In the test set there are approximately 10000 new users i.e. users not seen in train set. So the question is what products do we recommend for unseen users? This is called cold start problem in recommender systems.
\nReinforcement Learning does not offer a solution for this challenge. The path that I want to take to solve this problem is to study the first purchase behaviour of all past online customers. If you are interested see link notebook.
In the test set there are approximately 10000 new users i.e. users not seen in train set. So the question is what products do we recommend for unseen users? This is called cold start problem in recommender systems.
\nReinforcement Learning does not offer a solution for this challenge. The path that I want to take to solve this problem is to study the first purchase behaviour of all past online customers. If you are interested see link notebook.
I think it is also important to understand the behaviors of users who have fewer purchases( 1 or 2) as well. I don't think a single model could be used for all users.
", "rawMarkdown": "I think it is also important to understand the behaviors of users who have fewer purchases( 1 or 2) as well. I don't think a single model could be used for all users.", "votes": null }, { "id": "1702475", "postDate": "02/23/2022 16:28:35", "content": "One of the common solutions to cold start problem is called General TOP. To users without any purchases (or with just a couple of purchases) we recommend items that were bought the most number of times.
\nHere is a code for it:
general_top = transactions['article_id'].value_counts().head(12)\nand this is what sample_submission.csv contains for every user. General TOP considered a good baseline in recommender system task.
I agree with you @nroman that General Top is a reasonable baseline. What I am suggesting is that for new users it is even better to look at the General Top of the first purchase of the current customer base. In this way we exclude the preferences of the returning customers.
Adding to this topic approx. 1k items also suffer from cold start (articles that are provided in articles.csv but no transaction history).
Do you have any insights on the cold items that you can share? As stated by the competition data description those new articles are indeed available in the test week.
", "rawMarkdown": "Adding to this topic approx. 1k items also suffer from cold start (articles that are provided in `articles.csv` but no transaction history).\n\nDo you have any insights on the cold items that you can share? As stated by the competition data description those new articles are indeed available in the test week.", "votes": null }, { "id": "2020356", "postDate": "11/07/2022 11:46:44", "content": "Reinforcement learning is not the solution. You can use clustering and association rule
", "rawMarkdown": "Reinforcement learning is not the solution. You can use clustering and association rule", "votes": null } ], "comments": [ { "id": 1701924, "author_name": "ravi2493", "author_url": "", "post_date": "02/23/2022 07:07:16", "content": "I think it is also important to understand the behaviors of users who have fewer purchases( 1 or 2) as well. I don't think a single model could be used for all users.
", "votes": null, "replies": [] }, { "id": 1702475, "author_name": "nroman", "author_url": "", "post_date": "02/23/2022 16:28:35", "content": "One of the common solutions to cold start problem is called General TOP. To users without any purchases (or with just a couple of purchases) we recommend items that were bought the most number of times.
\nHere is a code for it:
general_top = transactions['article_id'].value_counts().head(12)\nand this is what sample_submission.csv contains for every user. General TOP considered a good baseline in recommender system task.
I agree with you @nroman that General Top is a reasonable baseline. What I am suggesting is that for new users it is even better to look at the General Top of the first purchase of the current customer base. In this way we exclude the preferences of the returning customers.
Adding to this topic approx. 1k items also suffer from cold start (articles that are provided in articles.csv but no transaction history).
Do you have any insights on the cold items that you can share? As stated by the competition data description those new articles are indeed available in the test week.
", "votes": null, "replies": [] }, { "id": 2020356, "author_name": "syednasirhasan", "author_url": "", "post_date": "11/07/2022 11:46:44", "content": "Reinforcement learning is not the solution. You can use clustering and association rule
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1701047": "In the `test` set there are approximately `10000` new users i.e. users not seen in `train` set. So the question is what products do we recommend for unseen users? This is called [`cold start`](https://en.wikipedia.org/wiki/Cold_start_(recommender_systems)) problem in recommender systems.\n`Reinforcement Learning` does not offer a solution for this challenge. The path that I want to take to solve this problem is to study the **first** purchase behaviour of all past online customers. If you are interested see link [notebook](https://www.kaggle.com/wti200/cold-start?scriptVersionId=88655405&cellId=12).", "1701924": "I think it is also important to understand the behaviors of users who have fewer purchases( 1 or 2) as well. I don't think a single model could be used for all users.", "1702475": "One of the common solutions to `cold start` problem is called `General TOP`. To users without any purchases (or with just a couple of purchases) we recommend items that were bought the most number of times.\nHere is a code for it:\n```\ngeneral_top = transactions['article_id'].value_counts().head(12)\n```\n\nand this is what `sample_submission.csv` contains for every user. General TOP considered a good baseline in recommender system task.", "1702639": "I agree with you @nroman that `General Top` is a reasonable baseline. What I am suggesting is that for new users it is even better to look at the `General Top` of the `first purchase` of the current customer base. In this way we exclude the preferences of the returning customers.", "1726007": "Adding to this topic approx. 1k items also suffer from cold start (articles that are provided in `articles.csv` but no transaction history).\n\nDo you have any insights on the cold items that you can share? As stated by the competition data description those new articles are indeed available in the test week.", "2020356": "Reinforcement learning is not the solution. You can use clustering and association rule" }, "source": "meta" }