Meet Adetomiwa Adedeji (tomy4reel), Adeyinka Michael Sotunde (MICADEE), and Artem Gusenkov (J4LLZ), winners of the Alvin Smart Money Classification Challenge. The challenge attracted 455 participants, all vying for a $3000 prize pool. The objective of this challenge was to classify purchases recorded on the Alvin money management app into different categories.
Adetomiwa Adedeji, Nigeria
Please introduce yourself.
My name is Adetomiwa Adedeji (tomy4reel), I am a data scientist and ML engineer from Nigeria.
Tell us a bit about your solution, and the approach you took.
My approach was to use fuzzy-matching of keywords/phrases in merchant names, and pass them to CatBoostClassifier which enabled the model to properly classify new merchants. I handled class imbalance by oversampling the dataset.
What set your winning solution apart from others?
I did not make use of the additional datasets provided.
Words of encouragement for others, or advice that has helped you?
I try to research and understand the problem better. BIY: Believe in yourself!
Artem Gusenkov, Georgia
Please introduce yourself.
My name is Artem Gusenkov (J4LLZ). I am an ML engineer at Badal.io and am based in Georgia.
Tell us a bit about your solution, and the approach you took.
The idea was to use all the data available (including unlabeled) and build word embeddings. I performed custom labeling based on the meaning of merchant names.
What set your winning solution apart from others?
Unlabeled training of word embeddings, fastText module, and usage of all the data available. Using some pre-built text features from NLP library spaCy.
How do you prepare for a challenge?
Reading blogposts about NLP.
Words of encouragement for others, or advice that has helped you?
Make sure you are reading and taking advice from the competition discussion sections.
Adeyinka Michael Sotunde, Nigeria
Please introduce yourself.
I am Adeyinka Michael Sotunde (MICADEE) from Lagos, Nigeria. I am currently working as Senior Planning Officer & Data Scientist at LAHASCOM.
Tell us a bit about your solution, and the approach you took.
Actually, I will start by saying that we experimented with a lot of ideas in this wonderful project. However, a few of the ideas we implemented were:
- After reading in the provided datasets, we started by creating new features based on the name of the vendor, such as bundles, safaricom, mall, shop etc.
- We renamed some vendor names after our exploratory data analysis (EDA) found several naming errors and contradictions.
- Based on the observation of some vendor names and naming similarity, we added some more data from the extra data provided for training.
- We also created time-based features, only from the purchase timing and never from the categorising timing provided.
- Lastly, we created a robust CatBoostClassifier model with 5kFolds CV after deriving the right set/combination of the model parameters, using our own customized catboost hyperparameter tuning setup.
What set your winning solution apart from others?
What gave us an edge is the thorough Exploratory Data Analysis (EDA) and thorough understanding of all datasets provided, which actually helped us a lot in creating some meaningful and also some very logical features that eventually gave rise to our winning solution.
How do you prepare for a challenge?
I prepare by first of all dissecting the given datasets into different parts as best as possible using any EDA tools available. By doing this, one will be able to fish out some meaningful insights from the given datasets that will definitely be helpful later in the course of the preprocessing stage.
Words of encouragement for others, or advice that has helped you?
Every data science enthusiast on Zindi must force themselves to start from scratch i.e. understanding the given datasets using EDA tools and any other useful tools before any ML implementation.
What do you like about Zindi?
Zindi is a great platform that has elevated a lot of data scientists in terms of knowledge, skills acquired etc., not only in this African continent but beyond by providing some great projects in the form of challenges to be tackled or solved by us all.