supasorn / linkprediction Goto Github PK

Test Parallelization

• Run Single SGD on C++ for 200-300 seconds
• Run Single SGD on python for 200-300 seconds
• Re-run GraphLab with unified=true [Ham]
• dynamic balancing [Aek]
• get the RMSE_test for the final run on normal data set [Aek]
• RMSE-test for predicting with avg rating (baseline) [Aek]

Coldstart

• run normal booom!!!! ---RMSE test 1.5 using unified [Aek]
• rewrite sampler [Aek]
• run it with different degree of sampling [Aek]

Writing

• Add a few summary sentence
• Experiment for Parallelization
  • Performance
  • Grid Search — Why do we run grid search on DSGD, not NOMAD?
  • Final Results - what’s the number for unified? I remember pure is 0.002 worse. Also, what’s the baseline if we predict all the rating = average rating.
• make and explain the final graph [Ham]
• Explain the NOMAD dynamic improvement
• Cold start explanation [Aek]
• Lessons learned [Aek, Ham]
• Conclusion

• Load Tags and genres
• stem tags (probably using nltk?)
• update SGD

Basic

• Load User, Movie into SGraph
• Write SGD

• output multiple files so it’s easier to parse
• log the matrix for every 25 iterations so we can plot RMSE_test against iteration for the optimal lambda

• print RMSE_test more often? — need to debate about this
• print norm of each matrix / vector at the end.

• merge tags with movies.csv
• remove symbol
• crawl data from IMDB