Reverie: A Recommender That Reads Watch History

Source codeem-ech/dl-reverie-recommender

Reverie reads a viewer's watch history and predicts the next title they will enjoy. A sequential GRU trained on MovieLens pairs with a neural collaborative-filtering model trained on about 9.9 million Letterboxd ratings, and both ship inside a web app with accounts, a swipe-based history builder, and a Blend that intersects two people's recommendations.

TensorFlowGRUFastAPIReactRecommender Systems

0.257

hit rate at 10, validation

0.242

hit rate at 10, test, three seeds

1.372

rating error (RMSE)

Problem

Understanding the challenge and user needs

Popularity lists and genre filters ignore order. What a person watched last week says more about tonight than what they watched last year, and a five-person team set out to prove that a model reading the sequence beats models that read only the totals.

Reverie landing screen asking What should I watch next — The landing screen takes an imported watch history and returns a ranked list.

Approach

Strategy and methods used to address the problem

A GRU learned from 1 million ratings by 6,040 MovieLens users, measured against three baselines on the same split: most-popular, recent-genre popularity, and item-kNN. Twenty-five logged runs tested embeddings, genre features, and regularization, and three seeds fixed the final test number. A FastAPI service and a React front end turned the model into a product, covered by 22 automated tests.

Grid of film posters used to build a watch history — A poster grid builds a history for viewers with nothing to import.

Architecture slide showing sources, data layer, models, and serving — The architecture runs from open datasets through a data layer and two models to a FastAPI service and a React interface.

Training and validation loss curves diverging after epoch 16 — Validation loss bottoms out while training loss keeps falling; early stopping restores the best weights.

Bar chart of hit rate for four feature combinations — The feature ablation: genre lifts the hit rate, while the rating signal on its own lowers it.

Final Solution

The final design and outcome

The GRU reached a hit rate at 10 of 0.257 on validation against 0.076 for item-kNN and 0.042 for most-popular, and 0.242 plus or minus 0.002 on the held-out test set across three seeds. The collaborative model cut rating error to an RMSE of 1.372 against 1.619 for the movie-mean baseline and beat that baseline for 94 percent of users.

The profile screen summarizes a viewer's taste as a radar chart across genres and ranks top picks beneath it.

Rows of recommended films grouped by reason — Recommendation rows group picks by the reason behind them, such as a film already watched or a preferred genre.

Bar chart comparing Reverie's hit rate with item-kNN and most-popular baselines — The sequential model's hit rate at 10 stands well clear of both baselines; the frozen test figure appears at right.

Bar chart of RMSE for the neural network against mean baselines — The collaborative model posts a lower rating error than the movie, user, and global means.

Next-title ranking on validation

Model	Hit rate at 10	MRR
Reverie (GRU)	0.257	0.120
Item-kNN	0.076	0.035
Recent-genre popularity	0.045	0.023
Most popular	0.042	0.022

Held-out test result across three seeds: 0.242 plus or minus 0.002.

Rating prediction on the test set

Model	RMSE
Neural collaborative filtering	1.372
Movie mean	1.619
User mean	1.963
Global mean	2.072

Lower is better. Ratings run from 1 to 10.

Impact & Validation

Results, learnings, and outcomes

Reverie shows the full path from a research question to a product: a model with honest baselines, an experiment log, a tested API, and an interface a person can use to find something to watch with a friend.

Watch together screen showing films two viewers would both enjoy — The Blend intersects two viewers' recommendations and lists what they already have in common.

Presentation

The deck, slide by slide

Title slide: REVERIE. What should I watch next? Deep Learning, IE University, Functional MVP. Amaly Attia, Em Echeverria, Lea Sarouphim Hochar, Stephan Pentchev, Cécile Tambey. A film strip icon sits above the title.

Slide 1 of 25