Our paper "Adapting Sequential Recommender Models to Content Recommendation in Chat Data using Non-Item Page-Models" has been accepted at the KARS workshop at RecSys 24

Abstract

Most research in sequential recommender models has focused on sequences that are purely made of items (e.g, movies, page clicks), excluding additional elements in the sequence that may provide more information for the next relevant item. Recently, it has been proposed to include non-item pages (e.g., list pages or blog posts), in order to represent the users' intent more clearly. In this paper, we transfer the same modelling principle to sequences made of items and text messages. This enables us to adapt arbitrary sequential recommender methods to a new application area: We can use any sequential recommendation model to recommend links to relevant content in a chat setting, using the history of previous messages and mentioned items as context. We evaluate our models on four different datasets and show that we can identify content relevant to the conversations well when using pre-trained embeddings for the messages in the conversations.