A Faster Path to Tailor-Made Voting Advice

Many voters find it hard to identify the specific politicians closest to their own stance out of the multitude of candidates for a political office. That’s where the platform Smartvote comes in: on the basis of a questionnaire, it recommends candidates who best match voters’ own views on political issues.

However, to get the recommendations, voters have to fill out a lengthy questionnaire with over 70 questions. Although there is an abbreviated version with around 30 questions, it doesn’t deliver the same quality of advice. “We wanted to discover a method that gives voters the best possible advice at any point in time regardless of how many questions they have already answered,” explains Professor Abraham Bernstein from the Department of Informatics at UZH.

To accomplish that, Bernstein and his PhD candidate assistant Fynn Bachmann tested multiple methods under which questions are not asked in a fixed order, but in a flexible sequence based on previous answers. The best of the investigated methods delivers much better-quality recommendations on the basis of half the questions than the ones employing short, static questionnaires do.

This means that the recommended candidates match 70% with the ones suggested upon completion of the full long questionnaire. “The model learns from prior answers given,” Bernstein says, “and that way is able to make good recommendations on the basis of fewer questions.”

The challenge was in finding out which question at any given point in time contributes the most to clarifying the respondent’s political positioning.

“Positioning” is meant literally here because under the best method, a political map of sorts with a left-right axis and a liberal–conservative axis is generated. All of the candidates are plotted on the map on the basis of their answers to the questionnaire, resulting in distinct clustering: candidates from the Social Democratic Party of Switzerland and from the Green Party tend to cluster in the middle left quadrant of the map, whereas candidates from the Swiss People’s Party tend to be found in the lower right quadrant.

Zoom (JPG, 208 KB)

The illustration shows the dispersion of candidates on a “political map.”  The dashed-dotted line marks the boundary at which the probability of a “yes” answer to the question “Should Switzerland exit the Schengen Treaty with the EU and reinstate more identity checks at border crossings?” is the same as the probability of a “no” response. The boundary runs distinctly between party clusters: the candidates below it are mainly representatives of the Swiss People’s Party and the Federal Democratic Union, and those left and above it are mostly representatives of Switzerland’s other political parties.

Uncertainty aids clarification

So, for each question, it then can be computed how high the probability is that a person located on a specific point on the map will answer it in the affirmative. The less confident the prediction, the more the question helps to narrow down that person’s political positioning and the more likely it is that the algorithm selects it as the next one to ask. “If the probability of a person answering ‘yes’ to the question stands at around 90%, that contributes little to clarifying their positioning,” Bachmann explains.

If someone, for example, answers “no” to the question of whether they endorse the bilateral agreements between Switzerland and the European Union, it makes little sense to ask if that person is in favor of Switzerland joining the EU. The probability of a “no” response is extremely high. If, however, it is unclear how the person would answer, the actual response given accordingly helps to adjust their positioning.

You can see the adaptive questionaire at work inthis demonstration.

The study was conducted in collaboration with Smartvote. The association that operates the platform is very interested in its findings. Daniel Schwarz, one of the founders of Smartvote, says that he and his team are convinced of the optimization that is achievable by employing an adaptive questionnaire. “It enables us to provide the best possible voting advice even to people who do not want to answer all 70 questions,” he adds.

Smartvote therefore plans to adapt the existing questionnaire during the next revision of the platform. Schwarz says that it is not yet certain when the adaptive questionaries will go live, as some technical implementation and performance issues still need to be cleared up. It is important to Schwarz that the computations required to select the questions run as efficiently as possible because the Smartvote server’s computing capacity is limited.

Smartvote also intends to adapt the method used to plot candidates from left to right and from conservative to liberal on a political map by applying the same algorithm utilized for the two-dimensional model employed for the adaptative questionnaire in the UZH study. Schwarz considers the improvements suggested by the study a textbook example of how scientific findings make their way into an actual application.

The method developed by Bernstein and Bachmann is premised on the system already possessing a large number of responses as a basis for generating the two-dimensional map. Smartvote has that because all candidates must fill out the questionnaire first. But the algorithm for the adaptive questionnaire can also be utilized for other questionnaires where such responses are missing and no “map” exists at the start as a basis for adaptively determining the most sensible next questions.

The suggestions for the order of questions, though, thus wouldn’t be very meaningful for the first people to fill out the questionnaire. Bernstein and Bachmann therefore investigated in a second study how this “cold start” problem could be solved with the aid of artificial intelligence.

To do that, they let ChatGPT fill out the Smartvote questionnaire with the stipulation that it behave like a sympathizer of a specific party. The result was clear-cut here as well: when the answers generated by the AI chatbot were plotted on the political map, clusters again formed in line with the respective party views. The dispersion of the FDP and SP voters simulated by the AI chatbot more or less resembled that of the parties’ actual voters in the real world. The UZH researchers obtained that result using the standard version of ChatGPT without training the chatbot on specific data.

To the authors of the study, this means that generative AI can be used to improve the quality of adaptive questionnaires. An initial meaningful group of any number of questionnaire takers can be simulated with the responses generated by AI. The algorithm for choosing the adaptive questions then can use that as a selection basis for picking a more meaningful next question each time for the first real people to fill out the questionnaire. “There is still a disadvantage for the first ones to answer the questionnaire,” Bernstein explains, “but it is much smaller than it would be if no initial data at all existed.”

Both studies show that adaptive questionnaires and the deployment of AI can be utilized to design surveys that yield results comparable to that of conventional static questionnaires but with much less time expenditure for survey participants.