An E-commerce Knowledge Graph Completion Method Based on Entity Importance and Similarity

Abstract

An e-commerce knowledge graph completion method based on entity importance and similarity represents the e-commerce knowledge graph in the form of triples, calculates the importance of e-commerce entities; divides entities into different categories through e-commerce entity similarity classification, and generates a negative triple set on this basis; initializes the entity and relationship embedding vector sets of the e-commerce knowledge graph; calculates the first-order importance of all relationships in the e-commerce knowledge graph; calculates the high-order importance of relationships in all positive triples, and calculates the preset high-order importance of relationships in all negative triples; calculates the comprehensive importance of relationships in all positive and negative triples; obtains the importance of all positive and negative triples; obtains the entity embedding vector set and the relationship embedding vector set by minimizing the loss function; obtains the completed triples. The present invention considers the importance of entities and relationships in the e-commerce knowledge graph, selects high-quality negative samples according to similarity classification, has a fast calculation speed and a high completion accuracy.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graphs, and particularly to a method for completing an e-commerce knowledge graph based on entity importance and similarity. Background Art

[0002] An e-commerce knowledge graph is essentially a semantic network that can model the relationships between e-commerce entities, and can formally describe the mutual relationships of e-commerce entities such as users, user attributes, products, and product attributes in the real e-commerce field. It can reflect the purchase relationship between users and products, the relationship between users, and the attributes of users and products, and is the core technology of e-commerce AI.

[0003] Completing an e-commerce knowledge graph means, through knowledge graph completion means, predicting the unknown purchase relationship between users and products in the knowledge graph based on the existing connection relationships between users, products, and product attributes, which is helpful for the convenience of users to purchase products, and also helps to increase the sales volume of products. The method for completing an e-commerce knowledge graph based on entity importance and similarity refers to, based on the network structure characteristics of the e-commerce knowledge graph and the similarity relationship between entities and relationships, inferring and predicting the probability ranking of missing entities and relationships, and then, according to the incomplete triple information to be predicted, selecting the entity or relationship with the highest ranking as the completion result of the triple. Since the data volume of the e-commerce knowledge graph is extremely large, it is particularly important for the accuracy of precise push to users. Since entity importance and similarity play a very important supplementary role in the process of completing the e-commerce knowledge graph, it is very important to consider entity importance and similarity in the completion process. Summary of the Invention

[0004] In order to overcome the deficiencies of the current method for completing an e-commerce knowledge graph, which lacks the distinction of the importance of entities and relationships and has a low accuracy, the present invention proposes a method for completing an e-commerce knowledge graph based on entity importance and similarity.

[0005] The specific technical steps for the present invention to solve its technical problems are as follows:

[0006] A method for completing an e-commerce knowledge graph based on entity importance and similarity includes the following steps:

[0007] Step 1: An e-commerce knowledge graph includes entities such as users, user attributes, products, and product attributes, as well as the relationships between entities. An e-commerce knowledge graph can be represented in the form of triples as G(E, R, S), where E represents a set of N e-commerce entities {e 1 , e 2 , e 3 ,..., e N}, and R represents a set of M relationships between entities {r 1, r 2 , r 3 , ..., r M}, where S is the set of all triples (h, r, t) in the e-commerce knowledge graph, where h ∈ E represents the head entity, t ∈ E represents the tail entity, and r ∈ R is the relationship between the head entity and the tail entity;

[0008] Step 2: In the e-commerce knowledge graph G, initialize the importance of each entity to 1. Arbitrarily select an entity u and calculate its entity importance:

[0009]

[0010] where

[0011]

[0012] where entity j is an entity connected to entity u, in(j) represents the set of entities connected to entity j, α ∈ (0, 1) is the random walk parameter, out(i) represents the out-degree of entity i. Traverse each entity in the e-commerce knowledge graph and repeat this step iteratively to obtain the importance of all entities in the e-commerce knowledge graph;

[0013] Step 3: In the e-commerce knowledge graph G, arbitrarily select two entities e i and e j , and calculate their similarity index:

[0014]

[0015] where ψ(e i ) represents the relationship-entity pairs connected to entity e i , |ψ(e i ) ∩ ψ(e j )| represents the number of common relationship-entity pairs of entity e i and e j , Γ(e i ) represents the set of triples connected to entity e i , |Γ(e i ) ∪ Γ(e j )| represents the number of all triples connected to entity e i , e j ; Traverse the e-commerce knowledge graph to obtain the similarity index between any two entities;

[0016] Step 4: In the e-commerce knowledge graph G, arbitrarily select an entity e i , and connect it to the entity e j with the highest similarity index, traverse the entire e-commerce knowledge graph, connect all entities with the entity having the highest similarity index, and several entity communities will be formed in the e-commerce knowledge graph; entities within a community belong to the same category and are connected to each other, while entities in different communities belong to different categories and are not connected to each other, and the number of communities is equal to the number of entity categories;

[0017] Step 5: Map the entities and relationships in the e-commerce knowledge graph into an n-dimensional vector space, where the embedding vector of entity i is represented as σ ei ={σ ei1 , σ ei2 , …, σ ein}, and the embedding vector of relationship i is represented as δ ri ={δ ri1 , δ ri2 ,..., δ rin}, where the set of entity embedding vectors is E v ={σ e1 , σ e2 ,..., σ eN}, and the set of relationship embedding vectors is R v ={δ r1 , δ r2 ,..., δ rM}, and randomly select n-dimensional data to initialize the entity and relationship embedding vector sets of the e-commerce knowledge graph G;

[0018] Step 6: In the e-commerce knowledge graph G, randomly select a relationship r i , and calculate the relationship importance between its first-order neighbor entities

[0019]

[0020] where h vi and t vi are the embedding vectors of the head entity h i and the tail entity t i of the relationship r i respectively, and ||·|| 2 represents the modulus of the vector. Traverse all relationships in the e-commerce knowledge graph G to obtain the first-order relationship importance of all relationships;

[0021] Step 7: In the e-commerce knowledge graph G, randomly select a relationship r i of a positive triple, and calculate its higher-order importance

[0022]

[0023] where, represents the tail entity of the relationship r i , represents the head entity of the relationship r i , Eimp (h i ) and E imp (t i ) respectively represent the entity importance of entities h i and t i , where where represents the number of triples containing relationship r i , represents the number of tail entities of the triples containing relationship r i , represents the number of head entities of the triples containing relationship r i . Traverse the e-commerce knowledge graph G to obtain the high-order importance of the relationships in all positive triples;

[0024] Step 8: In the e-commerce knowledge graph G, randomly select a positive triple (h, r, t). The head entity is replaced with any entity in the community where the entity is located with a probability of tph / (tph + hpt), or the tail entity is replaced with any entity in the community where the entity is located with a probability of hpt / (tph + hpt) to construct a negative triple (h′, r, t′). Traverse all the triples in the set S of positive triples in the e-commerce knowledge graph, and repeat Step 8 to obtain the set S′ of negative triples after replacing the entities;

[0025] Step 9: In the set S′ of negative triples, randomly select a relationship r i of a negative triple, and calculate its preset high-order importance

[0026]

[0027] . Traverse the set S′ of negative triples to obtain the preset high-order importance of the relationships in the negative triples;

[0028] Step 10: In the e-commerce knowledge graph G, randomly select a relationship r i of a positive triple, and calculate its comprehensive importance

[0029]

[0030] where α is a hyperparameter. Traverse the e-commerce knowledge graph G to obtain the comprehensive importance of the relationships in all positive triples;

[0031] Step 11: In the set S′ of negative triples, randomly select a relationship r i of a negative triple, and calculate its comprehensive importance

[0032]

[0033] . Traverse the set S′ of negative triples to obtain the comprehensive importance of the relationships in all negative triples;

[0034] Step 12: Arbitrarily select a positive triple pti in the e-commerce knowledge graph G and calculate its weight.

[0035]

[0036] Traverse the e-commerce knowledge graph G to obtain the importance of all positive triples.

[0037] Step 13: Arbitrarily select a negative triple nt in the negative triple set S′ and calculate its weight. i , calculate its weight

[0038]

[0039] Among them, E imp (h i ) and E imp (t i ′) represent the importance of entities h i and t i ′, R imp (r i ) represents the relationship importance of r i , and R imp ′(r i ) represents the relationship importance of r i ′; traverse the negative triple set S′ to obtain the importance of all negative triples.

[0040] Step 14: Calculate the loss function.

[0041] L = ∑ (h，r，t)∈S ∑ (h′，r，t′)∈S′ [γ + w + (t)d(h + r, t) - w - (t′)d(h′ + r, t′)] + (11)

[0042] Among them, γ is the margin hyperparameter, []+ represents the positive part operation, d is the distance between h + r and t, which is the modulus of the calculation result of the corresponding embedding vectors of h, r, and t.

[0043] Step 15: Repeat Step 12 to Step 14 to update the entity embedding vectors and relationship embedding vectors of the e-commerce knowledge graph G until L is less than the specified loss value, then terminate the iterative calculation to obtain the final entity embedding vector set and relationship embedding vector set of the e-commerce knowledge graph G.

[0044] Step 16: Arbitrarily select an incomplete triple (h, r,?) from the e-commerce knowledge graph, where? represents the missing e-commerce entity, traverse the entity set of the e-commerce knowledge graph G, calculate d(h + r, t), and select the entity t corresponding to the minimum d value as the completion result of the current triple to obtain the completed triple (h, r, t).

[0045] The technical concept of the present invention is to distinguish the importance of different entities and relationships in the knowledge graph, improve the quality of negative samples through similarity classification, and thereby improve the accuracy of e-commerce knowledge graph completion.

[0046] As described above, the specific implementation steps of the present invention make the present invention clearer. Any modifications and changes made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

[0047] The advantages of the present invention are: enhancing the accuracy of e-commerce knowledge graph completion by calculating the importance of entities and relationships, with a relatively small amount of calculation and high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of an e-commerce knowledge graph. The dots are entities in the knowledge graph, and each entity can abstractly represent information such as users, user attributes, products, and product attributes. There are lines connecting between entities, representing the relationships between entities. DETAILED DESCRIPTION OF THE INVENTION

[0049] The present invention will be further described below with reference to the accompanying drawings.

[0050] Refer to Figure 1 , a user recommendation method for e-commerce products applying an e-commerce knowledge graph completion method based on entity importance and similarity of the present invention, includes the following steps:

[0051] Step 1: An e-commerce knowledge graph includes entities such as users, user attributes, products, and product attributes, as well as the relationships between entities. An existing e-commerce knowledge graph is represented in the form of triples as G(E, R, S), where E represents a set of N e-commerce entities {e 1 , e 2 , e 3 ,..., e N}, R represents a set of M relationships between entities {r 1 , r 2 , r 3 ,..., r M}, and S is a set of all triples (h, r, t) in the e-commerce knowledge graph, where h ∈ E represents the head entity, t ∈ E represents the tail entity, and r ∈ R is the relationship between the head entity and the tail entity. As Figure 1 shown, the head entity represented by the dot h 1 represents the user h 1 , the straight line r 2 represents the purchase relationship, and the tail entity represented by the dot t 2 represents the product t 2 , forming a triple (h1 , r 2 , t 2 );

[0052] Step 2: In the e-commerce knowledge graph G, initialize the importance of each entity to 1. Arbitrarily select an entity u and calculate its entity importance

[0053]

[0054] where

[0055]

[0056] where entity j is a connected entity of entity u, in(j) represents the set of entities connected to entity j, α ∈ (0, 1) is a random walk parameter, out(i) represents the out-degree of entity i. Traverse each entity in the e-commerce knowledge graph and repeat this step iteratively to obtain the importance of all entities in the e-commerce knowledge graph;

[0057] Step 3: In the e-commerce knowledge graph G, arbitrarily select two entities e i and e j , and calculate their similarity index

[0058]

[0059] where ψ(e i ) represents the relation-entity pairs connected to entity e i , |ψ(e i ) ∩ ψ(e j )| represents the number of common relation-entity pairs of entities e i and e j , Γ(e i ) represents the set of triples connected to entity e i , |Γ(e i ) ∪ Γ(e j )| represents the number of all triples connected to entities e i , e j ; Traverse the e-commerce knowledge graph to obtain the similarity index between any two entities;

[0060] Step 4: In the e-commerce knowledge graph G, arbitrarily select an entity e i , and connect it to the entity e j with the highest similarity index. Traverse the entire knowledge graph and connect all entities to the entities with the highest similarity index. The e-commerce knowledge graph will form several entity communities; Entities within a community belong to the same category and are connected to each other, while entities in different communities belong to different categories and are not connected to each other. The number of communities is equal to the number of entity categories;

[0061] Step 5: Map the entities and relationships of the e-commerce knowledge graph into an n-dimensional vector space, where the embedding vector of entity i is represented as σ ei ={σ ei1 , σ ei2 , …, σ ein}, and the embedding vector of relationship i is represented as δ ri ={δ ri1 , δ ri2 ,..., δ rin}, where the set of entity embedding vectors is E v ={σ e1 , σ e2 ,..., σ eN}, and the set of relationship embedding vectors is R v ={δ r1 , δ r2 ,..., δ rM}; randomly select n-dimensional data to initialize the entity and relationship embedding vector sets of the e-commerce knowledge graph G;

[0062] Step 6: In the e-commerce knowledge graph G, randomly select a relationship r i , and calculate the relationship importance between its first-order neighbor entities

[0063]

[0064] where h vi and t vi are the embedding vectors of the head entity h i and the tail entity t i of the relationship r i respectively, and ||·|| 2 represents the modulus of the vector. Traverse all relationships in the e-commerce knowledge graph G to obtain the first-order relationship importance of all relationships;

[0065] Step 7: In the e-commerce knowledge graph G, randomly select a relationship r i of a positive triple, and calculate its higher-order importance

[0066]

[0067] where, represents the tail entity of the relationship r i , represents the head entity of the relationship r i , E imp (h i ) and E imp (t i ) represent the entity importance of entities h i and t i respectively, where Indicates the number of triples i that contain the relationship r i , Indicates the number of tail entities i that contain the relationship r i , Indicates the number of head entities i that contain the relationship r. Traverse the e-commerce knowledge graph G to obtain the high-order importance of the relationships in all positive triples; i ,

[0068] Step 8: In the e-commerce knowledge graph G, randomly select a positive triple (h, r, t). The head entity is replaced with any entity in the community where the entity is located with a probability of tph / (tph + hpt), or the tail entity is replaced with a probability of hpt / (tph + hpt) to construct a negative triple (h′, r, t′). Traverse all the triples in the set S of positive triples in the e-commerce knowledge graph and repeat Step 8 to obtain the set S′ of negative triples after replacing the entities;

[0069] Step 9: In the set S′ of negative triples, randomly select the relationship r of a negative triple i , and calculate its preset high-order importance

[0070]

[0071] Traverse the set S′ of negative triples to obtain the preset high-order importance of the relationships in the negative triples;

[0072] Step 10: In the e-commerce knowledge graph G, randomly select the relationship r of a positive triple i , and calculate its comprehensive importance

[0073]

[0074] where α is a hyperparameter. Traverse the e-commerce knowledge graph G to obtain the comprehensive importance of the relationships in all positive triples;

[0075] Step 11: In the set S′ of negative triples, randomly select the relationship r of a negative triple i , and calculate its comprehensive importance

[0076]

[0077] Traverse the set S′ of negative triples to obtain the comprehensive importance of the relationships in all negative triples;

[0078] Step 12: In the e-commerce knowledge graph G, randomly select a positive triple pti and calculate its weight

[0079]

[0080] Traverse the e-commerce knowledge graph G to obtain the importance of all positive triples;

[0081] Step 13: Arbitrarily select a negative triple nt from the set S' of negative triples i , and calculate its weight

[0082]

[0083] where E imp (h i ) and E imp (t i ′) represent the importance of entities h i and t i ′, R imp (r i ) represents the relation importance of r i , and R imp ′(r i ) represents the relation importance of r i ′; traverse the set S' of negative triples to obtain the importance of all negative triples;

[0084] Step 14: Calculate the loss function;

[0085] L = ∑ (h，r，t)∈S ∑ (h′，r，t′)∈S′ [γ + w + (t)d(h + r, t) - w - (t′)d(h′ + r, t′)] + (11)

[0086] where γ is the margin hyperparameter, []+ represents the positive part operation, d is the distance between h + r and t, which is the norm of the calculation result of the corresponding embedding vectors of h, r, and t;

[0087] Step 15: Repeat Steps 12 to 14 to update the entity embedding vectors and relation embedding vectors of the e-commerce knowledge graph G until L is less than the specified loss value, then terminate the iterative calculation to obtain the final entity embedding vector set and relation embedding vector set of the e-commerce knowledge graph G;

[0088] Step 16: Arbitrarily select an incomplete triple (h, r,?) from the e-commerce knowledge graph, where h represents the user, r represents purchase, and? represents the missing commodity entity. Traverse the entity set of the e-commerce knowledge graph G, calculate d(h + r, t), and select the entity t corresponding to the minimum d value as the completion result of the current triple to obtain the completed triple (h, r, t). As Figure 1 shown, the triple to be completed is (h 1 , r 1 ,?), and the entity corresponding to the minimum d(h 1 + r 1 , t 1 ) is t1 , the completed triple is (h 1 , r 1 , t 1 ), that is, it can be predicted that user h 1 may purchase product t 1 .

[0089] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art according to the inventive concept of the present invention.

Claims

1. An e-commerce knowledge graph completion method based on entity importance and similarity, comprising the following steps: Step 1: An e-commerce knowledge graph includes entities and the relationships between entities. The entities include users, user attributes, products, and product attributes. An e-commerce knowledge graph is represented in the form of triples as G(E, R, S), wherein, Let \(E\) denote the set of \(N\) e - commerce entities \(\{e 1 ,e 2 ,e 3 ,\cdots,e N \}\), and let \(R\) denote the set of \(M\) relationships between entities \(\{r 1 ,r 2 ,r 3 ,\cdots,r M \}\). Let \(S\) be the set of all triples \((h,r,t)\) in the e - commerce knowledge graph, where \(h\in E\) represents the head entity, \(t\in E\) represents the tail entity, and \(r\in R\) is the relationship between the head entity and the tail entity; Step 2: In the e-commerce knowledge graph G, initialize the importance of each entity to 1. Arbitrarily select an entity u and calculate its entity importance: where where entity j is an entity connected to entity u, in(j) represents the set of entities connected to entity j, α ∈ (0, 1) is a random walk parameter, out(i) represents the out-degree of entity i. Traverse each entity in the e-commerce knowledge graph and repeat this step iteratively to obtain the importance of all entities in the e-commerce knowledge graph; Step 3: In the e-commerce knowledge graph G, arbitrarily select two entities e i and e j , and calculate the similarity index between them: Among them, ψ(e i ) represents the relation-entity pairs connected to entity e i . |ψ(e i ) ∩ ψ(e j )| represents the number of common relation-entity pairs of entities e i and e j . Γ(e i ) represents the set of triples connected to entity e i . |Γ(e i ) ∪ Γ(e j )| represents the number of all triples connected to entities e i and e j ; Traverse the e-commerce knowledge graph to obtain the similarity index between any two entities; Step 4: In the e-commerce knowledge graph G, arbitrarily select an entity e i , and connect it to the entity e j with the highest similarity index. Traverse the entire e-commerce knowledge graph, connect all entities to the entities with the highest similarity index, and several entity communities will be formed in the e-commerce knowledge graph; the entities within a community belong to the same category and are connected to each other, while the entities in different communities belong to different categories and are not connected to each other. The number of communities is equal to the number of entity categories; Step 5: Map the entities and relationships of the e-commerce knowledge graph into an n-dimensional vector space, where the embedding vector of entity i is represented as σ ei ={σ ei1 , σ ei2 ,..., σ ein}, and the embedding vector of relationship i is represented as δ ri ={δ ri1 , δ ri2 ,..., δ rin}, where the set of entity embedding vectors is E v ={σ e1 , σ e2 ,..., σ eN}, and the set of relationship embedding vectors is R v ={δ r1 , δ r2 ,..., δ rM}; randomly select n-dimensional data to initialize the sets of entity and relationship embedding vectors of the e-commerce knowledge graph G; Step 6: In the e-commerce knowledge graph G, randomly select a relationship r i , and calculate the relationship importance among its first-order neighbor entities where h vi and t vi are the embedding vectors of the head entity h i and the tail entity t i of the relation r i respectively, ||·|| 2 represents the norm of the vector. Traverse all the relations in the e-commerce knowledge graph G to obtain the first-order relation importance of all relations; Step 7: Arbitrarily select a relationship r of a positive triple in the e-commerce knowledge graph G i , and calculate its high-order importance Among them, represents the tail entity of the relationship r i , represents the head entity of the relationship r i , E imp (h i ) and E imp (t i ) respectively represent the entity importance of entities h i and t i . Among them represents the number of triples containing the relationship r i , represents the number of tail entities containing the relationship r i , represents the number of head entities containing the relationship r i . Traverse the e-commerce knowledge graph G to obtain the high-order importance of relationships in all positive triples; Step 8: In the e-commerce knowledge graph G, arbitrarily select a positive triple (h, r, t). The head entity is replaced with any entity in the community where the entity is located with a probability of tph / (tph + htp), or the tail entity is replaced with a probability of hpt / (tph + hpt) to construct a negative triple (h, r, t'). Traverse all triples in the set S of positive triples in the e-commerce knowledge graph and repeat Step 8 to obtain a set S' of negative triples after entity replacement; Step 9: In the negative triple set S', randomly select a relation r of a negative triple i , and calculate its preset high-order importance Traverse the set S' of negative triples to obtain the preset high-order importance of the relationships in the negative triples; Step 10: In the e-commerce knowledge graph G, randomly select a relationship r of a positive triple and calculate its comprehensive importance i , and calculate its comprehensive importance where β is a hyperparameter. Traverse the e-commerce knowledge graph G to obtain the comprehensive importance of the relationships in all positive triples; Step 11: Arbitrarily select a relation r of a negative triple in the negative triple set S', and calculate its comprehensive importance i , and calculate its comprehensive importance Traverse the set S' of negative triples to obtain the comprehensive importance of the relationships in all negative triples; Step 12: Arbitrarily select a positive triple pt in the e-commerce knowledge graph G i , and calculate its weight Traverse the e-commerce knowledge graph G to obtain the importance of all positive triples; Step 13: Arbitrarily select a negative triple nt from the set S' of negative triples and calculate its weight i , and calculate its weight Among them, E imp (h i ) and E imp (t i ′) represent the importance of entities h i and t i ′, R imp (r i ) represents the relational importance of r i in the positive triple, and R imp ′(r i ) represents the relational importance of r i in the negative triple; traverse the negative triple set S', and obtain the importance of all negative triples; Step 14: Calculate the loss function: L = ∑ (h,r,t)∈S ∑ (h′,r,t′)∈S′ [γ + w + (t) d(h + r, t) - w - (t') d(h′ + r, t′)] + (11) where γ is the margin hyperparameter, + denotes the positive part operation, d is the distance between h + r and t, which is the norm of the calculation result of the corresponding embedding vectors of h, r, and t; Step 15: Repeat Steps 12 to 14 to update the entity embedding vectors and relationship embedding vectors of the e-commerce knowledge graph G until L is less than the specified loss value, then terminate the iterative calculation to obtain the final set of entity embedding vectors and relationship embedding vectors of the e-commerce knowledge graph G; Step 16: Arbitrarily select an incomplete triple (h, r,?) from the e-commerce knowledge graph, where? represents the missing e-commerce entity. Traverse the set of entities in the e-commerce knowledge graph G and calculate d(h + r, t). Select the entity t corresponding to the minimum d value as the completion result of the current triple to obtain the completed triple (h, r, t).

Images