Current research reveals that graph neural networks (GNNs) are prone to attribute inference attacks when applied to social network data. The graph isomorphism networks (GINs), a variant of the GNN model with a powerful message-passing scheme, are more susceptible to this issue due to the significantly high homophily property of the social media data. A naïve approach to handle this is to assume all users’ sensitive attributes to be known, and subsequently, remove these before employing the graph representation learning model. However, this disregards the diverse privacy preferences as observed in real life, where a user may or may not like to disclose some of his/her attributes. Moreover, simple removal of sensitive attributes may not always be effective due to their correlation with the non-sensitive attributes. In this research, we address the issue of partially observed sensitive attributes by introducing a disentanglement-incorporated privacy-preserving GIN model, termed as DP-GIN. The model first uses a module that performs disentangled representation learning to segregate the non-sensible user traits/attributes into distinctive sensible and non-sensible latent components which are supposed to be orthogonal. This ensures that the non-sensible latent attributes become independent on the sensible attributes. Subsequently, the model employs a GIN-based node classification module to classify the nodes using non-sensitive latent components. Extensive experimentation on four standard datasets validates DP-GIN’s efficacy in preserving privacy while maintaining classification performance.