Warning: I am a newb.
\n\nFor the past couple of days I have been messing around with AutoEncoders for Anomaly Detection, since they can sometimes produce good results with imbalanced data. But I can't get past the AUC=0.7 mark on my validation sets.
\n\nI use a simple architecture, with two Dense layers each for the Encoder/Decoder (pretty much this). I have used a couple of preprocessing schemes, like the one here.
\n\nHas anyone gotten it to work? Any tips you can share?
", "messages": [ { "id": "305433", "postDate": "03/28/2018 21:25:26", "content": "Warning: I am a newb.
\n\nFor the past couple of days I have been messing around with AutoEncoders for Anomaly Detection, since they can sometimes produce good results with imbalanced data. But I can't get past the AUC=0.7 mark on my validation sets.
\n\nI use a simple architecture, with two Dense layers each for the Encoder/Decoder (pretty much this). I have used a couple of preprocessing schemes, like the one here.
\n\nHas anyone gotten it to work? Any tips you can share?
", "rawMarkdown": "*Warning: I am a newb.*\n\nFor the past couple of days I have been messing around with AutoEncoders for Anomaly Detection, since they can sometimes produce good results with imbalanced data. But I can't get past the AUC=0.7 mark on my validation sets.\n\nI use a simple architecture, with two Dense layers each for the Encoder/Decoder (pretty much [this](https://medium.com/@curiousily/credit-card-fraud-detection-using-autoencoders-in-keras-tensorflow-for-hackers-part-vii-20e0c85301bd)). I have used a couple of preprocessing schemes, like the one [here](https://www.kaggle.com/joaopmpeinado/single-xgboost-lb-0-966/code).\n\nHas anyone gotten it to work? Any tips you can share?", "votes": null }, { "id": "305711", "postDate": "03/29/2018 09:55:31", "content": "Given Michael Jahrer is active here yo may see a great use of autoencoders. Check his winning solution in the Porto Seguro competition.
", "rawMarkdown": "Given Michael Jahrer is active here yo may see a great use of autoencoders. Check his winning solution in the Porto Seguro competition.", "votes": null }, { "id": "305728", "postDate": "03/29/2018 10:20:02", "content": "I dont think AE will work here. They contribute best for dense data with moderate featCnt (10 .. 1000). Sparse data makes a problem because AE need to densify the data internally, coz it need to reconstruct all feats. And too many #rows leads to huge runtime, which makes tuning impractical. But lets see
", "rawMarkdown": "I dont think AE will work here. They contribute best for dense data with moderate featCnt (10 .. 1000). Sparse data makes a problem because AE need to densify the data internally, coz it need to reconstruct all feats. And too many #rows leads to huge runtime, which makes tuning impractical. But lets see", "votes": null }, { "id": "305756", "postDate": "03/29/2018 10:55:18", "content": "Thanks for the responses, much appreciated! I will let the AutoEncoder go for now, if someone can make it work, it's not me. :)
", "rawMarkdown": "Thanks for the responses, much appreciated! I will let the AutoEncoder go for now, if someone can make it work, it's not me. :)", "votes": null } ], "comments": [ { "id": 305711, "author_name": "cpmpml", "author_url": "", "post_date": "03/29/2018 09:55:31", "content": "Given Michael Jahrer is active here yo may see a great use of autoencoders. Check his winning solution in the Porto Seguro competition.
", "votes": null, "replies": [ { "id": 305728, "author_name": "mjahrer", "author_url": "", "post_date": "03/29/2018 10:20:02", "content": "I dont think AE will work here. They contribute best for dense data with moderate featCnt (10 .. 1000). Sparse data makes a problem because AE need to densify the data internally, coz it need to reconstruct all feats. And too many #rows leads to huge runtime, which makes tuning impractical. But lets see
", "votes": null, "replies": [] }, { "id": 305756, "author_name": "antmarakis", "author_url": "", "post_date": "03/29/2018 10:55:18", "content": "Thanks for the responses, much appreciated! I will let the AutoEncoder go for now, if someone can make it work, it's not me. :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "305433": "*Warning: I am a newb.*\n\nFor the past couple of days I have been messing around with AutoEncoders for Anomaly Detection, since they can sometimes produce good results with imbalanced data. But I can't get past the AUC=0.7 mark on my validation sets.\n\nI use a simple architecture, with two Dense layers each for the Encoder/Decoder (pretty much [this](https://medium.com/@curiousily/credit-card-fraud-detection-using-autoencoders-in-keras-tensorflow-for-hackers-part-vii-20e0c85301bd)). I have used a couple of preprocessing schemes, like the one [here](https://www.kaggle.com/joaopmpeinado/single-xgboost-lb-0-966/code).\n\nHas anyone gotten it to work? Any tips you can share?", "305711": "Given Michael Jahrer is active here yo may see a great use of autoencoders. Check his winning solution in the Porto Seguro competition.", "305728": "I dont think AE will work here. They contribute best for dense data with moderate featCnt (10 .. 1000). Sparse data makes a problem because AE need to densify the data internally, coz it need to reconstruct all feats. And too many #rows leads to huge runtime, which makes tuning impractical. But lets see", "305756": "Thanks for the responses, much appreciated! I will let the AutoEncoder go for now, if someone can make it work, it's not me. :)" }, "source": "meta" }