IBM Know Me: a Twitter Experiment



Engadget: I'm a neurotic. IBM told me so


What Type of Experiment?

Know me was an experimental project designed to test the concept of matching personality traits to events happening around you. The particular type of event we focused on was volunteering. The team used cognitive computing software to create a psycholinguistic analysis of the user, based off of their Twitter profile.

This experiment was conducted on web and native IOS platforms.




Hypothesis

We are predicting that by using natural language processing the team will be able to collect data written by users through the content they create and post on their social network(s). With this information we will be able to use our tech to create a psycholinguistic profile of the user and generate personalized recommendations based on their personality.

By doing a qualitative analysis with our target users, we can begin to paint a picture of the consumers behavior while interacting with machine learning or bot-like systems. These findings could inform future products with similar AI-like behavior.

How Did You Get User Data?

Before we began to experiment, the engineering team gained access to an API of a volunteering website. From there, they were able to pull in the categories, titles, and descriptions of the volunteer opportunities.

What we are aiming to do is to determine what personality you have through natural language processing. On average we need about 200 tweets. That equals 2,000 words plus or minus.


"create a psycholinguistic profile of the user and generate personalized recommendations based on their personality."







More About Psycholinguistic Analysis

The system outputs a graph that has 90+ personality traits and percentages that create your psycholinguistic profile. The visualizations are mostly to show how the system interpreted the data. Engineers use a JSON file to extract the data points in order to build our front end.


Starting With the Web Illusion

The experiment started with a proof of concept. Candidate interviews were done to determine the likelihood of how our potential users might interact with our prototype. After the users granted us access to their Twitter data, we were able to manually (a human posing as a computer) send volunteer recommendations to our candidates. The candidates thought they were interacting with a bot. We needed to prove that there was interest and figure out what the technical constraints we needed to overcome were before we built a more robust prototype.

Our Target User

One of the team's goals was to test the recommendation engine. Given that the team is recommending volunteer opportunities, they thought it would be best to focus on college students and recent grads. The thought was that these types of people in the 19 - 27 age group would be working on building a resume or completing some type of college credit through volunteering. Thus, more likely to participate in our experiment.




Building the Reality: Web Hybrid

We want to be able to gather the data we need quickly and with the lowest friction possible for our testers. What we need is access to their Twitter account. From there, we can begin to send our users recommendations based off of their profile (this is where the AI magic comes in). Then, we simply have the users review the matches sent to them and rate them accordingly. This started with a thumbs up/thumbs down approach.


Data Ingestion

This is the stage where we collect the data to anonymously compare results to other user's responses with similar profile traits. Then, we send them back to the recommendation engine when the user gives feedback.

Psycholinguistic Profile

This is where the psycholinguistic analysis takes place. The user doesn’t see their psycholinguistic profile. We store that behind the scenes. When the user makes recommendations, we compare that to that other similar profiles that are fed back into the recommendation engine.

Home Screen

The home screen is where all the activity takes place. It’s an area where the user can evaluate recommendations we send to them from the volunteer match API we run through our recommendation engine based off the users psycholinguistic profile.







Insights We Learned at This Stage

We were able to successfully test our recommendation engine. We found as more users tested the system that our recommendation engine seemed to deliver back satisfactory results. At this point the team felt that we could begin to narrow the results by learning more about the user's interest. This lead us to construct our last experiment.

Developing the Final Experiment

The team felt confident enough to begin discussing what the next experiment would be. We wanted to narrow down the results by learning more about our users. How do we begin to build an interactive experience with a machine? Below are some design considerations we followed.

  • Allow the user to pick volunteer categories for narrowing.


  • Simple interactive experience.


  • Connect the user to the content.


  • Engaging.


  • Light gamification.



"How do we begin to build an interactive experience with a machine?"




Planning the Final Concept: Hybrid App

After we had enough profile data and events to match against, it was time to narrow down selections. We created quick prototypes that would allow users to adjust their interest levels, thus giving us more insight into what matters most to them.


Interactive UX

After we had enough profile data and events to match against, it was time to narrow down selections. We created quick prototypes that would allow users to adjust their interest levels, thus giving us more insight into what matters most to them.

The user should have the ability to scale the nodes up and down with a limit of three nodes. Each node will represent a category and the size will determine the users interest in that category on scale of 0 - 100. 10% being the smallest size, with the largest size being 80%. When the user starts scaling nodes larger than others, it will force some nodes down to a smaller size. We are posing these actions as a “machine interaction”. These help to gather feedback from our users on how and why the machine is helping them make decisions on their category selections.






What Happened Next?

Unfortunately, we only began to build the interactive portion of our experiment when resourcing on the Watson Group was needed elsewhere and our experiment was never fully completed. However, we were able to to run our concepts by testers and gather impressions and feedback on the new experience we designed. This allowed us to make an informed decision and gather valuable insights to measure against our hypothesis.