← Amelia Zhang
03·VLM·In progress · solo

Multimodal cold-start recommendation

Fine-tuning a vision-language model for product recommendation when there is no interaction history to lean on — the cold-start case that classical collaborative filtering fails.

In progress — no results claimed
Problem

Collaborative filtering breaks the moment a user or product has no history — the cold-start problem. Content-based approaches recover some ground using item features, but they under-use the signal that is actually available at cold-start: the item's images and text description, together.

The question is whether a fine-tuned VLM can turn that multimodal item representation into a useful recommendation signal without any interaction data at all.

Approach

Fine-tuning a vision-language model to produce joint image-text embeddings tuned for recommendation, rather than for the model's original pre-training objective. Evaluation against content-based baselines under strict cold-start splits — items unseen at training time, no leakage from historical interactions.

Method details, baselines, and dataset choices are still being finalized. Workshop paper in preparation.

Status

Active. No headline metrics to report yet — writeup and evaluation are ongoing. Check back, or reach out if the specific problem is relevant to what you're working on.

Stack
VLM (fine-tune)·PyTorch·HuggingFace