When a user or item has little interaction history, a recommender has to work with weaker evidence than collaborative filtering normally relies on. Start by identifying which side is cold, then use available content, relationships, or domain knowledge to form initial recommendations—and test those recommendations against a collaborative-filtering baseline once enough interactions exist. Large language models (LLMs) can help supply or interpret signals, but the available survey evidence does not show that they universally outperform conventional recommenders.
What is cold start, and which side is new?
Cold start is the lack of behavioral evidence needed to model a user’s preferences or an item’s likely audience. The challenge may involve a new user, a new item, or both. Zhang et al.’s January 2025 survey frames the problem around accurately modeling new or interaction-limited users and items.
As an Amazon Associate I earn from qualifying purchases.
New user
A new user has little or no recorded activity, so the system cannot yet infer preferences reliably from that person’s clicks, ratings, or purchases. There may still be useful user-provided information, such as stated interests, and the catalog may have descriptive text or metadata. Those signals can support a first set of recommendations, but they are not the same as observed behavior.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →New item
A new item has little or no interaction history, even if its description, category, or other attributes are known. The system may be able to compare those attributes with items that users have engaged with, but it has not yet observed how people respond to the new item itself.
#1 Best Overall
Both are new or interaction-limited
When both the user and the item lack behavioral evidence, there is less basis for estimating their fit from interactions. Content, graph relationships, domain information, or an LLM’s world knowledge may offer additional signals, as Zhang et al. describe, but each signal must be relevant and dependable for the domain.
Choose a starting strategy from the signals you have
There is no single cold-start method to apply in every case. First inventory the evidence available for the user and for candidate items. Then choose how to use it, keeping behavioral evidence distinct from descriptive or inferred signals.
Rank #2
- Use scikit-learn to track an example ML project end to end
- Explore several models, including support vector machines, decision trees, random forests, and ensemble methods
- Exploit unsupervised learning techniques such as dimensionality reduction, clustering, and anomaly detection
- Dive into neural net architectures, including convolutional nets, recurrent nets, generative adversarial networks, autoencoders, diffusion models, and transformers
- Use TensorFlow and Keras to build and train neural nets for computer vision, natural language processing, generative models, and deep reinforcement learning
| Available evidence | Practical starting point | What to check |
|---|---|---|
| Item descriptions or metadata, but little user history | As a design implication, use content-led discovery or ask the user for a small amount of preference information before personalizing further. | Whether descriptions are complete, current, and specific enough to distinguish relevant items. |
| Descriptive text for a new item, but few interactions | As a design implication, represent the item from its content and use that representation to find candidates or compare it with items that have interaction history. | Whether the representation captures meaningful distinctions and whether recommendations change as real interaction evidence arrives. |
| Graph relationships or domain information | Use these as additional signals when their links or rules are relevant to the user, item, and recommendation task. | Whether the relationships are reliable and sufficiently current; a connection alone does not establish preference. |
| LLM-generated or LLM-interpreted information | Use an LLM to generate recommendations directly or as one component in a larger retrieval and ranking pipeline. | Whether the output is grounded in eligible items and whether its quality holds up against a suitable baseline. |
| Ample interaction history | Compare the generative approach with supervised collaborative filtering trained on that history. | Whether the new method improves the outcomes that matter in the intended setting, rather than merely producing fluent explanations. |
These are design implications, not results from a controlled head-to-head comparison. The reviewed surveys do not establish a universally best method for each signal combination.
Decide what role the LLM should play
“Generative recommendation” can describe different system designs. Li, Zhang, Liu, and Chen’s LREC-COLING 2024 survey describes one paradigm that generates recommendations directly from the complete pool of items, collapsing stages such as score computation and reranking into one LLM step. That is a description of an architecture, not evidence that a single-stage system is operationally preferable in every setting.
Rank #3
Direct recommendation generation
An LLM can generate recommendations from an item pool rather than leaving each recommendation stage to a separate model. This makes the model responsible for producing the recommendation output, so the design should make clear how it is constrained to the eligible catalog and how output quality is evaluated.
Feature extraction or representation
An LLM can also help interpret user or item information and turn it into features or representations used by other parts of a recommender. This keeps the LLM in a supporting role rather than treating it as the entire recommendation system.
Rank #4
Retrieval-augmented generation or reranking
A retrieval-augmented design retrieves relevant information or candidate items and gives that material to the LLM to use when generating or ordering recommendations. Deldjoo et al.’s KDD 2024 review describes retrieval augmentation as a way to externalize knowledge rather than place all of it in model parameters. The review reports that this can facilitate online updates and reduce hallucinations, and may require fewer LLM parameters because knowledge is externalized. These are reported advantages, not guarantees for every implementation.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →The same review reports that untuned LLMs generally underperform supervised collaborative-filtering methods trained with sufficient data, while they can be competitive in near-cold-start settings. It also reports that few-shot prompting typically improves on zero-shot prompting. These findings support testing LLMs where interaction evidence is scarce; they do not establish that prompting alone solves cold start or that an LLM should replace a well-trained collaborative-filtering model.
Best Value
Build an evaluation that accounts for changing evidence
Cold-start performance should be assessed in the setting the system is meant to serve. In particular, do not treat a single ranking score as proof that recommendations are useful or harmless. Deldjoo et al.’s Gen-RecSys survey identifies evaluation of recommendation impact and potential harm as necessary and an open research challenge; the reviewed evidence does not establish a universal metric threshold.
- Separate new-user, new-item, and near-cold-start cases so results for one do not conceal weaknesses in another.
- Compare an LLM-based design with supervised collaborative filtering when sufficient interaction data is available, and state clearly when the comparison instead concerns a sparse-history setting.
- Check whether the system uses the intended item pool and whether its recommendations are supported by the available content, relationships, or domain information.
- Reassess recommendations as new interactions and item information arrive. For retrieval-augmented systems, verify that external knowledge can be updated as intended rather than assuming that online updates happen automatically.
- Evaluate the impact and potential harms of recommendations alongside relevance or ranking quality. The reviewed survey identifies this as important but does not supply a settled, universal evaluation recipe.
The evidence here is based on surveys published through January 2025: Li et al.’s LREC-COLING survey from May 2024, Deldjoo et al.’s KDD review from August 2024, and Zhang et al.’s arXiv survey dated January 3, 2025. It supports qualitative comparisons, not a numeric performance guarantee or a current state-of-the-art ranking.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
