Assignment 2: Analyzing Bluesky

Due Date: Monday 2/16

Overview

In this assignment, you’ll analyze several aspects of Bluesky. You’ll construct reverse-chronological feeds. You’ll make simple follow recommendations. And you’ll examine homophily — the tendency for people to associate with similar others — in Bluesky replies. You’ll collect data using Bluesky’s public API, analyze what senators see in their feeds, and measure whether reply patterns exhibit gender homophily.

The assignment has two parts:

Background

Bluesky is a social media platform built on the open AT Protocol. Unlike Twitter/X, Bluesky’s API is publicly accessible without authentication for read-only operations, making it valuable for studying social network dynamics.

We focus on U.S. Senators with verified Bluesky accounts. As of January 2026, 42 senators have verified accounts: 41 Democrats and 1 Independent (Bernie Sanders). No Republican senators have verified Bluesky accounts. This partisan skew is itself a phenomenon worth reflecting on (see Part III).

Provided Materials

Download all materials from the course website:

The name files contain historical Social Security Administration baby name registration data in tab-separated format (name\tyear\tcount). These are useful for inferring gender from first names based on historical naming patterns.

Part I: Feed Analysis

When a senator opens Bluesky, they see a reverse-chronological feed of posts from accounts they follow. This feed represents their information environment on the platform.

I.1 Data Collection

Write code to collect the feed for each senator:

  1. For each senator, retrieve the list of accounts they follow using getFollows
  2. For each account in that follow list, fetch that account’s posts from the last 24 hours using getAuthorFeed
  3. Combine these posts into a single reverse-chronological feed for each senator
  4. Save the data as JSON (one file per senator, or one combined file)

Note: Look ahead at Section I.3. You’ll need to track both the follow relationships and the posts (for computing Jaccard similarities). Consider saving follow data and feed data separately.

API endpoints (see bluesky_helpers.py for wrapper functions):

For API documentation see: Bluesky API documentation

Rate limits: Add time.sleep(0.1) between requests. The public API allows ~3,000 requests per 5 minutes.

I.2 “Senators You May Know” Recommendations

Build a simple recommendation system to suggest which senators each senator should follow, based on the follow graph among senators. Use a triangle-counting approach: for each senator X who doesn’t follow senator Y, count how many senators that X does follow also follow Y. This is the “mutual follows” count — the more senators in X’s network who follow Y, the stronger the recommendation.

recommendation_score(A, B) = |{C : A follows C and C follows B}|

This is analogous to main feature in “people you may know” recommendations on various social networks.

Task

For each senator:

  1. Identify which other senators they do not currently follow
  2. For each non-followed senator, compute the recommendation score (number of followed senators who follow them)
  3. Report the top 3 recommendations with their scores

If a senator already follows all other senators, note this in your output. If a senator follows no other senators, note this in your output.

Output

Generate a table in your report (programmatically, not by hand) with the following format:

Senator Recommendation 1 Recommendation 2 Recommendation 3
Bernie Sanders Chuck Schumer (8) Amy Klobuchar (7) …
Elizabeth Warren follows all senators    
… … … …

The number in parentheses is the recommendation score (mutual follows count).

Additionally, identify:

(Note: patterns in recommendations, network centrality, and disconnected senators are the subject of questions in Section I.4 below.)

I.3 Echo Chamber Analysis

Examine the degree to which senators’ feeds overlap using two complementary measures:

Follow Jaccard Similarity

For each pair of senators, compute the Jaccard similarity of the accounts they follow (all accounts, not just senators):

Jaccard(A, B) = |follows_A ∩ follows_B| / |follows_A ∪ follows_B|

This measures overlap in terms of who the senators follow, treating all followed accounts equally regardless of how active they are.

Post Jaccard Similarity

The follow Jaccard treats a dormant account the same as one posting 50 times per day, and doesn’t necessarily reflect similarities in their user experience. To capture similarity in what senators actually see, define and compute a suitable Jaccard coefficient over post in their feed in the last 24 hours.

Each of these measures (follow Jaccard, post Jaccard) can be interpreted as similarity measures between the senators. Computing these measures for all the senators in the dataset forms two similarity matrices.

Visualization and analysis

  1. Create two heatmaps: one for follow Jaccard similarity, one for post Jaccard similarity. You may present these as two separate figures or side-by-side in a single figure for easy comparison.

  2. Sort the rows and columns so that similar senators are adjacent. Use hierarchical clustering to determine the ordering (see scipy.cluster.hierarchy.linkage and dendrogram), or (if you want to explore alternative visualization techniques) explore some other sorting algorithm. An unsorted heatmap with arbitrary row/column order is much harder to interpret — clusters appear scattered rather than as visible blocks along the diagonal. Use the same ordering for both heatmaps to enable direct comparison.

  3. Compare and contrast the two matrices: What do they show? Interpret any clustering patterns you see. Where do the two matrices differ? Senators with similar follows but different posts might follow the same accounts but experience very different information environments based on posting frequency.

Suggested questions to consider (you don’t need to answer all or any of these, but they may guide your interpretation): Do senators from the same state follow similar accounts? Do female senators follow systematically different accounts than male senators? Are there identifiable “information bubbles”? How does weighting by post volume change the similarity structure?

I.4 Reflection Questions

Include brief responses to these questions in your report:

  1. What does the overlap structure of senators’ feeds (from the Jaccard similarity analysis) suggest about shared versus divergent information within the Democratic caucus? Are there identifiable clusters or outliers?

  2. A senator following 50 high-volume accounts might see 500 posts/day; one following 500 accounts might see 5,000. What are the implications of this variation for how senators use the platform or stay informed about constituent concerns?

  3. Based on your “Senators You May Know” analysis, which senators appear most frequently as recommendations? What does this suggest about their position in the senate follow network? Are there senators who are notably disconnected from their colleagues on Bluesky?

Part II: Reply Homophily Analysis

In this section we will look at the posts by senators themselves, and who replies to these posts. We are going to specifically investigate homphily in the replies, meaning that women disproportionately reply to female senators, and men disproportionately reply to male senators. The key question here is how to think about “disproportionately”.

II.1 Data Collection

Write code to collect replies to senator posts. To ensure sufficient data for gender comparisons, collect reply data for at least 5 female senators and at least 5 male senators (the dataset contains 14 female and 28 male senators), uniformly at random. If you can, take the time to run your code long enough to collect data for all the senators.

For each senator in your sample:

  1. Fetch their posts from the last 7 days using getAuthorFeed
  2. For each post with replies, fetch the reply thread using getPostThread
  3. Extract replier information (handle, display name, timestamp) and post metadata (reply count, for understanding sampling)
  4. Save the data (in JSON) for each senator

Important: Save the post’s total replyCount field along with the replies you captured. This lets you determine what fraction of replies you collected for each post, which is essential for the timing analysis.

Important limitation: The API returns at most ~200 replies per post, focused on the earliest replies. For posts with 1,000+ replies, you will see only a fraction of the conversation. You must consider this limitation in your analysis.

II.2 Gender Inference

Make an inference about the gender of repliers from their display names. Using the SSA baby name data. For instance:

  1. Extract a likely first name from each display name
  2. Look up the name’s historical gender distribution
  3. Classify as female if >60% of registrations are female, male if >60% are male, otherwise unknown

Report what fraction of repliers you can classify and discuss the limitations of this approach. Feel free to try to improve on this method (query an LLM?), but it is not a required part of the assignment.

II.3 Homophily Measurement

Measure gender patterns in replies:

  1. Baseline: Across all senators, what fraction of classifiable repliers are female vs. male? Call these p_female and p_male = 1 - p_female.

  2. By senator gender:
    • For female senators, what fraction of their repliers are female? Call this obs_F.
    • For male senators, what fraction of their repliers are male? Call this obs_M.

  3. Homophily coefficient: Compute a measure of same-gender homophily for each senator gender:

    H_female = obs_F - p_female
H_male   = obs_M - p_male

    Interpretation:

    • H > 0 indicates homophily: same-gender repliers are over-represented relative to baseline
    • H < 0 indicates heterophily: same-gender repliers are under-represented
    • H = 0 indicates no preference: replier gender matches the overall baseline

    Example: If the baseline is 40% female (p_female = 0.40) and female senators receive 45% female repliers (obs_F = 0.45), then H_female = 0.45 - 0.40 = +0.05, indicating mild homophily.

Create a visualization showing the gender breakdown of repliers for female vs. male senators. Is there evidence of homophily?

  1. Statistical significance: Assess whether the observed homophily is statistically significant or could plausibly arise by chance. There are many ways to structure this inquiry. Explain your approach. As a guiding principle, the null hypothesis is that replier gender is independent of senator gender (i.e., all senators draw from the same baseline distribution)

    Report whether your homophily findings are statistically significant and discuss what this means for your conclusions.

II.4 Reply Timing Analysis

Since the API preferentially returns early replies, investigate whether early and late replies differ.

Focus on posts with 50–200 total replies (where you likely captured most replies):

  1. Split replies into “early” (first 25% by timestamp) and “late” (last 25%)
  2. Compare the gender composition of early vs. late replies
  3. Look for other differences: reply length, engagement received (likes), or any other quantifiable patterns you observe

This analysis helps you understand the bias introduced by the API’s ~200-reply limit and reason about what you might be missing for high-engagement posts. What senators are often getting more than 200 replies per post? Given what this timing analysis shows, do you have any caveats you would add to your previous analysis? Are they minor or major caveats?

API limits frequently require analysis like this one to include caveats/hedges, so becoming comfortable with interpreting the limitations of the data you collected is very good practice.

II.5 Reflection Questions

Include brief responses to these questions in your report:

  1. What are the key limitations of inferring gender from first names?

  2. We focused on gender homophily. What other forms of homophily might be interesting to study on Bluesky (e.g., geographic, topical, temporal), and how would you operationalize them?

  3. For a senator whose posts routinely receive 1,000+ replies, roughly what fraction of replies are we capturing? What might be systematically different about the replies we miss? Based on your reply timing analysis, do early replies appear different from late replies? How should these limitations affect our confidence in the homophily estimates?

Part III: Platform biases

  1. Our dataset contains 42 senators: 41 Democrats and 1 Independent. No Republican senators have verified Bluesky accounts. This complete partisan skew raises important questions about what we can learn from this analysis. What does this tell us about the generalizability of any findings about “senators” or “political figures”? With zero Republicans, can we make any claims about partisan patterns?

  2. If we wanted to study partisan homophily (do Democrats engage more with Democrats?), what alternative approaches or data sources might we consider? Briefly research senator activity on other platforms and sketch out a plan for how you’d study partisan homophily.

  3. How might the partisan composition of Bluesky’s overall user base affect the homophily patterns we observe, even when studying only gender?

Deliverables

Submit the following:

  1. Report (PDF, 6 pages max) containing:
    • “Senators You May Know” recommendation table and discussion
    • Feed overlap analysis (both heatmaps)
    • Homophily results with visualization
    • Reply timing analysis
    • Responses to reflection questions in Parts I, II, and III

  2. Any code you wrote, either as python scripts or notebooks

Grading Rubric

Component Weight Criteria
Data Collection 15% Correct API usage, rate limiting, handles edge cases
Feed Analysis (Recommendations + similarity) 20% Correct algorithm, clear table/heatmaps, interpretation
Homophily Analysis 20% Sound methodology, appropriate statistics, visualization
Reply Timing Analysis 10% Thoughtful comparison, addresses API limitations
Written Questions 20% Depth of reflection, honest engagement with limitations
Code & Report Quality 15% Clean code, clear writing, effective figures

Technical Reference

See bluesky_helpers.py for:

Run python bluesky_helpers.py to test the API functions and see example usage.

For official API documentation, see: Bluesky API Reference

Academic Integrity

You may work in groups of up to 3 students. List all group members on your report and code files. You may use the provided helper code and consult API documentation. Do not share code with other groups. Describe your use of AI tools (AI is permitted).

Good luck!