DIGITAL WELLNESS THROUGH DATA

Data mining a decade of my own social media • 26,720 likes, 4,571 videos, 500+ connections • CMU data mining final

← Back to Case Studies

Why screen time isn't enough

We're told to care about screen time, but screen time is a terrible metric. Research has struggled to find consistent relationships between hours on screens and mental health outcomes; a 2023 systematic review in BMC Psychology concluded that "screen time" may no longer even be the appropriate thing to measure, because it says nothing about content or how people interact with it.

"Social network sites are not 'good' or 'bad'. Their mental health consequences critically depend on how these sites are used." (World Psychiatry, 2021)

The literature's best framework, passive scrolling versus active participation, still treats reading hateful comments, watching a funny cat video, and seeing an advertisement as equally relevant to well-being. Nobody is looking at the actual content. Your phone can tell you that you spent four hours on Twitter yesterday. It cannot tell you what emotional content you gravitated toward, whether your consumption shifted during stressful life events, or what your algorithm thinks you want.

Digital wellness tools give us calories without nutrition labels: we know how much we consumed, but not what, or how it's shaping us.

The dataset is me

This was my final project for CMU's 45-851 Data Mining, completed December 2025, and the question behind it had been nagging me for years: what if you could see your algorithm the way your algorithm sees you? I'm a typical Gen Z social media addict, but I'm also a computer scientist, and I've always wondered what the algorithm is actually doing to me. Not in theory. To me, specifically.

A decade of activity turns out to be a lot of evidence: 26,720 liked tweets, 4,571 saved YouTube videos, and 500+ LinkedIn connections, a digital paper trail I had never once read back.

The motivation comes from behavioral science: environmental modification. It's easier to not eat chocolate if you don't keep it in the house, but you can't curate an algorithm you can't observe. So the project's real goal was to make the invisible visible: who am I online, and does that person match who I think I am?

What the data revealed

My content diet (Twitter)

Zero-shot classification sorted 26,720 likes into ten clean categories: news, memes, VTubers, art, and friends. Whatever the platform's problems, my feed is finely tuned; it delivers exactly what I want. Emotion classification went a layer deeper, mapping where I find joy and where I find sadness in my own feed, the emotional nutrition I feed my mind daily.

My attention patterns (YouTube)

Clustering 4,500 unsorted saved videos surfaced a multitasking library I had organized without ever realizing it: different content for different tasks, background noise for work, full-attention watches for downtime.

My career story (LinkedIn)

Connection growth over time reads like a memoir. Every spike tells a story: the Stanford summer program, each internship search, the first job, the layoff. Career anxiety has a signature, and it's written in connection requests.

None of it changed who I am. It showed me how I engage: how I move through fandoms, how I navigate the tech industry, how anxieties and interests leave breadcrumbs in likes, saves, and connections. The data doesn't redefine me. It just makes visible what was always there.

How it works

Everything starts with the exports every platform is required to hand over: the Twitter archive, Google Takeout, and LinkedIn's data export. From there, one pipeline:

FROM MESSY GDPR EXPORTS TO A JOURNAL OF INSIGHTS raw exports twitter · takeout · linkedin clean + merge python · pandas analyze huggingface · sklearn · stats journal react · recharts

The analysis layer is where the data mining lives: zero-shot topic classification with BART-large-MNLI, emotion classification with DistilRoBERTa, sentiment with twitter-roberta, K-means clustering over TF-IDF vectors of video titles, and chi-square tests, t-tests, regression, and time-series analysis to check that the patterns were real. Claude orchestrated the processing and documentation; Gemini turned the results into the interactive scrollytelling site.

The war stories were half the education. Twitter had revoked API access, so the analysis parses the personal archive's messy JSON instead. LinkedIn's basic export omits connection timestamps entirely, so connection dates had to be inferred from shared work history. Real exports are hostile; that's part of the lesson.

The hard line

Every project I ship has a line it doesn't cross. This one: only my own data, from official exports, and the only person profiled is me. Nothing was scraped from anyone else, no one else's behavior was classified, and the one test subject gave enthusiastic consent on account of being the author.

That's also why this is reproducible by anyone. The whole project exists because of privacy law: GDPR's Article 20 right to data portability (EU, 2018) and the CCPA (California, 2020) were written to give users control over their data, but the same right of access creates an opportunity. If you can download your data, you can analyze it yourself.

Where it is now

The full study lives as an interactive scrollytelling site, with every chart, chapter, and citation: digital-wellness-through-data.vercel.app.

It ends with a "Your Turn" section: reflection questions plus copy-paste prompts for exploring your own exports with the AI assistant of your choice, and instructions for downloading your data from Twitter/X (Settings → Your Account → Download an archive), Google (takeout.google.com), and LinkedIn (Settings → Data Privacy → Get a copy of your data).

Project completed December 2025 for CMU 45-851 Data Mining, by one human chaos coordinator and an AI team: Claude on processing and documentation, Gemini on the interactive UI, ChatGPT on proofreading and portraits.

Explore the live study

References

More projects