Transcript

Kaggle Solution Walkthroughs: UM - Game-Playing Strength of MCTS Variants with Team Senkin13

From Kaggle · Episode page · Watch on YouTube

· 15 min · 1,319 words · 8 timestamped sections

This is the full, timestamped transcript of "Kaggle Solution Walkthroughs: UM - Game-Playing Strength of MCTS Variants with Team Senkin13", an episode of Kaggle published August 22, 2025 (15 min, about 1,319 words). It mentions 17 people, works and ideas, including Data Science, CatBoost, Two-stage Modeling, Leaderboard, Ensemble Methods.

What is discussed, and when

  1. – 3:00: CatBoost, Leaderboard, Cross Validation, Post-processing
  2. – 6:00: Two-stage Modeling, Leaderboard, LightGBM, Augmentation
  3. – 9:00: Leaderboard, Feature Selection, TF-IDF
  4. – 12:00: OOF Predictions
  5. – 14:32: Two-stage Modeling, Leaderboard, Augmentation, Cross Validation

Notable quotes from "Kaggle Solution Walkthroughs: UM - Game-Playing Strength of MCTS Variants with Team Senkin13"

  • Previous speakers so first let me introduce my background I have Applied Mathematics bachelor's degree now I work as a data scientist for AI software startup I have lots of experience in C competition especially table data competion.
    about Data Science · at
  • The post processing because I found the prediction values are smaller than targets so both in cross local cross validation and the public leaderboard so I tried some different coefficient here I use the 1.2.
    about Cross Validation · at
  • Train data it only boost their 01 so from SE seventh version I changes one stage to two stage the two stage means the first stage I just use augmented data as train data and the O predictions as the second stages feature.
    about Two-stage Modeling · at
  • Version I just use light gbn as stage one model and the cat booster as stage two model but nice version I use three gbdt models as stage one and the two models as stage two so it also can improved and my final.
    about CatBoost · at
  • So include some noise I will reduce these features and here I list up the top 20 feature importance from cat boost models so you can see the top three are from stage one's prediction features because stage one I use three.
    about Feature Engineering · at
  • Caboost final blender use weighted average booster is better than l gbn so gave C booster 0.7 L gbn 0.3 and yeah Truth very common truth and because I think I my.
    about GBDT · at
  • Finding I think same as the second place solution and other golden Mets ranker I use flip augmentation flip agent one to agent two pa and changes Advantage P1 to one minor Advantage P1 the target is the.
    about Augmentation · at
  • Best local Closs validation coent because finally simple model I think even I use Ensemble but my simple model is also can get a third place of this competition so I didn't get too much improvement.
    about Ensemble Methods · at

Who and what is mentioned in this episode?

Products & tech

  • CatBoost — Product · Primary model · 1 mention · first at
    Used as the main model in the speaker's competition solution, described as the best performing model and given highest weight (0.7) in the final ensemble
  • LightGBM — Product · Secondary model · 1 mention · first at
    Used as a secondary model in the ensemble approach, given lower weight (0.3) compared to CatBoost in the final solution
    Ensemble models being both in stage one and stage two the eight version I just use light gbn as stage one model and the cat booster as stage two model but nice version I use…
  • XGBoost — Product · Referenced model · 1 mention · first at
    Mentioned once as part of the GBDT model family, but not actively used in the speaker's solution
    Version I just use light gbn as stage one model and the cat booster as stage two model but nice version I use three gbdt models as stage one and the two models as stage two so it also can improved and my final.

Concepts & topics

  • Data Science — Concept · Field · 3 mentions
    The overarching field encompassing all discussed techniques and models.
  • Two-stage Modeling — Concept · Strategy · 2 mentions · first at
    A key methodology discussed for improving model performance in data science competitions.
  • Leaderboard — Concept · Framework · 4 mentions · first at
    Used to evaluate the performance of models in competitions.
    I use Ensemble and postprocessing multiply 1.12 get the best public leader board and also best private leaderboard I think this competition the public leaderboard and private leaderboard match very well.
  • Ensemble Methods — Concept · Strategy · 4 mentions
    Combining multiple models to improve predictions and reduce variance.
  • Feature Engineering — Concept · Strategy · 4 mentions
    Critical for improving model accuracy by creating new features from existing data.
  • Cross Validation — Concept · Framework · 3 mentions · first at
    Technique to assess how the results of a statistical analysis will generalize to an independent data set.
  • Augmentation — Concept · Strategy · 2 mentions · first at
    Technique used to enhance the training data for better model performance.
  • GBDT — Concept · Model · 1 mention · first at
    Gradient Boosted Decision Trees, a machine learning technique.
  • SVD — Concept · Technique · 4 mentions
    Singular Value Decomposition, used for dimensionality reduction.
    Most important then some features from agent name and then Advantage P1 then the English name Engish English lure the TF valf SVD features and after that like some B balance or complexity playout this.
  • OOF Predictions — Concept · Technique · 2 mentions · first at
    Out-of-fold predictions used in cross-validation to prevent overfitting.
    Models so you can see the top three are from stage one's prediction features because stage one I use three models just to predict the train data and the test data and the prediction values will be a new feature for stage two so they are.
  • Post-processing — Concept · Strategy · 1 mention · first at
    Adjusting model outputs after prediction to improve results.
    About 0.002 both in public and private leaderboard the fifth version I used the post processing because I found the prediction values are smaller than targets so both in cross local cross validation and the public leaderboard so.
  • TF-IDF — Concept · Technique · 1 mention · first at
    A numerical statistic that reflects how important a word is to a document in a collection.
    SVD to reduce the Dimensions I think the most important point is the you shouldn't choose the Max features this parameter of TF IDF because your train data and test data are different games so you will use more.
  • Feature Selection — Concept · Strategy · 1 mention · first at
    Process of selecting a subset of relevant features for model construction.
    Use 300 that can give me a best leaderboard score and I Al finally I also redu some features with the new importance future selection method this method like shffle Shuffle the target then.
  • Applied Mathematics — Concept · Field · 1 mention · first at
    The speaker's academic background influencing their approach to data science.
    Previous speakers so first let me introduce my background I have Applied Mathematics bachelor's degree now I work as a data scientist for AI software startup I have lots of experience in C competition especially table data competion.

Full transcript with timestamps

so let me share my third PR solution my agent uh agenda is same as previous speakers so first let me introduce my background I have Applied Mathematics bachelor's degree now I work as a data scientist for AI software startup I have lots of experience in C competition especially table data competion now I I have 13 gold medals and first let me uh introduce my summary I here I list up 10 major versions of my Solutions of course I have more uh mirror versions but uh I think the 10 major versions uh are most important the first version is a like a like basic feature uh with light gbn uh the scad in public is 0 0.444 the basic features including the roow features and some uh some features extracted from engine name so it's just a very simple solution second version I added the t uh English name English Ru names tfidf features uh so it it uh uh improved a lot and the third version I changed light gbn to cab boost models I found it

uh uh improved much so this competition Catabus is the best model and the first version uh I added some interaction features uh like uh uh Ral feature between uh top important features so uh it also improved uh uh about 0 0.002 both in public and private leaderboard the fifth version uh I used the post processing because I found the prediction values uh are smaller than uh targets so both in uh cross local cross validation and the public leaderboard so uh I tried some different uh coefficient here I I I use the 1.2 multiply 1.2 to they can uh boost 0.003 and the six uh version uh I found flip augumentation is very uh useful but here I after I uh use the orentation data to train to original train data it only boost their 01 so from SE uh seventh uh version I changes one stage to two stage the two stage means uh the first stage I just use augmented data as train data and uh the O uh o predictions as the second

stages feature so it uh uh improved uh very much from 0.424 to 0.420 of public leader and also I tried the uh because the seventh uh version uh the stage one and stage two I use same features uh for a a ace version uh I uh remove the tfidf features from stage one so it also improved uh much uh 0.002 and the nin uh version I started to Ensemble uh models being both in stage one and stage two the eight the eight version I just use light gbn as stage one model and the cat booster as stage two model but nice version uh I use three gbdt models as stage one and the two models as stage two so it also can improved uh and my final solution uh I use Ensemble and uh uh postprocessing multiply 1.12 uh get the best uh uh public leader board and uh also uh best private leaderboard I think this competition uh the public leaderboard and private leaderboard match very well and the my local CV also matched very

well okay so let me show some feature uh selection and uh engineering so I I just mentioned uh some basic features interaction features it's uh very simple uh basic features just uh use row features and drop some constant features and the agent property extracted from engine name interaction features I I choose the top uh top 10 St features and uh test the interaction just uh use almost uh raal features uh and the English rules tfidf first uh I use tfidf then I use SVD to reduce the Dimensions I think the most important point is the you shouldn't uh choose the Max features this parameter of TF IDF because uh your train data and test data are different games so you you will use more max large number of Max features it will are overfeat the train data but we wanted to uh make uh the train and test data have same have have more same features so I reduce the Max features finally I use 300 that can give me a best leaderboard

score and I Al finally I also redu some features with the new importance future selection method this this method like uh uh shffle Shuffle the target then then run one run many times with the shuffle Target it can uh output the different features importance distribution and uh uh know which features uh are have more noise uh because uh you will sh Shuffle the target uh there some features uh some features or original Target it it very important but you will Sher Shuffle the the Target that be become unimportant it means these features uh are very related to uh right target but some features um maybe you will Shuffle the targets it's still very important that mean these features uh not not related to original targets so include some noise I will reduce these features and uh here I list up the top 20 feature importance from cat boost models so you can see the top three are from stage one's uh prediction features

because uh stage one uh I use three models just to predict uh the train data and the test data and the prediction values will be a new feature for stage two so they are most important then uh some features uh from agent name and then Advantage P1 then uh the English name Engish English lure the TF valf SVD features and after that uh like some B balance or complexity playout this feature is uh like interaction feature and others uh are same like interaction features okay so my training methods here I use two stage modeling uh first stage just use the flip orentation as train data use three models and predict uh the here uh the augmented data also use same Target and the same group partition as stage two so the oof predictions will be the new future of stage two uh finally uh stage two I use two models like GBM and caboost uh final final blender uh use weighted average booster is better than l gbn so gave C

booster 0.7 L gbn 0.3 and uh uh yeah Truth uh very common truth and uh because I think I my solution is very lightweight so it only take 1 hour 30 minutes and the the prediction time in C is about 30 minutes important and interesting finding finding I think uh same as the second place solution and other golden golden Mets ranker I use flip augmentation uh flip agent one to agent two pa uh and changes Advantage P1 to one minor Advantage P1 the target is the the inverse the target others keep same also I found my local models prediction value are smaller than Target so uh yeah we can calibrate predictions with M multiplying a CO efficient like uh 1.1 uh I I found uh my local best Co efficient not same as the public so we can submit we can select two final submissions so one I use the best leaderboard coefficient and another is best local uh Closs validation coent because finally uh simple

model I think uh even I use Ensemble but uh uh my simple model is also uh can get a third place uh of this competition so I didn't get too much improvement from uh Ensemble uh different C average so this this this simple model is enough to get third place solution thir place but yes my unial solution is the two stage modeling so two stage is the most important most uh uh improved M model Solutions okay any questions okay

Questions about this episode

What is discussed in "Kaggle Solution Walkthroughs: UM - Game-Playing Strength of MCTS Variants with Team Senkin13"?

In order: 0:00 CatBoost, Leaderboard, Cross Validation, Post-processing; 3:00 Two-stage Modeling, Leaderboard, LightGBM, Augmentation; 6:00 Leaderboard, Feature Selection, TF-IDF; 9:00 OOF Predictions; 12:00 Two-stage Modeling, Leaderboard, Augmentation, Cross Validation.

What is mentioned in this Kaggle episode?

17 people, works and ideas are mentioned, including Data Science, CatBoost, Two-stage Modeling, Leaderboard, Ensemble Methods, Feature Engineering. Each one on this page links to its own page with every episode that mentions it.

How long is "Kaggle Solution Walkthroughs: UM - Game-Playing Strength of MCTS Variants with Team Senkin13"?

The episode runs 15 min; the transcript is about 1,319 words in 8 timestamped sections.

More from Kaggle

About this transcript: the episode's spoken words as captioned on YouTube, split into timestamped sections. Reclip.Me locates every person, work and idea in the text, keeps the time it is first named, and quotes the words around it verbatim. Last updated February 16, 2026.