Will DGL 0.5 have model zoo for recommender systems

Hi

DGL 0.5 is released recently.
There are some articles stating that recommender systems might be included in this version.

While there is no explicit documentation about DGL rec-sys, will DGL 0.5 include model zoo for recommender systems in release?

Thank you.

1 Like

Hi @Lyman, unfortunately due to the bandwidth of the team, we prioritize other enhancement such as documentation, code refactoring and distributed training over DGL RecSys in 0.5. DGL-RecSys is still on our watch but may need more community engagement to drive the effort. That said, we may develop DGL-RecSys in a different way than DGL-KE and DGL-LifeSci, but the actual model is still under debate. If you have any suggestions, feel free to leave it here.

1 Like

Hi, thank you for the explanation.

I would suggest to have a simple tutorial for building bipartite graph based recommendation system by DGL. A simple tutorial will be helpful, and we can extend this simple model by integrating other DGL modules.

Thank you very much.
Lyman

Hi,

I would love to refine our tutorials. Which model do you suggest to have a tutorial? And what is the scale of the dataset you would expect (like MovieLens-100K or something much larger than that)?

1 Like

Hi,

Some well-known models (e.g., GAT or GraphSage) for recommendation system based on bipartite graph are all welcome.
We can try to integrate other modules by our own.

The scale in our scenario would only be around 10K~100K.
We may reach to the scale in millions, but still have a long way to go.

Thank you.
Lyman

In this case, we do have a tutorial in WWW’20 online for recommendation:

Please feel free to follow up. Thanks.

1 Like

Hi,

Does this tutorial uses GNMC as referred in the below paper


using Bilinear decoding for link prediction or not sure on the approach.

For simplicity it uses a dot product between user and item embedding and train with L2 loss, instead of the bilinear decoding (per rating value) and cross entropy loss. The GCMC example here shows an instance of a bilinear decoder which loyally reproduces the GCMC paper, which produces a likelihood for each rating value.

Hi Barclayll

Is there any workaround to avoid performance issues with the existing GCMC implementation with node features.

What kind of performance issue do you have?

I’m planning to use DGL with side-channel features in a heterogeneous graph as node features and it was mentioned in GCMC readme section that it has some performance issues while using node features.

That performance issue is talking about accuracy, which depends on the feature importance of your dataset. For instance, if your feature is noisy, including them might result in overfitting. You may need to try different inclusions of the features or different architectures to find the optimal solution.