This article is for members only. If you register as a member (free), you can see everything.
Table of Contents
In this series, four Kaggle Masters from Recruit will explain how Kaggle works, how to work with it, and how to go one step ahead from a beginner. In Part 1, for those who are thinking of starting Kaggle, I explained how Kaggle works, how to work after actually participating in the competition, and recommended content for beginners. In Part 2, the author explains "what he actually did" to raise his rank, using actual examples.
And in this 3rd installment, we will report on the round-table discussion by four Kaggle Masters, with the theme of "Different from other teams - Winning medals, the path to becoming a Kaggle Master". What kind of points and strategies should we be aware of in order to achieve top prizes? As a gold medalist, I would like to give you practical advice for becoming a Kaggle Master, as well as a message for those who want to become a Kaggle Master.
The first theme is about 'walls and realizations that should be overcome' for the top prizes. Mr. Hirotake Auchi started the discussion. “In most competitions, high-scoring models are published as baselines in discussions and notebooks, so if you use that as a reference, the model will score about 80 points out of 100. However, 80 points. In order to win the gold medal with 100 points, I think that it is essential to add "your own ingenuity" in addition to the information that has already been published.Personally, ideas that other data scientists are unlikely to do I believe that the difference between gold and silver is whether you can build your own original model.”
Receiving that, Mr. Fuyusei Hatori said, "Indeed, I don't think there are many original elements that are the 'last push' in discussions and notebooks. Only those who noticed it It can be said that the phenomenon in which the score is increasing by one head is common in competitions,” he clearly states the elements necessary for winning high ranks based on his participation experience.
Kento Obata points out that trying everything that other data scientists are doing is the minimum requirement for aiming for top prizes. “The stance of checking all the information published in discussions and notebooks, researching everything, and then trying everything that comes to mind is common to those who win gold medals. However, on rare occasions, I see a model that uses an approach that is full of originality, but I believe that reproducibility is difficult.”
About the Kaggle-sponsored Mechanisms of Action Prediction (MoA) competition, which won a gold medal this time, Mr. Auchi said, "All the members rationally accumulated the results of 'If you do this, this will happen.' As a result, I think it led to the acquisition of the gold medal.In particular, the point that demonstrated the original element was the use of a model that was well analyzed in the previous competition, which was not posted in the discussion," said Mr. Hatori. The result of steadily testing all past models has come to fruition,” he recalls.
Fig. 1 Requirements for winning a gold medal in the Kaggle competition (cited from https://www.kaggle.com/progression) I asked about "Tsu)". Mr. Akira Sasaki, who is in charge of improving the recommendation system and search system in the human resources area within the company, says, "First of all, be aware of trying to secure time to touch data even a little during the day. ``You need to have the enthusiasm to concentrate and work on data analysis with about the same momentum,'' he answered. On top of that, participating as a team gave them the feeling of wanting to contribute to the team, which became a great motivation to persevere through the competition. In response to this, Mr. Hatori said, "Participating as a team is good because the excitement of the day before the school festival lasts for a while."Mr. Auchi shared an episode in which he was able to concentrate on preparations using the company system. "I was able to fully commit to the Kaggle competition by utilizing the company's STEP leave system, which allows me to take up to 28 days of leave once every three years of service."
Mr. Obata used to incorporate the knowledge gained from competitions into his company's group as a mission in the past, but he said that he did not incorporate it in this Kaggle competition. The reason for this is, "I want to clearly separate what I can do at Kaggle and what I can do at work. If you like data analysis, you can enjoy it even outside of work and commit to the Kaggle competition. It will end up."
Obata says that many people in the group he belongs to participate in Kaggle competitions, and he felt a sense of danger that he would not be able to keep up with his work if he didn't participate, so he started participating. Mr. Miss. “Because I started participating with everyone, I was able to build a certain amount of groundwork, and little by little I was able to keep up with conversations with project members. I'm confident.
Following this story, Mr. Hatori said that two months after the employee who joined the company as a new graduate in April 2021 deployed it to an actual online service using a solution that he had tried in a past Kaggle competition. Presenting examples of significant successes in On top of that, he feels that the experience and knowledge he has cultivated at Kaggle has been put to good use in his work. “In data analysis work, it is important to have an analysis design that presents what kind of analysis to proceed with what kind of population based on what kind of plan. There is an increasing number of cases in which we apply this design to actual work and discuss it based on it. There is a difference, and by participating in the Kaggle competition, I strongly feel that such 'points' have been strengthened."
The four Kaggle Masters have stated that they will continue to actively participate in Kaggle, which has realized these benefits, and also said that they would like to aim for the highest rank of Kaggle Grandmaster. . Hatori mentioned two Kaggle Grandmasters who joined the company as new graduates in 2021. “There are only a few hundred Grandmasters in the world, and it is no exaggeration to say that they are geniuses at the highest level.”
Mr. Auchi said, "When you become a Grandmaster, the level of how you approach the competition is completely different. Therefore, if you aim to be a Grandmaster, be sure to look back after the competition and make use of it yourself. I think it is necessary to make various efforts, such as properly organizing and summarizing the models published in notebooks and discussions and the methods that have actually been tried,” he said, presenting the “road to Grandmaster”. On top of that, ``To become a Grandmaster, you need 5 gold medals (one of which must be obtained solo), and even if you have plenty of time to devote to becoming a Grandmaster, it won't be possible with just that. It will be difficult,” he adds.
On top of that, Mr. Auchi introduced an episode in which the two Grandmasters played a central role in holding an in-house “Recruit Kaggle Competition.” Four Kaggle Masters also participated in the competition, competing in the accuracy of a model that predicts whether the content of word of mouth is positive or negative from the word of mouth log of the in-house service. “Even though I was not involved in this service, it was refreshing to be involved in building the model, and I had a very meaningful experience. If someone in the company was pleased with the model I presented, I would be happy. So, it made me feel excited,” Hatori said happily.
Fig. 2 Conditions for becoming Grandmaster of the Kaggle competition (cited from https://www.kaggle.com/progression) To view the rest, you need to enable JavaScript in your browser settings.Copyright© Digital Advantage Corp. All Rights Reserved.