KG-based Recommendations

Recommender systems have become particularly popular with the rise of the Web and e-commerce. These systems aim to reduce the information overload for users by filtering out content based on predictions of user preferences. Over the years, several techniques have evolved for recommender systems. Nonetheless, collaborative and content-based filtering are the most dominant techniques.

In this blog post, KG-based recommendation approaches are investigated on the Pokémon KG (version a).

Software Setup

Hyperparameters

Some of the models have hyperparameters that influence their performance and are required for better reproducibility.

Linked Data Semantic Distance (LDSD)

This model has no hyperparameters.

RDF2Vec

TransE

Evaluation

A brief evaluation of the models were performed for the first generation of Pokémons, which were in total 151 monsters. No dataset was available for an utility/relevance ranking of these Pokémons, which is why we used the scrapped infobox recommendations from Google instead (see post and dataset). However, an expert evaluation of relevance would lead to a more sophisticated evaluation.

P@N (i.e. precision at N) was computed for all the models.

Recommendation - Sample

In the recommendations, we focused on the 151 Pokémons in the first generation.

Jigglypuff

Google (CF)	RDF2Vec	TransE	LDSD
1. Wigglytuff	1. Lickitung	1. Wigglytuff
2. Snorlax	2. Kadabra	2. Clefairy
3. Squirtle	3. Lapras	3. Clefable
4. Pikachu	4. Rapidash	4. Snorlax
5. Charmander	5. Poliwag	5. Lickitung
6. Evee	6. Muk	6. Chansey
7. Bulbasaur	7. Jynx	7. Jynx
8. Psyduck	8. Hypno	8. Hypno
9. Charizard	9. Geodude	9. Poliwhirl