Effectiveness of discrete weight matrices in GODE-CF

Determine whether the discrete per-node weight matrix W included in Graph Neural Ordinary Differential Equations-based Collaborative Filtering (GODE-CF), which parameterizes the ODE derivative using one or two LightGCN layers, is beneficial for recommendation performance compared to using no weight matrix.

Background

The paper discusses ODE-based approaches to collaborative filtering, contrasting LightGCN (which removes weight matrices) with GODE-CF, a method that parameterizes the ODE derivative function using one or two GCN layers and includes a discrete weight matrix for each node embedding. While GODE-CF reports improvements in some scenarios, its use of discrete weights conflicts with the continuous nature of ODE modeling.

Motivated by this ambiguity, the authors propose CDE-CF, which uses MLP-based continuous weight generation integrated into the ODE function. The explicit uncertainty concerns whether the discrete weight matrix in GODE-CF actually helps, prompting the development of a continuous alternative.