{ "id": 460911, "title": "34th Place Solution Writeup for Open Problems – Single-Cell Perturbations", "url": "/competitions/open-problems-single-cell-perturbations/discussion/460911", "author_name": "Octopus210", "post_date": "2023-12-11T17:58:53.632000", "votes": 4, "comment_count": 0, "views": 0, "content": "

First of all, I would like to express my gratitude to the organizers and kagglers of this competition. I learned a lot from the open notebooks and discussions.
\nEspecially, I would like to express my big thanks to the following people.
\n@mehrankazeminia
\n@somayyehgholami
\n@alexandervc
\n@antoninadolgorukova
\n@kishanvavdara
\n@pablormier

\n

34th Solution

\n

We blended the results of three tasks to find one result. Being different approaches, blending these models was very effective.

\n

The three approaches
\nTask1 : test set as categorical variable
\nTask2 : test set as continuous variable ( genes as sampled)
\nTask3 : test set as continuous variable ( compounds as sample)

\n

\"34th_solution\"

\n

Basically, it is a simple solution that blends public notebook's score and sklearn's model. If you combine it with a better performing model (like PY-BOOST), the performance will be even better.

\n
    \n
  1. integration of biological knowledge
    \nIn Task 3, we incorporated the decoupler. A small improvement in public scores was obtained.
    2. \n
\n

2.Exploration of the problem
\nIn Task 2 and Task 3, CD8 data were excluded. op2-about-positive-control

\n

3.Model design
\nI used sklearn's MLP, lightgbm, and Ridge.

\n

4.Robustness
\nRobustness is considered to be high, because I sought a single result from three different perspectives.

\n

5.Documentation & code style
\nIn preparation

\n

6.Reproducibility
\nIn preparation

", "messages": [ { "id": 2557896, "postDate": "2023-12-11T17:58:53.633Z", "content": "

First of all, I would like to express my gratitude to the organizers and kagglers of this competition. I learned a lot from the open notebooks and discussions.
\nEspecially, I would like to express my big thanks to the following people.
\n@mehrankazeminia
\n@somayyehgholami
\n@alexandervc
\n@antoninadolgorukova
\n@kishanvavdara
\n@pablormier

\n

34th Solution

\n

We blended the results of three tasks to find one result. Being different approaches, blending these models was very effective.

\n

The three approaches
\nTask1 : test set as categorical variable
\nTask2 : test set as continuous variable ( genes as sampled)
\nTask3 : test set as continuous variable ( compounds as sample)

\n

\"34th_solution\"

\n

Basically, it is a simple solution that blends public notebook's score and sklearn's model. If you combine it with a better performing model (like PY-BOOST), the performance will be even better.

\n
    \n
  1. integration of biological knowledge
    \nIn Task 3, we incorporated the decoupler. A small improvement in public scores was obtained.
    2. \n
\n

2.Exploration of the problem
\nIn Task 2 and Task 3, CD8 data were excluded. op2-about-positive-control

\n

3.Model design
\nI used sklearn's MLP, lightgbm, and Ridge.

\n

4.Robustness
\nRobustness is considered to be high, because I sought a single result from three different perspectives.

\n

5.Documentation & code style
\nIn preparation

\n

6.Reproducibility
\nIn preparation

", "rawMarkdown": "First of all, I would like to express my gratitude to the organizers and kagglers of this competition. I learned a lot from the open notebooks and discussions.\nEspecially, I would like to express my big thanks to the following people.\n@mehrankazeminia\n@somayyehgholami\n@alexandervc\n@antoninadolgorukova\n@kishanvavdara\n@pablormier\n\n\n\n**34th Solution**\n\nWe blended the results of three tasks to find one result. Being different approaches, blending these models was very effective.\n\nThe three approaches\nTask1 : test set as categorical variable\nTask2 : test set as continuous variable ( genes as sampled)\nTask3 : test set as continuous variable ( compounds as sample)\n\n![34th_solution](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2646279%2Fb2df8d249afa3fa8af194ba5d14f3438%2FOC2_34th.png?generation=1702317451136015&alt=media)\n\nBasically, it is a simple solution that blends public notebook's score and sklearn's model. If you combine it with a better performing model (like PY-BOOST), the performance will be even better.\n\n1. integration of biological knowledge\nIn Task 3, we incorporated the decoupler. A small improvement in public scores was obtained. \n\n2.Exploration of the problem\nIn Task 2 and Task 3, CD8 data were excluded. [op2-about-positive-control](https://www.kaggle.com/code/yoshifumimiya/op2-about-positive-control)\n\n3.Model design\nI used sklearn's MLP, lightgbm, and Ridge. \n\n4.Robustness\nRobustness is considered to be high, because I sought a single result from three different perspectives. \n\n5.Documentation & code style\nIn preparation\n\n6.Reproducibility\nIn preparation", "votes": 4 } ], "comments": [], "raw_markdown_by_id": { "2557896": "First of all, I would like to express my gratitude to the organizers and kagglers of this competition. I learned a lot from the open notebooks and discussions.\nEspecially, I would like to express my big thanks to the following people.\n@mehrankazeminia\n@somayyehgholami\n@alexandervc\n@antoninadolgorukova\n@kishanvavdara\n@pablormier\n\n\n\n**34th Solution**\n\nWe blended the results of three tasks to find one result. Being different approaches, blending these models was very effective.\n\nThe three approaches\nTask1 : test set as categorical variable\nTask2 : test set as continuous variable ( genes as sampled)\nTask3 : test set as continuous variable ( compounds as sample)\n\n![34th_solution](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2646279%2Fb2df8d249afa3fa8af194ba5d14f3438%2FOC2_34th.png?generation=1702317451136015&alt=media)\n\nBasically, it is a simple solution that blends public notebook's score and sklearn's model. If you combine it with a better performing model (like PY-BOOST), the performance will be even better.\n\n1. integration of biological knowledge\nIn Task 3, we incorporated the decoupler. A small improvement in public scores was obtained. \n\n2.Exploration of the problem\nIn Task 2 and Task 3, CD8 data were excluded. [op2-about-positive-control](https://www.kaggle.com/code/yoshifumimiya/op2-about-positive-control)\n\n3.Model design\nI used sklearn's MLP, lightgbm, and Ridge. \n\n4.Robustness\nRobustness is considered to be high, because I sought a single result from three different perspectives. \n\n5.Documentation & code style\nIn preparation\n\n6.Reproducibility\nIn preparation" } }