Kaggle Winners Walkthroughs: Child Mind Institute—Problematic Internet Use with Team Peyman Armaghan

What this episode is about

This walkthrough details Team Peyman Armaghan's innovative approach to analyzing problematic internet use through time series data in a competition organized by the Child Mind Institute. The team employed clustering algorithms to extract valuable features from user activity data, leveraging their background in telecommunications and economics to enhance their understanding of AI and machine learning, ultimately transforming time series insights into meaningful representations of user behavior.

Key takeaways

• Switching from engineering to finance isn't just a pivot; it's an Olympic-level leap through intellectual gymnastics.
• Using semi-supervised learning is like inviting uninvited guests: they can either add to the party or wreak havoc.
• Optimization thresholds can feel like using a squint: miss the nuances of data, focusing blindly on relative positions.
• Noisy data is like an adventurous toddler—chaotic yet surprisingly capable of leading to groundbreaking insights with the right nudges.
• Leaving feature engineering until the last minute is like cramming for an exam: you might pass, but don't expect an A+.

Who speaks in this episode?

• Peyman Armaghan (Host) · 13 min speaking

What is discussed, and when

1. 0:00 – 3:00: Cayman's clustering algorithm
2. 3:00 – 6:00: PCIT total
3. 9:00 – 12:00: light GBM

Notable quotes from "Kaggle Winners Walkthroughs: Child Mind Institute—Problematic Internet Use with Team Peyman Armaghan"

• Competition to extract useful information from time series I used the Cayman's clustering algorithm with 15 cluster to extract useful information from time series data for that reason I picked six perk files two most.

  about Cayman's clustering algorithm · at 1:57

• Led to overfitting and inconsistent result. So I decided to ignore optimization completely and use PCIT total as target label. And I applied a fixed threshold to change the target variable to the competition target variable.

  about PCIT total · at 3:43

• To simplify my model I removed the time series feature and I just used a single light GBM model for final prediction and with this simplified approach I got the score of 0.573 in my private leader board which I could assume a solid.

  about light GBM · at 11:40

Who and what is mentioned in this episode?

Concepts & topics

• Cayman's clustering algorithm — Concept · Strategy · 1 mention · first at 2:02
  Used to extract useful information from time series data in the competition, contributing to the success of the participant.
• PCIT total — Concept · Strategy · 1 mention · first at 3:51
  Used as a target label in the competition to avoid overfitting and inconsistent results from threshold optimization.
• light GBM — Concept · Strategy · 1 mention · first at 11:49
  Used as a simplified final model for prediction in the competition, contributing to a solid score.

Full transcript with timestamps

0:00 Let me by start thanking the competition organizer. Uh uh it was very uh amazing journey that we compete on and learn at the same time and it was very valuable for me. uh uh I began uh my presentation with an brief overview of myself and then uh as uh uh as you can see I go to the details of my approach and my models. Uh I have a bachelor and master degree in telecommunication engineering and uh I have a PhD in economics. Uh after uh two or three years that I work in the communication engineering field uh I changed my course and uh I begin to work uh during past 15 years primarily as a analyst and investment manager.

1:06 um recent development on AI and also this big hype on LLM uh got me exciting uh to go to more details about AI and machine learning and uh with my background in C programming language and my engineering background uh during past I think 10 or nine months I put hard work to learn Python and machine learning algorithm and I think that's hard work in past 10 months helped me a lot to succeed in this competition. uh to begin with my approach I think it's good to start with time series uh it was a challenging part of this competition uh uh to extract useful information from time series I used the Cayman Cayman's clustering algorithm with uh 15 cluster to uh extract useful information from time series data uh for that reason uh I picked six perk files uh two most variant files from each classes and then I combined this data set together and then uh fit uh C means cluster algorithm uh then I use this uh fitted model to uh

2:38 extract cluster from uh other time series data sets. uh I assume that each cluster could represent a movement in park data. Uh so uh I average each cluster value over the time duration and then I use this average as a feature for each user. uh by doing that at the end uh I have uh 15 feature uh features per user that uh I believe could uh represents the average of some kind of activity during the time that the user use the actively uh watch.

3:28 Uh another challenging part of this competition I believe was choosing the right target. Uh after some initial submission I found that uh using threshold of optimization could simply led to overfitting and inconsistent result. So I decided to uh ignore optimization completely and use PCIT total uh as target label. And I applied a fixed threshold to uh uh change the target variable to the competition target variable. uh since the exact value between this threshold was less important than the relative position to the the threshold for final prediction. I uh further bean the PCI uh total into 10 beans and uh adjusted the threshold according. A big portion of data uh in the training data set was on label data.

4:38 So I decided to use a semi-supervised learning approach or pseudo labeling to predict the label of this unlabelled data and to use them with uh other uh or other data to train my final model. So I decided to use an assemble of models. I used five models for this purpose. three gradient boost and decision tree one neural network and one loss or regression model. Uh I then uh combine and assemble this model models using hill climbing and then uh I use this whole uh models to predict the label of unlabelled data. uh and then uh for the my final p prediction I train my model use both uh the uh this uh predicted label and the original label that we had in data sets.

5:49 uh for my final prediction I alo I'll also use in symbol of models again uh three gradient decision gradient boosting decision tree uh one simple decision tree theory uh two regressor model and one uh neural network uh I train uh the models on the entire data set and validate it only on original label in my cross validation approach uh I use sample weight proportional to frequency of every sample in my regressor and uh I should admit that I got that idea from Leonard from the notebook that he publish uh in discussion format.

6:38 uh I also uh applied hard boing to insemble this models and uh use the data which I was derived from my semi-supervised learning. Uh in this picture you can see the summary of uh what I've done for my models and the workflow of data from the beginning to the final prediction. Uh so uh I joined uh to the competition I think uh three weeks before the deadline and uh I noticed that the data are noisy with a lot of missing values. Uh so I postponed the feature engineering to the end of the to the end of last day of the competition. But unfortunately at the end I ran off run out of time and don't have uh enough time for feature engineering and I simply mapped season cate categorical feature to numerical values and uh I used KN&N imputer for uh filling the missing data in features and that was it. uh I didn't do any other feature engineering. Uh here you can see the feature importance of uh my uh large GBM

8:19 uh model in my base models. Uh as you can see uh that uh the age and physical heads are among the most important features in my model. uh this uh ranking is by gain and we can see if uh we want to rank the feature importance by the contribution to the uh split of uh trees in gradient boosting. Again you can see that the uh sleep disturbance scales is uh the most important feature and also the time spended on the computer uh on the computer is the second most important feature. Uh here I plotted some partial dependence uh between uh most important features and my target variable and uh as it was I think obvious from the domain knowledge. You can see that there are some direct relation with uh features like uh uh a sleep uh disturbance scale and uh times spended on the computer with uh target variable uh that we predicted.

9:47 uh these are also some other partial independ independence uh dependence uh variable that uh I will skip them uh for my training methods uh I used a five-fold class uh cross validation uh with consistent uh random state across all folds uh for my semi-supervised model I use hill coingming approach to uh ensemble models based on their mean square error uh matrix and uh for final model I applied hard hard voting to insemble the base models with uh equal weights. uh uh the key approaches that I believe has contributed to my success was uh the way of uh I extract information from time series by using uh clustering methods uh using uh PCAT total as a target label. I think helped me to uh ignore completely the uh uh threshold optimization and put my effort on my models. Uh and uh also using sample weight in training regressor I think was a uh important uh uh important part of my job. uh as I said it wasn't my idea

11:26 and uh using semi-supervised learning for unlabelled data uh helped me to improve my cross validation score uh at the end uh to simplify my model I removed the time series feature and I just used a single uh light GBM model for final prediction and with this uh simplified approach uh I got the score of uh uh 0.573 in my private leader board which I could assume a solid score to uh have a overview of the whole competition. uh I believe that uh when the when we have evaluation metrics like coins copper and where the selecting the right threshold could significantly impact the result. Uh I think uh the winning approach could often involves redu reducing the randomness of metrics.

12:42 uh I saw a lot of participants uh uh use uh different seeds uh to uh reduce this randomness but I tried to use as much as as much model I could to uh reduce this randomness and because that advantage I had in this competition I was I think was my prayer experience in insemble method uh which I use here again and uh I think that uh ignoring the threshold optimization also was a big advantage for me. uh most fun aspect of this competition uh uh which I believe that was very challenging and I think still there is a lot of room to work on it was the time series data and uh how we can uh extract useful information useful information from that data uh I think it was very interesting um so thank you for your attention And I am uh ready to answer any questions.

Questions about this episode

Who speaks in "Kaggle Winners Walkthroughs: Child Mind Institute—Problematic Internet Use with Team Peyman Armaghan"?

Peyman Armaghan (Host).

What is "Kaggle Winners Walkthroughs: Child Mind Institute—Problematic Internet Use with Team Peyman Armaghan" about?

This walkthrough details Team Peyman Armaghan's innovative approach to analyzing problematic internet use through time series data in a competition organized by the Child Mind Institute. The team employed clustering algorithms to extract valuable features from user activity data, leveraging their background in telecommunications and economics to enhance their understanding of AI and machine…

What is mentioned in this Kaggle episode?

3 people, works and ideas are mentioned, including Cayman's clustering algorithm, PCIT total, light GBM. Each one on this page links to its own page with every episode that mentions it.

How long is "Kaggle Winners Walkthroughs: Child Mind Institute—Problematic Internet Use with Team Peyman Armaghan"?

The episode runs 14 min; the transcript is about 1,524 words in 11 timestamped sections.

More from Kaggle

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