csv` but watched no improve so you’re able to regional Curriculum vitae. In addition experimented with undertaking aggregations oriented merely to the Bare has the benefit of and Terminated has the benefit of, however, watched no increase in local Cv.
Automatic teller machine withdrawals, installments) to see if the consumer is broadening Automatic teller machine distributions as go out continued, or if client was decreasing the minimum repayment just like the installment loans online in Virginia day ran towards the, etcetera
I happened to be getting together with a wall. For the July thirteen, We paid down my personal studying speed so you’re able to 0.005, and you will my personal local Cv went to 0.7967. The general public Lb was 0.797, and the private Pound are 0.795. It was the greatest local Curriculum vitae I was able to find which have just one model.
After that model, I spent plenty big date seeking to tweak the hyperparameters right here there. I attempted lowering the reading rate, choosing better 700 otherwise eight hundred features, I attempted using `method=dart` to apply, decrease certain articles, changed certain beliefs having NaN. My get never ever increased. In addition checked dos,step three,cuatro,5,six,7,8 season aggregations, however, not one helped.
To the July 18 I composed a different sort of dataset with additional keeps to try to raise my personal get. There are they of the clicking here, as well as the password to create it by clicking right here.
Toward July 20 I grabbed the typical off two habits one was indeed educated on some other date lengths for aggregations and had public Lb 0.801 and personal Pound 0.796. I did even more blends next, and lots of had high into the personal Lb, however, none previously beat the general public Pound. I attempted along with Genetic Programming provides, target security, altering hyperparameters, but absolutely nothing aided. I tried by using the based-for the `lightgbm.cv` so you can lso are-train with the full dataset and that did not let possibly. I tried enhancing the regularization because I thought that i got so many has nevertheless failed to let. I tried tuning `scale_pos_weight` and discovered it don’t help; indeed, often broadening pounds of non-self-confident examples do boost the regional Curriculum vitae more than increasing pounds from positive advice (restrict intuitive)!
In addition thought of Cash Finance and Individual Financing once the same, therefore i were able to lose many the large cardinality
While this is taking place, I happened to be fooling doing a lot that have Neural Companies because the We had plans to create it a blend to my design to find out if my personal score improved. I am grateful Used to do, given that We provided certain neural systems back at my team later. I need to give thanks to Andy Harless having promising everybody in the competition to develop Neural Systems, along with his very easy-to-go after kernel you to driven us to say, “Hey, I’m able to accomplish that also!” He only made use of a rss submit neural system, however, I experienced plans to play with an organization embedded neural network with another normalization system.
My higher individual Lb get performing by yourself was 0.79676. This would have earned me score #247, sufficient to own a silver medal nonetheless most respected.
August 13 I written a different sort of updated dataset that had a bunch of brand new has actually that i is actually in hopes create take me even higher. The newest dataset is available by pressing right here, while the password to generate it can be located because of the clicking here.
Brand new featureset had features which i consider was indeed really book. It’s got categorical cardinality prevention, conversion process out of bought classes to help you numerics, cosine/sine conversion process of hour out-of application (so 0 is virtually 23), proportion between the said income and you may average income for your work (in the event the stated earnings is a lot high, maybe you are lying making it appear to be the job is perfect!), money separated of the complete part of domestic. We grabbed the total `AMT_ANNUITY` you only pay away per month of one’s energetic past apps, immediately after which divided you to definitely by the earnings, to see if their ratio try adequate to consider a different sort of financing. We got velocities and you may accelerations off specific articles (age.g. This might let you know if customer is actually beginning to rating brief into the currency and that more likely to standard. In addition checked velocities and accelerations from those times due and you can amount overpaid/underpaid to see if these people were with previous fashion. Unlike anyone else, I was thinking the fresh new `bureau_balance` table are very helpful. We re-mapped brand new `STATUS` line in order to numeric, deleted all of the `C` rows (because they contained no additional advice, these were just spammy rows) and you can from this I became able to find out hence bureau applications was effective, which were defaulted for the, an such like. This also assisted during the cardinality protection. It actually was getting regional Cv regarding 0.794 even when, so maybe We threw out too much advice. If i had longer, I would not have reduced cardinality much and would have just remaining one other beneficial have I created. Howver, they probably aided too much to the new diversity of cluster heap.
Recent Comments