GitHub link: github.com/brege/image-activity

Overview

• Generate heatmaps and histograms of image saving activity over hours, days, and months
• Use file timestamps, modified-times, EXIF, and regex parsing for refined image discovery
• Add bands and markers for major life events

Background

I wanted to determine if my image activity is dependent on major events and device purchases in my life.

• do I tend to take more pictures during certain times of year?
• how has my screenshot usage evolved over the last 15 years?
• do I have “honeymoon” periods after a device purchase?
• in what ways has my camera and screenshot usage changed between being an academic, chef, and developer?

I’m not a social media person, although my mastodon did see an uptick of usage following my hip surgery, where I began hiking and foraging a lot.

My image activity fits in three main categories:

1. camera: storage of camera photos from my phone
2. screenshots: screenshots on both my laptop and phone
3. internet: pictures downloaded from the internet

Gallery

I’ve marked in these first line charts, Camera Usage and Image Capture Concurrency, times when I’ve purchased a major device (a new phone or laptop) and a couple key periods of my life. These plots have all been normalized to a 0-100 photo count scale.

Camera Usage

From 2010 to 2017 I was a Physics TA and, following my 2014 physics prelims, a computational astrophysics doctoral researcher. I began attending conferences in 2015, exploring places around Pullman, WA during these researcher years there.

At the end of 2017, I left that life. I embraced my love of food and cooking and became a professional chef for a number of years thereafter, including the Covid-19 pandemic. This period of my life saw a greater number of photos taken: pictures of plates, menus, schedules, etc. My camera photos before this time were mostly non-work-related: travel, events, and pets drove image origination.

Image Capture Concurrency

Heatmaps

I only have one experience with online coursework: the data science bootcamp I attended in the fall of 2023. This period did not have a major impact on my screenshotting habits. There are three principal areas in which screenshot usage was more frequent:

1. The creation of my website brege.org around August 2016.
2. As an executive chef, screenshotting is recurrent for scheduling, text message records, receipts/purchase dates, etc.
3. Agentic-driven coding workflows, beginning midway through 2025, saw a surge in screenshot usage. Screenshots have become a large part of my front-end debugging workflow for web app development–extending well beyond data-structured Cypress end-to-end tests.

I did not find my screenshot usage noticeably change during my brief stint with online coursework.

In general, it appears that I take more screenshots on desktop earlier in the week and in the afternoon (averaged over the last ~15 years). To my surprise, the heatmap for screenshots on my phone have nearly identical densities. I assumed this would be biased toward the weekend and closer to 17:00 because of sports and restaurant dinner service.

Camera usage frequency, on the other hand, is made distinct by day of week only on density during Thursday evening and Saturday afternoon. It’s especially featured in both my Chef days and post-op mobility.

Histograms

By device and source, then binned on hours of the day, day of the week, and month of the year, histograms provide a finer distribution in one dimension.

For the hourly concentration of all three photo habits, my activity roughly follows a Boltzmann distribution.

These distributions generally peak at two distinct hours:

• camera photos and screenshots center around 15:00
• internet photos are generally concentrated around 20:00

Each bin is averaged for each picture type over the last 15 years, regardless of timezone.

Image activity generally increases at the beginning and end of standard university semesters, which also include the height of summer and the holiday period when I am always always travelling. Screenshotting is highest in the fall to mid-winter.

In my experience, restaurants are historically busier between, roughly, Friendsgiving and Father’s Day. Camera usage also largest during high summer. Beach. Hiking. Produce selection during chef years.

I have survey fatigue

Surveys are marketing emails. I can’t believe I used to take the time to respond to some of them. An analysis on my email history has shown that my hunch on survey spam is correct. Around 2023, I began marking all surveys as spam, and I’ve got the data to prove just how rampant companies have used surveys to get their brand in your inbox.

Emails that matter

My main goal in exploring my emails in-depth was to build a predictor of whether an email had any future usefulness.

Mail from friends and family, correspondence with students, receipts and financial records, etc. fit the binary of keep. After I manually processed this massive backlog of email (with great help of Thunderbird’s filters), what I found I discarded most, besides spam, were surveys, newsletters, and other mass mailers.

What I found, qualitatively, was:

• imperfect spelling, capitalization, and grammar
• little-to-no HTML markup
• all emails meant only for me sans phishing and spam

These traits defined true keepsakes.

Introducing sanoma

en.wiktionary.org/wiki/sanoma

sanoma (noun) Finnish 
  message, communication (a communication or the content of a physical
  message; also the message contained in some act or expression such as a
  work of art)

sanoma (github.com/brege/sanoma) uses YAML workflows to define multi-step analysis pipelines. The workflow runner automatically discovers and executes tools from the sanoma/analysis/ and sanoma/plot/ directories, making it easy to chain data extraction, filtering, analysis, and visualization into reproducible pipelines.

I developed this YAML workflow method in my Markdown-to-PDF project–oshea–where I realized comprehensive end-to-end tests were just manifest workflows. It’s an intuitive way to string command line sequences together. The pipeline term in machine learning/data science is congruent to this system.

Data Mining

While much of this can be done in a Jupyter notebook (far easier to refresh plots this way, although :MarkdownPreview in Neovim is sufficient), I built this project as a way to data-mine my own activity. I also want to create a visualization harness for many things on my computer:

• text message history
• email history
• screenshot frequency
• browser history and bookmarks

Because email is text-based, and because my first concept of “AI” was the need for combative spam filters that have been built over the last thirty years, email felt like a good starting point.

Grad-school Emails

The monthly timeline reveals the academic year rhythm: high volume during active semesters with dramatic drops during summer breaks and winter holidays. The 2016-2017 dip corresponds to the dissertation defense period, where militant email sanitation was a reprieve from LaTeX and simulation monitoring–hence the dip.

My personal dataset has about 35K emails between my grad-school emails and my current website’s personal email. Not included are my Gmail and undergrad email(s). I plan on synchronizing those at a later date.

Grad-school Timeline Seasonality

WSU’s Okta system required changing passwords every 6 months, and some time after my defense my account died. I am thankful that I had a Thunderbird profile tucked away on a drive that allowed me to recover all of my university emails.

Grad-school and onward Histogram

The year-over-year histogram demonstrates consistent academic seasonality, with September-April peaks and May – mid-August valleys across all years of graduate study. Even with teaching summer labs, the bureaucratic pressure in the summertime dies. I loved teaching in the summer.

Spam, Marketing, and Surveys

The spam timeline shows minimal marketing emails pre-2010, followed by a sharp increase around university enrollment. By 2015, spam reached 60-80% of all emails and has remained consistently high. The GDPR implementation around 2018 created a spike in unsubscribe language as companies scrambled to comply with new regulations.

Marketing Spam Trends

The tail in the beginning of this timeline is presented for context. It only includes a “purified” hotmail account mailbox from my teenage years that extended a bit into my undergrad years. Those years overlap with Gmail usage (not integrated into this data) and my GVSU university email.

Keyword Buckets

Another useful filter for spam emails is checking for keywords like unsubscribe in the message body.

unsubscribe_bait dominates with over 12,500 matches, followed by satisfaction surveys (~8k) and direct “survey” requests (~4k). This reveals how modern marketing shifted from direct sales to engagement-focused tactics requesting feedback and reviews.

Conclusion: Satisfaction Surveys are the new email cancer

The heatmap (filtered to post-2010) shows “satisfaction” spam as the most persistent threat, maintaining 20-25% frequency from 2012 onwards. Survey-based spam shows steady growth, intensifying after 2020, when both GDPR constraints pressured companies to invent new angles of attack, becoming increasingly desperate for customer “feedback” (attention) during the pandemic. Satisfaction feedback surveys are advertisements.