New Strategies for Learning Knowledge Graph Embeddings: The Recommendation Case

Abstract

Knowledge graph embedding models encode elements of a graph into a low-dimensional space that supports several downstream tasks. This work is concerned with the recommendation task, which we approach as a link prediction task on a single target relation performed in the embedding space. Training an embedding model requires negative sampling, which consists in corrupting the head or the tail of positive triples to generate negative ones. Although knowledge graph embedding models and negative sampling have extensively been investigated for link prediction, their combined use for performing recommendations over knowledge graphs remains largely unexplored in the literature. In this work, we propose two specialization strategies for training embedding models and performing knowledge graph-based recommendations. Both strategies first train an embedding model on the whole knowledge graph. Then, during a specialization phase, a dedicated negative sampling scheme is applied to refine the pre-trained model. Experimental results on two public datasets demonstrate that a simple strategy which refines a pre-trained model by sampling random negative tails for the target relation proves to be very effective. This strategy significantly improves performance with respect to traditional rank-based evaluation metrics as well as a newly introduced metric that reflects the semantic validity of the top-ranked candidate entities.

Sem@$K$: Is my knowledge graph embedding model semantic-aware?

“…using our proposed metric called Sem@K [ Hubert et al , 2022a , b ]. In [ Hubert et al , 2022a ], Sem@K was specifically defined for the recommendation task which was seen as predicting tails for a unique target relation. Sem@K was then extended in [ Hubert et al , 2022b ] to the more generic LP task, where not only tails but also heads are corrupted and all relations are considered.…”

“…More specifically, in this work, our goal is to assess the ability of popular KGEMs to capture the semantic profile (i.e., domain and range) of relations in a LP task. To do so, we build on Sem@K, the semantic-oriented metric that we introduced in [ Hubert et al , 2022a ] for a recommendation task. This metric was originally used to evaluate the ability of KGEMs to recommend items that are of the expected type.…”

Enhancing Knowledge Graph Embedding Models with Semantic-driven Loss Functions

“…However, semantic-driven approaches would benefit from a semantic-oriented evaluation. To the best of our knowledge, the work around Sem@K [Hubert et al, 2022a[Hubert et al, ,b, 2023 is the only one to provide appropriate tools for measuring KGEM semantic awareness. Hence, our experiments will also be evaluated with this metric.…”

“…Besides, precedent work pointed out the performance gain of incorporating ontological information as measured by rank-based metrics such as MRR and Hits@K [Cao et al, 2022, d'Amato et al, 2021, Guo et al, 2015, Minervini et al, 2017. However, while such approaches include semantic information as KGEM inputs, the semantic capabilities of the resulting KGEM are left unassessed, even though this would provide a fuller picture of its performance [Hubert et al, 2022a[Hubert et al, , 2023. Hence, our second research question:…”

“…This makes our approach widely usable in practice. Furthermore, the impact of loss functions is evaluated using both rank-based metrics and Sem@K [Hubert et al, 2022a[Hubert et al, , 2023 -a metric that measures the semantic awareness of KGEMs for the LP task.…”