Comparing Scalable Predictive Analysis using Spark XGBoost Platforms
Download as PPTX, PDF0 likes25 views
The document discusses a comparative analysis of scalable predictive analysis methods using XGBoost on various big data platforms for Airbnb listing predictions. It highlights the efficiency and accuracy of XGBoost, particularly when utilized in conjunction with GPU-accelerated environments, showcasing improved performance over traditional models. The findings indicate that platforms leveraging XGBoost, especially with RAPIDS and H2O, offer significant advantages in model training time and accuracy metrics.
![Big Data Artificial Intelligence Center (BigDAI)
Jongwook Woo
CalStateLA
Why XGBoost
XGBoost stands for Extreme Gradient Boosting,
Famous as it has won many Kaggle competitions.
Ensemble tree with distributed gradient-boosting based on [4]
– hardware optimization
– parallelized tree building
– tree pruning using ‘depth-first’ approach
– regularization through both LASSO (L1) and Ridge (L2) for avoiding
overfitting
– efficient handling of missing data
– built-in cross-validation capability (at each iteration)](https://image.slidesharecdn.com/airbnbratingprediconi-221212001506-d47b7315/85/Comparing-Scalable-Predictive-Analysis-using-Spark-XGBoost-Platforms-3-320.jpg)
![Big Data Artificial Intelligence Center (BigDAI)
Jongwook Woo
CalStateLA
Big Data AI Center: Regression Models [1-3]
Big Data: distributed parallel computing
Regression models to predict price of Airbnb listings
RAPIDS XGBoost models
74 % higher accuracy and up to 90 % faster than the
traditional models.
In Spark EMR, XGBoost using GPU
–is 17 % faster than using only CPU.](https://image.slidesharecdn.com/airbnbratingprediconi-221212001506-d47b7315/85/Comparing-Scalable-Predictive-Analysis-using-Spark-XGBoost-Platforms-4-320.jpg)
![Big Data Artificial Intelligence Center (BigDAI)
Jongwook Woo
CalStateLA
Big Data AI Center: Performance of Models [1-3]
0
100
200
300
400
500
600
XGBoost
CPU
XGBoost
GPU
XGBoost
GPU
XGBoost
CPU
DT RF DF RF GBT GBT
Linux File
Systems
HDFS S3 S3 S3 S3 Linux File
Systems
Linux File
Systems
S3 Linux File
Systems
RayDP EMR EMR EMR EMR EMR RayDP RayDP EMR RayDP
Train Time (sec)](https://image.slidesharecdn.com/airbnbratingprediconi-221212001506-d47b7315/85/Comparing-Scalable-Predictive-Analysis-using-Spark-XGBoost-Platforms-5-320.jpg)
![Big Data Artificial Intelligence Center (BigDAI)
Jongwook Woo
CalStateLA
Big Data AI Center: Performance of Models [1 – 3]
Cluster Models Training
Time (sec)
RMSE R2
HDFS XGBoost GPU 15.4 33.42 0.731
S3 DT 31.1 130.4 0.238
S3 GBT 173.2 128.7 0.248
S3 RF 48.2 130.1 0.247
S3 XGBoost CPU 21.5 33.4 0.731
RayDP
EXT4
XGBoost CPU 4.3 31.1 NA
Muralidharan, S., Yadav, S., Huh, J., Lee, S., & Woo, J. (June 2022). “Scalable Prediction Models for Airbnb Listing in
Spark Big Data Cluster using GPU-accelerated RAPIDS”, Journal of Information and Communication Convergence
Engineering, 20(2), 96–102. https://doi.org/10.6109/JICCE.2022.20.2.96](https://image.slidesharecdn.com/airbnbratingprediconi-221212001506-d47b7315/85/Comparing-Scalable-Predictive-Analysis-using-Spark-XGBoost-Platforms-6-320.jpg)
![Big Data Artificial Intelligence Center (BigDAI)
Jongwook Woo
CalStateLA
XGBoost: H2O Sparkling Water [8]](https://image.slidesharecdn.com/airbnbratingprediconi-221212001506-d47b7315/85/Comparing-Scalable-Predictive-Analysis-using-Spark-XGBoost-Platforms-17-320.jpg)
Comparing Scalable Predictive Analysis using Spark XGBoost Platforms
- 1. Jongwook Woo BigDAI CalStateLA ICONI 2022 Jeju Island, Korea December 12, 2022 Jongwook Woo ([email protected]) Savita Yadav, Samyuktha Muralidharan, Big Data AI Center (BigDAI) California State University, Los Angeles Comparing Scalable Predictive Analysis using Spark XGBoost Platforms
- 2. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Contents Why XGBoost Dataset Big Data Platforms Comparison of Experimental Results Conclusion
- 3. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Why XGBoost XGBoost stands for Extreme Gradient Boosting, Famous as it has won many Kaggle competitions. Ensemble tree with distributed gradient-boosting based on [4] – hardware optimization – parallelized tree building – tree pruning using ‘depth-first’ approach – regularization through both LASSO (L1) and Ridge (L2) for avoiding overfitting – efficient handling of missing data – built-in cross-validation capability (at each iteration)
- 4. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Big Data AI Center: Regression Models [1-3] Big Data: distributed parallel computing Regression models to predict price of Airbnb listings RAPIDS XGBoost models 74 % higher accuracy and up to 90 % faster than the traditional models. In Spark EMR, XGBoost using GPU –is 17 % faster than using only CPU.
- 5. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Big Data AI Center: Performance of Models [1-3] 0 100 200 300 400 500 600 XGBoost CPU XGBoost GPU XGBoost GPU XGBoost CPU DT RF DF RF GBT GBT Linux File Systems HDFS S3 S3 S3 S3 Linux File Systems Linux File Systems S3 Linux File Systems RayDP EMR EMR EMR EMR EMR RayDP RayDP EMR RayDP Train Time (sec)
- 6. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Big Data AI Center: Performance of Models [1 – 3] Cluster Models Training Time (sec) RMSE R2 HDFS XGBoost GPU 15.4 33.42 0.731 S3 DT 31.1 130.4 0.238 S3 GBT 173.2 128.7 0.248 S3 RF 48.2 130.1 0.247 S3 XGBoost CPU 21.5 33.4 0.731 RayDP EXT4 XGBoost CPU 4.3 31.1 NA Muralidharan, S., Yadav, S., Huh, J., Lee, S., & Woo, J. (June 2022). “Scalable Prediction Models for Airbnb Listing in Spark Big Data Cluster using GPU-accelerated RAPIDS”, Journal of Information and Communication Convergence Engineering, 20(2), 96–102. https://doi.org/10.6109/JICCE.2022.20.2.96
- 7. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Contents Why XGBoost Dataset Big Data Platforms Comparison of Experimental Results Conclusion
- 8. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Dataset Details Dataset: Airbnb Listings Dataset URLs: – https://public.opendatasoft.com/explore/dataset/airbnb- listings/table/?disjunctive.host_verifications&disjunctive.amenities&disjunctiv e.features Total Dataset size: 4 GB Dataset Format: CSV Not easy to handle with the traditional systems – In Data Engineering, Analysis, and Science Predicting whether the listing has a good rating or not. Using two-class classification algorithms to build a model – to classify the listings as high rated and low rated, based on the features of the listing.
- 9. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Dataset Details (Cont’d) Data Transformation Label: Review Scores Rating – Converting the Review Scores Rating column to binary class • Review Scores Rating >= 80 -> High Rating • Review Scores Rating < 80 -> Low Rating Features: 25 columns – "Host Listings Count", "Host Total Listings Count", "Calculated host listings count", "Security Deposit", "Cleaning Fee" , "Host Response Time", "Host Response Rate", "Host Acceptance Rate", "Property Type", "Room Type", "Weekly Price", "Monthly Price", "Maximum Nights", "Review Scores Accuracy", "Review Scores Cleanliness", "Review Scores Checkin", "Review Scores Communication", "Review Scores Location", "Review Scores Value", "Cancellation Policy", "Bedrooms", "Bathrooms", "Beds", "Extra People", "Minimum Nights"
- 10. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Need for Big Data Predictive Analysis Performance issue using traditional predictive models The size of the data set is 4 Giga-Bytes. traditional systems mostly generate – a memory error – or takes several hours or days to build them. Need Big Data Non-expensive platform, which is distributed parallel computing systems – and that can store a large-scale data and process it in parallel Apache Hadoop and Spark since 2006 – Non-expensive Super Computer – Any small companies or university labs can own it
- 11. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Contents Why XGBoost Dataset Big Data Platforms Comparison of Experimental Results Conclusion
- 12. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Big Data: Distributed Computing Systems Hadoop Spark cluster in-memory computing engine – built by AMPLab at UC Berkeley in 2012. – Spark is 100 times faster than Hadoop MapReduce in theory – Supports Machine Learning Algorithms More Platforms can leverage Spark Cluster for XGBoost Nvidia RAPIDS Intel Big DL H2O Sparkling Water How is performance of the platform?
- 13. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA XGBoost AWS EMR EMR Hadoop Spark Big Data cluster – EMR 6.7.0: • Hadoop 3.2.1, Spark 3.2.1 • Does not have XGBoost – 3 nodes cluster XGBoost support – NVidia RAPIDS • set up Rapids into EMR Spark – to build an XGBoost model with GPU and CPU – Big DL – Sparkling Water
- 14. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA AWS EMR Big Data Cluster PySpark, RAPIDS, Big DL, and Sparkling Water installed AWS EMR Big Data Cluster with GPU: 2 x g4dn.2xlarge, m3.xlarge 1. Airbnb files: HDFS 2. Data Read & Engineering 3. Train and Test Models
- 15. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Intel Big DL Intel provides BigDL for distributed Deep Learning applications in the Spark cluster. And, provides XG DDL DDL lib DDL lib Deep Learning in Spark
- 16. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA XGBoost (Cont’d) H2O The open-source platform H2O.ai focuses on bringing AI to businesses through software. –H2O is its flagship product H2O Sparkling Water integrates H2O's scalable machine learning engine with Spark –with XGBoost library
- 17. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA XGBoost: H2O Sparkling Water [8]
- 18. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Contents Why XGBoost Dataset Big Data Platforms Comparison of Experimental Results Conclusion
- 19. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Evaluation Measurement Accuracy Precision –Intend to reduce the number of False Positives AUC Computing Time Parallel Computing to train models –With GPU
- 20. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Comparing Computing Time of Spark Platforms
- 21. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Comparing Accuracy of Spark Platforms
- 22. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Comparing the performance XGBoost models Big DL with GPU –is 25 – 50% faster for model training time than other platforms. H2O Sparkling has 5 - 7% better AUC –and 0.7% better Precision than others.
- 23. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Contents Why XGBoost Dataset Big Data Platforms Comparison of Experimental Results Conclusion
- 24. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Conclusion Comparing the performance of the platforms in Big Data Predictive Analysis for large-scale data XGBoost in RAPIDS, The Big DL, and Sparkling Water Big DL with GPU –fastest for model training time. H2O Sparkling has better AUC and Precision than others
- 25. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Questions?
- 26. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA References 1. Muralidharan, S., Yadav, S., Huh, J., Lee, S., & Woo, J. (June 2022). Scalable Prediction Models for Airbnb Listing in Spark Big Data Cluster using GPU-accelerated RAPIDS. Journal of Information and Communication Convergence Engineering, 20(2), 96– 102. https://doi.org/10.6109/JICCE.2022.20.2.96 2. Samyuktha Muralidharan, Savita Yadav, Sanghoon Lee, and, Jongwook Woo, "Scalable Price Prediction Models of Hosting Business Leveraging Big Data with GPU", KSII The 17th Asia Pacific International Conference on Information Science and Technology (APIC-IST) June 19-21 2022, pp95-97, ISSN 2093-0542 3. Savita Yadav, Samyuktha Muralidharan, Sanghoon Lee, Jongwook Woo, "Scalable Predictive Analysis for Airbnb Listing Rating", KSII The 13th International Conference on Internet (ICONI) 2021, Dec 12-14 2021, Jeju Island, Korea, pp370-372, ISSN 2093-0542 4. Why does XGBoost work so well?, https://www.kaggle.com/general/196541 5. Sparkling Water: Run on Hadoop, https://docs.h2o.ai/sparkling-water/3.2/latest- stable/doc/install/install_and_start.html#run-on-hadoop 6. Sparkling Water Booklet, https://docs.h2o.ai/sparkling-water/3.0/latest-stable/booklet/SparklingWaterBooklet.pdf 7. NVidia Rapids, https://github.com/NVIDIA/spark-rapids 8. Lesson 1, Sparkling Water Training, https://aquarium.h2o.ai/lab/8
- 27. Big Data Artificial Intelligence Center (BigDAI) Jongwook Woo CalStateLA Big Data Cluster Linearly Scalable for more storage and computing power Big Data Cluster