Assignment 1

Problem and Dataset

The problem I picked is an ordinal regression task, namely predicting review rating for books on Goodreads (Kaggle competition)
The data is composed of approximately 0.9M book reviews, containing the book, the author, the review text, and its stats. For a full description, check out the information on Kaggle

EDA Raw Data

The book_id feature is ordinal instead of categorical
The date columns (i.e. date_added, date_updated, read_at, started_at) are strings instead of datetime
The features read_at and started at have many missing entries

Rating 4 is the most common => the naive baseline is setting every review's rating to 4 and computing the F1 score
People tend to leave a comment when the review is either really bad or really good

Data Processing

Absolute values for votes and comments (cannot be negative)
Fill in read_at missing values with added_at
Convert date strings to pandas datetime
Fill in missing values with the mode of the respective column
Derive additional features (missing_started_at, reading_duration, review_length, spoiler, hour, month, dayofweek, year, ...)
Drop review text
Convert id features to category type

Naïve Baseline

Set all ratings in the validation set to 4
F1 score - 0.08

Baseline Model

LightGBM Classifier (default parameters)
Below is the key metric (F1 score), training charts, predictions and feature importance tables.
Conclusions:
F1 score improve drastically in comparison with the naïve baseline, even though the model was trained using the default hyperparameters
Looking at the evolution of the F1 score and multi logloss over the iterations, one can see that they did not reach a plateau yet. Hence, increasing the number of iterations would likely increase the performance of the model
The feature importance bar plot, suggests that the book and user id are the most important, implying that certain users and books are reviewed either really well or really poorly
F1 score - 0.38