pKozaneba2025-08-26~27
Consideration of changing Kozaneba to a Canvas implementation
- Basically, it's almost a reworking of the original.
How do you feel about changing from the paper-based KJ method to Kozaneba?
- It's knee-swing (gymnastics), not KJ method.
- Grouping is used, but only to make it easier to move them together.
- I didn't do the KJ legal "bundle with nameplate".
- Is the UI so bad that you're unconsciously avoiding it?
- Folding it up creates space, but that doesn't make me happy.
- Using "smooth folding" to make it harder to read when zoomed out.
- I use "Large sticky note" as a sticky note that can be read even when zoomed out.
Now when I import the text, it's all lined up in a square.
Should I use embedding and place it in 2D?
Implement as a selection menu?
public information AI is embedding, UMAP two-dimensional, and then agglomerative clustering.
Wait, what? Maybe we don't need to reduce it to 2 dimensions with dimensionality reduction if we eventually drop it into a sticky note arrangement?
- Or, in effect, what we're doing with graph-based dimensionality reduction is going to be almost the same.
Let's put off building this place from scratch for now.
- I feel like I'm going to make a degraded UMAP.
Making the sticky coordinates a grid point means putting a float in the bin.
- I don't want the number in one bin to be too large.
- I don't want it to be too large in scale.
- Here's the parameter tradeoff.
- Experiment with variable parameters once
GPT5 is a one-shot deal for the functions I once achieved with Kozaneba...
- My implementation breaks lines mercilessly, even for English words, so maybe this is better.
- And you're making this sample data itself for demonstration purposes.
- I'm noticing traps like skin color change emojis and super long one-words.
- I don't have the punctuation forbidding process in there, but I could do it just by saying "do it".
- This code: https://chatgpt.com/c/68ad5ac8-f398-8321-89c7-b4617c143f00
Test of 10,000 sticky notes
- At any rate, I've found that I can zoom in and out of 10,000 sheets with no performance issues.
- This code: https://chatgpt.com/c/68ad3ab0-defc-8332-a311-5c8191d70f0e
- This is what happens when you simply put a sticky out in the UMAP position
Changed the position of stickies to snap to integer grid points by dividing by NOTE_SIZE (120px), counting the number of stickies at each grid point and displaying the top 10 at the bottom left of the screen.
- This is SCALE=600, 90~60 overlapping cases
- Even with SCALE=2000, that's about 50~26 overlaps.
- SCALE=4000 would be almost a point, still 35~15 overlap.
When there are multiple stickies on the same grid, we implemented a logic to find a free grid close to the spiral and distribute the stickies.
I've got a great gripping motion.
- opening screen
I was able to deploy. https://canvas-kozaneba-prototype.vercel.app/
Data that has been collapsed to a single point in a scatter plot that basically maintains the coordinates where the square is lumped.
- I am able to visualize density in terms of size.
- I've done it before to visualize density by making the points of the scatterplot translucent.
- I was under the impression that it was "easy to understand the internal structure," but it was not adopted because it looked blurry and unappealing.
About "Visualize density in terms of size"
Q: How does this compare to [Kernel density estimation
- When you draw a scatter plot, the public doesn't understand that "each point is an opinion," so you have to tell them every single time, and that's why you're trying to make them understand without explanation in the first place. - I'm sure many of you have had the experience of writing sticky notes and arranging them in a workshop. - So, the main objective was to generate a "diagram of stickies in a row," and in the process of doing so, we discovered that we could visualize the density of the stickies. - Kernel Density Estimation and contour plots are fine for data scientists, but for layman's perception, the level of abstraction will be even higher than scatter plots and harder to understand. - Tried. -  - I made it a heat map instead of contour lines. - Right: resolution=800, bandwidth=0.05 - If the drawing is detailed enough to do this, it will have similar cluster finding ability. -
next pKozaneba2025-08-29
This page is auto-translated from /nishio/pKozaneba2025-08-26~27 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.