Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FELRec: Efficient Handling of Item Cold-Start With Dynamic Representation Learning in Recommender Systems

Kuba Weimann, Tim O. F. Conrad

TL;DR

FELRec tackles item cold-start in dynamic recommender systems by replacing learnable item embeddings with a parameter-free dynamic cache, enabling gradient-free updates of user/item representations through a single forward pass. It leverages MoCo- or BYOL-inspired representation learning within a Transformer-based encoder to update cache entries, supporting zero-shot generalization to unseen data. On MovieLens and Twitch, FELRec achieves substantial gains in item cold-start performance compared with strong baselines, while maintaining competitive results for existing items and exhibiting strong generalization across datasets. The approach offers a scalable, low-parameter solution for real-time recommendations in rapidly evolving catalogs, with clear pathways for extension when side information or debiasing strategies are available.

Abstract

Recommender systems suffer from the cold-start problem whenever a new user joins the platform or a new item is added to the catalog. To address item cold-start, we propose to replace the embedding layer in sequential recommenders with a dynamic storage that has no learnable weights and can keep an arbitrary number of representations. In this paper, we present FELRec, a large embedding network that refines the existing representations of users and items in a recursive manner, as new information becomes available. In contrast to similar approaches, our model represents new users and items without side information and time-consuming finetuning, instead it runs a single forward pass over a sequence of existing representations. During item cold-start, our method outperforms similar method by 29.50%-47.45%. Further, our proposed model generalizes well to previously unseen datasets in zero-shot settings. The source code is publicly available at https://github.com/kweimann/FELRec .

FELRec: Efficient Handling of Item Cold-Start With Dynamic Representation Learning in Recommender Systems

TL;DR

FELRec tackles item cold-start in dynamic recommender systems by replacing learnable item embeddings with a parameter-free dynamic cache, enabling gradient-free updates of user/item representations through a single forward pass. It leverages MoCo- or BYOL-inspired representation learning within a Transformer-based encoder to update cache entries, supporting zero-shot generalization to unseen data. On MovieLens and Twitch, FELRec achieves substantial gains in item cold-start performance compared with strong baselines, while maintaining competitive results for existing items and exhibiting strong generalization across datasets. The approach offers a scalable, low-parameter solution for real-time recommendations in rapidly evolving catalogs, with clear pathways for extension when side information or debiasing strategies are available.

Abstract

Recommender systems suffer from the cold-start problem whenever a new user joins the platform or a new item is added to the catalog. To address item cold-start, we propose to replace the embedding layer in sequential recommenders with a dynamic storage that has no learnable weights and can keep an arbitrary number of representations. In this paper, we present FELRec, a large embedding network that refines the existing representations of users and items in a recursive manner, as new information becomes available. In contrast to similar approaches, our model represents new users and items without side information and time-consuming finetuning, instead it runs a single forward pass over a sequence of existing representations. During item cold-start, our method outperforms similar method by 29.50%-47.45%. Further, our proposed model generalizes well to previously unseen datasets in zero-shot settings. The source code is publicly available at https://github.com/kweimann/FELRec .
Paper Structure (26 sections, 4 equations, 3 figures, 3 tables)

This paper contains 26 sections, 4 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. FELRec
  4. Problem statement
  5. Cache
  6. Model architecture
  7. Encoder
  8. Projection & prediction networks
  9. Model training
  10. Model evaluation
  11. Experimental setup
  12. Datasets
  13. Baseline methods
  14. Evaluation
  15. Results
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Architectures of FELRec-Q (a) and FELRec-P (b). In a user-item interaction, the user is represented as a sequence of items, and the item as a sequence of users. First, FELRec fetches the vector representations of items and users from the cache. Next, our model encodes the sequences into new vector representations of the user and the item, which are then stored in the cache. FELRec is trained to make the representations of user and item similar using contrastive learning (a) adopted from MoCohe2020momentum, or a similarity loss (b) adopted from BYOLgrill2020bootstrap. $\mathbin{ /\mkern-6mu/}$ means a stop-gradient operation.
  • Figure 2: Recommendation performance of FELRec and SASRec on MovieLens 25M (a) and Twitch (b) depending on the number of users that have interacted with the ground-truth item. Lower average rank is better. "I" is the new items partition, "O+U" are the observed and new users partitions. Performance improves over time.
  • Figure 3: T-SNE vandermaaten2008tsne embedding of users and items represented by FELRec (a) and SASRec (b). Red dots are user (left) and item (right) in a single interaction. The embedding is coloured based on how well the users or items match the red dot on the opposing plot; bright colors mean a better match and dark colors a worse match. The ground-truth item (right dot) lies in the top $10\%$ recommended items.