Privacy-sensitive training of user interaction prediction models

A method performed by one or more data processing apparatus on a user device, the method comprising:

• maintaining a local instance of a machine learning model;
• receiving, by the user device, data defining current parameter values of the machine learning model over a data communication network, wherein the machine learning model is configured to process an input comprising:
• (i) a search query, and (ii) a data element, to generate an output which characterizes a likelihood that a given user would interact with the data element if the data element were presented to the given user on a webpage identified by a search result responsive to the search query;
• generating, by the user device, training examples for training the machine learning model based on interaction of a user of the user device with data elements presented to the user on webpages identified by search results responsive to search queries submitted by the user, wherein generating a training example for training the local instance of the machine learning model comprises:
• determining that the user of the user device:
• submits a particular search query to a search system, receives a set of search results responsive to the particular search query from the search system, navigates to a webpage identified by a particular search result, and interacts with a particular data element presented on the webpage;
• and
• generating a training example that comprises:
• (i) a training input, and (ii) a target output that should be generated by the machine learning model by processing the training input, wherein the training input comprises the particular search query and the particular data element, and wherein the target output comprises data indicating that the user interacted with the particular data element;
• training, by the user device, the local instance of the machine learning model on the training examples using machine learning training techniques to determine an update to the current parameter values of the machine learning model;
• and
• transmitting, by the user device and to a global training system, parameter update data defining the update to the current parameter values of the machine learning model over the data communication network;
• wherein the global training system is configured to:
• receive the parameter update data transmitted by the user device over the data communication network;
• and
• update the current parameter values of the machine learning model using the parameter update data received from the user device;
• wherein the user device is one of a plurality of user devices that receives data defining the current parameter values of the machine learning model and determines respective updates to the current parameter values of the machine learning model.

A system comprising:

• one or more computers;
• and
• one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:
• transmitting, by the one or more computers, data defining current parameter values of a machine learning model to each of a plurality of user devices over a data communication network, wherein:
• the machine learning model is configured to process an input comprising:
• (i) a search query, and (ii) a data element, to generate an output which characterizes a likelihood that a given user would interact with the data element if the data element were presented to the given user on a webpage identified by a search result responsive to the search query;
• and
• each of the plurality of user devices is configured to:
• maintain a local instance of the machine learning model;
• generate training examples for training the machine learning model based on interaction of a user of the user device with data elements presented to the user on webpages identified by search results responsive to search queries submitted by the user, wherein generating a training example for training the local instance of the machine learning model comprises:
• determining that the user of the user device:
• submits a particular search query to a search system, receives a set of search results responsive to the particular search query from the search system, navigates to a webpage identified by a particular search result, and interacts with a particular data element presented on the webpage;
• and
• generating a training example that comprises:
• (i) a training input, and (ii) a target output that should be generated by the machine learning model by processing the training input, wherein the training input comprises the particular search query and the particular data element, and wherein the target output comprises data indicating that the user interacted with the particular data element;
• train the local instance of the machine learning model on the training examples, using machine learning training techniques, to determine an update to the current parameter values of the machine learning model received from the one or more computers;
• and
• transmit, to the one or more computers, parameter update data defining the update to the current parameter values of the machine learning model over the data communication network;
• receiving, from each of the plurality of user devices and by the one or more computers, the parameter update data transmitted by the user device over the data communication network;
• and
• updating, by the one or more computers, the current parameter values of the machine learning model using the respective parameter update data received from each of the plurality of user devices.