Explaining and Suggesting
Relatedness
in Knowledge Graphs
--刘达欣
Contents
• Introduction
• Promblem Formalization
• The Recap Approch
• Implementation&Evaluation
• Concluding&Future work
• Perspective of RECAP
Introduction
• Background：Browsing,Searching,and Discoverying
knowledge in KG(Knowledge Graph).
• Problem:
• A.How to build relatedness explanations
• B.How to query KGs
• Related Work: (O output,F filter,Q query,L local)
Promblem Formalization
• Input:
K=<G,O,A>.(Kownledge,Graph,Ontology,Endpoint)
• Assumptions
• Support Different KB(KGs in the LOD cloud).That’s to the
the computations are reduced to a set of queries against A.
• Desired Output:
Given K=<G,O,A> and a pairs of (ws,wt) return
• Definition of KG
𝐺𝑒 (𝑤𝑠 , 𝑤𝑡 ) ⊆ 𝐺
• T:a set of RDF triple.
• KG: G=<V,𝜀 >. V is subject or object of instance in T.
𝜀 = {(𝑠, 𝑝, 𝑜) ∈ 𝑇
The Recap Approch
• Definition Explanation
• Given a knowledge base K=<G,O,A> and a pair
of entities (ws, wt) where ws, wt∈ G, an explanation is a
tuple of the form E=(ws, wt,Ge), where ws, wt ∈ Ge, Ge ⊆ G,
and Ge is connected
• Definition Necessary Edge
• 边e ∈ G在E=(ws, wt,Ge)上是必须的当且仅当e位于ws,
wt间的简单路径上
• Definition Minimal Explanation
E=(ws, wt,Ge)是最小的当且仅当E由ws, wt间的所有简单
路径合并得到
The Recap Approch
• Solution of Problem A：
• Algorithm 1: Building Relatedness Explanations
•
•
•
•
•
输入:一对实体(ws, wt),整数k，查询端点A的地址
输出:一个表示Explanation的图Ge
(1)Find paths.
(2)Rank paths
(3) Select and merge top-m paths
The Recap Approch
-- Solution of Problem A
• Find Paths:
• Definition k-connectivity Pattern
• 给定K=<G,O,A>,实体对(ws, wt),整数k。一个k连通模式由4元式
组成𝛱 =< 𝑤𝑠 , 𝑤𝑡 , 𝑄, 𝑘 > ，其中Q为一个由k个三元组模式连接的
查询集合。
• Examples:一个2连通模式(实体对为(:FL,:TvH ),k=2)
•
•
•
•
SELECT DISTINCT * WHERE{:FL ?p1 ?n1. ?n1 ?p2 :TvH}
SELECT DISTINCT * WHERE{:FL ?p1 ?n1. :TvH ?p2 ?n1}
SELECT DISTINCT * WHERE{?n1 ?p1 :FL. :TvH ?p2 ?n1}
SELECT DISTINCT * WHERE{?n1 ?p1 :FL. ?n1 ?p2 :TvH}
• Definition Path
Given K=<G,O,A> and a k-connectivity pattern
Π=<ws, wt,Q, k>, a path π is a set of edges:
𝜋(𝑤𝑠 , 𝑤𝑡 ) = 𝑤𝑠
𝑝1
𝑛1
𝑝2
其中ni为实体，pi为(逆)属性
𝑛2
𝑝3
𝑛3 . . . 𝑛𝑞
𝑝𝑘
𝑤𝑡
The Recap Approch
-- Solution of Problem A
• Rank Paths
• Ranking by Path Informativeness
• Ranking by Pattern Informativeness
• Ranking by Path Diversity
• 1 Path Informativeness
• Predicate Frequency Inverse Triple Frequency(pfitf)
The Recap Approch
-- Solution of Problem A
• Rank Paths
• 2 Ranking by Pattern Informativeness
• Pattern
•路
𝜋(𝑤𝑠 , 𝑤𝑡 ) = 𝑤𝑠
𝑝1
𝑛1
𝑝2
𝑛2 . . . 𝑛𝑞
𝑝𝑞
𝑤𝑡 的patten表示为
𝑝𝑞
𝑝2
? 𝑣1 ? 𝑣2 . . . ? 𝑣𝑞 𝑤𝑡
另𝑃𝜋 表示从路径集合𝑃𝜋中抽取的patten集合，那么
𝜋(𝑤𝑠 , 𝑤𝑡 ) = 𝑤𝑠
𝑝1
对于一个特定的pattern
𝜋 ∈ 𝑃𝜋
的信息量有：
The Recap Approch
-- Solution of Problem A
• Rank Paths
• 3 Ranking by Path Diversity
• Definition Path Diversity
• 其中π1，π2为ws,wt间的两条路径
• Labels(π)表示路径π上的RDF谓语集合
The Recap Approch
-- Solution of Problem A
• Selection and Mergeing Strategies
The Recap Approch
-- Solution of Problem B
• Query KG Algorithm：
• Input :一对实体(ws,wt),整数k,查询端点A
• Output:一个排序后的实体集合
(1)
(2)
(3)
(4)
使用算法1找到一个Explanation E=<ws,wt,Ge>
Build 实体查询模式Qe
通过端口A，以模式Qe查询 KG,获取一个实体集合
对实体集合排序
The Recap Approch
-- Solution of Problem B
• Build Explanation Pattern
• Definition Explanation Pattern
The Recap Approch
-- Solution of Problem B
• Rank Result
• PageRank
Implementation&Evaluation
• Recap Based on JavaFX(GUI) and Jena
• Experiment 1:Performance
Concluding&Future work
• Solution of two problem:
• A.How to build relatedness explanations
• B.How to query KGs
• For A: Find pathsRank pathsSelect&merge top-m
paths
• For B： Build Query Pattern of Query Entities via Query
Pattern Rank the Entities(PageRank)
• Future Work:
• Find how to accelerate and Explore better rank algorithm.
Perspective of RECAP:
Thank you!
Q&A