I have seen that some people treat this task as a classification problem and other as a regression.
\n\nSo I wonder if it would be a good idea to use keras functional api and make one model with an output being a multilabel classifier and the other output as a regression, each output would have its appropriate loss function.
\n\nI am new to this field, so I don't know if it will improve the predictions.
\n\nIf someone know something about this, I would be glad to read his/her opinion.
", "messages": [ { "id": "600144", "postDate": "08/15/2019 17:44:02", "content": "I have seen that some people treat this task as a classification problem and other as a regression.
\n\nSo I wonder if it would be a good idea to use keras functional api and make one model with an output being a multilabel classifier and the other output as a regression, each output would have its appropriate loss function.
\n\nI am new to this field, so I don't know if it will improve the predictions.
\n\nIf someone know something about this, I would be glad to read his/her opinion.
", "rawMarkdown": "I have seen that some people treat this task as a classification problem and other as a regression.\n\nSo I wonder if it would be a good idea to use keras functional api and make one model with an output being a multilabel classifier and the other output as a regression, each output would have its appropriate loss function.\n\nI am new to this field, so I don't know if it will improve the predictions.\n\nIf someone know something about this, I would be glad to read his/her opinion.", "votes": null }, { "id": "600600", "postDate": "08/16/2019 10:38:55", "content": "I found this https://arxiv.org/pdf/1808.10564.pdf. \nIt may be helpful for you.
", "rawMarkdown": "I found this https://arxiv.org/pdf/1808.10564.pdf. \nIt may be helpful for you.", "votes": null }, { "id": "600671", "postDate": "08/16/2019 12:47:22", "content": "Did you try this joint learning? It works for you?
", "rawMarkdown": "Did you try this joint learning? It works for you?", "votes": null }, { "id": "600774", "postDate": "08/16/2019 14:39:31", "content": "Very interesting paper, this will definitely help, thank you.
", "rawMarkdown": "Very interesting paper, this will definitely help, thank you.", "votes": null }, { "id": "600822", "postDate": "08/16/2019 15:55:51", "content": "I've tried what that paper did for the bottle neck layers and fair number of experiments to go with. At one point I used five contributions, four losses for the classification head, and a single MSE for the regression. I found it didn't offer any significant advantages to score boost over either a regression or classification head independently with single losses for each, and results were within the uncertainty of changing different seeds to add to that uncertainty.
\n\n\"Further, we find that scores based evaluation usually corresponds to better performance, so we use scores based evaluation.\"
\n\nFor some of my experiments I found the above statement true, other experiments false. I could also prove or disprove either depending on what approach I used for processing/combining scores. In conclusion, the majority of large score boosts are likely coming from other approaches and procedures.
", "rawMarkdown": "I've tried what that paper did for the bottle neck layers and fair number of experiments to go with. At one point I used five contributions, four losses for the classification head, and a single MSE for the regression. I found it didn't offer any significant advantages to score boost over either a regression or classification head independently with single losses for each, and results were within the uncertainty of changing different seeds to add to that uncertainty. \n\n*\"Further, we find that scores based evaluation usually corresponds to better performance, so we use scores based evaluation.\"*\n\nFor some of my experiments I found the above statement true, other experiments false. I could also prove or disprove either depending on what approach I used for processing/combining scores. In conclusion, the majority of large score boosts are likely coming from other approaches and procedures.", "votes": null }, { "id": "600826", "postDate": "08/16/2019 15:59:37", "content": "Thank you for your complete response which will save me time, I will keep the idea for an other project. And keep it simple for this project
", "rawMarkdown": "Thank you for your complete response which will save me time, I will keep the idea for an other project. And keep it simple for this project", "votes": null }, { "id": "602174", "postDate": "08/18/2019 17:49:39", "content": "same result with gabriel. when I tried it , it hurts my model's performance. :(\nthough there are some techniques for multi task learning, I'm not sure it worth.\nbut, I'm curious high score of Chanhu is come from this.
", "rawMarkdown": "same result with gabriel. when I tried it , it hurts my model's performance. :(\nthough there are some techniques for multi task learning, I'm not sure it worth.\nbut, I'm curious high score of Chanhu is come from this.", "votes": null } ], "comments": [ { "id": 600600, "author_name": "chanhu", "author_url": "", "post_date": "08/16/2019 10:38:55", "content": "I found this https://arxiv.org/pdf/1808.10564.pdf. \nIt may be helpful for you.
", "votes": null, "replies": [ { "id": 600671, "author_name": "vanche", "author_url": "", "post_date": "08/16/2019 12:47:22", "content": "Did you try this joint learning? It works for you?
", "votes": null, "replies": [] }, { "id": 600774, "author_name": "jonasfreibs", "author_url": "", "post_date": "08/16/2019 14:39:31", "content": "Very interesting paper, this will definitely help, thank you.
", "votes": null, "replies": [] }, { "id": 600822, "author_name": "gabrielhae", "author_url": "", "post_date": "08/16/2019 15:55:51", "content": "I've tried what that paper did for the bottle neck layers and fair number of experiments to go with. At one point I used five contributions, four losses for the classification head, and a single MSE for the regression. I found it didn't offer any significant advantages to score boost over either a regression or classification head independently with single losses for each, and results were within the uncertainty of changing different seeds to add to that uncertainty.
\n\n\"Further, we find that scores based evaluation usually corresponds to better performance, so we use scores based evaluation.\"
\n\nFor some of my experiments I found the above statement true, other experiments false. I could also prove or disprove either depending on what approach I used for processing/combining scores. In conclusion, the majority of large score boosts are likely coming from other approaches and procedures.
", "votes": null, "replies": [] }, { "id": 600826, "author_name": "jonasfreibs", "author_url": "", "post_date": "08/16/2019 15:59:37", "content": "Thank you for your complete response which will save me time, I will keep the idea for an other project. And keep it simple for this project
", "votes": null, "replies": [] }, { "id": 602174, "author_name": "vanche", "author_url": "", "post_date": "08/18/2019 17:49:39", "content": "same result with gabriel. when I tried it , it hurts my model's performance. :(\nthough there are some techniques for multi task learning, I'm not sure it worth.\nbut, I'm curious high score of Chanhu is come from this.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "600144": "I have seen that some people treat this task as a classification problem and other as a regression.\n\nSo I wonder if it would be a good idea to use keras functional api and make one model with an output being a multilabel classifier and the other output as a regression, each output would have its appropriate loss function.\n\nI am new to this field, so I don't know if it will improve the predictions.\n\nIf someone know something about this, I would be glad to read his/her opinion.", "600600": "I found this https://arxiv.org/pdf/1808.10564.pdf. \nIt may be helpful for you.", "600671": "Did you try this joint learning? It works for you?", "600774": "Very interesting paper, this will definitely help, thank you.", "600822": "I've tried what that paper did for the bottle neck layers and fair number of experiments to go with. At one point I used five contributions, four losses for the classification head, and a single MSE for the regression. I found it didn't offer any significant advantages to score boost over either a regression or classification head independently with single losses for each, and results were within the uncertainty of changing different seeds to add to that uncertainty. \n\n*\"Further, we find that scores based evaluation usually corresponds to better performance, so we use scores based evaluation.\"*\n\nFor some of my experiments I found the above statement true, other experiments false. I could also prove or disprove either depending on what approach I used for processing/combining scores. In conclusion, the majority of large score boosts are likely coming from other approaches and procedures.", "600826": "Thank you for your complete response which will save me time, I will keep the idea for an other project. And keep it simple for this project", "602174": "same result with gabriel. when I tried it , it hurts my model's performance. :(\nthough there are some techniques for multi task learning, I'm not sure it worth.\nbut, I'm curious high score of Chanhu is come from this." }, "source": "meta" }