Reduce Waiting Time With MLP Neural Networks

Benefit for Customers and Businesses

Customers

Many customers fear and face long waiting time when it comes to visit postal offices, ticket counters, shops or restaurants. They would even happily change their visiting time if they only knew in advance how long the waiting would be.

Businesses

The business owners on the other side would welcome a wider distribution of customers over their opening time period.
In this short blog I like to discuss a proposal using neural networks.

Goal: Waiting Time Prediction

The goal of the network is to get a waiting time prediction for a certain time slot like:
"The expected waiting time between 11:45 and 12:50 is 7 minutes." or "Low waiting time at 12:30."

First: Collecting Data

The first thing is data, data, data. It is important that reasonable amount of data is collected weeks, month or even years before a prediction can be reasonably be done. For  waiting time prediction network the following data features can be recommended:

Time slot (key)

The time slot for the feature data might be about 1  up to 5 minutes.

Location (key)

Often, multiple location (such as counter 1, counter 2, office A, office B) are involved.

Average waiting time

The average waiting time can be used for input and output data for the training and prediction of the network. Average waiting time could also be deduced if individuals can be track directly.

Number of customers, queue size

The number of customers or queue size per time slot is one of the primary data inputs and outputs. Next to using this data as network input it can be used to estimate the average waiting time.

Secondary Data

Using the primary data only would not yet lead to a high prediction in a trained network. The following secondary data the network allows to reach much better predictions.

Weekday

Weekdays (Monday, Tuesday, ...) have significant influence in most waiting situations such as Museum, Restaurant etc.

Holiday calendar

An encoded holiday calendar that affects the majority of customers helps the network to adjust properly. Holiday calendar includes school free time, local religious events, sport events.

Weather

Weather data correlates with many situations in both direction (longer or shorter waiting time) depending on the business

Closed customer data

In case of a closed customer community such as company staff or hotel guests more data may be available like age, gender, ma

Network

As network a multi layer neural perceptron network (MLP) can be thought of:

Data preparation

The following data preparation has to be done:

1. Cluster Data
2. Standardise Data
3. Standard Deviation Normalisation
4. 'One Hot' preparation for week days (e.g. Weekdays as single columns)
5. 'One Hot' for time slots

Setup Network

Multi-dimensional input
One or more hidden layers
Error function sigmoid, 'Adam', cross entry tests

Train

90% of data can be used for training
10% of data can be used for prediction verification

Hosting

OOIT.com AG provides a secured low cost hosting for preparing, training and querying predictions.