1 of 14

1

RECRUIT RESTAURANT VISITOR FORECASTING

Jingjing (Olivia) Liang | Samuel Musch | Xiangke Chen | Xuan Ji | Xue Ni | Yassine Manane

2 of 14

2

Today we deliver and report on:

A model to predict daily visitors for Japanese restaurants

An improved understanding of the factors based on prediction results

Novel features for the manager to be aware of

3 of 14

A Restaurant Owner’s Dilemma

What will happen if you don’t know how many people will visit?

​

  • Reviews
  • Food Supply
  • Staff Allocation
  • Customer Experience

3

4 of 14

Manager Intuition

4

BEFORE VS AFTER

5 of 14

Manager Intuition

5

BEFORE VS AFTER

6 of 14

Manager Intuition

6

BEFORE VS AFTER

7 of 14

Data Sources

7

Original Datasets From Kaggle:

Visit & Reservations

Store Information

Calendar Date Information

​

External Dataset:

Population

Weather data

​

​

​

​

​

​

​

​

8 of 14

Factors Considered

8

Geographic Features:

  • Location
  • Population / Density

Time Series Features:

  • Prior Year Mapping
  • Days to Previous 25th
  • (Consecutive) Holidays

​

​

​

​

​

​

​

​

​

​

​

Visitors Features:

  • Exponential Moving Average
  • Aggregation (min, max, mean, median)
  • Which days stores are closed

9 of 14

9

Implementation of different models

Baseline Time Series

External factors

RMSLE - 0.561

No data pre-processing

Trends + seasonality

RMSLE - 0.542

ARIMA

FACEBOOK PROPHET

LIGHT GBM

LSTM

SEQ2SEQ

SEQ2SEQ

+

LGBM

Feature importance

No statistical assumptions

RMSLE - 0.5214

State of art for sequence data

RMSLE - 0.50277

Ensemble approach

RMSLE - 0.50002

10 of 14

10

Implementation of different models

Baseline Time Series

External factors

RMSLE - 0.561

No data pre-processing

Trends + seasonality

RMSLE - 0.542

ARIMA

FACEBOOK PROPHET

LIGHT GBM

LSTM

SEQ2SEQ

SEQ2SEQ

+

LGBM

Feature importance

No statistical assumptions

RMSLE - 0.5214

State of art for sequence data

RMSLE - 0.50277

Ensemble approach

RMSLE - 0.50002

11 of 14

Leverage time factors to make better customer experience

11

Previous year mapping

Aggregated number of visitors

for the same day of week

Days to last paycheck

Lagged # of visitors in last 7 days

Holiday Flag

12 of 14

Impact and Value

  1. Cut manager intuition error in half
  2. Optimize staff allocation and inventory management
  3. Save expenditures on salary payment
    1. $132 = Minimum wage * 2 waiters * 8 hours
    2. Minimum wage is roughly $8.26 per hour
  4. Generalizable Algorithm

12

13 of 14

Next Steps

  1. Feature improvement
    1. Weather
    2. Distance from Busiest Area
  2. Modeling improvement
    • Add Attention Mechanism to RNN model
    • More diverse model ensembling
  3. Expand the timespan

13

14 of 14

References

14

  1. https://blog.keras.io/a-ten-minute-introduction-to-sequence-to-sequence-learning-in-keras.html
  2. https://github.com/Arturus/kaggle-web-traffic
  3. https://github.com/MaxHalford/kaggle-recruit-restaurant/blob/master/Solution.ipynb
  4. https://www.kaggle.com/pureheart/1st-place-lgb-model-public-0-470-private-0-502
  5. https://www.kaggle.com/plantsgo/solution-public-0-471-private-0-505
  6. https://www.kaggle.com/h4211819/holiday-trick
  7. https://www.kaggle.com/c/recruit-restaurant-visitor-forecasting/discussion/49100

​