Feedback Loop

A recommender only learns from what it shows, so it keeps showing what it already knows. Watch 400 simulated readers' feeds narrow over 30 rounds, then switch on exploration and see the catalogue open back up. Rate a dozen things first and follow your own feed through the loop.

1Tell it what you like

Thumbs up or down on twelve things from eight topics. They become your first clicks, and the system learns from them like it does from everyone else.

Loading the catalogue…

2Run the loop

Each round, every reader gets a feed of 10 items, clicks on some, and the recommender retrains on all the clicks so far. Items it never shows get no data at all.

One-click presets

Policy

Rounds Pace

Add exploration

Switch at round to Release a good new item at round

Runs start from the same seed and the same random clicks, so a switched run follows the plain one exactly until the switch, then bends.

While a slow run is going:

Simulation settings

Readers Items Feed size k Seed ε (ε-greedy) c (UCB) c (Thompson) λ (MMR) Factors Ridge λ ALS sweeps Repeat decay Avoid repeats

Repeat decay is how much less likely a reader is to click something the n-th time it is shown (0.7ⁿ). “Avoid repeats” lets the recommender discount items it has already shown that reader by the same factor. Changing readers, items, k or seed builds a new world; the ratings above are kept.

Same code as the Python package, run in your browser, with the same random numbers: a run here reproduces digit for digit with the command below.

3Your feed, run by run

Each column is one round of your feed, coloured by topic; a dot is a click. Pick two runs to compare side by side and click a column to read that round's feed.

Why click data is censored, collaborative filtering in a few lines, and what UCB and Thompson sampling have to do with it: How it works →