Influence Beyond Similarity: A Contrastive Learning Approach to Object Influence Retrieval

Innovative art or fashion trends do not spring out of nowhere: they are products of societal contexts, movements and economic turning points. To understand the dynamics of innovation, it is necessary to understand influence relations between agents (e.g. artists, designers, creatives) and between the objects (e.g. clothes, paintings) that these agents produce. However, acquiring knowledge about these connections is challenging given that they are frequently undocumented. Recent literature has focused on discovering influence relations between agents, utilizing either object similarity or social network information. However, these methods often overlook the importance of direct relations between objects or oversimplify the complex nature of influence by approximating it with similarity.

To overcome this gap, we introduce Object Influence Retrieval (OIR), a task aimed at retrieving objects that potentially influenced a given object. To measure task performance, we describe two datasets for OIR: WikiartINFL (paintings) and iDesignerINFL (fashion items), both enriched with agent influence information. Additionally, we present CLOIR, a Contrastive Learning approach leveraging transfer learning from a pre-trained model to represent objects, incorporating agent influence information through contrastive learning. CLOIR shows up to a 30% improvement in Precision@k and Mean Reciprocal Rank in the OIR task compared to a baseline based on similarity between objects.