The Best Feed
Once you start rating articles, a Best link appears in the sidebar. This algorithmic feed shows all your unread entries sorted by predicted score, surfacing the content you’re most likely to enjoy. It updates automatically as new articles arrive and get scored.
Rate Your Reading
Lion Reader tracks your preferences through both explicit voting and implicit behavior. Voting uses a 5-point scale from −2 to +2 with upvote and downvote controls. Tap the up arrow to vote +1, tap again to boost to +2, and tap a third time to clear. Downvoting works the same way in the other direction.
Your implicit behavior also contributes to scoring. Starring an entry signals strong interest (+2), saving an article for later indicates moderate interest (+1), and marking an entry as read from the list without opening it indicates low interest (−1). Explicit votes always take priority over implicit signals.
Server-Side Machine Learning
Once you’ve rated at least 20 entries, Lion Reader trains a machine learning model to predict scores for new content. The model uses TF-IDF text vectorization with bigrams, combined with Ridge Regression. Titles are weighted 2x during feature extraction, and per-feed features capture your source-level preferences — so the model learns not just what topics you like, but which publications you trust.
The model is cross-validated using Mean Absolute Error (MAE) and Pearson correlation metrics to ensure prediction quality. Predictions are tempered by a confidence score based on how well the model recognizes an entry’s vocabulary and feed — uncertain predictions are pulled toward zero so they don’t dominate your Best feed. New entries are automatically scored right after feed fetches, and the model retrains weekly as you continue rating.
Scoring features:
- Best feed — All unread entries sorted by predicted score
- Explicit voting — −2 to +2 scale with cycling up/down controls
- Implicit signals — Starring (+2), saving (+1), quick-mark-read (−1)
- Server-side ML — TF-IDF with bigrams + Ridge Regression
- Per-feed features — Learns source-level preferences
- Confidence-based shrinkage — Uncertain predictions move toward zero
- Automatic scoring — New entries scored after feed fetches, model retrains weekly