20 Aug 2019, 07:43

## Meet the winners of the Mtoto News and Childline Call Volume Prediction Challenge

Zindi is excited to announce the winners of the recently closed challenge that focussed on child protection in Kenya.

The objective of the challenge was to forecast the number of incoming calls that Childline Kenya will receive per hour, per day. The solution is intended to help Childline manage its resources and staff their call center more efficiently, enabling them to support as many children as possible.

The challenge attracted 180 data scientists from across Africa, of whom 41 were on the leaderboard. The winner of this challenge is Lawrence Moruye from Kenya, and the runners-up are Olaleye Eniola and Olayinka Fadahunsi, who are both from Nigeria. We caught up with the winners to hear about their experience in participating in the challenge.

### 1st place: Lawrence Moruye

Zindi handle: Lawrence_Moruye

Where are you from? Kenya

Tell us a bit about yourself.

I'm a student at Multimedia University Of Kenya, taking mathematics and computer science; with interests in the fields of data science and AI.

Tell us a bit about your solution and the approach you took.

Approach:
Since observations were made sequentially with time, I treated the challenge as a time series problem.
1. Data visualisations: Checked assumptions for time series analysis
I had to plot the series and check for stationarity (constant mean, constant variance and whether it's autocovariance was not dependent on time.
2. Data transformations: Trend reduction and data splitting
At first I dropped values which were above a certain standard deviation thinking that they were outliers but I realized most of this values occurred in July and the series had just started taking a new trend (upward). This affected my score(with RMSE of about 30). I was looking for a way to reduce the trend. I did log transformation in order to penalize this higher values more than small values, there was a slight improvement in my score.
I then started looking for a way to work with this series without eliminating these higher values, which was data splitting. I took this higher values (which occurred in July and split them into 2). Then, I chose my training dataset to be all values from January, plus one split from July. Then I tested on the remaining split.
3. Data modelling: i) ARIMA, ii) Random Forest model, and iii) XGBOOST
The main models were the ARIMA, Random Forest, and an XGBOOST. However throughout the competition an xgboost remained robust! Probably because it allows cross-validation at each iteration of the boosting process.

What were the things that made the difference for you that you think others can learn from?

Data visualizations and data splitting. Visualizations brought out a clear picture of the components of a time series like trend. My training sample was from the beginning to some recent point in time and then test set came from that point till the end. Also the features I chose had an impact.

What are the biggest areas of opportunity you see in AI in Africa over the next few years?

AI is going to be the most revolutionary technology that has the potential to solve some of the most pressing challenges that impact Africa and drive growth and development in various sectors. With AI, agriculture will be done more efficiently thus raising yields; healthcare will be better-tailored and more accessible thus improving lives; financial services will be more secure and reach more Africans who need them hence expanding accessibility and promoting development.

What are you looking forward to most about the Zindi community?

Looking forward to learning from best of the best and sharing knowledge with other members

### 2nd place: Olaleye Eniola

Zindi handle: OLALEYE_ENIOLA_DSN

Where are you from? Nigeria

Tell us a bit about yourself.

A 3rd-year student of Systems Engineering at the University of Lagos, who is passionate about artificial intelligence.

Tell us about the approach you took.

It was kind of a tough project for me; I had to read through some post on how to work with time series data which was a bit of Anova and the rest.

What were the things that made the difference for you that you think others can learn from?

I ensembled different weak models which performed well in some part of the data than others.

Where are you from? Nigeria

Tell us a bit about yourself.

I am a Data Scientist at a financial institution in Nigeria who is passionate about solving real world problems while expanding my skill sets.

Tell us about the approach you took.

The Childline call prediction challenge presented an exciting problem. My solution started with exploratory data analysis which pointed me in the right direction to creating new features - Feature Engineering.
Feature engineering led to the creation of multiple features. However, I used feature selection to select the most important features. The most important feature was the mean number of calls per/hour/holiday/weekday.
A model was then built using a boosting algorithm. Finally two diverse models were selected and stacked via ensemble methods.

What were the things that made the difference for you that you think others can learn from?

Creating new features made the most important impact to the performance of the models.

What are the biggest areas of opportunity you see in AI in Africa over the next few years?

Satellite image analysis to help eliminate hunger and powerty in Africa. Its application could be seen in farm yield predictions.

What are you looking forward to most about the Zindi community?

A community that encourages and rewards sharing of code.

### Make a difference in the life of a child and help Childline Kenya improve their services

We encourage all participants to share their code with Zindi and the competition hosts as well as on GitHub as a public good to the sector.

This competition was hosted by Mtoto News (mtotonews.com) and Childline Kenya (childlinekenya.co.ke).