BIZNIQUE / LINEAR ALGEBRA / 40
Combine latent factors with useful context
Combine latent factors with useful context
A model can combine a low-rank signal with information that the matrix alone does not contain.
Embedded Files
THE IDEA
Read the example.
Read the example.
A latent factor model summarises relationships through a limited rank. A hybrid model can add item, user or other contextual features. The illustration shows the rank-limited part; it does not claim to implement a complete hybrid recommendation system.
Prediction = factor contribution + context contribution
Use it well
Use it well
Evaluate whether context improves held-out predictions, especially for sparse or new entities. More inputs also create more opportunities for leakage and unstable effects.
TRY IT YOURSELF
A quick check.
A quick check.
Why might item metadata help when a new item has no interaction history?
Check your reasoning
Check your reasoning
Its known attributes can provide a representation before enough interactions exist to estimate an item-specific latent factor.
Page updated
Google Sites
Report abuse