Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Speaker Deck
PRO
Sign in
Sign up
for free
Recommendation Engine for wide transactions
harjinder-hari
June 09, 2017
Programming
0
82
Recommendation Engine for wide transactions
harjinder-hari
June 09, 2017
Tweet
Share
More Decks by harjinder-hari
See All by harjinder-hari
harjinderhari
0
59
harjinderhari
0
120
harjinderhari
0
79
harjinderhari
0
17
Other Decks in Programming
See All in Programming
rukiadia
3
940
yattom
32
11k
ganchan11
0
180
fr0gger
2
2.9k
yusuke57
0
120
daipresents
0
360
bkuhlmann
4
300
takuyaa
4
490
ajstarks
2
760
kilometer
2
300
kentatada
0
120
tkmnzm
1
230
Featured
See All Featured
rocio
155
11k
zenorocha
298
40k
reverentgeek
167
7.3k
productmarketing
7
790
robhawkes
53
2.9k
eitanlees
116
10k
scottboms
252
11k
jlugia
217
16k
bermonpainter
343
26k
andyhume
64
3.8k
jacobian
257
20k
bryan
32
3.5k
Transcript
Rec Sys - wide transactions Harjinder Mistry Red Hat |
@hmistry
Agenda 1. RecSys - 2 min primer 2. Problem -
Definition 3. Challenges in Standard Approaches 4. Our approach & architecture
RecSys examples
Basic terminologies user-item matrix explicit vs implicit feedback — user-user
— user-item — item-item image source
Frequent Pa!ern mining Applications — Customer Analysis — Brick-and-mortar retail
— Handling cold-start situation — Retrieval
Frequent Pa!ern mining Algorithms — apriori — FP Growth
openshi!.io
Helping developers become more efficient recommendations on packages recommendations on
the stack
Input data Projects/stacks - from code repositories — Java (pom.xml)
— Node.js (packages.json) — Python (requirements.txt)
spark, elastic cloud compute.... cool - let's rock
developers are amazing - but, of course
Wide transactions - challenges — existing methods didn't work —
time to train was huge — memory issues
As a self-serve platform, turnaround time as important as accuracy
Matrix Factorization is fast image source
Let's use matrix factorization (ALS) to generate frequent pa!erns
Step 1: Train ALS model
Step 2: Generate initial seed: random candidate set
Step 3: Find recommended product(package)
Step 4: Add it to the frequent pa!ern list and
continue
None
Why not deep learning?
Code, Slides and Contact ____ Code will be open-sourced soon!
Harjinder Mistry email:hmistry@redhat.com