I recently experimented with Data Augmentation. I created pairs of sessions with the same label values for each level_group, then cut each session at random level points and combined them. With this method, I doubled the training data and trained my NN model, but the score wasn't improved.
\nIn creating pairs for Data Augmentation, I generated several features such as elapsed_time_diff aggregation feature and clustering for each session, and creating pairs within the same cluster. However, there is still a difference in distribution between the augmented data and the original training data, so I am using Adversarial Validation to select samples close to the training data. But, It did not work.
\nHas anyone had success with a similar method for Data Augmentation?
", "messages": [ { "id": "2235911", "postDate": "04/26/2023 12:20:47", "content": "I recently experimented with Data Augmentation. I created pairs of sessions with the same label values for each level_group, then cut each session at random level points and combined them. With this method, I doubled the training data and trained my NN model, but the score wasn't improved.
\nIn creating pairs for Data Augmentation, I generated several features such as elapsed_time_diff aggregation feature and clustering for each session, and creating pairs within the same cluster. However, there is still a difference in distribution between the augmented data and the original training data, so I am using Adversarial Validation to select samples close to the training data. But, It did not work.
\nHas anyone had success with a similar method for Data Augmentation?
", "rawMarkdown": "I recently experimented with Data Augmentation. I created pairs of sessions with the same label values for each level_group, then cut each session at random level points and combined them. With this method, I doubled the training data and trained my NN model, but the score wasn't improved.\n\nIn creating pairs for Data Augmentation, I generated several features such as elapsed_time_diff aggregation feature and clustering for each session, and creating pairs within the same cluster. However, there is still a difference in distribution between the augmented data and the original training data, so I am using Adversarial Validation to select samples close to the training data. But, It did not work.\n\nHas anyone had success with a similar method for Data Augmentation?", "votes": null }, { "id": "2238208", "postDate": "04/28/2023 09:59:54", "content": "For me, I simply concatenate the data from sessions. E.g., to predict questions 1 to 3, I use level 0-4, to predict questions 4-13, I use level 0-4 and 5-12 (not only 5-12), and to predict question 14-18, I use all levels. Of course, model with the last level should take longer sequences. But, as like you, it doesn't seem to help in my case (CV: 0.6937 -> 0.6938 for 1 fold with a transformer, I think it's just noise)
", "rawMarkdown": "For me, I simply concatenate the data from sessions. E.g., to predict questions 1 to 3, I use level 0-4, to predict questions 4-13, I use level 0-4 and 5-12 (not only 5-12), and to predict question 14-18, I use all levels. Of course, model with the last level should take longer sequences. But, as like you, it doesn't seem to help in my case (CV: 0.6937 -> 0.6938 for 1 fold with a transformer, I think it's just noise)", "votes": null }, { "id": "2258237", "postDate": "05/14/2023 03:18:10", "content": "This is a very interesting topic. I think the big problem is that the number of sessions is only about 20,000, especially when training with NNs.
", "rawMarkdown": "This is a very interesting topic. I think the big problem is that the number of sessions is only about 20,000, especially when training with NNs.", "votes": null }, { "id": "2260665", "postDate": "05/15/2023 19:35:45", "content": "I tried a somewhat similar approach doing linear interpolation between sessions with similar answers and consistently got a worse score.
", "rawMarkdown": "I tried a somewhat similar approach doing linear interpolation between sessions with similar answers and consistently got a worse score.", "votes": null } ], "comments": [ { "id": 2238208, "author_name": "shinomoriaoshi", "author_url": "", "post_date": "04/28/2023 09:59:54", "content": "For me, I simply concatenate the data from sessions. E.g., to predict questions 1 to 3, I use level 0-4, to predict questions 4-13, I use level 0-4 and 5-12 (not only 5-12), and to predict question 14-18, I use all levels. Of course, model with the last level should take longer sequences. But, as like you, it doesn't seem to help in my case (CV: 0.6937 -> 0.6938 for 1 fold with a transformer, I think it's just noise)
", "votes": null, "replies": [] }, { "id": 2258237, "author_name": "jinmiyashita", "author_url": "", "post_date": "05/14/2023 03:18:10", "content": "This is a very interesting topic. I think the big problem is that the number of sessions is only about 20,000, especially when training with NNs.
", "votes": null, "replies": [] }, { "id": 2260665, "author_name": "sakvaua", "author_url": "", "post_date": "05/15/2023 19:35:45", "content": "I tried a somewhat similar approach doing linear interpolation between sessions with similar answers and consistently got a worse score.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2235911": "I recently experimented with Data Augmentation. I created pairs of sessions with the same label values for each level_group, then cut each session at random level points and combined them. With this method, I doubled the training data and trained my NN model, but the score wasn't improved.\n\nIn creating pairs for Data Augmentation, I generated several features such as elapsed_time_diff aggregation feature and clustering for each session, and creating pairs within the same cluster. However, there is still a difference in distribution between the augmented data and the original training data, so I am using Adversarial Validation to select samples close to the training data. But, It did not work.\n\nHas anyone had success with a similar method for Data Augmentation?", "2238208": "For me, I simply concatenate the data from sessions. E.g., to predict questions 1 to 3, I use level 0-4, to predict questions 4-13, I use level 0-4 and 5-12 (not only 5-12), and to predict question 14-18, I use all levels. Of course, model with the last level should take longer sequences. But, as like you, it doesn't seem to help in my case (CV: 0.6937 -> 0.6938 for 1 fold with a transformer, I think it's just noise)", "2258237": "This is a very interesting topic. I think the big problem is that the number of sessions is only about 20,000, especially when training with NNs.", "2260665": "I tried a somewhat similar approach doing linear interpolation between sessions with similar answers and consistently got a worse score." }, "source": "meta" }