Club Mahindra Data Olympics

Public Leaderboard: 3rd Place : 95.001
Private Leaderboard: 3rd Place : 95.8484

Training

There is file Final_solution.ipynb . Run each cell and at the end you will get total 6 submission files generated by following models ->

1. Single LGB model
2. Single Catboost model
3. 5 fold LightGBM model
4. 5 fold Catboost model.
5. Stacking of[xgb,lgb,catboost]
6. Ensemgle of[1,2,3,4,5,6] using (model_1 X 0.3+model_2 X *0.2+model_3 X *0.3 +model_4 X *0.2) X *0.8 + model_5 X *0.2

Problem Statement

Food & Beverages Spend Prediction in Club Mahindra Resorts

Club Mahindra (Club M) makes significant revenue from Food and Beverages (F&B) sales in their resorts. The members of Club M are offered a wide variety of items and our taks was to make prediction of amount spend by a mamber per night which could help them to plan inventory accordingly.

Brief approach-

Thanks Analytics Vidhya and Club Mahindra for organising such a wonderful hackathon,The competition was quite intense and dataset was very clean to work.

Approach :

Step-1 :

I started my problem with very basic approach changing all the features (resort id , persontravellingID , main_product_code and other ordinal features to category . Made some common for each date columns [booking date , checkin_date, checkout_date ]:

1. Weekday
2. Month
3. Day
4. Day of year
5. Week of year
6. Is month end
7. Year

Step-2: Intuitive features

1. In_out : Checkout_Date - Checkin_Date
2. book_in:Checkout_date - booking_date
3. Roomnights per stay : roomnights/in_out
4. Roomnights per book span : roomnights / book_out

Step - 3: Time Based Features :

1. Prev_resort_time = Time when the resort was previously booked.
2. Prev_resort_member_time = Time when the resort was previously booked by a particular member.
3. Next_resort_time = Time when the resort will Next booked.
4. Next_resort_member_time = Time when the resort will next booked by a particular member.

Step-4 : Groupby Features

S.No.	TYPE	Value_column	ON
1.	COUNT	_	RESORT_ID
2	COUNT	_	RESORT_ID,MemberID
3.	COUNT	_	['resort_id','checkout_dateyear','checkout_datemonth']
4.	COUNT	_	['memberid','checkout_dateyear']

5	VAR	roomnights	RESORT_ID
6	Median	roomnights	RESORT_ID,MemberID
7.	MAX	roomnights	[resort_id,checkout_dateyear,checkout_datemonth]
8.	MIN	roomnights	[memberid','checkout_dateyear']

9	VAR	in_out	RESORT_ID
10	Median	in_out	RESORT_ID,MemberID
11.	MAX	in_out	['resort_id','checkout_dateyear','checkout_datemonth']
12.	MIN	in_out	['memberid','checkout_dateyear']

13	VAR	total_pax	RESORT_ID
14	Median	total_pax	RESORT_ID,MemberID
15	MAX	total_pax	['resort_id','checkout_dateyear','checkout_datemonth']
16	MIN	total_pax	['memberid','checkout_dateyear']

…… in Similar fashion approx ~ 72 combinations were tried which gave a boost of rmse from 96 to 95.3 on LB and nearly same change in Local CV.

Modeling:

My final model consist of ensemble of

[ lightGBM , Catboost , 5_fold_Light GBM , 5_fold_Catboost and stacking of [xgb,catboost,lightGBM]

)