Human decisions in Machine Learning

The results of any Machine Learning project are affected by a range of very human, not-so-autonomous decisions. Based on our experiences in both research projects and large-scale industry settings, Jenn Thom and I presented six types of those decisions at the AAAI Spring Symposium on the UX of ML. This includes your goals and the metrics you optimize for, implicit and explicit user feedback mechanisms and how you collect your training data. We also specifically call out a group that sometimes gets overlooked: the data editors and curators involved in selecting and annotating the data that machines learn from.

Read our position paper here, or check out the visual summary of my talk by the fabulous Chris Noessel:

Just published: functions of 😀 in messaging.

If you’re processing user-generated datasets, you better understand the language that people use to communicate with each other – and that includes emojis. This MobileHCI’16 paper, a collaboration with fabulous natural language processing ex-colleagues at Yahoo, Paloma de Juan  and Joel Tetreault, outlines the variety of meanings and sentiment that emojis are meant to express. Beyond the variety of meanings, different linguistic functions, and combinations with text, we also discuss some of the pragmatic Unicode support issues. Happy texting and happy processing!

Paper: Henriette Cramer, Paloma de Juan, Joel Tetreault. Sender-intended functions of emojis in US-based messaging. MobileHCI’16, Florence, Italy, Sept 2016.  Full pdf here

Autonomous things, humans, data & design [Dutch]

Last weekend, I authored a piece for the Dutch about the importance of the human element for systems that autonomously learn or adapt. Feedback loops between systems and the people around them, social expectations of their behavior, the ecosystems of systems organizations and different (non-)users around the autonomous ‘thing’, and the optimization targets they have – they all have human consequences. In the end, the more we take advantage of autonomous systems and large-scale data, the more pressing it becomes to consider all that matters on a human-scale as well. Data & people-focused design: both are crucial. Article in Dutch.newscientistscreengrab22jan2015

Jan-March: design studio at Stanford

During January to March 2016, I’ll be an industry Studio Instructor in Stanford’s CS 247: HCI design studio course, coordinated & taught by Stanford’s Michael Bernstein. My studio focuses on the theme of ‘autonomous things’, and asks students to explore either the social or physical context in which autonomous agents interact with people. Looking forward to creative student work! For those in the area, I’ll be on campus Monday and Friday afternoons.


Recently presented & published

  • Mohamed Kafsi, Henriette Cramer, Bart Thomee, Ayman Shamma. Neighborhood characteristics through social media. Full paper WWW’15. (14.1% acceptance rate, brutal!). We present a Geographic Hierarchical Model that helps decide whether social media content is not just local, but also locally descriptive. We validate and apply the model using flickr data and distinguish between neighborhood, city and country-level content, and discuss people’s reasoning in whether content is (not) locally descriptive. pdf
  • Katie O’Donnell, Henriette Cramer. People’s Perceptions of Personalized Ads, TargetAd workshop at WWW’15. Survey & interviews looking into which personalized ads are (not) appreciated, as well as people’s concerns surrounding advertising content. Life events, deals as well as (teen) considerations of matching style and what is inspirational rather than ‘in the way’. pdf
  • Henriette Cramer, Effects of ad quality & content-relevance on perceived content quality. Note CHI’15 (with honorable mention). Evaluating ads for quality in isolation of their publishing context has its limitations. This study shows that the distinction between ads and content is crucial in perceptions of site quality, going beyond quality of ads themselves. pdf
  • Henriette Cramer, Maia Jacobs, Couples’ Communication Channels: What, When & Why? Note CHI’15. Couple’s don’t stick to just one communication channels, they pick and choose. Household coordination, playful expression, relationship upkeep and enjoying the emotional reaction of the other to your messages all play a role. Channel choice is not necessarily an all-or-nothing game, and using multiple channels can add meaning. Pdf. Georgia Tech news here.

Especially great to have the work with past interns Mohamed, Maia and Katie out there. Thanks for working with us!

Upcoming talks

Two talks related to location & personalization coming up:

RecSys CrowdRec Workshop, Oct 6th, Foster City, CA
Invited keynote: A community-sourced view of people’s surroundings

Berkeley Institute of Design (BiD) lunch seminar, Oct 14th.
Personalizing, while socializing.

Personalizing also means understanding your limits – people’s experiences are a whole lot richer than just their clicks and updates. Not all aspects of people’s experiences are captured by systems’ representations, and models based on aggregates of user-generated data from social media present additional challenges. Using examples from research projects at Yahoo Labs, I’ll address both opportunities and challenges in creating personalized location-based experiences; and in using mobile data to characterize local surroundings.

There? Say hi!