Geolocaton Predictor

This is the code for predicting geolocaton of tweets trainning on token frequencies using Decision Tree and Naïve Bayes.

Implementation

Feature selection

In util/preprocessing/merge.py,

• feature_filter shows it drops single character features like [a, b, ..., n]

• merge shows it intuitively merges similar features like [aha, ahah, ..., ahahahaha] and [taco, tacos]

• merge also shows it uses Recursive Feature Elimination to rank features and select best 300 features.

Classifier Combination

In preprocess/merge.py,

• result_combination shows it uses results of Decision Tree, Random Forest, Bernoulli Naïve Bayes, Complement Naïve Bayes and Multinomial Naïve Bayes to vote out the majority predictions while the following shows the models can be slightly different every time training.

Instance manipulation

In util/train.py,

• complement_nb shows it uses bagging to generate multiple training datasets.
• complement_nb also shows it uses 42-Fold Cross Validation to generate multiple training datasets.

Algorithm manipulation

In util/train.py,

• complement_nb also shows it uses GridSearchCV to generate multiple classifiers and select the best based on accuracy.

Dataset

• See Eisenstein, Jacob, et al.

  Although some file's size in datasets is greater than 50.00 MB, it was still added in datasets for convenience. (See, http://git.io/iEPt8g )

Requirements

• python3+

pip install -r requirements.txt  

Usage

Note: The code will remove the old models and results every time running. MAKE SURE you have saved your satisfying models..

Train

python run.py -t datasets/train-best200.csv datasets/dev-best200.csv  

the output would be like:

INFO:root:[*] Merging datasets/train-best200.csv   
 42%|████████         | 1006/2396 [00:05<00:20, 92.03 users/s]  
...  
...  
[*] Saved models/0.8126_2019-10-02_20:02  
[*] Accuracy: 0.8125955095803455  
 precision    recall   f_scoreCalifornia   0.618944  0.835128  0.710966  
NewYork      0.899371  0.854647  0.876439  
Georgia      0.788070  0.622080  0.695305  
weighted     0.827448  0.812596  0.814974  

Predict

python run.py -p models/ datasets/dev-best200.csv   

the output would be like:

...  
INFO:root:[*] Saved results/final_results.csv  
INFO:root:[*] Time costs in seconds:  
 PredictTime_cost  11.98s  

Score

python run.py -s results/final_results.csv  datasets/dev-best200.csv  

the output would be like:

[*] Accuracy: 0.8224697308099213  
 precision    recall   f_scoreCalifornia   0.653035  0.852199  0.739441  
NewYork      0.747993  0.647940  0.694381  
Georgia      0.909456  0.858296  0.883136  
weighted     0.833854  0.822470  0.824577  
INFO:root:[*] Time costs in seconds:  
 ScoreTime_cost  1.48s  
  

Train&Predict&Score

python run.py \
 -t datasets/train-best200.csv datasets/dev-best200.csv \
 -p models/ datasets/dev-best200.csv \
 -s results/final_results.csv datasets/dev-best200.csv

Help

python run.py -h  

Used libraries

• sklearn for easily using Complement Naive Bayes, some feature selectors and other learning tools.
• pandas, numpy for easily handling data.
• tqdm for showing the process of loop.
• joblib for dumping/loading memory to/from disk.
• nltk for capturing word types on the purpose of feature filtering

License

See LICENSE file.