Invent Analytics
Published:
Location
Istanbul, Turkey
Duration
From July 2020 to September 2020
About the Company
Invent Analytics is a company specialized in providing analytics solutions to retailers by using different machine learning and data science techniques. Their main activities include inventory management and sales forecasting.
My Role
Throughout the internship period, I developed a machine learning pipeline using the Walmart sales data provided in the famous M5 Forecasting Challenge.
After conducting a thorough study about the top solution that scored the highest in the challenge, I worked on devising a model which utilizes different interesting ideas and observations from different solutions. Then, I started spent the rest of the time implementing, testing, and refining the model’s performance by tuning different parameters and doing further feature engineering.
Skills Learned
- Handling time series data.
- Feature Engineering and Extraction.
- Building full Machine Learning pipelines.
- Estimating the distribution of the data to decide on how to model it.
- Usage of Gradient Boosting algorithms such as XGBoost and Light GBM.
- Hyperparameter optimization.
- Creating and evaluating Machine Learning ensemble models.
- Improved my python skills.
Challenges Faced
There were two tasks that I spent the most time on and where challenging to handle.
1. Choosing features
There was a wide range of features that could be extracted and appended to the training data, such as lags and rolling features, but these features had different impacts and effect levels on the model. Also, having many features would significantly slow down the training process, so I had to choose the least but most effective ones. Therefore, I had to keep trying different values and different techniques to decide on the best feature set, which was more time consuming than I expected.
2. Hyperparameter optimization
Deciding on the parameters for gradient boosting algorithms was the most difficult part. There were numerous important parameters to manipulate. I attempted using some packages that can help in optimizing the parameters instead of doing a brute force search. I tried using hyperopt package in python, which uses Bayesian optimization algorithms to find the best set of parameters. The library was not as fast as required but it was better than the other options I have attempted.