Machine Learning with Knowledge Graphs

Volker Tresp gave a nice keynote talk at ESWC 2014 on leveraging the matrix factorization approach for prediction tasks on knowledge graphs (or ontologies, taxonomies, thesauri, RDF graphs, Web of Data, Linked Open Data... you name it).

Because I've been involved in the topic since 2012, for those who are interested in RESCAL model (the main avenue of his research group) and related work, I can complement his talk with a few (new and old) resources, he did not mention.

RESCAL Extensions

Exploiting type constraints while factorization:

[1] D. Krompass, M. Nickel, V. Tresp. Large-Scale Factorization of Type-Constrained Multi-Relational Data. 2014 [PDF]

[2] C. Kai-Wei, W. Yih, B. Yang, C. Meek. Typed Tensor Decomposition of Knowledge Bases for Relation Extraction // Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP 2013). - 2014 [PDF]

Using RESCAL in information retrieval (entity search), i.e., my visiting project at Emory University.

[3] N. Zhiltsov, E. Agichtein. Improving Entity Search over Linked Data by Modeling Latent Semantics. Proceedings of the International Conference on Information and Knowledge Management (CIKM 2013). - 2013 [PDF]

RESCAL Implementations

Original implementation:

[1] scikit-tensor. https://github.com/mnick/scikit-tensor, 2013

RESCAL in sparse matrices and coupled tensor-matrix factorization:

[2] Ext-RESCAL. https://github.com/nzhiltsov/Ext-RESCAL, 2012

Fond of machine learning and remote control helicopters and RC quadcopters? Buy one on AliExpress!

***

My Visiting Project at Emory University: Entity Search over Linked Data

In this post, I would like to summarize some research output of my recent work at Emory University as a visiting graduate student, which was funded via the Fulbright Program. The work in collaboration with professor Eugene Agichtein is devoted to entity search over linked data, a very hot topic nowadays. To put that into the context, Google Knowledge Graph, Facebook Graph Search and WolframAlpha are likely the best examples of products that empower entity-centric user experience. That is, this work is about how to resolve search queries soliciting information about certain people, organizations, locations, events, scientific concepts and other entities, assuming that such information has been somehow extracted, structured and shared ("linked data"). For example, you may think of entity descriptions available at Wikipedia, Freebase, GeoNames, Last.FM, CiteSeer and other compelling web resources.

Our main research contribution is a novel mathematical model that incorporates the semantics of similarity links in the entity graph into the ranking mechanism in a scalable way. A part of this work is presented in the 4-page poster paper Zhiltsov, N., Agichtein, E. Improving Entity Search over Linked Data by Modeling Latent Semantics accepted for CIKM 2013. The remaining part will be published (hopefully!) as a full paper early next year.

From the engineering perspective, this project gave birth to a bunch of open source software spin-offs that might be interesting for researchers/developers from related communities (i.e., machine learning, information retrieval, Semantic Web):

• Ext-RESCAL is a memory efficient and scalable implementation of the sparse tensor factorization algorithm RESCAL
• Anduin is for processing RDF/N-Quads data on Apache Hadoop
• Lucene-MLM adds support of mixture of language models to Apache Lucene 4.0
• a large data set of labeled entity search queries and entities from Yahoo! SemSearch Challenge. 

PS. I would like to thank all the people who have made this my first visit to the United States not only possible but wonderful (deliberately hiding most of their names to protect their privacy): the Fulbright representatives in Moscow and Atlanta/Georgia, professor Eugene Agichtein & his colleagues from the Intelligent Information Access lab at Emory University, people from Math & CS department and ISSS office at Emory University, my new friends in Atlanta, kickball teammates, and Fulbrighters in Georgia I met during our terrific Fall Trip. You're great!