I'm fairly new to this topic and ML in general, but I have submitted multiple attempts. What I'm kind of annoyed by, is the fact that my best score is just by grouping certain features together and calculate the median. When I apply actual ML-algorithms on this, my score is not nearly as good.
\n\nI'm kind of curious as to how many of the top scores uses actual ML, and not just tweaking the baseline, and what concepts I should look further into, in order to improve my score.
", "messages": [ { "id": "236850", "postDate": "10/28/2017 12:14:11", "content": "I'm fairly new to this topic and ML in general, but I have submitted multiple attempts. What I'm kind of annoyed by, is the fact that my best score is just by grouping certain features together and calculate the median. When I apply actual ML-algorithms on this, my score is not nearly as good.
\n\nI'm kind of curious as to how many of the top scores uses actual ML, and not just tweaking the baseline, and what concepts I should look further into, in order to improve my score.
", "rawMarkdown": "I'm fairly new to this topic and ML in general, but I have submitted multiple attempts. What I'm kind of annoyed by, is the fact that my best score is just by grouping certain features together and calculate the median. When I apply actual ML-algorithms on this, my score is not nearly as good.\n\nI'm kind of curious as to how many of the top scores uses actual ML, and not just tweaking the baseline, and what concepts I should look further into, in order to improve my score.", "votes": null }, { "id": "236851", "postDate": "10/28/2017 12:17:05", "content": "FWIW, my ML attempts (xgboost on all series jointly) has not been successful at all - my best attempt so far is based on time series. I am experimenting with Kalman filter.
", "rawMarkdown": "FWIW, my ML attempts (xgboost on all series jointly) has not been successful at all - my best attempt so far is based on time series. I am experimenting with Kalman filter.", "votes": null }, { "id": "236899", "postDate": "10/28/2017 16:05:55", "content": "Same here - XGboost is behaving weirdly - even when I input lags and averages as features it seems to underpredict sales.
\n\nAlthough as a side comment - grouping by store, item and promotion and taking the average is in fact an ML model - you're basically growing a decision tree as deep as possible on the three features :)
", "rawMarkdown": "Same here - XGboost is behaving weirdly - even when I input lags and averages as features it seems to underpredict sales.\n\nAlthough as a side comment - grouping by store, item and promotion and taking the average is in fact an ML model - you're basically growing a decision tree as deep as possible on the three features :)", "votes": null }, { "id": "236912", "postDate": "10/28/2017 16:42:15", "content": "I had the same problem, when I tried Random Forest almost directly on the data, the NWRMSLE was about ~1.1. However my problem was that the train data doesn't contain samples with zero unit sales, and therefore random forest was very biased, and predicted higher amount of sales.
\n\nThe solution was to generate more training data with unit sales is zero. After this, the NWRMSLE dropped to ~0.64.
", "rawMarkdown": "I had the same problem, when I tried Random Forest almost directly on the data, the NWRMSLE was about ~1.1. However my problem was that the train data doesn't contain samples with zero unit sales, and therefore random forest was very biased, and predicted higher amount of sales.\n\nThe solution was to generate more training data with unit sales is zero. After this, the NWRMSLE dropped to ~0.64.", "votes": null }, { "id": "236940", "postDate": "10/28/2017 20:08:46", "content": "Brilliant observation - I wish I could upvote you more than once :-) 'Elementary, my dear Watson' - then again, most things are once smb else has pointed it out. Thanks!
", "rawMarkdown": "Brilliant observation - I wish I could upvote you more than once :-) 'Elementary, my dear Watson' - then again, most things are once smb else has pointed it out. Thanks!", "votes": null }, { "id": "236947", "postDate": "10/28/2017 20:24:36", "content": "Very interesting! My ML-score was also ~1 when I first tried. I will keep this in mind when I try again tomorrow as I've used up my 5 tries for today.
", "rawMarkdown": "Very interesting! My ML-score was also ~1 when I first tried. I will keep this in mind when I try again tomorrow as I've used up my 5 tries for today.", "votes": null }, { "id": "237095", "postDate": "10/29/2017 13:05:56", "content": "What do you mean by generate more training-data with unit_sales = 0? When I load the training data, everything below 0, I set to exactly 0.
\n\nIt's totally fine if you don't want to say. It's a competition after all :)
", "rawMarkdown": "What do you mean by generate more training-data with unit_sales = 0? When I load the training data, everything below 0, I set to exactly 0. \n\nIt's totally fine if you don't want to say. It's a competition after all :)", "votes": null }, { "id": "237100", "postDate": "10/29/2017 13:13:02", "content": "Some combinations are not present in the training data. If a combination is not present, it means that unit_sales is zero for this combination.
\n\nSo, I implicitly generated this configuration before training a model.
", "rawMarkdown": "Some combinations", "votes": null } ], "comments": [ { "id": 236851, "author_name": "konradb", "author_url": "", "post_date": "10/28/2017 12:17:05", "content": "FWIW, my ML attempts (xgboost on all series jointly) has not been successful at all - my best attempt so far is based on time series. I am experimenting with Kalman filter.
", "votes": null, "replies": [] }, { "id": 236899, "author_name": "apapiu", "author_url": "", "post_date": "10/28/2017 16:05:55", "content": "Same here - XGboost is behaving weirdly - even when I input lags and averages as features it seems to underpredict sales.
\n\nAlthough as a side comment - grouping by store, item and promotion and taking the average is in fact an ML model - you're basically growing a decision tree as deep as possible on the three features :)
", "votes": null, "replies": [] }, { "id": 236912, "author_name": "deyury", "author_url": "", "post_date": "10/28/2017 16:42:15", "content": "I had the same problem, when I tried Random Forest almost directly on the data, the NWRMSLE was about ~1.1. However my problem was that the train data doesn't contain samples with zero unit sales, and therefore random forest was very biased, and predicted higher amount of sales.
\n\nThe solution was to generate more training data with unit sales is zero. After this, the NWRMSLE dropped to ~0.64.
", "votes": null, "replies": [ { "id": 236940, "author_name": "konradb", "author_url": "", "post_date": "10/28/2017 20:08:46", "content": "Brilliant observation - I wish I could upvote you more than once :-) 'Elementary, my dear Watson' - then again, most things are once smb else has pointed it out. Thanks!
", "votes": null, "replies": [] }, { "id": 236947, "author_name": "nilzone", "author_url": "", "post_date": "10/28/2017 20:24:36", "content": "Very interesting! My ML-score was also ~1 when I first tried. I will keep this in mind when I try again tomorrow as I've used up my 5 tries for today.
", "votes": null, "replies": [] }, { "id": 237095, "author_name": "nilzone", "author_url": "", "post_date": "10/29/2017 13:05:56", "content": "What do you mean by generate more training-data with unit_sales = 0? When I load the training data, everything below 0, I set to exactly 0.
\n\nIt's totally fine if you don't want to say. It's a competition after all :)
", "votes": null, "replies": [] }, { "id": 237100, "author_name": "deyury", "author_url": "", "post_date": "10/29/2017 13:13:02", "content": "Some combinations are not present in the training data. If a combination is not present, it means that unit_sales is zero for this combination.
\n\nSo, I implicitly generated this configuration before training a model.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "236850": "I'm fairly new to this topic and ML in general, but I have submitted multiple attempts. What I'm kind of annoyed by, is the fact that my best score is just by grouping certain features together and calculate the median. When I apply actual ML-algorithms on this, my score is not nearly as good.\n\nI'm kind of curious as to how many of the top scores uses actual ML, and not just tweaking the baseline, and what concepts I should look further into, in order to improve my score.", "236851": "FWIW, my ML attempts (xgboost on all series jointly) has not been successful at all - my best attempt so far is based on time series. I am experimenting with Kalman filter.", "236899": "Same here - XGboost is behaving weirdly - even when I input lags and averages as features it seems to underpredict sales.\n\nAlthough as a side comment - grouping by store, item and promotion and taking the average is in fact an ML model - you're basically growing a decision tree as deep as possible on the three features :)", "236912": "I had the same problem, when I tried Random Forest almost directly on the data, the NWRMSLE was about ~1.1. However my problem was that the train data doesn't contain samples with zero unit sales, and therefore random forest was very biased, and predicted higher amount of sales.\n\nThe solution was to generate more training data with unit sales is zero. After this, the NWRMSLE dropped to ~0.64.", "236940": "Brilliant observation - I wish I could upvote you more than once :-) 'Elementary, my dear Watson' - then again, most things are once smb else has pointed it out. Thanks!", "236947": "Very interesting! My ML-score was also ~1 when I first tried. I will keep this in mind when I try again tomorrow as I've used up my 5 tries for today.", "237095": "What do you mean by generate more training-data with unit_sales = 0? When I load the training data, everything below 0, I set to exactly 0. \n\nIt's totally fine if you don't want to say. It's a competition after all :)", "237100": "Some combinations" }, "source": "meta" }