Prediction and feature importance analysis for wordle strategy based on machine learning models

Abstract

The word guessing game, Wordle, has attracted considerable attention from researchers in computer studies and mathematics due to its significant relevance. Many studies have been taken on Wordle to explore its potential in simulating human behaviors during gameplay and devising optimal strategies for players. In this paper, strategies based on mathematics theories or machine learning technologies are compared by their feature importance in order to explore whether there is a loss of feature meaning in machine learning methods. In the whole study, the effect of features and their relations to the expectations of tries of game Wordle are also explored. More specifically, this paper introduces machine learning models such as linear, random forest, bagging as well as gradient boosting decision trees to visually provide a relation between words features (such as repeat frequency, vowels, consonants) and the expectation of attempts in a game. After establishing machine learning models, the feature importance is derived by feature engineering techniques. The importance is then compared with spearman statistical correlations based on the dataset to hence draw the conclusion of the change of feature meaning in machine learning methods. Study results indicate that there is loss of meaning and effect of features in the better fitted prediction models (gradient boosting decision trees, random forest) compared to the statistic approaches.