Riiid! Answer Correctness Prediction Kaggle Challenge: 4th Place Solution Summary

TL;DR

This paper describes a transformer-based solution for the Riiid! Answer Correctness Prediction Kaggle challenge, achieving high accuracy by incorporating novel time-aware attention and embedding techniques.

Contribution

It introduces a unique transformer model with time-aware attention and combined embeddings, enhancing prediction performance over previous methods.

Findings

Achieved 0.817 AUC score, ranking 4th on Kaggle leaderboard.

Demonstrated the effectiveness of time-aware attention in sequence modeling.

Showed that embedding concatenation improves model input representation.

Abstract

This paper presents my solution to the challenge "Riiid! Answer Correctness Prediction" on Kaggle hosted by Riiid Labs (2020), which scores 0.817 (AUC) and ranks 4th on the final private leaderboard. It is a single transformer-based model heavily inspired from previous works such as SAKT, SAINT and SAINT+. Novel ingredients that I believed to have made a difference are the time-aware attention mechanism, the concatenation of the embeddings of the input sequences and the embedding of continuous features.