Riccardo Cappuzzo
@riccardocappuzzo.com
Research engineer at Inria Saclay, working on the Skrub library. PhD in computer science.
Python, data preparation, ML, tabular learning.
ORCID: 0000-0002-4448-2959
Hoshiyomi ☄️
https://www.riccardocappuzzo.com
https://github.com/rcap107
Python, data preparation, ML, tabular learning.
ORCID: 0000-0002-4448-2959
Hoshiyomi ☄️
https://www.riccardocappuzzo.com
https://github.com/rcap107
Funny bug of the day: if you try to use pandas' "guess_datetime_format" with datetimes where the hour and minute are the same as the year (like 1959 and 19:59), the parser will fail and return None.
This bug is present in pandas 2.3.3, but has been fixed in the dev version.
This bug is present in pandas 2.3.3, but has been fixed in the dev version.
January 12, 2026 at 10:17 PM
Funny bug of the day: if you try to use pandas' "guess_datetime_format" with datetimes where the hour and minute are the same as the year (like 1959 and 19:59), the parser will fail and return None.
This bug is present in pandas 2.3.3, but has been fixed in the dev version.
This bug is present in pandas 2.3.3, but has been fixed in the dev version.
"ok the test run is done, let's see"
...
"this will be hard to debug"
...
"this will be hard to debug"
October 9, 2025 at 1:29 PM
"ok the test run is done, let's see"
...
"this will be hard to debug"
...
"this will be hard to debug"
Working hard on the next @skrub-data.bsky.social slide deck...
September 4, 2025 at 10:19 PM
Working hard on the next @skrub-data.bsky.social slide deck...
Really cool graffiti I spotted while walking around in the town where I live
June 13, 2025 at 9:10 PM
Really cool graffiti I spotted while walking around in the town where I live
Now that the paper is out, I can finally share the totally-not-confusing script/plot/table map I made to track which scripts prepare which figures and tables and from what data.
If it wasn't clear, don't do this. If you *really* have to, I used the @obsidian.md canvas for this.
If it wasn't clear, don't do this. If you *really* have to, I used the @obsidian.md canvas for this.
May 20, 2025 at 8:26 AM
Now that the paper is out, I can finally share the totally-not-confusing script/plot/table map I made to track which scripts prepare which figures and tables and from what data.
If it wasn't clear, don't do this. If you *really* have to, I used the @obsidian.md canvas for this.
If it wasn't clear, don't do this. If you *really* have to, I used the @obsidian.md canvas for this.
A bit of a mess up with this figure! This is what it's supposed to look like 🙈
May 19, 2025 at 4:01 PM
A bit of a mess up with this figure! This is what it's supposed to look like 🙈
⏱️ Complex aggregation methods are slower and don't significantly boost prediction performance.
6/
6/
May 19, 2025 at 3:43 PM
⏱️ Complex aggregation methods are slower and don't significantly boost prediction performance.
6/
6/
⚖️ Beware of diminishing returns! Performance plateaus as more candidates are retrieved, while resource costs (time and RAM) keep rising.
5/
5/
May 19, 2025 at 3:43 PM
⚖️ Beware of diminishing returns! Performance plateaus as more candidates are retrieved, while resource costs (time and RAM) keep rising.
5/
5/
🎯 Simple metric-based retrieval and candidate selection methods often outperform complex methods and are more efficient.
4/
4/
May 19, 2025 at 3:43 PM
🎯 Simple metric-based retrieval and candidate selection methods often outperform complex methods and are more efficient.
4/
4/
🔍 Good table retrieval is crucial helps finding candidates with useful features and fewer missing values. Jaccard containment is helpful but has its limits.
3/
3/
May 19, 2025 at 3:43 PM
🔍 Good table retrieval is crucial helps finding candidates with useful features and fewer missing values. Jaccard containment is helpful but has its limits.
3/
3/
🌳 Tree-based models offer better prediction and computational performance than deep learning-based methods in our setting, which involves training models over features that contain a large fraction of missing values.
2/
2/
May 19, 2025 at 3:43 PM
🌳 Tree-based models offer better prediction and computational performance than deep learning-based methods in our setting, which involves training models over features that contain a large fraction of missing values.
2/
2/
🌟 New paper alert! 🌟
Our paper, "Retrieve, Merge, Predict: Augmenting Tables with Data Lakes", has been published in TMLR!
In this work, we created YADL (a semi-synthetic data lake), and we benchmarked methods for augmenting user-provided tables given information found in data lakes.
1/
Our paper, "Retrieve, Merge, Predict: Augmenting Tables with Data Lakes", has been published in TMLR!
In this work, we created YADL (a semi-synthetic data lake), and we benchmarked methods for augmenting user-provided tables given information found in data lakes.
1/
May 19, 2025 at 3:43 PM
🌟 New paper alert! 🌟
Our paper, "Retrieve, Merge, Predict: Augmenting Tables with Data Lakes", has been published in TMLR!
In this work, we created YADL (a semi-synthetic data lake), and we benchmarked methods for augmenting user-provided tables given information found in data lakes.
1/
Our paper, "Retrieve, Merge, Predict: Augmenting Tables with Data Lakes", has been published in TMLR!
In this work, we created YADL (a semi-synthetic data lake), and we benchmarked methods for augmenting user-provided tables given information found in data lakes.
1/
More fun digging around my @last.fm scrobbles using with @matplotlib.org
I had no idea how much of a difference changing fonts and background color could make
I had no idea how much of a difference changing fonts and background color could make
May 1, 2025 at 11:00 PM
More fun digging around my @last.fm scrobbles using with @matplotlib.org
I had no idea how much of a difference changing fonts and background color could make
I had no idea how much of a difference changing fonts and background color could make
I haven't been using a lot of Copilot until very recently, so I'm still learning what it can do.
It just blew my mind by autocompleting the dictionary "release_dates" with the correct dates for Muse albums based on the fact I am looking at data about Muse in the script.
wow
It just blew my mind by autocompleting the dictionary "release_dates" with the correct dates for Muse albums based on the fact I am looking at data about Muse in the script.
wow
May 1, 2025 at 6:51 PM
I haven't been using a lot of Copilot until very recently, so I'm still learning what it can do.
It just blew my mind by autocompleting the dictionary "release_dates" with the correct dates for Muse albums based on the fact I am looking at data about Muse in the script.
wow
It just blew my mind by autocompleting the dictionary "release_dates" with the correct dates for Muse albums based on the fact I am looking at data about Muse in the script.
wow
First experiment plotting my Last.fm scrobbles
With 10 years worth of data, I'll be working on this for a while.
Also first time working with @matplotlib.org stackplots, much finagling was involved
With 10 years worth of data, I'll be working on this for a while.
Also first time working with @matplotlib.org stackplots, much finagling was involved
April 30, 2025 at 10:33 PM
First experiment plotting my Last.fm scrobbles
With 10 years worth of data, I'll be working on this for a while.
Also first time working with @matplotlib.org stackplots, much finagling was involved
With 10 years worth of data, I'll be working on this for a while.
Also first time working with @matplotlib.org stackplots, much finagling was involved
The similarity is uncanny
April 10, 2025 at 2:07 PM
The similarity is uncanny
One good thing about living in Paris is that, well, you're living in Paris.
December 14, 2024 at 11:51 PM
One good thing about living in Paris is that, well, you're living in Paris.
The good thing about running experiments in France is that I can feel slightly less guilty about my emissions, but still
yikes
yikes
December 12, 2024 at 10:45 AM
The good thing about running experiments in France is that I can feel slightly less guilty about my emissions, but still
yikes
yikes
Sister's Christmas cat
November 28, 2024 at 8:01 PM
Sister's Christmas cat
November 27, 2024 at 11:25 AM
This is a very simple example of what I am working with, only I have potentially thousands of lines like this.
Looking at the documentation, it does look like I wouldn't need a lot of the features of SSSOM (and it might just add overhead in my scenario).
Still, thanks for the clarification 👍
Looking at the documentation, it does look like I wouldn't need a lot of the features of SSSOM (and it might just add overhead in my scenario).
Still, thanks for the clarification 👍
November 19, 2024 at 9:38 AM
This is a very simple example of what I am working with, only I have potentially thousands of lines like this.
Looking at the documentation, it does look like I wouldn't need a lot of the features of SSSOM (and it might just add overhead in my scenario).
Still, thanks for the clarification 👍
Looking at the documentation, it does look like I wouldn't need a lot of the features of SSSOM (and it might just add overhead in my scenario).
Still, thanks for the clarification 👍