GroupLens Research Projects
This page describes several of the active research projects within GroupLens.
GroupLens has a long history of research on recommender systems, starting with the original GroupLens USENET article recommender and the development of automatic collaborative filtering. That work continues today, as we run multiple recommendation services and use them to advance the art of recommendation.
MovieLens (http://movielens.org) is a web site that helps people find movies to watch. It has hundreds of thousands of registered users. We conduct online field experiments in MovieLens in the areas of automated content recommendation, recommendation interfaces, tagging-based recommenders and interfaces, member-maintained databases, and intelligent user interface design.
LensKit is an open source toolkit for building, researching, and studying recommender systems. It is intended to support reproducible research on recommender systems, and provide recommendation technology for integration into research or production systems. More detailed information and documentation are available on the project page and BitBucket.
The rating interfaces project is investigating different interfaces for rating movies. We are currently doing this by looking for interfaces that yield consistent ratings using re-rating experiments. We are also pursuing theoretic work that shows both how to measure ratings consistency, and explains why consistent ratings are probably the most efficient at gathering information about user preferences. Of course, we are also looking for rating interfaces that users will like.
Users annotate content online with tags - short, descriptive words or phrases. We are currently pursuing ideas organized around the idea of a tagging system as a garden: tag seeding, tag weeding, and tag architecture. For example, we are developing algorithmic techniques for identifying high-quality "seed" tags, and will be conducting user studies to determine which algorithms and interfaces have the best potential for increasing high-value tag applications across the system.
The BookLens project aims to be a book recommendation service. Similar to MovieLens, we hope that BookLens will help people find books to read. What makes BookLens different is that we aim to be a backend service for many different book communities. We currently are working with the Saint Paul Public Library as a recommender service and are looking to expand to libraries throughout the Twin Cities.
GroupLens has several ongoing research projects related to understanding and improving Wikipedia. We have employed data mining and analysis to better understand the differences in contribution value between Wikipedia contributors and insight into Wikipedia's gender imbalance. We have also worked to understand the health of English Wikipedia's user community and how it has changed over time.
SuggestBot (http://en.wikipedia.org/wiki/User:SuggestBot) does Intelligent Task Routing by matching Wikipedia articles in need of improvement with contributors who have shown interest in similar articles. It currently runs in four languages: English, Portuguese, Swedish, and Norwegian. While running as a service to the Wikipedia community it is also used as a platform for live experiments as we work on improving its performance.
This research focuses on understanding the differences in Wikipedia user behavior over time across different languages. We are working to understand the patterns of migration of users and content between different language versions of Wikipedia. We are also working to build predictive models of language proficiency.
Cyclopath is a geowiki: an editable map where anyone can share notes about roads and trails, enter tags about special locations, and fix map problems - like missing trails. Hundreds of Twin Cities cyclists are already doing this, making Cyclopath the most comprehensive and up-to-date bicycle information resource in the world.
Cycloplan is Cyclopath... for planners. We've added tools for traffic planners, engineers, and analysts, so they can analyze bicycle connectivity by creating and processing models. Planners can analyze how users currently are routed around the network, and they can model hypothetical infrastructure to determine how new (or even removed) links affect routing. Cycloplan is also access-controlled, so planners can create and maintain their own private copies of the map. And Cycloplan integrates with third-party GIS tools, making it easy to export and import data to and from the industry-standard Shapefile format, which allows users to edit data in their favorite third-party GIS applications.
The aim of this project is to make Cyclopath available to users on the go and to make use of location and sensing data on mobile devices to improve the map and the user experience. We are currently working on using GPS cycling tracks to improve map data (such as missing streets) and on using location to elicit context-aware user contributions such as landmarks. The Android app is available for download at Google Play.
Cyclopath in Greater MN
On a grant from the Minnesota Department of Transportation, GroupLens is expanding Cyclopath, which is currently limited to the 7-county metro area around Minneapolis and St. Paul, to the entire state of Minnesota. Currently in progress, this project has three main aspects: (a) Merging all the state's road and trail data and connecting it appropriately with the existing metro area data within Cyclopath; (b) Scaling and tuning the route-planner algorithm so that it is able to serve long routes (such as from Rochester to International Falls) in quick time; and (c) Improving the user interface (the Flash client that runs within the browser) to make it easier for new users to get adjusted to and start using and contributing to Cyclopath. The Cyclopath team collaborates chiefly with Greta Alquist from MN/DoT for this project.
We are developing and studying platforms that support citizen science, crowdsourcing, and volunteered geographic information. These projects largely are concerned with processing the submissions of simple geographic data (e.g., GPS locations or photos) by on-location volunteers from mobile devices.
CitizenSense is a platform targeted at groups such as local municipalities or non-profits. It supports a wide array of civic-minded public engagement opportunities by using the sensors in smartphones (e.g., GPS, camera, microphone) to crowdsource location-specific data. Organizations define campaigns, or data collection goals, that anyone with a smartphone can participate in. Technologically, we are still building out the infrastructure to realize our research goals, both in the mobile app, and in the back-end infrastructure. We're interested in a number of research areas within this system, such as motivating participation, mechanisms for ensuring or judging data quality, and understanding campaign organization workflows. As an early trial, we've been collaborating with the Humphrey School of Public Affairs to build a bicycle and pedestrian traffic counting campaign.
River Watch (http://riverwatch.umn.edu) is an established community of volunteers monitoring water quality on the Red River of the North and its tributaries. We are working with River Watch to streamline their data collection and quality assurance workflow, in part by introducing a mobile app to facilitate field data entry. We are also developing new apps to allow untrained volunteers not currently affiliated with River Watch to easily submit photos and observations during periodic flood events in the region. The new applications are based on the modular wq framework (http://wq.io), a platform for custom citizen science and volunteered geographic information apps we have made available for general use.
Questions and Answers
This research focuses on understanding patterns of behavior and content generation in existing online communities such as Yahoo Answers and Stack Overflow. We are currently building a new question and answer site to support online field studies. We expect this research infrastructure will allow us to address research questions relating to intelligent task routing and question recommendation algorithms and interfaces.
UMarvel is a joint project with the academic health center // Fairview hospital. This group builds interventions for physicians, medical specialists and patients that are aimed at improving patient outcomes and increasing provider efficiency. Our first project was an intervention technique to bridge communication gaps between physicians and radiologists by developing a system that annotates radiology reports to provide useful reference information to the receiving physician (and possibly for the patient). This project was spun off as a startup company which is currently running a pilot study at Mayo. Our current project is looking at ways to organize information around a particular disease or medical trial (such as an experimental cancer treatment), primarily for the benefit of patients.
Gaming and Social Psychology (WoWLens)
This is a joint project with Carnegie Mellon, the University of Pittsburgh and the University of Minnesota. We aim to study theories of social psychology, primarily small group dynamics, in online games including World of Warcraft and League of Legends. We have explored the impact of various rating dimensions in group selection, and we have compared social and skill-based factors in the success of small group cooperation challenges. Most recently we have been examining a large dataset of competitive match data and the impact of social feedback options on player behavior.
We are studying the behavior of users in social networking sites like Twitter and Pinterest with the goals of better understanding the pattern of interactions and developing improved methods for personalization. For example, we have developed methods for distinguishing informational and conversational messages in Twitter, and we have performed a statistical study of Pinterest to understand who participates and what types of content are shared.
Precision crowdsourcing is an emerging project supported by Google in which we will be exploring personalized and context-sensitive ways to more effectively turn crowdsourced data consumers into contributors.
We gratefully acknowledge the support of the National Science Foundation, under grants IIS 11-11201, IIS 10-17697, IIS 09-68483, IIS 09-64695, IIS 08-08692, IIS 08-12148, IIS 07-29344, IIS 05-34692, and IIS 05-34939.