The goal of this assignment is to get some experience implementing the simple but effective machine learning technique, naïve Bayes classification, and applying it to a binary text classification task (i.e., spam detection). This assignment will describe the details of a particular implementation of a naïve Bayes classifier with the categories spam and ham (i.e., not spam). The teaching assistants have created a solution following these same directions, and your work will be evaluated based on its performance on an unseen test set in comparison to this solution. In addition, we ask you to submit a short report including the following information:
Your performance on the development data.
Your performance on the development data with a model built from only 10% of the training data.
Your performance on the development data after trying a different feature set, or approach to smoothing, and/or unknown words.
This is also an introduction to the Vocareum platform. Vocareum is like BlackBoard for programming assignments. You will use Vocareum to submit your assignment, and later to check your grade. You can submit your assignment as many times as you like, although we will grade the last version submitted applying any appropriate late penalties. So only submit after the deadline if you think your new version will result in a better grade after the late penalty. Each time you submit your code, Vocareum will run it, perform some basic checks, and tell you the results.
We have uploaded the email data for this assignment to Vocareum. There is training data, development data and test data. When you submit your assignment, Vocareum will run your nblearn.py script (see below) on the entire training set, and then run your nbclassify.py script (also described below) on the development data. Vocareum will alert you to any errors and report your performance on the development data. The training and development data are also available for download via Blackboard, and you will need to work locally on your own computer to run some of the experiments described here. However, the majority of your grade will depend on how your code performs on Vocareum not your computer. You will be writing your code in Python3. Vocareum is currently using version 3.6.4 of Python.
You will write three programs: nblearn.py will learn a naïve Bayes model from labeled data, nbclassify.py will use the model to classify new data and nbevaluate.py will print precision, recall and F1 scores based on the output of nbclassify.py on development (i.e., labeled) data. In addition, you will run experiments using these programs and report the results using this template: http://nld.ict.usc.edu/cs544-spring2020/assignment1/Report.txt
1. Write nblearn.py
nblearn.py will be invoked in the following way:
>python3 nblearn.py /path/to/input
And output a model file called nbmodel.txt
1.1 Reading data
The argument is a data directory. The script should search through the directory recursively looking for subdirectories containing the folders: “ham” and “spam”. Note that there can be multiple “ham” and “spam” folders in the data directory. Emails are stored in files with the extension “.txt” under these directories. The file structure of the training and development data on Blackboard is exactly the same as the data on Vocareum although these directories (i.e., “dev” and “train”) will be located in different places on Vocareum and on your personal computer.
ham and spam folders contain emails failing into the category of the folder name (i.e., a spam folder will contain only spam emails and a ham folder will contain only ham emails). Each email is stored in a separate text file. The emails have been preprocessed removing HTML tags, and leaving only the body and the subject. The files have been tokenized such that white space always separates tokens. Note, in the subject line, “Subject:” is considered as one token. Because of special characters in the corpus, you should use the following command to read files:
open(filename, “r”, encoding=”latin1″)
1.2 Learning the model
You will need to estimate and store P(spam) and P(ham) as well as conditional probabilities P(token|spam) and P(token|ham) for all unique tokens. These probabilities should be stored in the model file nbmodel.txt. The format of the file is up to you but your nbclassify.py program must be able to read it.
You are free to use any smoothing method, but you should at least use add-one smoothing. The official solution will use add-one smoothing. During testing, you can simply ignore unknown tokens not seen in training (i.e., pretend they did not occur).
nbclassify.py will be invoked in the following way:
>python3 nbclassify.py /path/to/input
The argument is again a data directory but you should not make any assumptions about the structure of the directory. Instead, you should search the directory for files with the extension “.txt”. nbclassify.py should read the parameters of a naïve Bayes model from the file nbmodel.txt, and classify each “.txt” file in the data directory as “ham” or “spam”, and write the result to a text file called nboutput.txt in the format below:
In the above format, LABEL is either “spam” or “ham” and path is the path to the file, and the filename (e.g., on Windows a path might be: “C:\dev\4\ham\0001.2000-01-17.beck.ham.txt”).
If you are taking the default approach to unknown tokens not seen in training, then nbclassify.py should simply ignore them (i.e., pretend they did not occur).
3. Write nbevaluate.py
nbevaluate.py will be invoked in the following way:
>python3 nbevaluate.py nboutput_filename
nboutput_filename is the output file of nbclassify.py described above. For each line in the file, nbevaluate.py will split the line into the guessed label and file path. nbevaluate.py will search for ham or spam in the path to determine the true label of the example (i.e., “spam” or “ham”). If neither is found, then it will skip to the next line in the file. Otherwise, the true label will be compared to the guessed label. You will need to maintain counts allowing you to calculate precision, recall and F1 score for both spam and ham and print them once all output has been processed. You will include these values in your report. Note that your program should not crash if no labeled examples are seen and you should be careful to avoid dividing by zero.
The first experiment is simply to
run nblearn.py on the entire training data,
run nbclassify.py on the dev data using the resulting model
use nbevaluate.py to measure performance and add the results to your report.
The second experiment requires you to train a model with 10% of the training data. You need to be careful how you divide the data (e.g., it would be a disaster if your training data only had one class). We recommend calculating a target number of files (i.e., 10% of the total) and selecting half of those files to be ham examples, and half to be spam examples. You can do this manually (e.g., copying files to create a smaller training directory) or by writing Python code (i.e., separate scripts or modifying nblearn.py). If you modify nblearn.py, it must still work as shown above so that we can test your code (i.e., we should be able to run it by simply including the data directory as the sole command line argument, and by default it should use all the data for training the model). After training a model with 10% of the data, classify the dev data, evaluate the results, and add the numbers to your report.
Your third task is to attempt at least one modification of your approach such as a different approach to smoothing, a different way to handle words never seen during training, or modifying the bag of words features. As an example of modifying the features, you could experiment with dropping features based on frequency or composition (e.g., tokens with no alphabetic characters).
Note (important): We strongly recommend that the modification is NOT the default behavior of your code as the modification could hurt performance on the unseen test set. Also, remember that we need to be able to run your code as described above with the data directory being the only command line argument.
What to turn in and Grading
You need to submit the following on Vocareum:
Report.txt. Download and fill in the template (http://nld.ict.usc.edu/cs544-spring2020/assignment1/Report.txt). Include the results from your experiments as well as a description of the modifications you used to try to improve performance on the development set. Make sure that none of your code writes to a file called Report.txt. We don’t want to accidently damage the file when we test your code.
All your code: nblearn.py, nbclassify.py, nbevaluate.py and any additional code you created for the assignment (e.g., to sample from the training set). This is required and the academic honesty policy (see rules below) applies to all your code.
10% of your grade (10 points) is based on the report. Make sure to answer every question. The other 90% is based on the performance of your code compared to the official solution written by the teaching assistants following the directions above. If your performance is the same or better then you receive full credit. Otherwise, your score is proportional to your relative performance:
90 points * ( 1 – ( performance_of_solution – your_performance ) / performance_of_solution )
On time (no penalty): before February 23rd, 4pm.
One day late (10 point penalty): before February 24th, 4pm.
Two days late (20 point penalty): before February 25th, 4pm.
Three days late (30 point penalty): before February 26th, 4pm.
Zero credit after February 26th, 4pm.
DO NOT look for any kind of help on the web outside of the Python documentation.
DO NOT use any external libraries other than the default Python libraries.
Questions should go to Piazza first not email. This lets any of the TAs or instructors answer, and lets all students see the answer.
When posting to Piazza, please check first that your question has not been answered in another thread.
DO NOT wait until the last minute to work on the assignment. You may get stuck and last minute Piazza posts may not be answered in time.
This is an individual assignment. DO NOT work in teams or collaborate with others. You must be the sole author of 100% of the code and writing that you turn in.
DO NOT use code you find online or anywhere else.
DO NOT turn in material you did not create.
All cases of cheating or academic dishonesty will be dealt with according to University policy.
How will the TAs/Professors grade my HW?
We will be writing scripts which will run on Vocareum. Since we automate the process of grading, make sure that you write your code to match the output format “exactly”. If you do not do this there will be a penalty.
I found a piece of code on the web. Is it ok to use it?
No! We run plagiarism checks on the code you write. The plagiarism detector is a smart algorithm which can tell if you’ve copied the code from elsewhere/others.
Instead please contact TAs/Professor during the office hours with your questions.
Vocareum terminates my code before it finishes, but it finishes executing on my computer. This usually means that your implementation is inefficient. Keep an eye out for places where you iterate. Run time issues can be solved by using efficient data structures (Hashmap, Hashset etc.).
The late policy above still applies, and is automatically enforced by the system. Submit your code incrementally to make sure it functions on Vocareum.
When are the office hours ?
Please refer to the course website. http://nld.ict.usc.edu/cs544-spring2020/