Run final classifier on all unlabeled tweets

[aronwc]

Using best classifier from #2, classify remaining tweets and report a number of statistics, such as:

• number of positive/negative/other
• fraction of positive/negative/other per month
• fraction of positive by location (where available)

[aronwc]

• classify all organic tweets
• add new column pr_pos (probability of positive)
  • clf.predict_proba(X)[1]
• vec.transform(new_tweets)

Then, for each month, we can compute:

• % tweets classified as 1
• sum of probabilities of class 1
• compare to total # tweet

[ElaineResende]

All tweets = [30025, 36995, 46271, 62883, 65971, 90870, 110735, 84692, 96447, 94180, 114894, 158669] .Total: 992632 Positive tweets = [8493, 10135, 13116, 17242, 18652, 22375, 28034, 24310, 23276, 26574, 31461, 42163] .Total: 265831 Negative tweets = [21532, 26860, 33155, 45641, 47319, 68495, 82701, 60382, 73171, 67606, 83433, 116506] .Total: 726801

26.78% of tweets were classified as positive 73.22% of tweets were classified as negative/neutral

[aronwc]

This looks good! Could you please do an additional version where you only count each user once? That is, if a user appears more than once, only consider their first tweet.

On Jun 4, 2015, at 2:07 PM, ElaineResende notifications@github.com wrote:

All tweets = [30025, 36995, 46271, 62883, 65971, 90870, 110735, 84692, 96447, 94180, 114894, 158669] .Total: 992632 Positive tweets = [8493, 10135, 13116, 17242, 18652, 22375, 28034, 24310, 23276, 26574, 31461, 42163] .Total: 265831 Negative tweets = [21532, 26860, 33155, 45641, 47319, 68495, 82701, 60382, 73171, 67606, 83433, 116506] .Total: 726801

— Reply to this email directly or view it on GitHub.

[ElaineResende]

Considering one tweet per user

Below is correct!!

[ElaineResende]

Not considering retweets

[aronwc]

Great -- it looks like the "spike" switched from Apr'13 to Mar'13. Is this due to the filtering, or was there a problem in the original graph? Perhaps re-generate the original graph?

[ElaineResende]

I think it is due to the filtering, but I will double check.

[aronwc]

I see -- the March spike still exists in the first plot. Perhaps the April spike was due to the same users tweeting many times in April.

[ElaineResende]

Sorry I just saw your last comment now. Yes, I double checked and I believe the same.

[aronwc]

Couple last things

• Can you please add plots for the total number of tweets (nonorganic + organic) as well as the number of nonorganic tweets? No need to run the classifier on these
• Please use 'o' linestyle (e.g., 'bo-') so we can see the dots as well as the lines.

Thanks!

[aronwc]

• the new graph:
  • plot % change (nt - n{t-1}) / n_{t-1}
  • lines for : all tweets, organic, negative, positive
• We'll look for places where "all" increases, but "positive" decreases, or vice versa
• Figure out why the last point does not go up much for the "all" tweets, but it does for the other categories
• Figure out what people are talking about in Mar 2013

[aronwc]

To deal with users who tweet more than once:

• For each month, each user contributes 1 if at least one of their tweets is positive, otherwise 0

So, the y-axis is number of users who have tweeted at least one positive tweet this month.

[ElaineResende]

It follows what I have done:

1) First, I got 574318 tweets from users who tweeted more than once, they represent 57.8% of total tweets. 2) After that I have created a dictionary with all of those users as key and the class label and posted time as the values. The size of this dictionary is 113340 users (considering RT's). 3) Then I got just the keys which have positive class as value for the users. From that, I got 35145 tweets that were positively classified. 4) At the end I counted the number of tweets by month

The graph of those tweets is shown below.

Not considering RT's

[ElaineResende]

For March 2013 I have done n-grams

bigrams = [(('?', '?'), 30949), (('http', ':'), 23120), (('RT', '@'), 17071), (('electronic', 'cigarette'), 5975), (('...', 'http'), 5903), (('!', '!'), 3918), (('an', 'electronic'), 3189), (('?', 'http'), 3179), (('green', 'smoke'), 2848), (('electronic', 'cigarettes'), 2809)]

trigrams = [(('?', '?', '?'), 24862), (('...', 'http', ':'), 5898), (('?', 'http', ':'), 3169), (('an', 'electronic', 'cigarette'), 2957), (('smoking', 'an', 'electronic'), 2171), (('electronic', 'cigarette', '?'), 2039), ((':', 'http', ':'), 1952), (('!', '!', '!'), 1877), (('&', 'amp', ';'), 1747), (('caught', 'smoking', 'an'), 1727)]

4-grams = [(('?', '?', '?', '?'), 21847), (('smoking', 'an', 'electronic', 'cigarette'), 2124), (('an', 'electronic', 'cigarette', '?'), 1810), (('caught', 'smoking', 'an', 'electronic'), 1713), (('Onew', 'caught', 'smoking', 'an'), 1695), (('cigarette', '?', 'http', ':'), 1609), (('electronic', 'cigarette', '?', 'http'), 1606), ((':', 'Onew', 'caught', 'smoking'), 1440), (('allkpop', ':', 'Onew', 'caught'), 1379), (('@', 'allkpop', ':', 'Onew'), 1346)]

People talked a lot about Onew being caught vaping.

Also, I have generated a word cloud on the internet.

[ElaineResende]

September has the greatest number of retweets which is 22196 .

Bigrams = [(('?', '?'), 62409), (('http', ':'), 35810), (('RT', '@'), 23660), (('!', '!'), 11628), (('#', 'Ecigs'), 6395), (('.', '#'), 5641), (('quit', 'smoking'), 4938), (('?', 'http'), 4912), (('...', 'http'), 4784), (('amp', ';'), 4608)]

Trigrams = [(('?', '?', '?'), 47450), (('!', '!', '!'), 5684), (('?', 'http', ':'), 4908), (('...', 'http', ':'), 4733), (('&', 'amp', ';'), 4604), (('to', 'quit', 'smoking'), 3061), (('.', '#', 'Ecigs'), 2926), (('the', '#', 'FDA'), 2621), (('quit', 'smoking', '.'), 2445), (('.', 'http', ':'), 2428)]

4grams = [(('?', '?', '?', '?'), 40319), (('smoking', '.', '#', 'Ecigs'), 2341), (('the', '#', 'FDA', 'to'), 2137), (('to', 'quit', 'smoking', '.'), 2130), (('#', 'FDA', 'to', 'limit'), 2128), (('millions', 'to', 'quit', 'smoking'), 2119), (('a', 'product', 'that', 'is'), 2117), (('that', 'is', 'getting', 'millions'), 2115), (('is', 'getting', 'millions', 'to'), 2115), (('product', 'that', 'is', 'getting'), 2115)]

[aronwc]

Great work, thanks!

Based on your ngrams, the September spike appears to be strongly influenced by a push by states Attorneys General to encourage the FDA to regulate ecigarettes. See this story: http://articles.latimes.com/2013/sep/24/business/la-fi-mo-electronic-cigarette-attorney-general-20130924

It appears that people who are pro-ecigs launched a Twitter campaign to discourage this effort, by having people send messages to the Attorneys General saying, e.g.,

@INATTYGENERAL Why did you sign a letter asking the #FDA to limit a product that is getting millions to quit smoking. #Ecigs save lives!!!

I also found a discussion on Reddit where pro-ecig users are discussing how to best conduct their campaign: http://www.reddit.com/r/electronic_cigarette/comments/1n3k7r/fight_for_your_right_to_vape_daily_action_plan/

Are those n-grams only for those tweets classified as positive?

[aronwc]

If a user tweets 10 times in a month, does that contribute 10 or 1 to the y-axis in your graph?

Also, can you please regenerate the graph without limiting to users who have tweeted more than once? (i.e., include all users).

Finally, please create a separate graph showing the percentage of positive users per month. (i.e., of all the users who posted tweets about ecigs this month, what percentage posted at least one positive tweet?)

[ElaineResende]

I am sorry for n-grams I had considered all tweets. Now, I consider just positive tweets. So, we have total 15799 tweets for March.

bigrams = [(('?', '?'), 4336), (('RT', '@'), 1769), (('http', ':'), 1638), (('I', "'m"), 1300), (('!', '!'), 1225), (('electronic', 'cigarette'), 1071), (('e', 'cig'), 988), (('.', 'I'), 777), (('.', '#'), 729), (('electric', 'cigarette'), 703)]

trigrams = [(('?', '?', '?'), 2871), (('!', '!', '!'), 578), (('&', 'amp', ';'), 483), (('...', 'http', ':'), 414), (('good', 'for', 'me'), 354), (('too', 'good', 'for'), 354), (('because', 'i', 'used'), 353), (('electronic', 'cigarette', 'which'), 353), (('BlackFriday', 'is', 'too'), 353), (('i', 'used', 'electronic'), 353)]

4grams = [(('?', '?', '?', '?'), 2184), (('BlackFriday', '#', 'BlackFriday', 'is'), 353), (('good', 'for', 'me', 'because'), 353), (('BlackFriday', 'is', 'too', 'good'), 353), (('too', 'good', 'for', 'me'), 353), (('because', 'i', 'used', 'electronic'), 353), (('me', 'because', 'i', 'used'), 353), (('electronic', 'cigarette', 'which', 'i'), 353), (('for', 'me', 'because', 'i'), 353), (('which', 'i', 'missed', 'a'), 353)]

And for September we have 29222 total positive tweets

bigrams = [(('?', '?'), 9916), (('!', '!'), 6066), (('#', 'Ecigs'), 4373), (('quit', 'smoking'), 4223), (('RT', '@'), 3942), (('.', '#'), 3441), (('the', '#'), 3221), (('to', 'quit'), 2931), (('#', 'FDA'), 2803), (('I', "'m"), 2594)]

trigrams = [(('?', '?', '?'), 6491), (('!', '!', '!'), 2841), (('to', 'quit', 'smoking'), 2764), (('.', '#', 'Ecigs'), 2645), (('the', '#', 'FDA'), 2593), (('quit', 'smoking', '.'), 2308), (('#', 'FDA', 'to'), 2130), (('FDA', 'to', 'limit'), 2128), (('millions', 'to', 'quit'), 2125), (('a', 'product', 'that'), 2117)]

4grams = [(('?', '?', '?', '?'), 4579), (('#', 'FDA', 'to', 'limit'), 2128), (('the', '#', 'FDA', 'to'), 2128), (('millions', 'to', 'quit', 'smoking'), 2119), (('a', 'product', 'that', 'is'), 2116), (('that', 'is', 'getting', 'millions'), 2115), (('getting', 'millions', 'to', 'quit'), 2115), (('product', 'that', 'is', 'getting'), 2115), (('is', 'getting', 'millions', 'to'), 2115), (('a', 'letter', 'asking', 'the'), 2102)]

[ElaineResende]

Q: If a user tweets 10 times in a month, does that contribute 10 or 1 to the y-axis in your graph? A: It considers 10. I am going to fix it for 1.

Edit:

What I have done to fix it is: Before I was creating a dictionary like this: d[u] = label, date, where u is the 'username', label is the class, 'date is the posted time'. And then I created like this: d[u,label,date] As I understand I have one key for each user by month. So, after that I deleted the keys which had label equal to 0. At the end the size of the dictionary is 199382, and I sort by month and plot the graph. 

[ElaineResende]

[aronwc]

Beautiful! Can you please make the following small changes:

• change the title to "Twitter sentiment toward e-cigarettes, 10/2012-09/2013"
• change the ylabel to "Number of Users"
• Remove the "positive percentile..." legend
• change the color of both the "positive percentage" line and the ylabel "percentage of users who tweet..." to green

[ElaineResende]