Machine Learning for Ranked Retrieval

Notes:

Why?

Getting field weights right is hard
Title match > Meta keyword match?
Query "iphone": "iPhone cover" > "iPhone X"?

Just rank by click!

But what about new items?

Notes:

How?

Calculate weights
Apply new weights
Compare actual rank to expected rank (nDCG, MAP, ERR, …)
Try harder

Notes:

From

Notes:

To

Notes: