DevsNearMe
•Download as PPTX, PDF•
0 likes•1,887 views
This document discusses a model for predicting attendance at events based on multiple data sources and factors. It considers using absolute numbers from check-ins, weights from context like weather and paid events, user past behavior patterns, and classifications based on age, gender and other user attributes. The goal is to separate true signals from noise and provide a probability distribution of potential attendance outcomes.
Convert to study guideBETA
Transform any presentation into a summarized study guide, highlighting the most important points and key insights.
DevsNearMe
- 1. DISCOVER REAL TIME KNOWLEDGE CLUSTERS DevsNearMe
- 2. Priorities ï‚— Separate signal from noise ï‚— Can we at least predict better than others ï‚— Ideally a probability distribution model giving us ideas for best-case, expected value or worst-case behavior
- 3. Model/Algorithm  Prediction based on  Absolute number + Prediction (Context –based information + Learning from User past behavior)
- 4. Absolute numbers ï‚— Foursquare check-ins are fairly reliable, as are MTA and TSA swipes ï‚— This just gets added to the prediction, no weighing applied currently but may be modify if there is a trend of fake data being generated
- 5. Context based Prediction  Context based – Use Decision Tree Learning to generate weights to apply to event rsvp counts for eventbrite, meetup, facebook.  E.g A meetup event rsvp has a higher weight if it is a paid event, has free giveaways and if the weather is nice  Weights are in range 0-1 and we multiply each event rsvp count by their weight and divide by 3 to get the weighted average rsvp count.  Events of similar nature in a nearby radius will downgrade the potential attendance
- 6. User Learning based prediction ï‚— A persons likelihood of attending an event can be modelled in a Bayesian manner ï‚— Past event attendance/rsvp ratio , history of attending a series of events of a particular nature ï‚— Item based classification is another factor e.g if person a,b,c,d attend events X and Y and we know that b,c, and d are attending event Z, there is a higher chance for a to attend event Z
- 7. Age based classification ï‚— ï‚— Sharing peaks at teenage, early adulthood and then falls down ï‚— Influence of social data needs inversely weighing to infer total count of people at an event
- 8. Gender based classification ï‚— Social sharing can weigh in gender for better classification
- 9. Miscelleneous ï‚— Chart data sources : appdata, beevolve, appdata, quora, statista