We are developing a model to predict US flight delays using data from 2019 found on kaggle here. We classify delays as any flight that leaves >15 mins after its departure time. The columns we selected/created as features are:
- Day
- Month
- Hour
- Carrier code
- Flight number
- Origin
- Destination
- Distance
In progress.
Daniel - Data analysis
George - Model development
No blockers
- Python
- Apache Spark
- Hive
- Neo4j
The following algorithms are tested, using pyspark's machine learning packages, to see which best fit the data:
- Random Forest
- Decision Tree
- Multilayer Perceptron
- Naïve Bayes
We also tested the use of undersampling as the dataset was extremely imbalanced.
To execute the ML code, download spark and from your spark home execute the following:
bin/spark-submit --master local[4] ~path_to_file/ca4022-spark/ML\ models/Multilayer_Perceptron_Classifier.py
Daniel Sammon, 18364071
George Dockrell, 18745115
| Model | Accuracy | Area Under ROC Curve | Area under PR curve |
|---|---|---|---|
| Random Forest | 0.82 | 0.67 | 0.32 |
| Decision Tree | 0.83 | 0.60 | 0.22 |
| Multilayer Perceptron | 0.82 | 0.52 | 0.18 |
| Naïve Bayes | 0.53 | 0.48 | 0.17 |
| Model with undersampling | Accuracy | Area Under ROC Curve | Area under PR curve |
|---|---|---|---|
| Random Forest | 0.63 | 0.68 | 0.65 |
| Decision Tree | 0.63 | 0.62 | 0.61 |
| Multilayer Perceptron | 0.51 | 0.51 | 0.51 |
| Naïve Bayes | 0.52 | 0.48 | 0.48 |