System and method for multi-modal graph-based personalization
A method for learning latent representations of individual users in a personalization system uses a graph-based machine learning framework. A graph representation is generated based on input data in which the individual users are each represented by a node. The nodes are associated with labels. Node vector representations are learned by combining label latent representations from a vertex and neighboring nodes so as to reconstruct the label latent representation of the vertex and updating the label latent representations of the neighboring nodes using gradients resulting from application of a reconstruction loss. A classifier/regressor is trained using the node vector representations and the node vector representations are mapped to personalizations. Actions associated with the personalizations are then initiated.
1. A method for learning latent representations of nodes in a personalization system using a graph-based machine learning framework, the method comprising:
generating a graph representation containing the nodes based on input data;
associating the nodes with labels;
learning node vector representations by combining label latent representations from a vertex and neighboring nodes so as to reconstruct the label latent representation of the vertex and updating the label latent representations of the neighboring nodes using gradients resulting from application of a reconstruction loss, wherein the node vector representations are learned using message passing, wherein the label latent representations of the neighboring nodes are sent to the vertex from the neighboring nodes and are combined with a current label latent representation of the vertex so as to reconstruct the label latent representation of the vertex, and wherein the gradients are sent from the vertex to the neighboring nodes;
training a classifier/regressor using the node vector representations and mapping the node vector representations to personalizations; and
initiating actions associated with the personalizations.
2. The method according to claim 1 , wherein the steps of combining and updating in the learning of the node vector representations are repeated until a convergence threshold is reached.
3. The method according to claim 1 , further comprising computing a new node vector representation for a new node being added to the graph representation using label latent representations of neighboring nodes.
4. The method according to claim 1 , wherein the nodes represent patients, the graph representation is of a patient-disease network, the labels include symptoms of the patients, the personalizations include diseases detected by the classifier/regressor and the actions include generating a treatment plan based on the detected diseases.
5. The method according to claim 4 , further comprising computing a new node vector representation for a new patient added to the patient-disease network and creating edges with existing nodes that share at least some of the same symptoms as the new patient.
6. The method according to claim 1 , wherein the nodes represent geographical locations, the graph representation provides the geographical locations that are close to each other as neighboring nodes, the labels include public safety information of the geographical locations, the personalizations include a prediction of one or more of the geographical locations that has a higher cost derived from public safety incidents and the actions include sending an electronic alert to safety officials so that additional prevention and protection measures are deployed in the one or more of the geographical locations.
7. The method according to claim 1 , wherein the nodes represent online consumers, the graph representation provides the online consumers having similar online behavior as neighboring nodes, the labels include attributes of the online consumers, the personalizations include a product offered for sale online and the actions include generating an online advertisement for the product that is directed to a computing device of a respective one of the online consumers for which the classifier/regressor determined the personalization.
8. The method according to claim 1 , wherein the gradients are computed based on a loss between the reconstructed label latent representation of the vertex and the current label latent representation of the vertex.
9. The method according to claim 1 , wherein label latent representations of non-neighboring nodes are not used to reconstruct the label latent representation of the vertex, and wherein the gradients are not used for updating the label latent representations of the non-neighboring nodes.
10. The method according to claim 1 , wherein the gradients are derived from distances between outputs of functions applied to the label latent representations of the neighboring nodes and outputs of functions applied to the label latent representation of the vertex such that the gradients move the label latent representations in a Euclidean space of the graph representation until the reconstruction loss reaches a minimum.
11. A personalization system which learns latent representations of nodes using a graph-based machine learning framework, the personalization system consisting of one or more computer processors which, alone or in combination, provide for the execution of the following steps:
generating a graph representation containing the nodes based on input data;
associating the nodes with labels;
learning node vector representations by combining label latent representations from a vertex and neighboring nodes so as to reconstruct the label latent representation of the vertex and updating the label latent representations of the neighboring nodes using gradients resulting from application of a reconstruction loss, wherein the node vector representations are learned using message passing, wherein the label latent representations of the neighboring nodes are sent to the vertex from the neighboring nodes and are combined with a current label latent representation of the vertex so as to reconstruct the label latent representation of the vertex, and wherein the gradients are sent from the vertex to the neighboring nodes;
training a classifier/regressor using the node vector representations and mapping the node vector representations to personalizations; and
initiating actions associated with the personalizations.
12. The personalization system according to claim 11 , wherein the steps of combining and updating in the learning of the node vector representations are repeated until a convergence threshold is reached.
13. The personalization system according to claim 11 , wherein the one or more computer processors, alone or in combination, are further configured to provide for the execution of the step of computing a new node vector representation for a new node being added to the graph representation using label latent representations of neighboring nodes.
14. The personalization system according to claim 11 , wherein the nodes represent patients, the graph representation is of a patient-disease network, the labels include symptoms of the patients, the personalizations include diseases detected by the classifier/regressor and the actions include generating a treatment plan based on the detected diseases, and wherein the one or more computer processors, alone or in combination, are further configured to provide for the execution of the step of computing a new node vector representation for a new patient added to the patient-disease network and creating edges with existing nodes that share at least some of the same symptoms as the new patient.
15. The personalization system according to claim 11 , wherein the nodes represent geographical locations, the graph representation provides the geographical locations that are close to each other as neighboring nodes, the labels include public safety information of the geographical locations, the personalizations include a prediction of one or more of the geographical locations that has a higher cost derived from public safety incidents and the actions include generating an electronic alert to safety officials so that additional prevention and protection measures are deployed in the one or more of the geographical locations.
16. The personalization system according to claim 11 , wherein the nodes represent online consumers, the graph representation provides the online consumers having similar online behavior as neighboring nodes, the labels include attributes of the online consumers, the personalizations include a product offered for sale online and the actions include generating an online advertisement for the product that is directed to a computing device of a respective one of the online consumers for which the classifier/regressor determined the personalization.
17. The personalization system according to claim 11 , wherein the gradients are computed based on a loss between the reconstructed label latent representation of the vertex and the current label latent representation of the vertex.