7/31/2019 Docs Slides Lecture16
1/22
Recomme
Systems
Problem
formula4oMachineLearning
7/31/2019 Docs Slides Lecture16
2/22
Example:Predic/ngmoviera/ngs
Userratesmoviesusingonetofivestars
Movie Alice(1) Bob(2) Carol(3) Dave(4)
Loveatlast
Romanceforever
Cutepuppiesoflove
NonstopcarchasesSwordsvs.karate
=
=
=
ra
=
u
(d
7/31/2019 Docs Slides Lecture16
3/22
Recomme
SystemsContent-based
recommenda4
MachineLearning
7/31/2019 Docs Slides Lecture16
4/22
7/31/2019 Docs Slides Lecture16
5/22
Problemformula/on
ifuserhasratedmovie(0otherwise)
ra4ngbyuseronmovie(ifdefined)
=parametervectorforuser
=featurevectorformovie
Foruser,movie,predictedra4ng:
=no.ofmoviesratedbyuserTolearn:
7/31/2019 Docs Slides Lecture16
6/22
Op/miza/onobjec/ve:
Tolearn(parameterforuser):
Tolearn :
7/31/2019 Docs Slides Lecture16
7/22
Op/miza/onalgorithm:
Gradientdescentupdate:
7/31/2019 Docs Slides Lecture16
8/22
Recomme
Systems
MachineLearning
Collabora4
filtering
7/31/2019 Docs Slides Lecture16
9/22
Problemmo/va/on
Movie Alice(1) Bob(2) Carol(3) Dave(4)
(romance) (a
Loveatlast 5 5 0 0 0.9Romanceforever 5 ? ? 0 1.0
Cutepuppiesof
love
? 0 ? 0.99
Nonstopcar
chases
0 0 5 0.1
Swordsvs.karate 0 0 5 ? 0
7/31/2019 Docs Slides Lecture16
10/22
Problemmo/va/on
Movie Alice(1) Bob(2) Carol(3) Dave(4)
(romance) (a
Loveatlast 5 5 0 0 ?Romanceforever 5 ? ? 0 ?
Cutepuppiesof
love
? 0 ? ?
Nonstopcar
chases
0 0 5 ?
Swordsvs.karate 0 0 5 ? ?
7/31/2019 Docs Slides Lecture16
11/22
Op/miza/onalgorithm
Given ,tolearn:
Given ,tolearn :
7/31/2019 Docs Slides Lecture16
12/22
Collabora/vefiltering
Given (andmoviera4ngs),
canes4mate
Given ,
canes4mate
7/31/2019 Docs Slides Lecture16
13/22
Recomme
Systems
MachineLearning
Collabora4ve
filteringalgori
7/31/2019 Docs Slides Lecture16
14/22
Collabora/vefilteringop/miza/onobjec/ve
Given ,es4mate :
Given ,es4mate :
Minimizing and simultane
7/31/2019 Docs Slides Lecture16
15/22
Collabora/vefilteringalgorithm
1. Ini4alize tosmallrandom2. Minimize usinggradi
descent(oranadvancedop4miza4onalgorithm).E.every :
3. Forauserwithparametersandamoviewith(leafeatures,predictastarra4ngof.
7/31/2019 Docs Slides Lecture16
16/22
Recomme
Systems
MachineLearning
Vectoriza4on
Lowrankmafactoriza4on
7/31/2019 Docs Slides Lecture16
17/22
Collabora/vefiltering
Movie Alice(1) Bob(2) Carol(3) Dave(4)
Loveatlast 5 5 0 0
Romanceforever 5 ? ? 0
Cutepuppiesof
love
? 0 ?
Nonstopcar
chases
0 0 5
Swordsvs.karate 0 0 5 ?
7/31/2019 Docs Slides Lecture16
18/22
7/31/2019 Docs Slides Lecture16
19/22
Findingrelatedmovies
Foreachproduct,welearnafeaturevector.
Howtofindmoviesrelatedtomovie?
5mostsimilarmoviestomovie:
Findthe5movieswiththesmallest .
7/31/2019 Docs Slides Lecture16
20/22
Recomme
Systems
MachineLearning
Implementa4o
detail:Meannormaliza4on
7/31/2019 Docs Slides Lecture16
21/22
Userswhohavenotratedanymovies
Movie Alice(1) Bob(2) Carol(3) Dave(4) Eve(5)
Loveatlast 5 5 0 0 ?
Romanceforever 5 ? ? 0 ?Cutepuppiesoflove ? 0 ? ?
Nonstopcarchases 0 0 5 ?
Swordsvs.karate 0 0 5 ? ?
7/31/2019 Docs Slides Lecture16
22/22
MeanNormaliza/on:
Foruser,onmoviepredict:
User5(Eve):