2. Exploration of data sets
In this part, we use SQL to gather information from the dataset. Given spend_usd and some other variables that we are going to engineer, we will try to answer what makes an influential political campaign. For our purpose, we define influential as a single campaign with a large total number of impressions.
To investigate what contributes most to the impressions of a campaign, we are going to create plots for the following variables to see how they related to the impressions. These variables have been grouped into different categories for an easier understanding. Some of them could be directly extracted from the tables, others need to be engineered.
In summary, we have found that,
• Targeting the states where the most money was spent was hugely related to impressions. This was likely due to the races in these states being especially competitive.
• Age targeting made for campaigns with more impressions (the non targeted by age campaigns did worst), but targeting young people did not seem to work.
• Advertiser data was not useful, poorly correlated to impressions, so we decided not to use these features.
• BLACK WEEKs were a good predictor, but did anomalously high with 7 black weeks.
• Also campaigns using video and image ads did better than text ads.Lastly the close a campaign was to the election the better it did.
If you are interested in knowing how we engineer the features and which features we feed into prediction models, click here.
To investigate what contributes most to the impressions of a campaign, we are going to create plots for the following variables to see how they related to the impressions. These variables have been grouped into different categories for an easier understanding. Some of them could be directly extracted from the tables, others need to be engineered.
In summary, we have found that,
• Targeting the states where the most money was spent was hugely related to impressions. This was likely due to the races in these states being especially competitive.
• Age targeting made for campaigns with more impressions (the non targeted by age campaigns did worst), but targeting young people did not seem to work.
• Advertiser data was not useful, poorly correlated to impressions, so we decided not to use these features.
• BLACK WEEKs were a good predictor, but did anomalously high with 7 black weeks.
• Also campaigns using video and image ads did better than text ads.Lastly the close a campaign was to the election the better it did.
3. Predictions based on explorations
After we have engineered several features in the previous part, we establish a logistic regression model using BigQuery to make predictions about impressions of a political campaign given these features. We randomly separate the dataset into 3 sets using a hash function, which are training set (80%), validation set (10%) and testing set (10%). Our prediction results on test set show that the model has a 40.0% precision score, 38.9% recall score, 37.4% accuracy and 36.3% f1 score.
4. Conclusions
What have we learned?
We learned that in this political ad dataset, the advertiser launching the campaign doesn't have much to do with how many people their ad campaign touches. The number of ads, total money spent, and average ad cost of an advertiser do not let us know much at all about how influential their campaign will be.
We've learned that the duration, money spent, and number of ads in a campaign lead to better results. The timing of a campaign can also influence whether or not people see it.
Additionally, we have learned that careful targeting of an ad campaign will lead to better results.We have also learned that while more difficult to work with, some denormalized databases can be used to draw insight from.Broader targeting, less return (everyone in age, gender and young people hard to reach)
What conclusions we can make or unable to make about the data set?
• Playing it safe doesn't pay: Ads not targeted at a specific group made less impressions. We saw this with gender and age.
• The youth isn't listening: We found ads targeted at young voters to have low amounts of impressions.
• People hate reading: Non text forms of advertisements were much more effective at reaching voters.
• Advertiser doesn't tell us much: this may have to do with the dataset only containing advertiser information from a single election cycle.
What is obvious, and what did we not expect to see?
It seemed obvious that targeted campaigns would do better, and that more ads can get you more views. I thought we were going to see the advertisers running a campaign have more influence, I figured the DNCCC and similar such groups that run a lot of ads, spending a ton of money having an edge over others in the amount of impressions created.
With more time we would have better gone over geo features. The irregular form of the data made it hard, and while we got close, came up short in the end.
We learned that in this political ad dataset, the advertiser launching the campaign doesn't have much to do with how many people their ad campaign touches. The number of ads, total money spent, and average ad cost of an advertiser do not let us know much at all about how influential their campaign will be.
We've learned that the duration, money spent, and number of ads in a campaign lead to better results. The timing of a campaign can also influence whether or not people see it.
Additionally, we have learned that careful targeting of an ad campaign will lead to better results.We have also learned that while more difficult to work with, some denormalized databases can be used to draw insight from.Broader targeting, less return (everyone in age, gender and young people hard to reach)
What conclusions we can make or unable to make about the data set?
• Playing it safe doesn't pay: Ads not targeted at a specific group made less impressions. We saw this with gender and age.
• The youth isn't listening: We found ads targeted at young voters to have low amounts of impressions.
• People hate reading: Non text forms of advertisements were much more effective at reaching voters.
• Advertiser doesn't tell us much: this may have to do with the dataset only containing advertiser information from a single election cycle.
What is obvious, and what did we not expect to see?
It seemed obvious that targeted campaigns would do better, and that more ads can get you more views. I thought we were going to see the advertisers running a campaign have more influence, I figured the DNCCC and similar such groups that run a lot of ads, spending a ton of money having an edge over others in the amount of impressions created.
With more time we would have better gone over geo features. The irregular form of the data made it hard, and while we got close, came up short in the end.