the second triple can be selected as all or part of the important triples in the business knowledge graph;

[0106] Learning the second context subgraph according to the attention-enhanced graph convolutional neural network to obtain a vector representation of the second context subgraph;

[0107] Inputting the vector representation of the second context subgraph and the second triple into the translation model for learning, to obtain the knowledge representation of entities and relationships in the business knowledge graph; understandably, the specific structure of the translation model may be as described above;

[0108] Business recommendations are made using the knowledge representation of the entities and relationships.

[0109] In this way, by learning the entire business knowledge graph, complete knowledge representation information can be obtained, and further by adding contextual information such as subgraphs, the learned knowledge representation can be made more accurate.

[0110] The following describes the embodiment of the present invention in detail by taking providing high-quality service recommendations to communication users as an example.

[0111] In order to recommend satisfactory services to users and improve service quality, communication operators usually have a service recommendation knowledge graph (or service knowledge graph) that describes the semantic relationship between users and between users and communication services, and recommend services to users based on the knowledge in the service knowledge graph. Figure 5 As shown in Figure 1, part of the content of a business recommendation knowledge graph is described.

[0112] In real business recommendation, when the knowledge graph is vectorized and learned through the existing knowledge representation technology, the triples in the knowledge graph are used as input, and the vectorized representation of the entities and relations in the triples is obtained through continuous iteration. For example, the triples such as <Wang Daming's son Wang Xiaoming><Wang Xiaoming handles the business package of 188 yuan><Zhang Xiaoli handles the business package of 288 yuan> are used to learn the knowledge representation based on the TransE translation model principle e1+r≈e2. However, this method does not make full use of the contextual information of the entities or relations in the knowledge graph, and only learns the information contained in the triples, which makes the accuracy of the existing knowledge representation low; at the same time, in real applications, the business knowledge graph is constantly changing with the addition, deletion and update of external information. For example, the age of users in the business knowledge graph is constantly changing, and the occupation of users also changes over time. These changes will affect the changes in users' demand for communication services. Therefore, the vectorized representation of the business knowledge graph needs to be continuously updated to cope with the changing scenarios in real applications. However, the traditional method of knowledge representation based on static knowledge graphs can only update the entire graph when the knowledge graph changes. Since the business recommendation knowledge graph contains a large amount of user and business information and is large in scale, it is not feasible to update the entire knowledge graph due to the large amount of computation and time-consuming. Therefore, the existing knowledge representation technology cannot provide a knowledge representation with high accuracy and real-time update for the business recommendation knowledge graph.

[0113] Based on the above content, this embodiment proposes a knowledge representation dynamic update mechanism based on subgraphs and translation models, which is mainly divided into two parts: offline knowledge representation learning and online knowledge representation update. Through these two parts, the vectorized representation of the knowledge graph can be realized. At the same time, the knowledge graph can be continuously updated according to the changes in actual applications, and finally a high-accuracy, high-reliability, real-time, and dynamically changing knowledge representation is obtained. Figures 6 to 10 , the specific content can be as follows:

[0114] (1) Offline knowledge representation learning: This part can be understood as the learning of static knowledge graphs, such as Fig.10 As shown in the figure, it is mainly divided into two steps: context subgraph embedding model training and translation model training.

[0115] 1) Contextual subgraph embedding model training

[0116] a) Extract context subgraph based on business knowledge graph: Extract context subgraph of entities and relationships in the business knowledge graph (here, subgraph information within 1 hop is selected as context information of knowledge graph elements (entities or relationships)).

[0117] For example, Figure 6 As shown in the figure, the contextual subgraph information of "Wang Xiaoming" is described. Not only the information of triples such as <Wang Daming's son Wang Xiaoming><Wang Xiaoming's professional student><Wang Xiaoming's business package of 188 yuan> is considered, but also the subgraph information of entities and relationships, that is, contextual information.

[0118] b) Learning contextual subgraph embedding: This can learn a graph embedding vector for each contextual subgraph based on the attention-enhanced graph convolutional neural network, that is, learning the vector representation of each contextual subgraph.

[0119] 2) Translation model training

[0120] In this 2), if Figure 7 As shown, the mapping function of the translation model, that is, the loss function, is The learned vector representation and triple information of each context subgraph can be used to continuously iterate and update through the stochastic gradient descent method, and finally the entity knowledge representation H(e) and relationship knowledge representation H(r) of the business knowledge graph can be learned.

[0121] After that, based on the learned knowledge representation, service recommendations are made. For example, we can calculate which services are most similar to the knowledge representation of "Wang Xiaoming" + the knowledge representation of "handling services". Here, "Wang Xiaoming" + "handling services" is most similar to the "188 yuan package", so we can infer that "Wang Xiaoming" may apply for the "188 yuan package", that is, the communication service related to students, and the 188 yuan package can be recommended to him.

[0122] (2) Online knowledge representation learning:

[0123] In online knowledge representation learning, for the updated business recommendation knowledge graph, instead of retraining the entire updated knowledge graph, only the subgraphs that have changed are updated, improving the efficiency of knowledge representation update. Example: Taking mobile business recommendation as an example, when Wang Xiaoming's occupation changes from "student" to "salesperson", Wang Xiaoming's mobile business should change accordingly. Because compared with students, the sales occupation requires more call time, etc. However, the traditional method cannot update the knowledge representation in real time, so it cannot capture the real-time changes in the knowledge graph. Therefore, it is necessary to update the knowledge representation online in real time.

[0124] The specific process of online knowledge representation update can be as follows: 1) Extract the updated context subgraph and triples; 2) Obtain the vector representation of the updated context subgraph; 3) Update the knowledge representation. The vector and knowledge representation of other subgraphs in the knowledge graph that have not been updated remain unchanged. Use the stochastic gradient descent algorithm to continuously iterate and update the updated part of the knowledge graph, so as to finally learn the knowledge representation of the updated knowledge graph.

[0125] For example, when Wang Xiaoming's occupation changes from "student" to "salesperson", the business knowledge graph is updated as Figure 8 shown. Extract the updated context subgraph. For example, Fig. 9A is the context subgraph of the updated entity "Wang Xiaoming", Fig. 9B is the context subgraph of the updated entity "salesperson". After Wang Xiaoming's occupation changes from "student" to "salesperson", through the continuous update of the knowledge representation of the knowledge graph, it can be calculated which services are most similar to "Wang Xiaoming" + "handle business". Here, "Wang Xiaoming" + "handle business" is most similar to the "288 yuan package", so it can be inferred that Wang Xiaoming may handle the "288 yuan package", that is, the mobile business related to sales. Thus, according to the change of the user's information, satisfactory mobile services are provided for the user, improving the service quality.

[0126] As Fig.10 shown, the specific steps of online knowledge representation update can be as follows:

[0127] S1: Extract the updated context subgraph and triples; for example, the update changes of the knowledge graph can be monitored to extract the updated subgraph information and triple information. For example, Figure 8 is the update graph of the business knowledge graph, Fig. 9A and Fig. 9B are the extracted updated context subgraphs.

[0128] S2: Obtain the vector representation of the updated context subgraph: Input the extracted updated subgraph into the context subgraph embedding learning network to obtain the vector representation of the updated subgraph;

[0129] S3: Knowledge representation update: Based on the learned vector representation of the updated subgraph, it is input into the translation model together with the updated triples for knowledge representation learning. The loss function is the same as the loss function in offline learning. The other subgraph vectors and knowledge representations in the knowledge graph that have not been updated remain unchanged, and the stochastic gradient descent algorithm is used to iteratively update the updated part of the knowledge graph, and finally the knowledge representation of the updated knowledge graph is learned.

[0130] It should be noted that, in addition to mobile communication service recommendations, the embodiments of the present invention can also be applied to service recommendations in other scenarios. For example, in the intelligent customer service scenario, the service content that meets the needs of the user can be recommended to the user through knowledge representation; or in the intelligent home gateway scenario, the knowledge representation of the knowledge graph can be updated in real time by monitoring the user's behavior, and the user's personalized information can be continuously learned, so that the device can be intelligently controlled according to the updated knowledge representation to meet the personalized needs of the user.

[0131] See also Fig.11 , Fig.11 is a schematic diagram of a business recommendation device provided by an embodiment of the present invention, and the device is applied to electronic devices such as Fig.11 As shown, the service recommendation device 110 includes:

[0132] An extraction module 111 is used to extract a first context subgraph updated in the business knowledge graph, and to extract a first triplet updated in the business knowledge graph;

[0133] A first learning module 112, configured to learn the first context subgraph to obtain a vector representation of the first context subgraph;

[0134] A second learning module 113 is used to input the vector representation of the first context subgraph and the first triple into a translation model for learning, so as to obtain an updated knowledge representation of the business knowledge graph;

[0135] The recommendation module 114 is used to make business recommendations using the updated knowledge representation.

[0136] Optionally, the loss function of the translation model is:

[0137]

[0138] Among them, (h, r, t) represents a triple; H(h) = E(h) + g(h), H(h) represents the knowledge representation of entity h, E(h) represents the vector representation of entity h, and g(h) represents the vector representation of the context subgraph of entity h; H(r) = E(r) + g(r), H(r) represents the knowledge representation of relation r, E(r) represents the vector representation of entity r, and g(r) represents the vector representation of the context subgraph of relation r; H(t) = E(t) + g(t), H(t) represents the knowledge representation of entity t, E(t) represents the vector representation of entity t, and g(h) represents the vector representation of the context subgraph of entity t.

[0139] Optionally, the extraction module 111 includes:

[0140] A first determining unit, used to determine a first entity that has changed in the business knowledge graph;

[0141] A first selection unit, configured to select an entity associated with the first entity within a preset hop from the business knowledge graph;

[0142] The first combining unit is used to combine the first entity and the selected entity to obtain an updated context subgraph of the first entity.

[0143] Optionally, the extraction module 111 includes:

[0144] A second determining unit, used to determine a first relationship that has changed in the business knowledge graph;

[0145] A third determining unit, configured to determine at least one entity pair associated with the first relationship;

[0146] A second selection unit is used to select, for each of the entity pairs, a relationship path within a preset length associated with the first relationship from the business knowledge graph;

[0147] The second combining unit is used to combine the first relationship and the selected relationship path to obtain an updated context subgraph of the first relationship.

[0148] Optionally, the first learning module 112 includes:

[0149] A learning unit, configured to input the first context subgraph into an attention-enhanced graph convolutional neural network for learning, and obtain a vector representation and a weight of each node in the first context subgraph;

[0150] A computing unit is used to compute the vector representation of the first context subgraph based on the vector representation and weight of each node.

[0151] Optionally, the extraction module 111 is further used to: extract a second context subgraph of entities and relationships in the business knowledge graph, and extract a second triple in the business knowledge graph;

[0152] The first learning module 112 is further used to: learn the second context subgraph according to the attention-enhanced graph convolutional neural network to obtain a vector representation of the second context subgraph;

[0153] The second learning module 113 is further used to: input the vector representation of the second context subgraph and the second triple into the translation model for learning, so as to obtain the knowledge representation of entities and relationships in the business knowledge graph;

[0154] The recommendation module 114 is further used to make business recommendations using the knowledge representation of the entities and relationships.

[0155] It is understandable that the service recommendation device 110 of the embodiment of the present invention can achieve the above Figure 1 The various processes of the method embodiment shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0156] In addition, an embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program can achieve the above-mentioned Figure 1 The various processes of the method embodiment shown can achieve the same technical effect, and to avoid repetition, they will not be described again here.

[0157] See also Fig.12 As shown, an embodiment of the present invention further provides an electronic device 120 , including a bus 121 , a transceiver 122 , an antenna 123 , a bus interface 124 , a processor 125 and a memory 126 .

[0158] In the embodiment of the present invention, the electronic device 120 further includes: a computer program stored in the memory 126 and executable on the processor 125. Optionally, when the computer program is executed by the processor 125, the following steps may be implemented:

[0159] Extracting a first context subgraph updated in the business knowledge graph, and extracting a first triplet updated in the business knowledge graph;

[0160] Learning the first context subgraph to obtain a vector representation of the first context subgraph;

[0161] Inputting the vector representation of the first context subgraph and the first triple into a translation model for learning to obtain an updated knowledge representation of the business knowledge graph;

[0162] Business recommendations are made using the updated knowledge representation.

[0163] It is understandable that when the computer program is executed by the processor 125, the above Figure 1 The various processes of the method embodiment shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0164] exist Fig.12 In the embodiment, a bus architecture (represented by bus 121) is provided, wherein bus 121 may include any number of interconnected buses and bridges, and bus 121 links together various circuits including one or more processors represented by processor 125 and memory represented by memory 126. Bus 121 may also link together various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. Bus interface 124 provides an interface between bus 121 and transceiver 122. Transceiver 122 may be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. Data processed by processor 125 is transmitted on a wireless medium via antenna 123, and further, antenna 123 also receives data and transmits the data to processor 125.

[0165] The processor 125 is responsible for managing the bus 121 and general processing, and may also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. The memory 126 may be used to store data used by the processor 125 when performing operations.

[0166] Optionally, the processor 125 may be a CPU, an ASIC, an FPGA or a CPLD.

[0167] The embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which can implement the above-mentioned Figure 1 The various processes of the method embodiment shown can achieve the same technical effect, and to avoid repetition, they will not be described again here.

[0168] Computer readable media include permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0169] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0170] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0171] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a service classification device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0172] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A business recommendation method, characterized in that: include: Extracting a first context subgraph updated in a business knowledge graph of a communication service, and extracting a first triplet updated in the business knowledge graph; The business knowledge graph is used to describe the relationship between different users and the relationship between users and communication services; The first context subgraph is context information of users and / or relationships updated in the business knowledge graph; The first triplet is information of users who have been updated in the business knowledge graph; Learning the context information of the updated users and / or relationships in the business knowledge graph to obtain a vector representation of the context information of the updated users and / or relationships in the business knowledge graph; Inputting the vector representation of the context information of the updated user and / or relationship in the business knowledge graph and the information of the updated user in the business knowledge graph into the translation model for learning, thereby obtaining the updated knowledge representation of the business knowledge graph; The translation model is a translation embedding model, which is used to project the vector representation of the context information of the updated user and / or relationship in the business knowledge graph and the information of the updated user in the business knowledge graph into a low-dimensional vector space to obtain the updated knowledge representation of the business knowledge graph; Using the updated knowledge representation of the business knowledge graph to make business recommendations; The learning of the context information of the updated users and / or relationships in the business knowledge graph to obtain the vector representation of the context information of the updated users and / or relationships in the business knowledge graph includes: Inputting context information of updated users and / or relationships in the business knowledge graph into the attention-enhanced graph convolutional neural network for learning to obtain vector representations and weights of the users and / or relationships; Based on the vector representation and weight of the user and / or relationship, a vector representation of the context information of the updated user and / or relationship in the business knowledge graph is calculated.

2. The method according to claim 1, characterized in that The loss function of the translation model is: Among them, (h, r, t) represents a triple; H(h) = E(h) + g(h), H(h) represents the knowledge representation of entity h, E(h) represents the vector representation of entity h, and g(h) represents the vector representation of the context subgraph of entity h; H(r) = E(r) + g(r), H(r) represents the knowledge representation of relation r, E(r) represents the vector representation of entity r, and g(r) represents the vector representation of the context subgraph of relation r; H(t) = E(t) + g(t), H(t) represents the knowledge representation of entity t, E(t) represents the vector representation of entity t, and g(h) represents the vector representation of the context subgraph of entity t.

3. The method according to claim 1, characterized in that The step of extracting the first context subgraph updated in the business knowledge graph of the communication business includes: Determine a first user who has changed in the business knowledge graph; A user and / or communication service associated with the first user within a preset jump is selected from the service knowledge graph, and the first user and the selected user and / or communication service are combined to obtain updated context information of the first user.

4. The method according to claim 1, characterized in that: The step of extracting the first context subgraph updated in the business knowledge graph of the communication business includes: Determine a first relationship that has changed in the business knowledge graph, where the first relationship includes a relationship between different users and / or a relationship between a user and a communication service; Determine at least one entity pair associated with the first relationship, the entity pair comprising at least one of the following: an entity pair consisting of different users, an entity pair consisting of a user and a communication service; For each of the entity pairs, a relationship path within a preset length associated with the first relationship is selected from the business knowledge graph, and the first relationship and the selected relationship path are combined to obtain context information of the occurrence update of the first relationship.

5. The method according to claim 1, characterized in that: The method further comprises: Extracting a second context subgraph of users and relationships in the business knowledge graph, and extracting a second triplet in the business knowledge graph; the second context subgraph is context information of users and relationships in the business knowledge graph, and the second triplet is information of users in the business knowledge graph; Learning the context information of users and relationships in the business knowledge graph according to the attention-enhanced graph convolutional neural network to obtain a vector representation of the context information of users and relationships in the business knowledge graph; Inputting the vector representation of the context information of users and relationships in the business knowledge graph and the information of users in the business knowledge graph into the translation model for learning, thereby obtaining the knowledge representation of users and relationships in the business knowledge graph; Business recommendations are made using the knowledge representation of users and relationships in the business knowledge graph.

6. A business recommendation device, characterized in that: include: An extraction module, used to extract a first context subgraph updated in a business knowledge graph of a communication service, and to extract a first triplet updated in the business knowledge graph; The business knowledge graph is used to describe the relationship between different users and the relationship between users and communication services; The first context subgraph is context information of users and / or relationships updated in the business knowledge graph; The first triplet is information of users who have been updated in the business knowledge graph; A first learning module is used to learn the context information of the updated users and / or relationships in the business knowledge graph to obtain a vector representation of the context information of the updated users and / or relationships in the business knowledge graph; A second learning module is used to input the vector representation of the context information of the updated user and / or relationship in the business knowledge graph and the information of the updated user in the business knowledge graph into the translation model for learning, so as to obtain the updated knowledge representation of the business knowledge graph; The translation model is a translation embedding model, which is used to project the vector representation of the context information of the updated user and / or relationship in the business knowledge graph and the information of the updated user in the business knowledge graph into a low-dimensional vector space to obtain the updated knowledge representation of the business knowledge graph; A recommendation module, used to make business recommendations using the updated knowledge representation of the business knowledge graph; Wherein, the first learning module includes: A learning unit, used for inputting context information of updated users and / or relationships in the business knowledge graph into the attention-enhanced graph convolutional neural network for learning, so as to obtain vector representations and weights of the users and / or relationships; A calculation unit is used to calculate the vector representation of the context information of the updated user and / or relationship in the business knowledge graph based on the vector representation and weight of the user and / or relationship.

7. The device according to claim 6, characterized in that The extraction module comprises: A first determining unit, configured to determine a first user who has changed in the business knowledge graph; A first selection unit, configured to select, from the business knowledge graph, users and / or communication services associated with the first user within a preset hop; The first combining unit is used to combine the first user and the selected user and / or communication service to obtain updated context information of the first user.

8. The device according to claim 6, characterized in that The extraction module comprises: A second determining unit, configured to determine a first relationship that has changed in the business knowledge graph, wherein the first relationship includes a relationship between different users and / or a relationship between a user and a communication service; A third determining unit is used to determine at least one entity pair associated with the first relationship, wherein the entity pair includes at least one of the following: an entity pair composed of different users, and an entity pair composed of a user and a communication service; A second selection unit is used to select, for each of the entity pairs, a relationship path within a preset length associated with the first relationship from the business knowledge graph; The second combining unit is used to combine the first relationship and the selected relationship path to obtain context information of the update of the first relationship.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the business recommendation method as claimed in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that: The readable storage medium stores a program or an instruction, and when the program or the instruction is executed by the processor, the steps of the service recommendation method according to any one of claims 1 to 5 are implemented.

Images