Personalizing Campaigns Though Machine Learning
Machine learning, the field upon which the vast majority of artificial intelligence systems depend on, has tremendous potential to do good if harnessed correctly. Machine learning can allow for more personal campaigns–and this need not be a bad thing. While it may be common to vilify the algorithm, when properly used, algorithms can allow for better timed phone calls, and conversations directly related to a voter’s interests, and hopefully, less robocalls in the middle of dinner.
Recently in the news
The firm then used algorithms, presumably some form of machine learning to target voters to personalize and target political advertisements. This is the bad side of machine learning for campaigns–data should be collected and used responsibly, and people should be given the right to opt-out.
For more details into the Cambridge Analytica situation, The Guardian put together a great summary.
The data, and the ethics of it
Campaigns have a lot of data. Addresses, phone numbers, household information, membership in various groups, party registration, and beyond. From this, central parties, and sometimes other groups use algorithms to predict the odds that a citizen will vote, donate, and if so, who they are likely to vote or donate to.
With the rise of social media, there is a wealth more data available, and it is not hard to collect. However, to do so ethically, becomes more difficult, especially as the companies owning the data are not heavily regulating usage of it, nor are there many laws (which may, in part, be due to the value of social media data to campaigns) or even ethical guidelines in the field.
Social media data is incredibly valuable, but as strange as it may seem, I would discourage campaigns from collecting it on a personal level. It is simply not viable at the moment to construct a way of getting the data in an ethical fashion, that is also feasible for a campaign. However, this isn’t to say that social media stats should be disregarded on the whole.
Facebook pages get to track certain demographic data, on the macro level about those who like the page–and that makes sense for campaigns to use. Using my home district of the MI-08 as an example, if a Democrat has a substantial amount of likes from the small town of Howell, it would be a positive sign for them in the election, whereas, if all their likes were from the stronghold of East Lansing, given the voter makeup, it would be discouraging news.
Issues with micro-targeting
There are, of course, issues when campaigns and politicians begin honing in and personalizing the message too much. In some ways, targeting the message on the micro level can be a good thing. However, it may also lead to an increased level of creating political bubbles, and less accountability in politics. A variety of others have explored this in some detail– here are some recommendations for reads depending on how long you have:
What we should be targeting
When I refer to personalizing politics, I don’t refer to changing stances when talking with different voters–this strikes me by and large as disingenous. But knowing in advance what a voter’s key issues are at a greater level of detail, plus knowing what time would be best to call a voter, based on their housing type, employment, age, etc, can be very powerful. It reduces the number of voicemails for the voter, and increases the odds of both the voter picking up, and the voter staying on the line.
We can use the data to predict things like if a voter is likely to be home when a volunteer is sent to their door, their preference for phone vs. an in person chat. We can predict in advance the number of people in the district, with a high level of precision, who would be willing to put up a yard sign for the candidate. Beyond that, we can also predict who is likely to volunteer–and even how to best make that pitch to them. Some of this builds on existing technologies–but some of this needs to be developed. We’ve got the data to probably do all of this (though we wouldn’t know until we were further in).
This just scratches the surface of the power of machine learning. So much has been done with it as of yet, ranging from the New York Times’ infamous needle, to various other projection softwares. Yet, there remains so much potential to reduce costs and improve staffing efficiencies of political campaigns that has not yet been harnessed.
How we can do it
Machine learning is becoming more accessible, to the point at which now, anyone who can code can use deep learning. However, to do machine learning well still takes a level of skill–as I am becoming acquainted with throughout my graduate-level machine learning course this semester. The number of people who have the talent to build incredibly detailed, highly accurate systems remains low. Plus, the median salary for qualified data scientists runs around $120,000– 3.2x more than your average field organizer, and 2x more than your average campaign manager. That is to say, most campaigns cannot afford a dedicated data scientist.
Fortunately, the average campaign probably doesn’t need a full-time machine learning person–although it wouldn’t hurt. It would make far more sense for the average state to have 2-3 people on their staff, to help build specific county models and maybe spend a week with the larger campaigns–but given the relatively low amount of supervision required for the algorithms, a dedicated machine learning specialist probably would not find themselves more than two weeks of machine learning specific work on an average Congressional campaign (though those two weeks could very well push thousands of votes).