CN106156488A - Knowledge graph based on Bayes's personalized ordering link Forecasting Methodology - Google Patents

Abstract

The present invention discloses a kind of knowledge graph based on Bayes's personalized ordering link Forecasting Methodology.The method considers the common factor of relation between prediction task subject and object by proposing a potential feature link forecast model, and utilizes Bayes's personalized ordering to optimize the forecast model of relation between subject and object.The inventive method can solve the problem that in knowledge graph link prediction the occuring simultaneously on the impact predicted the outcome of relation between subject and object, and avoids overfitting.

Description

Knowledge graph based on Bayes's personalized ordering link Forecasting Methodology

Technical field

The present invention relates to the Link Recommendation of knowledge graph, utilize Bayes's personalized ordering model to optimize the suitable of recommended links Sequence, belongs to computer technology, information network, data mining interleaving techniques application.

Background technology

Knowledge graph is an entity information storehouse that can deposit any data, often it is characterized in that being and other entity Relevant.Resource description framework (RDF) is the framework that a capture inter-entity is the most mutual.One RDF data collection is equivalent to One isomery figure, the most each summit and limit may belong to different entity class, and category information is trapped in various entity type and pass Taxonomical hierarchy structure between set type.From the natural language text of knowledge graph, set up the knowledge graph that can recommend need a large amount of Work because in extracting entity, relationship map and link prediction these three significant process, extracting data and semantic role mark Note is the most noisy.Therefore method for designing must reduce noise, this method to have considered below: (1) believable data Source；(2) probits obtained by a natural language processing engine, have expressed the correct probability analyzed；(3) want There is main body and the object of priori.Therefore the present invention uses the incorporation model of Bayes's personalization rank algorithm.

Link prediction in knowledge graph is a machine learning method, utilizes the priori of subject and object to assess three Trust value between key element.

Summary of the invention

Technical problem: it is an object of the invention to provide a kind of knowledge graph based on Bayes's personalized ordering link prediction side Method, in solution knowledge graph link prediction, between subject and object, the common factor of set of relationship, on the impact predicted the outcome, utilizes Bayes Personalized ordering optimizes the forecast model of relation between subject and object.

Technical scheme: knowledge graph link method based on Bayes's personalized ordering of the present invention propose one based on The link forecast model of feature carrys out the link in prediction knowledge figure knowledge graph, and utilizes Bayes's personalized ordering to optimize this mould Type.

Knowledge graph link method based on Bayes's personalized ordering of the present invention comprises the following steps:

Step 1) set up knowledge graph, specifically comprise the following steps that

Step 11) three elements figure G=(S, P, O) of definition knowledge graph, (s, p, o)=1 represents that main body S and object O exist to G Relation P, (s, p, o)=0 represents that main body S and object O do not exist relation P, wherein s ∈ S, p ∈ P, o ∈ O to G.Described three elements bag Containing the relation between main body, object and Subjective and Objective, described s is main body, and described o is object, and described p is the pass between subject and object System, described S is host complex, and described O is object set, and described P is the set of relation between subject and object.

Step 12) definition bipartite graph G_p(S_p,O_p), for relation p between subject and object, G_p(S_p,O_p)=0 represents main body s_p With object o_pBetween there is not relation p, G_p(s_p,o_p)=1 represents main body s_pWith object o_pBetween there is relation p, described s_p∈S_p,o_p∈ O_p, described bipartite graph is that the vertex set of figure may be partitioned into two mutually disjoint subsets, and two that in figure, each edge depends on Summit all belongs to the mutually disjoint subset of the two, and the summit in two subsets is non-conterminous.

Step 2) establish the link forecast model, specifically comprise the following steps that

Step 21) link prediction be conceptualized as, to each relation p with at G_pIn entity (s, o), target is to set up one Forecast model M_p(s, o), x_s,o=M_p(s, o), described p ∈ P, s ∈ S, o ∈ O, described x_s,oIt it is the result of forecast model.

Step 22) use logistic functionIt is used as standards of grading, describedIt is the probability of main body s and object o opening relationships p, describedIt is s Yu o opening relationships p.

Step 3) definition feature based forecast modelWherein s_pWith o_pIt is definition respectively Bipartite graph G_p(S_p,O_pSubject and object in), s ∈ S, p ∈ P, o ∈ O, described S are host complex, and described O is object set, Described P is the set of relation between subject and object, and described T is transposition operation.Described x_s,oIt is the result of forecast model,Value The biggest, s_pAnd o_pIn low gt, more for similar, the probability establishing the link relation is the biggest；DescribedDescribed U and V is low dimensional vector, and described IR is real number field, and described K is potential feature Number.Described vector space is such a set, and the most any two elements are added another yuan that may make up in this set Element, arbitrary element obtains another element in this set with Arbitrary Digit after being multiplied, described low-dimensional is that independent parameter number is less.

Step 4) utilize Bayes's personalized ordering algorithm to be optimized, step is as follows:

Step 41) set up order standardDescribed D_p It is training sample set, describedRepresent to exist between s with o and link, describedRepresent not exist between s with o and link, described σ (.) is logistic function, and described log istic function isDescribed a is function variable, describedFor the standardization of data, in order to avoid overfitting, described λ is a constant controlling matching speed, described Θ_p={ U^p,V^p,b^p, described U^p∈IR^|S|×K,V^p∈IR^|O|×K,b^p∈IR^|O|×1, described | S | represents host complex S interior element Number, described | O | represents host complex O interior element number.

Step 42) initialize U^pSized by for m × K value be all 0 matrix, initialization V^pSized by be all 0 for n × K value Matrix, b^pSized by for n × 1 value be all 0 vector.Described n is the element number in object set O, and described m is the theme set Element number in S, the number of the potential feature of described K.

Step 43) for eachIntegrating step 3) updateValue, finally return Return U^p, V^p, b^pValue, described renewalFormula be:

Described

Described

Described

Described

Described

Described α is a constant constant for Schistosomiasis control speed, and described λ is a constant controlling matching speed.

Step 5) according to step 2) and step 3) calculate the probability of the relation of establishing the linkDescribedIt it is the probability of main body s and object o opening relationships p；.

Step 6) definition prediction threshold values ε, whenTime, then there is relation p between prediction main body s and object o, whenTime, then there is not relation p between prediction main body s and object o.

Beneficial effect: the present invention proposes a kind of knowledge graph based on Bayes's personalized ordering link Forecasting Methodology, specifically Have the beneficial effect that:

1) we achieve one for the probability assessing knowledge graph three main points in link prediction.Specifically, I Used for reference successful method, in commending system field, use the algorithm of knowledge graph, and a prominent benchmark dataset entered Go and assessed thoroughly.

2) we have proposed the Link Recommendation model prediction task that a feature based embeds, and utilize Bayes personalized Optimisation technique based on rank algorithm sets up learning model for each predicate.

3) we apply a linear regression model (LRM) opening up for the induced subgraph of each predicate and original knowledge graph Flutter the relation between structure quantitative analysis precision of prediction.Our research indicate that, such as the index such as cluster coefficients and average degree, can Using use as the index affecting precision of prediction.

Accompanying drawing explanation

Fig. 1 is knowledge graph Link Recommendation method flow based on Bayes's personalization algorithm.

Detailed description of the invention

The link prediction specific embodiment that the present invention uses Bayes's personalized ordering algorithm below is retouched in more detail State.

With reference to the accompanying drawings 1, knowledge graph link problems is defined as bipartite graph, uses the following step that is embodied as:

(1) relation between the main body of Input knowledge figure, object and Subjective and Objective.

(2) bipartite graph G is built_p(S_p,O_pIn), with main body s_pWith object o_pCollect the vertex set as bipartite graph, main body and visitor Between body, relation p is as the limit collection of bipartite graph.G_p(S_p,O_p)=0 represents summit s_pWith summit o_pBetween there is not limit p, G_p(s_p,o_p)=1 Represent summit s_pWith summit o_pBetween there is limit p, described s_p∈S_p,o_p∈O_p。

Complete after knowledge graph link problems is defined as bipartite graph, forecast model to be established the link, specifically comprise the following steps that

(1) forecast model of feature based is set up

(2) logistic function is usedIt is used as standards of grading.

In being embodied as,It is worth the biggest, s_pAnd o_pIn low gt, more for similar, establish the link relation Probability is the biggest.DescribedDescribed U and V is low dimensional vector, and described IR is real number field, institute Stating the number that K is potential feature, described vector space is such a set, and the most any two elements are added and may make up this set Another interior element, arbitrary element obtains another element in this set with Arbitrary Digit after being multiplied, described low-dimensional is independent ginseng Keep count of less.DescribedIt is the probability of main body s and object o opening relationships p, describedIt is that s Yu o sets up pass It is p.

After link forecast model has been set up, Bayes's personalized ordering algorithm to be utilized is optimized, and is embodied as step Rapid as follows:

(1) U is initialized^pSized by for m × K value be all 0 matrix, initialization V^pSized by for n × K value be all 0 square Battle array, b^pSized by for n × 1 value be all 0 vector.Described n is the element number in object set O, and described m is the theme and gathers S Interior element number, the number of the potential feature of described K.

(2) for eachUpdateValue, finally return to U^p, V^p, b^pValue, Described renewalFormula be:

Described

Described

Described

Described

Described

Described α is a constant constant for Schistosomiasis control speed, and described λ is a constant controlling matching speed. DescribedIt it is order standard.

Described D_pIt is training sample set, describedRepresent to exist between s with o and link, describedRepresent between s and o There is not link, described σ (.) is logistic function, and described logistic function isDescribed a is that function becomes Amount, describedFor the standardization of data, in order to avoid overfitting, described λ is one and controls the normal of matching speed Number, described Θ_p={ U^p,V^p,b^p, described U^p∈IR^|S|×K,V^p∈IR^|O|×K,b^p∈IR^|O|×1, described | S | represents host complex S Interior element number, described | O | represents host complex O interior element number.

In being embodied as, according to the probability of linking relationshipSet prediction threshold values ε, when Time, then there is relation p between prediction main body s and object o, whenTime, then do not exist between prediction main body s and object o Relation p, describedIt it is the probability of main body s and object o opening relationships p.

Claims (4)

1. knowledge graph based on a Bayes's personalized ordering link Forecasting Methodology, it is characterised in that the method includes following step Rapid:

Step 1) set up knowledge graph；

Step 2) establish the link forecast model；

Step 3) definition feature based forecast modelWherein s_pWith o_pIt is two points of definition respectively Figure G_p(S_p,O_pSubject and object in), s ∈ S, p ∈ P, o ∈ O, described s are main bodys, and described o is object, described p be main body with Relation between object, described S is host complex, and described O is object set, and described P is the set of relation between subject and object, institute Stating T is transposition operation；Described x_s,oIt is the result of forecast model,It is worth the biggest, s_pAnd o_pIn low gt, more for phase Seemingly, the probability establishing the link relation is the biggest；DescribedDescribed U and V is low dimensional vector, Described IR is real number field, and described K is the number of potential feature；

Step 4) utilize Bayes's personalized ordering algorithm to be optimized；

Step 5) according to step 2) and step 3) calculate the probability of the relation of establishing the linkDescribedIt is Main body s and the probability of object o opening relationships p；

Step 6) definition prediction threshold values ε, whenTime, then there is relation p between prediction main body s and object o, whenTime, then there is not relation p between prediction main body s and object o.

Knowledge graph based on Bayes's personalized ordering the most according to claim 1 link Forecasting Methodology, it is characterised in that Described step 1) set up knowledge graph, specifically comprise the following steps that

Step 11) three elements figure G=(S, P, O) of definition knowledge graph, (s, p, o)=1 represents that main body S and object O exist relation to G (s, p, o)=0 represents that main body S and object O do not exist relation P, wherein s ∈ S, and p ∈ P, o ∈ O, described three elements comprise master for P, G Relation between body, object and Subjective and Objective, described S is host complex, and described O is object set, and described P is to close between subject and object The set of system,

Step 12) definition bipartite graph G_p(S_p,O_p), for relation p between subject and object, G_p(S_p,O_p)=0 represents main body s_pWith visitor Body o_pBetween there is not relation p, G_p(s_p,o_p)=1 represents main body s_pWith object o_pBetween there is relation p, described s_p∈S_p,o_p∈O_p, institute State the vertex set that bipartite graph is figure and may be partitioned into two mutually disjoint subsets, and two summits that in figure, each edge depends on are all Belonging to the mutually disjoint subset of the two, the summit in two subsets is non-conterminous.

Knowledge graph based on Bayes's personalized ordering the most according to claim 1 link Forecasting Methodology, it is characterised in that Described step 2) establish the link forecast model, specifically comprise the following steps that

Step 21) link prediction be conceptualized as, to each relation p with at G_pIn entity (s, o), target is to set up a prediction Model M_p(s, o), x_s,o=M_p(s, o), described p ∈ P, s ∈ S, o ∈ O, described x_s,oIt it is the result of forecast model；

Step 22) use logistic functionIt is used as standards of grading, described It is the probability of main body s and object o opening relationships p, describedIt is s Yu o opening relationships p.

Knowledge graph based on Bayes's personalized ordering the most according to claim 1 link Forecasting Methodology, it is characterised in that Described step 4) utilize Bayes's personalized ordering algorithm to be optimized, step is as follows:

Step 41) set up order standardDescribed D_pIt it is instruction Practice sample set, describedRepresent to exist between s with o and link, describedRepresent not exist between s with o and link, described σ (.) Being logistic function, described logistic function isDescribed a is function variable, describedWith In the standardization of data, in order to avoid overfitting, described λ is a constant controlling matching speed, described Θ_p={ U^p,V^p, b^p, described U^p∈IR^|S|×K,V^p∈IR^|O|×K,b^p∈IR^|O|×1, described | S | represents host complex S interior element number, described | O | Representing host complex O interior element number, described vector space is such a set, and the most any two elements are added and may make up this Another element in set, arbitrary element obtains another element in this set with Arbitrary Digit after being multiplied, described low-dimensional is only Vertical number of parameters is less；

Step 42) initialize U^pSized by for m × K value be all 0 matrix, initialization V^pSized by for n × K value be all 0 square Battle array, b^pSized by for n × 1 value be all 0 vector, described n is the element number in object set O, described m be the theme gather S Interior element number, the number of the potential feature of described K；

Step 43) for eachIntegrating step 3) updateValue, finally return to U^p, V^p, b^pValue, described renewalFormula be:

Described

Described

Described

Described

Described

Described α is a constant constant for Schistosomiasis control speed, and described λ is a constant controlling matching speed.