Whale Classification from echo-location clicks

This project primarily involves dealing with the development of a classifier system using under-water recordings of echo-location clicks. It currently looks to classify two types of Beaked Whales: The Cuvier's and the Gervais

Data Files:

---

• whale_data_15mb.np : Consists of a small subset of the data after filtering out the incomplete and misclassified data. It is used as the dataset for extracting and visualizing Eigen vectors.
• processed_data_15mb.np : The processed data files that consist of the data required for building the model i.e. Output of the Data_Processing_Whales.ipynb.
• processed_data_150mb.np : This is the same as processed_data_15mb.np , except that the Eigen Vector Analysis was done on a larger subset of Data.
• xgboost_output_full_data.pk : Pickle file that saves the classifcation predictions of the XGBoost model on the entire Test Dataset.

Data Pre-processing (Data_Processing_Whales.ipynb)

---

Input: whale_data_15mb.np
Output: processed_data_15mb.np

The following steps are carried out in this file:

• Calculating Mean and Covariance for the input data
• Extracting and Analyzing the Eigen Vectors
• Writing the processed data to the output file with following columns (zero-indexed):
  • Col 0-9: projections on first 10 eigen vectors
  • Col 10: rmse
  • Col 11: peak2peak
  • Col 12: label (0 if row.species==u'Gervais' else 1)

The code for the same can be found here. The preprocessed is a numpy array with 4175 rows (for the 15mb file)

XGBoost Analysis

---

Train/Validation/Test Split

The data is shuffled and divided as follow:

• Training: 70%
• Validation: 15%
• Testing: 15%

Parameters for XG Boost

• Maximum Depth of the Tree = 3
• Step size shrinkage used in update to prevents overfitting = 0.3
• Evaluation Criterion: Maximize Logliklihood
• Maximum Number of Iterations = 100, 1000

NOTE:

• Originally the data analysis was done on 16 GB with data for over 6 million clicks. The data files were preprocessed on PySpark (10 nodes) cluster.
• To avoid heavy computation time, smaller subsets of data have been provided along with the data generated after different stages.
• The file xgboost_output_full_data.pk has been provided so that visualization and analysis on the entire test dataset can be carried out.