Authors:
David Ralph
1
;
Yunjia Li
2
;
Gary Wills
1
and
Nicolas G. Green
1
Affiliations:
1
Electronics and Computer Sciences (ECS), University of Southampton, University Rd, Southampton, SO17 1BJ and U.K
;
2
Launch International LTD, 3000a Parkway, Whiteley, Fareham, PO15 7FX and U.K
Keyword(s):
Recommender Systems, Information Retrieval, Data Mining, Sparse Data, Partially Labelled Data.
Abstract:
In this paper, we introduce Transitive Semantic Relationships (TSR), a new technique for ranking recommendations from cold-starts in datasets with very sparse, partial labelling, by making use of semantic embeddings of auxiliary information, in this case, textual item descriptions. We also introduce a new dataset on the Isle of Wight Supply Chain (IWSC), which we use to demonstrate the new technique. We achieve a cold start hit rate @10 of 77% on a collection of 630 items with only 376 supply-chain supplier labels, and 67% with only 142 supply-chain consumer labels, demonstrating a high level of performance even with extremely few labels in challenging cold-start scenarios. The TSR technique is generalisable to any dataset where items with similar description text share similar relationships and has applications in speculatively expanding the number of relationships in partially labelled datasets and highlighting potential items of interest for human review. The technique is also app
ropriate for use as a recommendation algorithm, either standalone or supporting traditional recommender systems in difficult cold-start situations.
(More)