TU Wien:Data-intensive Computing VU (Knees)
|Lecturers||Peter Knees, Elmar Kiesling|
|Department||Information Systems Engineering|
|Links||tiss:194048 , Mattermost-Channel|
|Master Data Science||Pflichtmodul BDHPC/FD - Big Data and High Performance Computing - Foundations|
- MapReduce (Java)
- Spark (either scala or python): RDDs and DataFrame/DataSets, MLLib for Machine Learning
- Text processing: pre-processing, feature selection, text classification
Recommended/necessary: Java, Python (for pySpark) and/or scala (spark). Some basics of functional programming (basic lambda expressions) are useful as well. Having used MapReduce or Spark or knowing some of the concepts is certainly helpful. Some basic shell scripting and using shell for navigating (os and hadoop) file system. Some basic concepts of machine learning, e.g. train/validation/test splits or cross-validation, parameter tuning. Concepts of text processing / information retrieval, e.g. TF-IDF, feature selection using Chi-Square values.
Some of the material is also covered in the class "Advanced Database Systems", check material for that class if possible, it is partially more in-depth.
Entry exercise, 2 assignments (individual work) + group project (groups of 4)
- Entry exercise: Word count example using MapReduce on Hadoop cluster
- Assignment 1: Map Reduce, text pre-processing, Chi-Square feature selection. Write a MapReduce implementation.
- Assignment 2:
- Part 1: Spark RDDs - Part 2: Spark DataFrames, pipelines for MLLib . Part 3: ML pipepline for text classification
- Assignment 3: RecSys challenge for the given year
Dauer der Zeugnisausstellung
- start early, problems may be challenging if you're not somewhat fluent in java, python and/or scala.
- start early, the cluster may be used heavily and not working sometimes
- set up a spark environment on your own machine if possible, and start developping locally with smaller subsets of the data.
- test on the cluster at some point to make sure everything runs there (file io uisng hadoop, starting job from shell script, etc.).
Ass1 - Tipps for hadoop:
- you can run multiple map-reduce jobs
- pass simple variables via context Counter between jobs
- pass more complex data via (cached) files between jobs
- checkout the setup and cleanup function of mappers and reducers
- though they mention Avro, it created an unnecessary overhead for me and didn't work in the end. going with a text based solution is easier IMO
- check this link for chi-square calculation: http://www.learn4master.com/algorithms/chi-square-test-for-feature-selection
Ass2 - Tips:
Check accumulators and broadcaster
Verbesserungsvorschläge / Kritik