track 1 – matrix factorization final presentation · matrix factorization is a working model for...
TRANSCRIPT
![Page 1: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/1.jpg)
Track 1 – Matrix Factorization Final Presentation
Collaborative Filtering Markus Freitag, Jan-Felix Schwarz 7 July 2011
![Page 2: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/2.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
2
![Page 3: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/3.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
3
![Page 4: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/4.jpg)
€
5.1 10.0 21.34.7 9.2 1.90.0 21.9 14.77.9 8.5 40.210.1 0.2 2.99.1 8.1 8.716.6 20.1 4.17.8 1.0 0.1
⎛
⎝
⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟
T
•
1.9 20.1 9.423.1 0.1 4.210.2 4.0 1.91.2 0.7 12.27.3 9.3 13.76.3 28.1 7.29.0 5.3 3.25.2 11.1 12.05.7 3.9 2.70.3 0.0 0.16.7 21.2 0.06.4 7.9 3.2
⎛
⎝
⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟
=
10 50 0 0 250 100 0
100 800 50 30
20 020 100 0 50 60
0 9010 50 20
90 100 2020 100 0
0 0 4070 10 10
⎛
⎝
⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟
user features
item features
Recap: Matrix Factorization
4
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
![Page 5: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/5.jpg)
Interim Presentation | Markus Freitag, Jan-Felix Schwarz | 09.06.2011
€
5.1 10.0 21.34.7 9.2 1.90.0 21.9 14.77.9 8.5 40.210.1 0.2 2.99.1 8.1 8.716.6 20.1 4.17.8 1.0 0.1
⎛
⎝
⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟
T
•
1.9 20.1 9.423.1 0.1 4.210.2 4.0 1.91.2 0.7 12.27.3 9.3 13.76.3 28.1 7.29.0 5.3 3.25.2 11.1 12.05.7 3.9 2.70.3 0.0 0.16.7 21.2 0.06.4 7.9 3.2
⎛
⎝
⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟
=
10 50 80 90 0 0 100 2570 55 0 90 100 15 10 0100 76 80 90 10 30 20 00 90 10 50 30 90 100 100 10 100 70 40 20 10 020 100 100 0 10 50 60 9080 20 0 80 90 76 0 1010 50 90 20 10 90 100 100 0 10 50 90 100 40 2060 70 50 20 90 10 100 00 10 0 90 40 20 50 3040 80 70 10 100 0 10 10
⎛
⎝
⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜
⎞
⎠
⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟
user features
item features
Recap: Matrix Factorization
5
![Page 6: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/6.jpg)
Recap: SGD Algorithm
Stochastic Gradient Descent (SGD) ■ Approximation procedure for learning one feature
■ For each rating in the training set the feature values are modified relative to the prediction error □ User value += Learning Rate * Error * Item value □ Item value += Learning Rate * Error * User Value
■ Iterate over the training set until the sum of squared errors (SSE) converges
■ Training set split into 4 subsets (track, album, artist, genre) □ Don’t presume a common underlying model
6
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
![Page 7: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/7.jpg)
Roadmap & Implementation Status
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
30.6. 12.5.
First RMSE
+ Work on
performance KDD
submission
Get a better understanding
of the rating data +
Include Biases
(Include Temporal Effects)
Consider Item Relations
W1 W2 W3 W4 W5 W6 W7
Tweak
Tweak more
9.6.
7
![Page 8: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/8.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
8
![Page 9: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/9.jpg)
Better Biases
■ Naive Bias: Item Avg – Global Avg □ But: What if there is only one observed rating?
à See it as a draw from the true probability distribution
■ Best guess for actual mean: linear blend of the observed mean and the global mean □ Blending ratio equal to the ratio of variances
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
observed rating
true probability distribution
9
![Page 10: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/10.jpg)
Better Biases
■ Va: Variance of all the items‘ average ratings ■ Vb: Variance of individual item ratings
■ K = Va / Vb
■ S. Funk used a constant K ■ “But in fact K=25 seems to work well so I used that instead. :)“ ■ We also tried 25 and in fact it works
» http://sifter.org/~simon/journal/20061211.html
€
BetterItemAvg =K *GlobalAvg+sum(ObservedRatings)
K *count(ObservedRating)
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
10
![Page 11: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/11.jpg)
Better Biases
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
Validation Test 24.2480 27.3456 24.2540 26.8627
RMSE Improvement
11
![Page 12: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/12.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
12
![Page 13: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/13.jpg)
Post Processing
■ Idea: Identify bulk ratings □ Unify predictions in the bulk rating
■ Round to most frequent rating values (decades)
0 10 20 30 40 50 60 70 80 90 100
“There was also a bulk-rating feature where users could query the system [...] and rate them in one go. Lastly,
some users tried to influence the recommender by doing lots of ratings (possibly even semi-automatic)
using the latter option.“ http://tech.groups.yahoo.com/group/kddcup2011/message/68
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
13
![Page 14: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/14.jpg)
Post Processing
■ There are many sets of simultaneous ratings by one user □ Only few sets have a constant rating value!
□ But many sets have only one or two “outliers“
■ Identify users which tend to bulk rate using metrics and thresholds □ Average count of distinct rating values in potential bulks □ Average span between minimum and maximum rating value in
potential bulks
■ If a “bulk rating user” has simultaneous ratings in the test set assume a bulk rating
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
14
![Page 15: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/15.jpg)
Post Processing
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
Validation Test 24.9571 25.3100 24.7436 25.3909
RMSE Worsening
à Maybe tweaking parameters might have yielded a slight improvement
15
![Page 16: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/16.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
16
![Page 17: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/17.jpg)
Regularization
■ Regularization tries to prevent overfitting when learning a single feature
■ A certain factor K damps the learning effect
userValue += lrate * (err * movieValue - K * userValue)
movieValue += lrate * (err * userValue - K * movieValue)
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
17
![Page 18: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/18.jpg)
Regularization
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
Validation Test 24.2434 26.9824 24.9700 27.3153
RMSE Worsening
à Regularization probably only makes sense when more features are used
18
![Page 19: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/19.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
19
![Page 20: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/20.jpg)
Parameter Tweaking
■ Tested the effect of every parameter ■ Challenge: Parameters affect each other
■ Learning Rate set to 0.002 □ Lower: takes too long □ Higher: risk of shooting over, underfitting
■ Feature count set to 14 □ Lower: underfitting □ Higher: overfitting
■ Improvement threshold set to 0,01% (break condition)
□ Higher: single features get less “detailed” □ Lower: takes too long
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
20
![Page 21: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/21.jpg)
Submissions & RMSEs
Submission Results
# Description LR Features RMSE 1 Test with 50 - - 37.8262 2 First complete run 0.01 1 28.3295 7 Test with more f. 0.002 10 27.3462 12 Used validation set 0.002 10 26.5217
21 Better bias 0.002 10 26.8627
25 Regularization 0.002 14 27.3153
27 Best values 0.002 14 26.1261
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
21
![Page 22: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/22.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
22
![Page 23: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/23.jpg)
Overlaps of Good and Bad Predictions
item good item bad
matrix good 33% 2%
matrix bad <1% 65%
item good item bad
hierarchie good 58% <1%
hierarchie bad <1% 74%
matrix good matrix bad
hierarchie good 52% <1%
hierarchie bad 1% 59%
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
23
![Page 24: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/24.jpg)
Performance on Different Item Types
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
Matrix good bad
Item-based good bad
Hierarchy good bad
album -3% +3% -4% 0% -8% -1%
artist +12% -11% +16% -8% 34% -5%
genre 0% -5% 0% -5% -5% -6%
track -9% 14% -12% +13% -21% +12%
“portion of a type in the set” –
“portion of that type in the whole validation set”
24
![Page 25: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/25.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
25
![Page 26: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/26.jpg)
Contribution
■ Implemented a matrix factorization model for a collaborative filtering use case
■ Made the implementation work on a very large data set ■ Various extensions of the model
□ Biases □ Hierarchy □ Temporal Effects (Post Processing)
■ Analysis to get a better understanding of the data set
■ Combination with other approaches (Item-based, Hierarchy) ■ Close to the ladder: Best RMSE 25.3100 (~0.5 to #100)
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
26
![Page 27: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/27.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
27
![Page 28: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/28.jpg)
Future Work: Item Relations
■ We have 4 different prediction models ■ Combined these models to improve predictions
■ Blend prediction with associated predictions □ of the same user □ for related items
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
The Doors
Rock Psychedelic Rock Blues-Rock
1. The Changeling 2. Love Her Madly ... 10. Riders on the Storm
90
80 60
40
80
100
70
90
28
![Page 29: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/29.jpg)
Future Work: Temporal Effects
■ Express biases and vectors as functions over time ■ Model will be able to reflect trends in item popularity and user
preferences ■ Regression problem to learn functions
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011 0 2 4 6
29
![Page 30: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/30.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
30
![Page 31: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/31.jpg)
Lessons Learned
■ Matrix factorization is a working model for collaborative filtering ■ Squeezing out the last points of improvements for one model gets
harder and harder ■ Combination of different models is effective
■ Implement various approaches and combine them instead of optimizing one algorithm to extremes!
■ Be pragmatic! Hands-on experience ■ Collaborative Filtering, Machine Learning, and statistics
■ Efficient use of memory, performance optimization, profiling
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
31
![Page 32: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/32.jpg)
Agenda
1. Recap: Matrix Factorization 2. Latest Improvements
□ Better Biases □ Post Processing □ Regularization □ Parameter Tweaking
3. Statistical Analysis for Combination of Approaches 4. Contribution 5. Future Work 6. Lessons Learned
7. Summary
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
32
![Page 33: Track 1 – Matrix Factorization Final Presentation · Matrix factorization is a working model for collaborative filtering Squeezing out the last points of improvements for one model](https://reader030.vdocuments.net/reader030/viewer/2022040803/5e3e83c27705da4c1940aa0d/html5/thumbnails/33.jpg)
Summary
■ Implemented various extensions for the matrix factorization approach □ Not everything improved the RMSE
■ Optimized our algorithm by tweaking parameters, submissions for testing effects on the test set
■ Statistic analysis to compare our different approaches (Matrix, Item-based, Hierarchy)
■ Time was the limiting factor. There are still numerous extension possibilities.
Final Presentation | Markus Freitag, Jan-Felix Schwarz | 07.07.2011
Questions?
33