Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/taewan-p/spam_message_detector

Distinguish spam messages with Machine Learning.
https://github.com/taewan-p/spam_message_detector

Last synced: 6 months ago
JSON representation

Distinguish spam messages with Machine Learning.

Awesome Lists containing this project

README 

          
# spam_message_detector

Distinguish Korean spam messages with Machine Learning.



**Well, Right now this is not a detector, but close to an analysis report. Please read the whole README.md to see what is going on.😄**



## Beginning



Recently, I got annoyed with spam messages that keeps coming to my phone. I have seen several articles about how some data scientists put deep learning into distinguishing spam messages in Korean. I thought it would be cool to make your own spam detector since I recently seen some patterns of spam messages. So I decided to make one. However, this is my first time with Machine Learning. I do not know **anything** about it. I had no clue where to start before I read an article about **Naive Bayes classifier**. After reading this, I thought this is what I should use for my project.



## First Guess



In my phone, I had about 1500 received messages. I made a csv file of all my messages that were in my phone and classified spam/ham. And about 10% of them were spam messages. While looking at the spam messages in my phone, I found some patterns. Spam messages are likely to include:



1. Special Characters

2. Weird website address

3. To unsubscribe, call XXXXXXXXXX

4. [Sent from Web], (Advertisement), Sent from overseas

5. ID and PW for weird website



I would definitely say that Pattern number 4 is not typically true, not really all messages with number 4 is spam. But, my phone has been repetitively receiving messages with pattern number 4 whis is a spam message, so I decided to put this one too.



## First Try



First of all, I used [KRWordBank](https://github.com/lovit/kr-wordrank) Library to find out the words that are mostly in spam messages. The top 5 words that spam messages include:



1. ```"[Web 발신]" - meaning that it is sent from the web.```

2. ```"[국제발신]" - meaning that it is sent from overseas.```

3. ```"무료거부" - meaning "To stop subscribing this message(FREE)."```

4. ```"ld" - Tricky one. The robots would detect the word "ID", so the sender switched capital i to lowercase L.```

5. ```"(ê´‘ê³ )" - meaning that this is an advertisement.```



Based on this statistics, I came up with the code of counting messages that includes common words in spam messages.



```python

  def count_keywords(word): # returns [spam_count, ham_count]

    result = [0,0]

    for i in total: # i is ['message content', 'ham/spam']

      if word in i[0]:

        if i[1] == "spam":

          result[0] += 1

        else: # ham

          result[1] += 1

    return result

```



*I coded the old way though, I will change the code later*.



After running this code, I found out that the word that is common in spam messages **can** be common in ham messages too. 



```reStructuredText

1. "[Web발신]":[95, 942]

2. "[국제발신]":[54, 46]

3. "무료거부":[31, 5]

4. "lD":[9, 0]

5. "(ê´‘ê³ )":[45, 22]

```



The most common word "[Web발신]" is more common in regular ham messages. Well this is because the data I have for ham messages are way more than spam messages, and that phrase will be always in front of your message when you send a message via internet. Well, I decided to use the sixth most common word instead of  "[Web발신]" : ```"PW" - meaning password.```



### The Probability of P(x|c)



If I picked a spam message, The probability of the message including "A" = spam A word count / total spam.



For example, number 2, "[국제발신]". The probability of the picked spam message including that word being a spam is ```54/150 = 36%```.  If you include Laplace Smoothing, it would be ```54+0.5/150+1 = 36.09% ```. Pretty much the same. Let's calculate the rest of the probability.



```reStructuredText

1. "[국제발신]" : 0.3609271523178808

2. "무료거부" : 0.20860927152317882

3. "lD" : 0.06291390728476821

4. "(ê´‘ê³ )" : 0.30132450331125826

5. "PW" : 0.06291390728476821

```



### P(c)s



In my case, the P(c)s are the probability of a message being a spam.



```python

  # P(c) is the probability of the message being a spam

  # len(analysis) is the value of total spam messages out of the data extracted from my csv.

  # len(total) is the value of total messages including spam and ham message.

  prob_spam = probability_calc(len(analysis), len(total), 0)

  # P(~c)

  prob_not_spam = 1 - prob_spam

```



#### Result



```

P(c) : 0.1038781163434903

P(~c) : 0.8961218836565097

```



### The Mighty Formula



What is this title about? I am talking about the Naive Bayes Classifier formula!



```P(c|x) = P(x|c) P(c) / { P(x|c) P(c) + P(x|~c) P(~c) }```



I am using this formula because I did not assume ```P(c) = P(~c) = 0.5```. That's why the denominator is longer that usual.



### Calculation Finished.



Well... I calculated the results for the top 5 spam words that spam messages include. The probability of the message being a spam when a message includes this word is...



```

"[국제발신]" : 0.538144824492084

"무료거부" : 0.8506066969512376

"lD" : 0.949720339750663

"(ê´‘ê³ )" : 0.6678162925754492

"PW" : 0.8629433434994991

```



Pretty interesting huh? You can see that not all the top rank words have high probabilities.



### To Be Continued...



Well, I guess this is it for now. I do not have enough data to train the model I wanted to make. Not only there is no Korean Spam message dataset in the Internet, but also the dataset that exists is quite different from what I am looking for. The dataset I used for analysis is extracted from **MY PHONE**, so I guess I have to wait again to get some spam messages stacked. Haha



If you have spam messages dataset in Korean, feel free to open an issue or email me through swimingtw@gmail.com. If there is something wrong with my article, feel free to open an issue and correct me. As I mentioned above, I am new to Machine Learning and I am in a process of learning new things. Thank you for reading.