Bulle is opening its dataset and inviting researchers, students, engineers and non-profits to propose an evaluation metric and a recommendation algorithm that are at once performant, fair and non-divisive.
The recommendation algorithms of major platforms remain black boxes whose effects, addiction, polarisation, invisibilisation of small creators, are now well documented. Bulle made the opposite choice: a fully public algorithm, computed from a simple and auditable formula.
This challenge extends that logic. Rather than improving our algorithm internally, we open the data and invite the community to propose better. The selected proposal will be the subject of a deeper collaboration with Bulle to actually implement the algorithm into the platform.
These principles define what a good algorithm looks like for Bulle. Your evaluation metric and your algorithm must objectively meet these criteria.
Outside subscriptions and interests, the algorithm recommends the same content to everyone. No filter bubble, no behavioural personalisation.
Shares, reposts, quizzes, watchtime are the best quality signals. The algorithm uses them, but as proportions of views, never as raw numbers.
No bonus for shocking or provoking. Content generating many rejects or dislikes is penalised, even if it generates views.
A new or small creator must be able to surface. The algorithm rotates creators (round-robin) instead of saturating the feed with the most visible ones.
Today's publications get priority. The algorithm surfaces fresh content while it is still relevant, without freezing the feed on past hits.
Mass posting or clickbait should not pay off. Signals are normalised by views, so quantity never compensates for poor quality.
A publication loses relevance over time, without disappearing abruptly. No implicit shadow ban.
The algorithm must be able to re-evaluate a publication based on its recent views, not just on cumulative history. A publication with a bad start can climb back.
The chosen algorithm must be expressible in a few formulas, and auditable by anyone. No opaque model trained on user data.
A single CSV file, aggregated by day, containing all recommended publications (including those with zero views, to allow round-robin strategy evaluation).
| Column | Type | Description |
|---|---|---|
| publication_id | uuid | Anonymised (stable uuid) publication identifier. |
| creator_id | uuid | Anonymised (stable uuid) creator identifier. Useful for round-robin and fairness. |
| creator_status | text | Creator status: semipro or pro. |
| published_at | datetime | Publication date and time, rounded to the hour (UTC). |
| date | date | Evaluation day (recommendation day). |
| category_name | text | Top-level category (e.g. Nature and environment). |
| subcategory_name | text | Sub-category (e.g. Ecological awareness). |
| type | text | Format: video, carrousel (image or mixed image/video), message (text). |
| total_views | integer | Cumulative views at the evaluated date (may be 0). |
| prop_bookmarks | real β [0,1] | Bookmarks Γ· views. |
| prop_shares | real β [0,1] | Shares Γ· views. |
| prop_reposts | real β [0,1] | Reposts Γ· views. |
| prop_quiz | real β [0,1] | Quizzes completed Γ· views. |
| prop_love | real β [0,1] | "Love" reactions Γ· views. |
| prop_like | real β [0,1] | "Like" reactions Γ· views. |
| prop_dislike | real β [0,1] | "Dislike" reactions Γ· views. |
| prop_reject | real β [0,1] | Explicit rejects Γ· views. |
Chicken-and-egg effect. The proportions reflect views obtained under the current algorithm. A publication that was barely surfaced may have received fewer views. Always work in proportions, never in absolute values, and keep in mind that publications with very few views may be noisy.
Downloading requires a Bulle account. This restriction allows us to keep traceability of dataset accesses and to limit abusive usage. Only people who legitimately obtained the dataset through this page will be able to submit their work for evaluation.
A complete submission consists of two files: a description document and the ranking CSV produced by your algorithm.
A document of one A4 page maximum, in font size 12, explaining: the evaluation metric you propose, the corresponding algorithm, and the associated formulas. Formulas must be simple, explicit, and rely directly on the variables of the table (proportions, dates, status, etc.).
The output of your algorithm applied to the dataset, as a submission.csv file with three columns: date, publication_id, rank. One row per pair present in the dataset, unique ranks without gaps, starting at 1.
# submission.csv, three columns, header required date,publication_id,rank 2026-04-24,3f4b2490-8aad-5949-9fdc-e40e4566e5d3,1 2026-04-24,...,2 2026-04-25,...,1 # One row per (date, publication_id) pair in the dataset.
How to submit. Compress both files (PDF + CSV) into a .zip file and email it to contact@shabon.fr with subject "Bulle Algorithm Challenge", before June 1st, 2026 at 11:59 pm (Paris time). One submission per participant or team. In case of multiple submissions, only the latest one is retained.
A dedicated page will be published to showcase the rankings actually submitted, with display of the proposed publications for each given day. Participants will be able to visualise and compare their respective algorithms.
This page will be announced publicly after the submission deadline.
Bulle is a mission-driven project, run by SHABON. The reward of this challenge is first and foremost a public recognition of the work done.
The selected proposal will be featured in a public article on Bulle's blog, presenting the method, the author and the approach.
Individuals, students, teams, labs, non-profits, companies. One restriction only: the code and method must be publicly explainable.
The selected proposal will be the subject of a deeper collaboration with the Bulle team to actually implement the algorithm into the platform.
If the solution is integrated into Bulle, its author is publicly credited on the algorithm page.
For reference only, here are a few leads we have considered. They are neither imposed nor definitive: your work consists precisely in proposing better, or combining them differently.
We combine the eight engagement signals with weights reflecting what Bulle values. Bulle's current weighting favours reflection (quizzes) and voluntary sharing above plain "like", and penalises negative signals.
The exact weights remain at your discretion. The goal is for them to objectively reflect the principles stated above.
Rather than a short half-life that quickly extinguishes content, a simple exponential D(t) = exp(βt / Ο) with a time constant Ο to calibrate allows a more gradual decay. As an indicative range, Ο β 14 to 30 days gives quality content time to be discovered without freezing the feed.
The constant Ο should be calibrated to the target editorial pace. Too small and old information is extinguished, too large and the feed is saturated with old content.
To prevent a single creator from saturating the top of the ranking, a decreasing factor can be applied at each repetition: a publication's value is divided by 1 plus the number of publications by the same creator already ranked above.
Once each publication's value is estimated, the day's overall ranking can be evaluated through an nDCG (normalised discounted cumulative gain). The idea: heavily reward good decisions at the top of the ranking, much less at the bottom. The final score is averaged across all evaluated days.
Other aggregations are conceivable (weighted sum, Spearman, MAPβ¦). It is up to you to justify your choice.
Important. These formulas are only a starting point. The challenge is precisely to propose a public evaluation metric that objectively meets the principles stated above, along with a concrete algorithm that complies with it. You may reuse these leads, adapt them, or propose entirely new ones.