Distributed machine learning in edge computing

A system for distributed machine learning in an edge computing environment, comprising:

• a processing unit;
• and
• a memory coupled to the processing unit and storing instructions thereon, the instructions, when executed by the processing unit, performing actions comprising:
• monitoring a status of a connection between a computing system and a plurality of edge nodes, the plurality of edge nodes configured to receive a prediction model from the computing system via the connection, the prediction model being trained by the computing system on training data received from the plurality of edge nodes;
• determining the status of the connection between the computing system and the plurality of edge nodes has changed from a normal status to an abnormal status, with the abnormal status being a weak connection mode or an offline mode;
• in response to determining the status is an abnormal status, identifying according to a parameter server machine learning framework, at least one server node and a group of worker nodes from the plurality of edge nodes, the group of worker nodes configured to train a current prediction model using training data distributed from the plurality of edge nodes, the at least one server node configured to aggregate training results received from the group of worker nodes;
• determining a routing path for distributing the training data from the at least one server node to the group of worker nodes based on the abnormal status of the connection;
• distributing the training data from the at least one server node to the group of worker nodes via the routing path;
• training, by the group of worker nodes, a current prediction model received from the computing system using the training data received via the routing path to create a trained prediction model;
• and
• responsive to determining the status of the connection has returned to the normal status, causing the trained prediction model to be trained by the computing system on training data received from the plurality of edge nodes via the connection.

A computer program product for distributed machine learning in an edge computing environment, being tangibly stored on a non-transient machine-readable medium and comprising machine-executable instructions, the instructions, when executed on a device, causing the device to perform actions comprising:

• monitoring a status of a connection between a computing system and a plurality of edge nodes, the plurality of edge nodes configured to receive a prediction model from the computing system via the connection, the prediction model being trained by the computing system on training data received from the plurality of edge nodes;
• determining the status of the connection between the computing system and the plurality of edge nodes has changed from a normal status to an abnormal status, with the abnormal status being a weak connection mode or an offline mode;
• in response to determining the status is an abnormal status, identifying, according to a parameter server machine learning framework, at least one server node and a group of worker nodes from the plurality of edge nodes, the group of worker nodes configured to train a current prediction model using training data distributed from plurality of edge nodes, the at least one server node configured to aggregate training results received from the group of worker nodes;
• determining a routing path for distributing the training data from the at least one server node to the group of worker nodes based on the abnormal status of the connection;
• distributing the training data from the at least one server node to the group of worker nodes via the routing path;
• training, by the group of worker nodes, a current prediction model received from the computing system using the training data received via the routing path to create a trained prediction model;
• and
• responsive to determining the status of the connection has returned to the normal status, causing the trained prediction model to be trained by the computing system on training data received from the plurality of edge nodes via the connection.