Towards Analogy-based Recommendation: Benchmarking of Perceived Analogy Semantics

Requests for recommendation can be seen as a form of query for candidate items, ranked by relevance. Users are however o‰en unable to crisply de€ne what they are looking for. One of the core concepts of natural communication for describing and explaining complex information needs in an intuitive fashion are analogies: e.g., “What is to Christopher Nolan as is 2001: A Space Odyssey to Stanley Kubrick?”. Analogies allow users to explore the item space by formulating queries in terms of items rather than explicitly specifying the properties that they €nd aŠractive. One of the core challenges which hamper research on analogy-enabled queries is that analogy semantics rely on consensus on human perception, which is not well represented in current benchmark data sets. Œerefore, in this paper we introduce a new benchmark dataset focusing on the human aspects for analogy semantics. Furthermore, we evaluate a popular technique for analogy semantics (word2vec neuronal embeddings) using our dataset. Œe results show that current word embedding approaches are still not not suitable to suciently deal with deeper analogy semantics. We discuss future directions including hybrid algorithms also incorporating structural or crowd-based approaches, and the potential for analogy-based explanations.