What data does this notebook require as input?

A CSV of reviews that contains a precomputed embedding column per review, plus UserId, ProductId, and Score fields.

How are user and product embeddings created in this notebook?

User embeddings and product embeddings are computed by averaging all the embeddings of reviews written by the same user or about the same product within the training set.

What does the notebook's evaluation show about the relationship between embedding similarity and review scores?

There is a weak but visible trend where reviews with higher user-product embedding similarity tend to have higher review scores, even though the similarity is computed independently of the review text itself.

Is this notebook suitable as a production recommendation system?

No, it is not a production system—the correlation demonstrated is weak, and the simple averaging approach ignores users or products with only a single review, which is common in the sample data.

How does this embedding-based approach differ from traditional collaborative filtering?

This approach derives its signal from semantic review content embeddings rather than user-item interaction patterns, making it potentially useful as a supplementary feature alongside collaborative filtering.

Prompt Chain

Generate User and Product Embeddings for Recommendations

Name: User And Product Embeddings
Availability: OnlineOnly
Author: OpenAI Cookbook

Build user and product embeddings by averaging review embeddings, and use their similarity to weakly predict ratings.

Copy chain

Works with github

OpenAI Cookbook

Own this? Claim it

Spark score

out of 100

Updated last month

Version 1.0.0

Add to Favorites

Why it matters

Leverage user and product embeddings derived from review data to predict review scores and enhance recommendation systems. This asset provides a novel signal that can improve existing recommendation models.

Outcomes

What it gets done

Calculate user and product embeddings by averaging review data.

Evaluate embedding similarity against review scores in a test set.

Visualize the correlation between embedding similarity and review scores.

Utilize embeddings as an additional feature for recommendation improvements.

Install

Add it to your toolbox

Run in your project directory:

curl -fsSL https://spark.entire.vc/get/oai-userandproductembeddings | bash

Steps

Steps in the chain

Calculate user and product embeddings

Evaluate the embeddings

Visualize cosine similarity by review score

Overview

User And Product Embeddings

An OpenAI Cookbook notebook that derives user and product embeddings by averaging review embeddings, then shows their cosine similarity weakly correlates with review scores. Use it as an additional signal alongside collaborative filtering, not a replacement. Most users and products in the sample dataset appear only once, limiting representation quality.

What it does

This notebook builds on the companion Get_embeddings_from_dataset notebook to calculate a user embedding and a product embedding for every user and product in the training set, simply by averaging the embeddings of all reviews that user wrote or that product received. It then evaluates whether these embeddings carry predictive signal by measuring, on an unseen test set, the cosine similarity between each review's user and product embedding.

When to use - and when NOT to

Use this when you already have review-level embeddings and want to derive per-user and per-product representations from them as an additional recommendation signal - it's a lightweight complement to collaborative filtering, not a replacement for it, since it captures a different kind of signal. Most users and products in the underlying 50k-review sample appear only once, which caps how much a single-user or single-product average can really represent.

Inputs and outputs

Cosine similarity between a review's user and product embedding is calculated for every test-set review, then normalized to a 0-1 percentile scale so scores are evenly distributed rather than clustered. Grouping these normalized similarity scores by the review's actual star rating and plotting the distribution reveals a weak but real trend: higher user-product embedding similarity correlates with higher review scores.

Integrations

Because this predictive signal comes purely from averaged review embeddings rather than explicit user-item interaction patterns, it works differently from standard collaborative filtering and can be layered on top of it as an extra feature to nudge performance on existing recommendation or rating-prediction problems, rather than needing to replace whatever collaborative-filtering approach is already in place.

Who it's for

Developers working on recommendation or review-scoring systems who want a simple, embeddings-based signal - derived from existing review embeddings, not new interaction data - to complement collaborative filtering.

Source README

User and product embeddings

We calculate user and product embeddings based on the training set, and evaluate the results on the unseen test set. We will evaluate the results by plotting the user and product similarity versus the review score. The dataset is created in the Get_embeddings_from_dataset Notebook.

1. Calculate user and product embeddings

We calculate these embeddings simply by averaging all the reviews about the same product or written by the same user within the training set.

We can see that most of the users and products appear within the 50k examples only once.

2. Evaluate the embeddings

To evaluate the recommendations, we look at the similarity of the user and product embeddings amongst the reviews in the unseen test set. We calculate the cosine distance between the user and product embeddings, which gives us a similarity score between 0 and 1. We then normalize the scores to be evenly split between 0 and 1, by calculating the percentile of the similarity score amongst all predicted scores.

2.1 Visualize cosine similarity by review score

We group the cosine similarity scores by the review score, and plot the distribution of cosine similarity scores for each review score.

We can observe a weak trend, showing that the higher the similarity score between the user and the product embedding, the higher the review score. Therefore, the user and product embeddings can weakly predict the review score - even before the user receives the product!

Because this signal works in a different way than the more commonly used collaborative filtering, it can act as an additional feature to slightly improve the performance on existing problems.

FAQ

Common questions

Discussion