XMUDM/FETTLE

Who To Align With: Feedback-Oriented Multi-Modal Alignment in Recommendation Systems

This repository is the official implementation of "Who To Align With: Feedback-Oriented Multi-Modal Alignment in Recommendation Systems". Our paper has been accepted by SIGIR2024! 🎉

Framework

Quick Start

You only need to import the losses CLALoss and ILALoss into your code. e.g.

# <user, pos_item> are interacted user-item pairs.
ila_dt_loss = self.ila_dt_loss(u_emb, i_emb,
                            v_emb, t_emb, user, pos_item)

cla_loss = self.cla_loss(user_embeddings, item_embeddings,
                            v_emb, t_emb, user, pos_item)

loss +=  self.iladt_weight * ila_dt_loss + self.cla_weight * cla_loss

Performance of each component

Compared to the paper, results with the added +DT are included.