| Feature | Spektral | Stellargraph | Graph Nets |
| Tutorials availability | There’re about 20 examples that show different usages. | A lot of well documented demos in python notebooks. | Very few examples |
| Year of creation | 2020 | 2018 | 2018 |
| Paper | Spectral Clustering with Graph Neural Networks for Graph Pooling cited less than 100 times | No paper | Relational inductive biases, deep learning, and graph networks cited more than 1400 times |
| Author/s | danielegrattarola | CSIRO’s Data61 | DeepMind |
| Supports | Tensorflow | Tensorflow | Tensorflow and Sonnet |
| Input type | Simple datasets of multiple thousands single numerical value nodes | Dataframes | In the examples the node values were required to be arrays of float values. It’s also possible to choose whether to make a graph without edges, nodes without values, or edges without directions. |
| Usage | Examples show a case of citations network, molecular prediction model, and graph network classification. | Varied use cases of supervised and unsupervised learning. It also supports 20 different types of algorithms. | Examples show one physical example, one about the shortest path problem solution and another simplified demo. |
| Node prediction | Yes | Classification | Doesn’t predict graphs, nodes or edges according to the manual, but predicts states, shortest paths and other operations that are done over an existing graph. |
| Edge prediction | Not in the examples | Yes (link prediction) |
| Graph prediction | Yes | Classification |
| Support categorical data | | | Yes, in one hot vector. |
