Skip to content

Latest commit

 

History

History
108 lines (70 loc) · 4.95 KB

README.md

File metadata and controls

108 lines (70 loc) · 4.95 KB
Raw

Genpact Machine Learning Hackathon

competition link

Root of Mean Squared Logarithmic Error : 0.54

Private Leader Board Rank : 79/3872

Analysis and baseline model Notebook

Final model Notebook

Libraries used:

  • pandas, numpy, catboost, scikit learn, matplotlib, seaborn

Problem Statement

Your client is a meal delivery company which operates in multiple cities. They have various fulfillment centers in these cities for dispatching meal orders to their customers. The client wants you to help these centers with demand forecasting for upcoming weeks so that these centers will plan the stock of raw materials accordingly.

The replenishment of majority of raw materials is done on weekly basis and since the raw material is perishable, the procurement planning is of utmost importance. Secondly, staffing of the centers is also one area wherein accurate demand forecasts are really helpful. Given the following information, the task is to predict the demand for the next 10 weeks (Weeks: 146-155) for the center-meal combinations in the test set:

  • Historical data of demand for a product-center combination (Weeks: 1 to 145)
  • Product(Meal) features such as category, sub-category, current price and discount
  • Information for fulfillment center like center area, city information etc.

Data Dictionary

  1. Weekly Demand data (train.csv): Contains the historical demand data for all centers, test.csv contains all the following features except the target variable.
Variable Definition
id Unique ID
week Week No
center_id Unique ID for fulfillment center
meal_id Unique ID for Meal
checkout_price Final price including discount, taxes & delivery charges
base_price Base price of the meal
emailer_for_promotion Emailer sent for promotion of meal
homepage_featured Meal featured at homepage
num_orders (Target) Orders Count
  1. fulfilment_center_info.csv: Contains information for each fulfilment center
Variable Definition
center_id Unique ID for fulfillment center
city_code Unique code for city
region_code Unique code for region
center_type Anonymized center type
op_area Area of operation (in km^2)
  1. meal_info.csv: Contains information for each meal being served
Variable Definition
meal_id Unique ID for the meal
category Type of meal (beverages/snacks/soups….)
cuisine Meal cuisine (Indian/Italian/…)

Evaluation Metric

The evaluation metric for this competition is 100*RMSLE where RMSLE is Root of Mean Squared Logarithmic Error across all entries in the test set.

Test data is further randomly divided into Public (30%) and Private (70%) data.

Solution

  • Converted this time series problem to regression problem.

Data transformation:

  1. Here number of orders placed (target variable) is highly right skewd so that Log transformation is applied.
  2. Log transformation of base_price, checkout_price, and num_orders.

Feature engineering:

  1. For every record difference between base_price and checkout_price.
  2. Differenc of previous week checkout_price and current weeks checkout_price.
  3. Lag features of 10,11, and 12 week lagging features. Here I have used lag of last 10 weeks because we have to predict for 10 weeks in test dataset.
  4. Exponentially weighted mean over last 10, 11, and 12 weeks.

Cross validation strategy:

  • Last 10 weeks (136 - 145) of every center-meal pair data is used as a Validation dataset from train dataset.

Model

  1. One single CatBoost model which has RMSLE of 0.54.
  2. High regularization so it does not overfit because of new features made using target variable.

What didn't work:

  1. just using original data as it is and using catboost regressor gave RMSLE of 1.58
  2. Only using difference between base_price and checkout_price, difference between base_price and checkout_price as a features and not using any lag and exponentially weighed features didn't give good score.
  3. Rolling mean and median over last 26, 52, 104 weeks as features didn't work that well, feature importance was low.
  4. Geographical features had low feature importance, So didn't use them in final model.

TODO / Improvements:

  1. Extensive hyper parameter tuning and feature selection.
  2. Create more features based on Categorical Encoding methods like mean encoding, freqvency encoding, hash encoding etc.
  3. Try more algorithms like xgboost, LightGBM, Linear Regression etc.
  4. Try ARIMA , Prophet etc.
  5. Ensemble of different models.