{ "id": 368355, "title": "Graph Neural Netwok , is complex to model but efficient for this task ", "url": "/competitions/otto-recommender-system/discussion/368355", "author_name": "Youness EL BRAG", "post_date": "2022-11-24T19:31:01.799000", "votes": 15, "comment_count": 9, "views": 0, "content": "

The problem of session-based recommendation aims to predict user actions based on anonymous sessions. Previous methods model a session as a sequence and estimate user representations besides item representations to make recommendations. Though achieved promising results, they are insufficient to obtain accurate user vectors in sessions and neglect complex transitions of items. To obtain accurate item embed-
\nding and take complex transitions of items into account, method, i.e. Session-based Recommendation with Graph Neural Networks\n
\ncode

", "messages": [ { "id": 2042618, "postDate": "2022-11-24T19:31:01.800Z", "content": "

The problem of session-based recommendation aims to predict user actions based on anonymous sessions. Previous methods model a session as a sequence and estimate user representations besides item representations to make recommendations. Though achieved promising results, they are insufficient to obtain accurate user vectors in sessions and neglect complex transitions of items. To obtain accurate item embed-
\nding and take complex transitions of items into account, method, i.e. Session-based Recommendation with Graph Neural Networks\n
\ncode

", "rawMarkdown": "The problem of session-based recommendation aims to predict user actions based on anonymous sessions. Previous methods model a session as a sequence and estimate user representations besides item representations to make recommendations. Though achieved promising results, they are insufficient to obtain accurate user vectors in sessions and neglect complex transitions of items. To obtain accurate item embed-\nding and take complex transitions of items into account, method, i.e. [**Session-based Recommendation with Graph Neural Networks**\n](https://arxiv.org/abs/1811.00855)\n[code](https://github.com/CRIPAC-DIG/SR-GNN?utm_source=catalyzex.com)", "votes": 15 }, { "id": 2086414, "postDate": "2023-01-04T18:50:39.073Z", "content": "

With the SR-GNN model mentioned in the paper, each session with items [a1,a2,a3,…,an] will form the set (input, predict) = ([a1], [a2] ]), ([a1, a2], [a3]), ([a1,a2,a3],[a4]),… With the otto dataset, building such dataset is too large and close like can not load all on RAM. I tried a simpler construction, which is only interested in input with size<=2, eg ([a1,a2],[a3]),([a2,a3],[a4]) . Then with 100 GB RAM, hidden features = 100 and batchSize = 1024, I can train. But an epoch takes almost 12 hours, to train the model with the resources that Kaggle provides. That's some information I tested. Hope can be useful to everyone.

", "rawMarkdown": "With the SR-GNN model mentioned in the paper, each session with items [a1,a2,a3,...,an] will form the set (input, predict) = ([a1], [a2] ]), ([a1, a2], [a3]), ([a1,a2,a3],[a4]),... With the otto dataset, building such dataset is too large and close like can not load all on RAM. I tried a simpler construction, which is only interested in input with size<=2, eg ([a1,a2],[a3]),([a2,a3],[a4]) . Then with 100 GB RAM, hidden features = 100 and batchSize = 1024, I can train. But an epoch takes almost 12 hours, to train the model with the resources that Kaggle provides. That's some information I tested. Hope can be useful to everyone.", "votes": 3, "replies": [ { "id": 2087407, "postDate": "2023-01-05T15:18:58.080Z", "content": "

Thanks for your shaing, and how about the result?

", "rawMarkdown": "Thanks for your shaing, and how about the result?", "votes": 1 } ] }, { "id": 2085987, "postDate": "2023-01-04T13:46:47.580Z", "content": "

Hi Youness EL BRAG! I also use GNN but I cannot achieve good results, are you get good scores?

", "rawMarkdown": "Hi Youness EL BRAG! I also use GNN but I cannot achieve good results, are you get good scores?", "votes": 2, "replies": [ { "id": 2086375, "postDate": "2023-01-04T18:23:08.710Z", "content": "

Hello Barry, same for me, plus it's very slow to train but can be use to extract useful features I think

", "rawMarkdown": "Hello Barry, same for me, plus it's very slow to train but can be use to extract useful features I think", "votes": 2, "replies": [ { "id": 2086418, "postDate": "2023-01-04T18:52:58.073Z", "content": "

Totally agree with you. It takes a lot of time to train and configure the dataset to predict.

", "rawMarkdown": "Totally agree with you. It takes a lot of time to train and configure the dataset to predict.", "votes": 1 }, { "id": 2087403, "postDate": "2023-01-05T15:17:15.740Z", "content": "

Yes Rayan-aay, It also takes me a lot of time to train. I also agree with you.

", "rawMarkdown": "Yes Rayan-aay, It also takes me a lot of time to train. I also agree with you.", "votes": 1 } ] }, { "id": 2087516, "postDate": "2023-01-05T16:37:08.110Z", "content": "

actually i didn't implment itit yet , i just share the idea , but soon i will try to use Attention Graoh NN rather than GNN

", "rawMarkdown": "actually i didn't implment itit yet , i just share the idea , but soon i will try to use Attention Graoh NN rather than GNN " } ] }, { "id": 2057271, "postDate": "2022-12-06T23:15:08.197Z", "content": "

Great idea!

", "rawMarkdown": "Great idea!", "votes": 2 }, { "id": 2082491, "postDate": "2023-01-01T16:12:58.207Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2086414, "author_name": "Anh Bui", "author_url": "", "post_date": "2023-01-04T18:50:39.073000", "content": "

With the SR-GNN model mentioned in the paper, each session with items [a1,a2,a3,…,an] will form the set (input, predict) = ([a1], [a2] ]), ([a1, a2], [a3]), ([a1,a2,a3],[a4]),… With the otto dataset, building such dataset is too large and close like can not load all on RAM. I tried a simpler construction, which is only interested in input with size<=2, eg ([a1,a2],[a3]),([a2,a3],[a4]) . Then with 100 GB RAM, hidden features = 100 and batchSize = 1024, I can train. But an epoch takes almost 12 hours, to train the model with the resources that Kaggle provides. That's some information I tested. Hope can be useful to everyone.

", "votes": 3, "replies": [ { "id": 2087407, "author_name": "BarryZhou", "author_url": "", "post_date": "2023-01-05T15:18:58.080000", "content": "

Thanks for your shaing, and how about the result?

", "votes": 1, "replies": [] } ] }, { "id": 2085987, "author_name": "BarryZhou", "author_url": "", "post_date": "2023-01-04T13:46:47.580000", "content": "

Hi Youness EL BRAG! I also use GNN but I cannot achieve good results, are you get good scores?

", "votes": 2, "replies": [ { "id": 2086375, "author_name": "Rayan-aay", "author_url": "", "post_date": "2023-01-04T18:23:08.710000", "content": "

Hello Barry, same for me, plus it's very slow to train but can be use to extract useful features I think

", "votes": 2, "replies": [ { "id": 2086418, "author_name": "Anh Bui", "author_url": "", "post_date": "2023-01-04T18:52:58.073000", "content": "

Totally agree with you. It takes a lot of time to train and configure the dataset to predict.

", "votes": 1, "replies": [] }, { "id": 2087403, "author_name": "BarryZhou", "author_url": "", "post_date": "2023-01-05T15:17:15.740000", "content": "

Yes Rayan-aay, It also takes me a lot of time to train. I also agree with you.

", "votes": 1, "replies": [] } ] }, { "id": 2087516, "author_name": "Youness EL BRAG", "author_url": "", "post_date": "2023-01-05T16:37:08.110000", "content": "

actually i didn't implment itit yet , i just share the idea , but soon i will try to use Attention Graoh NN rather than GNN

", "votes": 0, "replies": [] } ] }, { "id": 2057271, "author_name": "Chris Deotte", "author_url": "", "post_date": "2022-12-06T23:15:08.197000", "content": "

Great idea!

", "votes": 2, "replies": [] }, { "id": 2082491, "author_name": "", "author_url": "", "post_date": "2023-01-01T16:12:58.207000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2042618": "The problem of session-based recommendation aims to predict user actions based on anonymous sessions. Previous methods model a session as a sequence and estimate user representations besides item representations to make recommendations. Though achieved promising results, they are insufficient to obtain accurate user vectors in sessions and neglect complex transitions of items. To obtain accurate item embed-\nding and take complex transitions of items into account, method, i.e. [**Session-based Recommendation with Graph Neural Networks**\n](https://arxiv.org/abs/1811.00855)\n[code](https://github.com/CRIPAC-DIG/SR-GNN?utm_source=catalyzex.com)", "2086414": "With the SR-GNN model mentioned in the paper, each session with items [a1,a2,a3,...,an] will form the set (input, predict) = ([a1], [a2] ]), ([a1, a2], [a3]), ([a1,a2,a3],[a4]),... With the otto dataset, building such dataset is too large and close like can not load all on RAM. I tried a simpler construction, which is only interested in input with size<=2, eg ([a1,a2],[a3]),([a2,a3],[a4]) . Then with 100 GB RAM, hidden features = 100 and batchSize = 1024, I can train. But an epoch takes almost 12 hours, to train the model with the resources that Kaggle provides. That's some information I tested. Hope can be useful to everyone.", "2085987": "Hi Youness EL BRAG! I also use GNN but I cannot achieve good results, are you get good scores?", "2057271": "Great idea!", "2082491": "" } }