An In-Depth Comparison of Neural and Probabilistic Tree Models for Learning-to-rank

Learning-to-rank has been intensively studied and has demonstrated significant value in several fields, such as web search and recommender systems. Over the learning-to-rank datasets given as vectors of feature values, LambdaMART proposed more than a decade ago, and its subsequent descendants based on gradient-boosted decision trees (GBDT), have demonstrated leading performance. Recently, different novel tree models have been developed, such as neural tree ensembles that utilize neural networks to emulate decision tree models and probabilistic gradient boosting machines (PGBM). However, the effectiveness of these tree models for learning-to-rank has not been comprehensively explored. Hence, this study bridges the gap by systematically comparing several representative neural tree ensembles (e.g., TabNet, NODE, and GANDALF), PGBM, and traditional learning-to-rank models on two benchmark datasets. The experimental results reveal that benefiting from end-to-end gradient-based optimization and the power of feature representation and adaptive feature selection, the neural tree ensemble does have its advantage for learning-to-rank over the conventional tree-based ranking model, such as LambdaMART. This finding is important as LambdaMART has achieved leading performance in a long period.