1 of 30

facebook.tracking.exposed

Final Presentations

2 of 30

User Uniqueness: How unique is my total post consumption compare to my friends?
Post Uniqueness: Which posts can be considered to have similar content?
Investigating Users’ Behavior: What correlation can we find?
User Emotional Reaction over time
Cross Bubble Surprise
Content Classification: can posts be classified under larger topics and how do these topics evolve over time?
Exposure and diversity: how often the algorithm change this? how many people are really appearing in your timeline?

3 of 30

User Uniqueness�

We can use Jaccard Similarity to measure uniqueness between any two sets

For any two users A and B, we can compare how unique they are across postIds or sources by using the complement of the similarity 1 - J(A, B)

4 of 30

User Uniqueness

Build a similarity matrix using Jaccard

Reduce to a single value for each user

Plot the runtime complexity

5 of 30

Posts Uniqueness

Data Exploration

The same post can have several source.

Scenario: Kialo posts on a user timeline and the user share it with their friend. The same post will appear as nature=organic and source= user_name even thou it is a reshare of Kialo’post.

6 of 30

Posts Uniqueness

Data Exploration

Looking at some post from business (The New York Time), the concatenate text can be slightly different, due to AB testing, but the posts are actually the same. We want to consider them as the same content. Even if me and my friend see a different variation of the post, we can consider we consumed the same content.

7 of 30

Posts Uniqueness

Measuring Post Similarity

8 of 30

Posts Uniqueness

Measuring Post Similarity

9 of 30

Investigating Users’ Behavior

Users are treated similarly by Facebook as long as they have a similar behavior on the platform

At the end, we could cluster users into:

frequent users (No. of Timelines per Day)
engaged users (Timeline length)

10 of 30

Users’ Emotional Reactions Over Time

14/03

25/03

11 of 30

Example: 14th of March - Real world event

12 of 30

Example: 14th of March - Random Event

13 of 30

26-27/02

14 of 30

26-27/02 ?

15 of 30

Anger Levels over time

16 of 30

Next Steps

Languages of the posts

Anger increase in the second part of the year

17 of 30

Cross bubble surprise

What news are important for Facebook?

Main idea is to look at number of impressions containing certain keyword. If a word “spikes” on a certain day, then something happened and we can visualize and analyze it.

To help with visual analysis, I’ve created a simple tool.

18 of 30

Cross bubble surprise

What news are important for Facebook?

At the moment, this is only an awareness tool. In future, there are multiple ways how to improve visualization and also analyze actual algorithm:

Are there different types of news? Are there ways to see organic and manufactured growth?
Right now I’ve looked at 1-day lag. It should be easy to add other summarization functions to help see the patterns better.
What news are actually seen by majority of users?
Should be interesting to look at ordering of the posts in timelines.
Cross-country leaks are not easy to explore since we don’t have country of residence for a user. Can be inferred.
Compare resulting trends with, e.g., Google Trends to see what news “make it” on Facebook.

Can also be a way to see what kind of biases are present in data.

Summarize by topic (e.g., LDA)/sentiment/etc., not by token.

19 of 30

Content Classification

20 of 30

Content Classification

Political Posts

21 of 30

Content Classification

Other

22 of 30

Content Classification

Sponsored Facebook Post Topics (Oct. - Dec. 2018)

23 of 30

Content Classification

Sponsored Facebook Post Topics (Oct. - Dec. 2018)

“Other” Ad

(e.g. Events)

Political Ad

Sales Ad

24 of 30

Exposure and diversity