[bids2015] when-to-post on social media

When-To-Post On Social NetworksOctober 9, 2015 @ Berkeley Institute for Data Science

*Nemanja Spasojevic, Zhisheng Li, *Adithya Rao, Prantik Bhattacharyya

● Klout is a social influence measurement tool.

● Users register on Klout.com and connect their social network accounts.

● Klout collects authorized/public information from connected networks.

● Klout derives influence scores and topics for users from collected data.

● Klout recommends:○ content to post○ times when to post.

What is Klout ?

What is Klout ?

What is Klout ?

● Maximize audience engagements:○ Reach friends ○ Better targeting by brands ○ Schedule campaign

● Personalized schedules vs. infographics

Motivation

?

...

Challenges● Data sparsity ● Lack of open data sets ● Unique audiences ● Specificity network dynamics

Problem Setting

For a user on a social network, find the best time to post a message in order to maximize the probability of receiving audience reactions.

Problem Setting

For a user on a social network, find the best time to post a message in order to maximize the probability of receiving audience reactions.

● Consider only: replies, retweets, favorites, likes, comments.● Weekly user behaviour cycle ● Observe only first 24hr of reactions● 15 min time bucket● Starting bucket is 00:00-00:15 Monday (relative to user’s timezone)

System Overview

Audience Behaviour

● Inherent delay ● Different networks have different engagement dynamics● 50% of first 24h reactions Twitter in 24 min while Facebook in 1h 42min● Estimate anticipated reactions over time

Post To Reaction Analysis: Network

● Reaction speed may depend on topic of the content

Post To Reaction Analysis: Topic

Twitter

● Depending on Actors in-degree they may react to posts faster or slower

Post To Reaction Analysis: Actor In-Degree

Twitter

Audience Behaviour - Network

Audience Behaviour - Location

Facebook

Twitter

Audience Behaviour - Location

Facebook

Twitter

Audience Behaviour - Location

Facebook

Twitter

Same City Audience Behaviour Correlation and Similarity

Audience Behaviour - Topics

Facebook

Twitter

Personalized Schedules

Personalized SchedulesAuthor Audience

● When do the users ai create posts?

● When does a specific audience member b

0 react to the posts

created by ai?

● What is the probability that b0

reacts to post in a certain time bucket t

k?

1st Degree Schedule 2nd Degree Schedule

Personalized Schedules

First-Degree Reaction Schedule

First-Degree Reaction Schedule

● When do the users ai create posts?

● When does a specific audience member b0

react to the posts created by ai?

Second-Degree Reaction Schedule

Second-Degree Reaction Schedule

● When do the users ai create posts?

● When does a specific audience member b0

react to the posts created by ai?

● What is the probability that b0

reacts to a post in a certain time bucket tk?

Weighted Schedules

● Unweighted:○ All audience members treated the

same

● Weighted○ Audience members weighted by past

engagement.○ A close friend may respond to your

posts more often than an acquaintance.

Personalized Schedules - Twitter Example

Personalized Schedules - EvaluationEvaluate on:● 56 days of unseen data● 0.5M active users

Baselines for a timezone :● Most Frequently Used (MFU)● Aggregate First-Degree (AFD)

Reaction gain of:● 17% on Facebook ● 4% on Twitter

Open Dataset

https://github.com/klout/opendatahttps://github.com/klout/opendata

● Anonymized Post-Reaction Timestamps ● 144+ million posts

○ Twitter 119M ○ Facebook 25M

● 1.1+ billion reactions○ Twitter 104M ○ Facebook 1B

● Anonymized user ID fingerprints map across networks

● Slightly perturbed timestamps

Future Work

https://github.com/klout/opendata● Personalized Post To Reaction Filter Functions ● More Sophisticated 2nd degree Model ● Topical Awareness ● Content Analysis (text, photo, video)● More Networks and Signals

Conclusion

https://github.com/klout/opendata● Reaction times are more than 4x faster on Twitter compared to other networks.● Audience behavior varies across different networks.● Users audiences across different cities exhibit different behavior patterns.● Using personalized schedules users can see reaction gain of up to :

○ 17% on Facebook ○ 4% on Twitter

● We hope open dataset can benefit future research on when-to-post problem.

KDD2015 http://arxiv.org/pdf/1506.02089v1

Q & A