Michael Roizner (@mroizner) profile picture
London, UK
Joined December 2008
Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)Photo by Michael Roizner (@mroizner)
“Why the hell did you recommend THIS to me? Your algorithm is trash!” If you’ve ever shipped a recommender, you know that moment: the metrics look great, but one furious exec can torpedo months of work. Read the “survival guide”: linkedin.com/pulse/why-hell-…
1
4
481
First time since 2019. #icml2026
1
22
1,059
| is not always associative in Python (even when no user-defined types are involved)
1
53
Wizz Air @wizzair sucks. Their website wouldn’t let us check in 3 of 4 passengers on our booking. We had to pay €144 at the airport, and they still refused to refund us even after we provided evidence. If you’re thinking of flying with them, think twice.
6
2
2
472
Don’t ever fly with @TurkishAirlines. It’s the worst flying experience of my life. And the worst customer support.
4
1
2
1,396
The importance of a good engagement model should not be underestimated.
In recommendations, distinguish 2 steps: 1) Training a model to predict user response. Simple case: p-click; more general: E(engagement | item, user, context). 2) Make recommendations with the model. Simple ranking by predictions is not the most optimal.
196
In recommendations, distinguish 2 steps: 1) Training a model to predict user response. Simple case: p-click; more general: E(engagement | item, user, context). 2) Make recommendations with the model. Simple ranking by predictions is not the most optimal.
Is popularity bias just a convenient excuse? linkedin.com/pulse/personali…
1
1
394
Much more often, popularity bias is just a popular myth that conceals system bugs.
Is popularity bias just a convenient excuse? linkedin.com/pulse/personali…
119
Is popularity bias just a convenient excuse? linkedin.com/pulse/personali…
1
4,380
We tend to focus on the final outputs of recommender systems, but as @MRoizner points out in his thoughtful article, it's crucial to pay attention to designing and evaluating candidate generation in the systems' early ranking stages. buff.ly/46H0G2I
2
1
7
4,641
The main goal of the early ranking stages is to find the best documents from the perspective of the final ranking
1
101
I delve deeply into the pros and cons of this approach, providing insights crucial for those working with recommender systems, in my latest article in @TDataScience 'The Principled Approach to Early Ranking Stages' towardsdatascience.com/the-p…
1
1
3
2,925
Additionally, I introduce a practical evaluation method for candidate generation, based on random candidates. It's a simple yet effective technique that I haven't seen mentioned elsewhere.
67
Getting "You've reached the current usage cap for GPT-4, please try again after xx:xx" more often lately.
1
142
For a deeper understanding of the key elements that power advanced recommendation engines, don't miss @MRoizner's explainer on two-tower networks and negative sampling. buff.ly/47pBA9C
5
16
5,141
What is the proper approach to implementing rules in recommender systems? @MRoizner's debut TDS post offers helpful guidance on how to balance the needs of developers and product managers. buff.ly/461QFh3
2
6
3,702