Kaggle Winners Walkthroughs: Santa 2024 - The Perplexity Permutation Puzzle with Team Santa Master

What is discussed, and when

1. 0:00 – 3:00: beam search, local optimization, simulated annealing
2. 3:00 – 6:00: B ranks, function words, content words
3. 6:00 – 9:00: function words, content words, alphabetical order
4. 9:00 – 11:30: function words

Notable quotes from "Kaggle Winners Walkthroughs: Santa 2024 - The Perplexity Permutation Puzzle with Team Santa Master"

• Actually I when I chimed up with him I was using simulated an earing made by myself but after she merged we found the IRS is quite better solution of this problem.

  about simulated annealing · at 1:13

• First one needs insert optimize. this der randomly selected nwords and insert them with beam search and next one is local optimize.

  about beam search · at 2:05

• Are not so easy also we found these words increases B ranks I will describe this B ranks later and we will get an advantage by putting these words at the beginning of the.

  about B ranks · at 3:28

• So we also found the consolidating these we called these words function words in one place and placing them at the beings tend to result in lower rows. Optimal answer.

  about function words · at 5:10

• It means we can decrease the loss and regarding ID5 we also found alphabetical order of content words decreases loss also splitting alphabetical content words decreases.

  about content words · at 7:31

Who and what is mentioned in this episode?

Concepts & topics

• function words — Concept · Key argument · 4 mentions · first at 5:22
  Identified as words that, when consolidated and placed at the beginning of a sentence, tend to result in lower rows in optimization problems.
• B ranks — Concept · Strategy · 2 mentions · first at 3:37
  A scoring metric where an increase indicates a decrease in loss, crucial for optimizing the solution.
• content words — Concept · Key argument · 2 mentions · first at 5:34
  Words that carry the main meaning in a sentence, whose alphabetical order and placement affect the optimization outcome.
• alphabetical order — Concept · Strategy · 1 mention · first at 7:46
  A method found to decrease loss when applied to content words in the optimization process.
  It means we can decrease the loss and regarding ID5 we also found alphabetical order of content words decreases loss also splitting alphabetical content words decreases.
• simulated annealing — Concept · Strategy · 1 mention · first at 1:20
  An optimization algorithm initially considered but later abandoned in favor of a better solution.
• beam search — Concept · Strategy · 1 mention · first at 2:16
  A search algorithm used in the insert optimization phase of the solution.
• local optimization — Concept · Strategy · 1 mention · first at 2:19
  A phase in the solution involving shuffling, swapping, and moving words to optimize the outcome.
  This der randomly selected nwords and insert them with beam search and next one is local optimize.

Full transcript with timestamps

0:00 Uh so let me start. Uh this is the agenda. Uh first I will introduce ourselves and next is main solution and uh last one is important findings. So teams um these teams consists of me and Dwakon and actually I I didn't know him but uh uh but when we che him with him um I thought he he is he he is familiar quite familiar with uh these optimizing competition. Um so I just decided to team up with him. Okay. Main solution. Um uh actually I when I chimed up with him I I I was using simulated an earing uh made by myself but uh after she merged uh we we found the IRS is quite better solution of this problem.

1:42 So, so I decided to uh so I decided not to use simulated onering. Uh yeah. So, so our main solution is IRS. Uh our IRS consists of uh first one needs insert optimize. Uh this der randomly selected nwords and insert them with beam search and next one is local optimize. Local optimize consists of shuffle subsection uh swap words and move words. So, so the kick was included in the first one inside optimize. This accept some randomness. So, so this would be the kick and this table stands for the result of the solution.

3:11 Um, as you can see, ID zero to two are quite easy to solve. uh but ID3 to five uh are not so easy also we we found these words increases B ranks I will describe this B ranks later and we will get an advantage by putting these words at the beginning of the sentence. Uh reind zingle s and so on. Uh the the board bold words stands for the stands for belonging to ID3. Uh ID3 uh if you start with MAGI the the optimal solution can be discovered in about three hours. ID zero and ID one start with rinder and ID two start with stray. So we don't need to uh we don't need to search 30 words uh for ID3 uh because uh yeah because these one two three four five six seven seven words uh has a big advantage uh regarding uh B ranks.

5:10 So we also found the consolidating these uh we we called uh these words function function words in one place and placing them at the beings tend to result in lower rows. Optimal answer of ID4 is function words plus content words. So if we don't have if we don't have this uh finding the the order of ID4 would be uh fractional of 50. But using this binding the the order of ID4 would be for 14 fraction fractional of 14 uh 14 plus uh 30 fractional of 36. Uh these yellow words are the function words. And uh regarding ID5, the order would be would be one fractional of 100 uh into fractional of 20 plus fractional of 80.

7:05 And this is the B drawings. Bangs is here. I will skip the detail of the of the scoring metric. But as you can see if the B rings increases we uh it means uh we can decrease the loss and uh regarding ID5 we also found alphabetical order of content words decreases loss also splitting alphabetical content words decreases loss. So we sort the optimal shape of ID ID5 is roughly function plus sorted content one plus sorted content two. So the order would be function functional of 20 uh plus uh 80 combination of uh 40 or so. And uh using this finding we could get uh better score uh 32.41 to uh 20 28 57.

8:55 This this means reader in reader board to 246.82 uh if I remember correctly a lot a lot of teams were struggling at this this score. Uh so it means uh it means we we also found these uh these same finding and uh it means we also have have the same mistakes. The mistakes means uh including uh the mistake means uh the function function words should not include and. Uh lastly we found and should be removed from function words and then uh we got the the optimal score. Uh yes, as you can see uh the right right figure um this is the 82 of uh 82 sentence. As you can see the and and and are including in in the function words. But if we if we removed these words from the beginning of the sentence and we we could get the the optimal score and also uh most of words uh not changed and the only the three and plus uh uride. Okay, this is the end of our solution.

11:19 Uh thank you for listening. Uh do you have any question?

Questions about this episode

What is discussed in "Kaggle Winners Walkthroughs: Santa 2024 - The Perplexity Permutation Puzzle with Team Santa Master"?

In order: 0:00 beam search, local optimization, simulated annealing; 3:00 B ranks, function words, content words; 6:00 function words, content words, alphabetical order; 9:00 function words.

What is mentioned in this Kaggle episode?

7 people, works and ideas are mentioned, including function words, B ranks, content words, alphabetical order, simulated annealing, beam search. Each one on this page links to its own page with every episode that mentions it.

How long is "Kaggle Winners Walkthroughs: Santa 2024 - The Perplexity Permutation Puzzle with Team Santa Master"?

The episode runs 11 min; the transcript is about 798 words in 7 timestamped sections.

More from Kaggle

• All analyzed episodes
• People mentioned on Kaggle
• Topics mentioned on Kaggle