CrowdAct: Achieving High-Quality Crowdsourced Datasets in Mobile Activity Recognition
summary
CrowdAct deploys an Android app and backend that combine active learning + gamification with an inaccuracy detection via supervised ML to produce higher-quality activity-recognition training data. A crowdsourced run with 120 workers collects 6,549 labels across 19 activity classes, yields one extra class per worker on average, increases label volume and recognition accuracy in both user-dependent and user-independent models, raises engagement (USE scores, button presses, CTR), and reduces detected inaccurate labels.
AI-written summary · about
related talks
06mA Multi-Sensor Approach to Automatically Recognize Breaks and Work Activities of Knowledge ...
19mCrowdsourcing GO: Effect of Worker Situation on Mobile Crowdsourcing Performance
29mAnnotating data the right way
25mGotta catch'em all: recognizing sloppy work in crowdsourcing tasks
06mA Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition
15mAI as a Service: Quick and Scalable Machine Learning POCs with No-Code/Low-Code Tools
25mCrowdsourcing platform engineering
23mManaging Small Datasets Without Losing Your Mind