![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2666506%2F776717d6d2c5f03f5cba80beacc8e936%2FCapture%20decran%202023-02-06%20a%202.33.37%20PM.png?generation=1675690426019369&alt=media\")
Hello everyone, first of all thanks for organizing this fun competition. I write this in order to introduce you my basic solution.
\nSpecial thanks to @cdeotte , @radek1, @buumoo and @gunesevitan for taking the time answering my multiples topics questions, I learned a lot from you all !
\n
As you can see I principally focused on a graph based solution. I can say that I've lost a big part of my time in this party because I over tuned it, while my mistake was a tippo I talked about here that I've noticed the last day of the competition. I also think that I didn't work that much on the candidate generation.
\nKaggle kernels
", "messages": [ { "id": "2131926", "postDate": "02/06/2023 13:42:33", "content": "Hello everyone, first of all thanks for organizing this fun competition. I write this in order to introduce you my basic solution.
\nSpecial thanks to @cdeotte , @radek1, @buumoo and @gunesevitan for taking the time answering my multiples topics questions, I learned a lot from you all !
\n
As you can see I principally focused on a graph based solution. I can say that I've lost a big part of my time in this party because I over tuned it, while my mistake was a tippo I talked about here that I've noticed the last day of the competition. I also think that I didn't work that much on the candidate generation.
\nKaggle kernels
", "rawMarkdown": "Hello everyone, first of all thanks for organizing this fun competition. I write this in order to introduce you my basic solution.\n\n### Credits\nSpecial thanks to @cdeotte , @radek1, @buumoo and @gunesevitan for taking the time answering my multiples topics questions, I learned a lot from you all !\n\n#### infos:\n- Extract 100 candidates per session. ( using only one co-visitation matrix )\n- Worked only on CARTS, I didn't run models for orders and clicks.\n\n### My pipeline\n\n\nAs you can see I principally focused on a graph based solution. I can say that I've lost a big part of my time in this party because I over tuned it, while my mistake was a tippo I talked about [here](https://www.kaggle.com/competitions/otto-recommender-system/discussion/382277) that I've noticed the last day of the competition. I also think that I didn't work that much on the candidate generation.\n## Some Remarks :\n\n- I noticed from my feature selection part that the graph features were at the top of the gain contribution of the model.\n- It's better to compute distances between embeddings, that throwing them as variables to the model.\n- I should use more co-visitation matrices ( here I used only one which were clicks2carts)\n- I should've run a model for orders and carts as well.\n\n## My mistakes\n\n- Extreme focus on the graph algorithms.\n- Didn't take much time on the candidate generation part.\n\n\n\n#### Setup:\nKaggle kernels", "votes": null }, { "id": "2131947", "postDate": "02/06/2023 13:54:26", "content": "Great job @rayanaay! And thanks for sharing your experience!
", "rawMarkdown": "Great job @rayanaay! And thanks for sharing your experience!", "votes": null } ], "comments": [ { "id": 2131947, "author_name": "tahamhaider", "author_url": "", "post_date": "02/06/2023 13:54:26", "content": "Great job @rayanaay! And thanks for sharing your experience!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2131926": "Hello everyone, first of all thanks for organizing this fun competition. I write this in order to introduce you my basic solution.\n\n### Credits\nSpecial thanks to @cdeotte , @radek1, @buumoo and @gunesevitan for taking the time answering my multiples topics questions, I learned a lot from you all !\n\n#### infos:\n- Extract 100 candidates per session. ( using only one co-visitation matrix )\n- Worked only on CARTS, I didn't run models for orders and clicks.\n\n### My pipeline\n\n\nAs you can see I principally focused on a graph based solution. I can say that I've lost a big part of my time in this party because I over tuned it, while my mistake was a tippo I talked about [here](https://www.kaggle.com/competitions/otto-recommender-system/discussion/382277) that I've noticed the last day of the competition. I also think that I didn't work that much on the candidate generation.\n## Some Remarks :\n\n- I noticed from my feature selection part that the graph features were at the top of the gain contribution of the model.\n- It's better to compute distances between embeddings, that throwing them as variables to the model.\n- I should use more co-visitation matrices ( here I used only one which were clicks2carts)\n- I should've run a model for orders and carts as well.\n\n## My mistakes\n\n- Extreme focus on the graph algorithms.\n- Didn't take much time on the candidate generation part.\n\n\n\n#### Setup:\nKaggle kernels", "2131947": "Great job @rayanaay! And thanks for sharing your experience!" }, "source": "meta" }