shahvanicaفارسی
Predicting attention

What predicts attention

A gradient-boosting model reads what is known when a story is published (tags, title length, body length, hour, weekday, year and month) and predicts its view percentile within its publication year. It is tested on 7,556 stories it never saw.

51%of the variance explained (test R²)
0.71rank correlation, predicted vs real
59%of predicted top stories really are (chance 20%)

Which inputs carry the signal

The strongest group, ⁨Tags⁩, reaches 43.8% alone; the whole model reaches 51.0%.

What each kind of clue can tell on its own

How much of the difference in reading a guesser can explain using only this kind of clue.

Test R² of a model that sees only this group.

Lost when the group is removed

What this group adds that no other group provides.

Full-model R² minus the R² of a model without the group. Negative values are shown as zero.

Predicted decile against what happened

Stories the model ranks higher really do get higher view percentiles.

Test stories sorted by predicted percentile and cut into ten equal groups; each bar is the real mean view percentile. 50% is the average story.

Tags that move the prediction

Effects are averaged over a fixed sample of stories: change one input, keep the rest, and see how the predicted view percentile moves. Tags overlap, so read them as hints.

Tags that push views up

Percentile points gained when the tag is added.

Change in predicted percentile (points) when the tag is added.

Tags that push views down

Percentile points lost when the tag is added.

Points of predicted percentile lost when the tag is added.

Adding ⁨مامان⁩ moves the predicted percentile by 30.5 points and adding ⁨عاشقی⁩ by -26.6. Tag choice is by far the strongest signal.

Effect of every tag on the predicted percentile
مامان74092%30.5
مادرزن24188%30.1
خاله45484%24.9
خواهر85685%24.8
عمه21180%23.9
خواهرزن55981%22.1
زن عمو27480%21.6
زن داداش19675%20.4
ضربدری33476%18.0
زندایی29074%17.2
دکتر16274%16.6
تابو1,16681%15.5
سکس خشن24564%14.8
زن میانسال19874%14.1
بیغیرتی1,92773%14.0

Other inputs

Pick an input to see how the average predicted percentile moves as it changes.

Year

Average predicted percentile by publication year.

Percentile is within publication year, so the year curve shows which eras held more of the stories that went on to be read, not time itself.

Explore

Search