System, method, and computer-readable media for leakage correction in graph neural network based recommender systems
Systems, methods, and computer-readable media provide a graph processing system that incorporates a graph neural network (GNN) based recommender system (RS), as well as a method for training a GNN based RS to address feature leakage that leads to overfitting of the trained GNN based RS. A message correction algorithm is used to modify a user node embedding and a positive item node embedding generated by the graph neural network when generating mini batches of training triples used to train the GNN based RS. The GNN message passing operations are performed on one graph only, in contrast to existing approaches which typically run GNN message passing operations on multiple adjusted input graphs constructed for multiple training triples.
1 . A method, comprising:
training a graph neural network based recommender system to make recommendations in respect of a bipartite graph, the training including:
(i) receiving a mini batch comprising a plurality of training triples, each training triple comprising a user node of the bipartite graph, a positive item node having an edge connecting the positive item node to the user node, and a negative item node not having an edge connecting the negative item node to the user node;
(ii) for each training triple, processing the training triple and the bipartite graph to generate, using a graph neural network:
a user node embedding;
a positive item node embedding; and
a negative item node embedding; and
(iii) for each training triple, modifying, using a message correction algorithm, the user node embedding and the positive item node embedding generated by the graph neural network, comprising:
processing the training triple, using the graph neural network, to further generate one or more intermediate user node embeddings, and one or more intermediate positive item node embeddings;
applying a first function to the one or more intermediate user node embeddings generated by the graph neural network to generate a first output of the first function;
applying a second function to a final user node embedding generated by the graph neural network and the first output of the first function to generate a modified user node embedding;
applying the first function to the one or more intermediate positive item node embeddings generated by the graph neural network to generate a second output of the first function; and
applying the second function to a final positive item node embedding generated by the graph neural network and the second output of the first function to generate the modified positive item node embedding;
wherein the first function is a linear combination function, and the second function is a subtraction function,
wherein the graph neural network including the modified user node embedding and the modified positive item node embedding is used to make one or more recommendations for a user based on positive items that the user has previously interacted with.
2 . The method of claim 1 , further comprising:
processing the modified user node embedding, the modified positive item node embedding, the negative item node embedding, and the bipartite graph to compute a loss function.
3 . The method of claim 2 , further comprising:
processing the loss function and the graph neural network to adjust a plurality of learnable parameters of the graph neural network.
4 . The method of claim 1 , wherein:
the message correction algorithm modifies the user node embedding and the positive item node embedding in parallel.
5 . The method of claim 1 , wherein:
the first function and second function are selected based on an architecture of the graph neural network.
6 . The method of claim 1 , wherein:
the message correction algorithm generates the modified user node embedding and the modified positive item node embedding by correcting for feature leakage.
7 . The method of claim 1 , wherein:
the message correction algorithm generates the modified user node embedding and the modified positive item node embedding by approximating a correction for feature leakage.
8 . A system comprising:
one or more processor devices; and
a memory storing instructions which, when executed by the one or more processor devices, cause the system to train a graph neural network based recommender system to make recommendations in respect of a bipartite graph, the training including:
(i) receiving a mini batch comprising a plurality of training triples, each training triple comprising a user node of the bipartite graph, a positive item node having an edge connecting the positive item node to the user node, and a negative item node not having an edge connecting the negative item node to the user node;
(ii) for each training triple, processing the training triple and the bipartite graph to generate, using a graph neural network:
a user node embedding;
a positive item node embedding; and
a negative item node embedding; and
(iii) for each training triple, modifying, using a message correction algorithm, the user node embedding and the positive item node embedding generated by the graph neural network, comprising:
processing the training triple, using the graph neural network, to further generate one or more intermediate user node embeddings, and one or more intermediate positive item node embeddings;
applying a first function to the one or more intermediate user node embeddings generated by the graph neural network to generate a first output of the first function;
applying a second function to a final user node embedding generated by the graph neural network and the first output of the first function to generate a modified user node embedding;
applying the first function to the one or more intermediate positive item node embeddings generated by the graph neural network to generate a second output of the first function; and
applying the second function to a final positive item node embedding generated by the graph neural network and the second output of the first function to generate the modified positive item node embedding;
wherein the first function is a linear combination function, and the second function is a subtraction function,
wherein the system uses the graph neural network including the modified user node embedding and the modified positive item node embedding to make one or more recommendations for a user based on positive items that the user has previously interacted with.
9 . The system of claim 8 , further comprising:
processing the modified user node embedding, the modified positive item node embedding, the negative item node embedding, and the bipartite graph to compute a loss function.
10 . The system of claim 9 , further comprising:
processing the loss function and the graph neural network to adjust a plurality of learnable parameters of the graph neural network.
11 . The system of claim 8 , wherein:
the message correction algorithm modifies the user node embedding and the positive item node embedding in parallel.
12 . The system of claim 8 , wherein:
the first function and second function are selected based on an architecture of the graph neural network.
13 . A non-transitory computer-readable medium having instructions tangibly stored thereon that, when executed by a processor, cause the processor to perform the method of claim 1 .
14 . The system of claim 8 , wherein:
the message correction algorithm generates the modified user node embedding and the modified positive item node embedding by correcting for feature leakage.
15 . The system of claim 8 , wherein:
the message correction algorithm generates the modified user node embedding and the modified positive item node embedding by approximating a correction for feature leakage.