My first exploration of basic statistical analysis techniques and a primitive neural network in order to classify messages utilizing sklearn and Keras. The messages used in the project represent reports or data that NATO forces would provide after conducting patrols in a given area. The messages are representative of operations in a counter insurgency environemnt (COIN) rather than a peer on peer force. The data set is further described in the paper by J. L. Graham, J. Rimland, and D. L. Hall, "A COIN-inspired synthetic data set for qualitative evaluation of hard and soft fusion systems".
The package is built using python3.5 .
pip install requirements.txt
The first file was an exploration and introduction to Sklearn's built in statsical packages and techniques in order to classify the messages. It will generate several graphics to represent the findings and accuracy of the techniques in an informative format.Additionally, it was the first foray into Stats and primitive data analysis that I have attempted.
python3 stats.py
The second file consists of my first neural network created in order to classify the provided messages.The file contains the first attempt at designing a network and the second iteration after additional research.The accuracy improved roughly four times between the two iterations.
python3 syncoin.py
This project exposed myself to many different libraries , clustering techniques and visual aides available in python.
The Y-axis is the actual label of the given message and the X-axis is the predicted label from the network.The network performed on par or
slightly better than the basic stats techniques from sklearn.
Gaussian Naive Bayes was employed and then compared to Multinomial Naive Bayes.
In the begining phases of the project, I utilized kmeans clustering in order to explore the data and investigate any possible relations or overlap in the set.A cursory glance was given to Agglomerative and Affinity Clustering.The big take away from these techniques are the hyper-parameters of each function can heavily influence the outcome of the analysis.Wattenberg, et al has a great article that visualizes the common issues that can occur with t-SNE.
With minimal tuning of the hyper-parameters, the clustering supports the notion that message types can overlap such as ISG and SUN/ISG.The black X represents the center of each cluster.The Dendrogram is based off of the same K-means clustering output and it visualizes the merging of the clusters and the distance between each iteration.
These two techniques weren't explored as much as k-means due to my unfamiliarity with the purpose of the hyper-parameters. and sklearn at the time.
- Klettgau - Initial work
- Joernhees - Execellent resource for sklearn as well as the fancy dendrogram visual.
This project is licensed under the MIT License - see the LICENSE.md file for details