[Paper Reading]--Exploiting Relevance Feedback in Knowledge Graph

《Exploiting Relevance Feedback in Knowledge Graph》

Publication: KDD 2015

Authors: Yu Su， Shengqi Yang, etc.

Affiliation: UCSB...

1. Short description:

p { margin-bottom: 0.1in; line-height: 120% }
a:link { }

This paper formulate the novice graph relevance feedback problem,  which applies relevance feedback in information retrieval area to graph query. User positive and negative feedback to inversely input the original graph query and improve the query result.

2. Focus: graph query,  subgraph matching

3. Novelty:  user relevance feedback; binary classifier to decide the trade-off to re-rank or re-search from graph

4. Motivation:

the new thing about this paper is it consider the ambigous of user input query.

users who do not need to understand the complexity of the schema of data graph,  so the input node name, type or keywords are generally ambigous or even not in the data graph.

5. Algorithms:

the query-specific function is based on the previous paper in the same group -- SLQ "schemaless and structureless graph querying "

the new graph matching function after tuning is $g(\theta^{*} )$

The framework is as follows:

It explored the two types of inferences:

Type inference: Infer the implicit type of each query node

Context Inference: neighborhood of the entity

The cons: 

In my opinion:

(1) It only explored the simple two node and three node star query

(2) The ground truth for deciding the re-rank and re-search was not clearly stated, which I think it is important to decide the runtime trade-off of the re-rank and re-search

(3) In reality, it is also not reliable and challenging to construct the ground truth for a new data graph to decide the runtime trade-off.

Reference:

Su, Yu, et al. "Exploiting relevance feedback in knowledge graph search." Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2015.