sanjidnet / san-fransisco-crime Goto Github PK

This project contains and demonstrates my contribution to the San Fransisco Crime Classification problem of Kaggle. The task is carried out using R.

DATASET:

For this report, only the training data, train.csv is used for quick evaluation of the approaches. Nonetheless, the models that are used can most certainly be tried on the test.csv file of the San Fransisco Crime Classification. The data can be retrieved from the link below: https://www.kaggle.com/c/sf-crime/data

NOTE: please keep train.csv in the working directory while running the code(s)

The packages used are: dplyr, lubridate, nnet, caret, doMC, data.table and phyclust

Brief descriptions of the files are given below:

• report.pdf: Description of the approaches that are taken.
• report.Rmd: This is the rmarkdown file that generated the pdf above. Given all the packages are installed, knitting this file in R should generate the same pdf. The process takes less than 5 minutes in a Unix machine of RAM 4GB.
• experiment.Rmd: This file is the same as report.Rmd. Any changes made and executed to the code chunk of this file will generate another report experiment.pdf.

Possible changes might be, changing the size of subset used for evaluation. The original report produces and shows results of a subset of 2,000 observations from the full 878,049. This is indeed done for convenient report generation, but as a result, the accuracy of the models suffer. The amount of subset to be considered can be easily changed by simple changing the value of m in the code. If, m is very large (i.e. close the 878,049), then computation will take longer and might even require bigger RAM.

Other possible changes include chaning the tuning parameters of the learning models. But once again, these changes must be done with patience and caution, since longer time and bigger RAM might be required.