Architecture
The following figure shows the overall framework of cold start recommendation.
Cold start recommendation process
An app requests the recommendation service for recommendations.
The recommendation service calls the candidate generation module specific for cold start to obtain matched candidate items.
The recommendation service calls the ranking service specific for cold start algorithms and passes parameters such as the list of candidate items, user and item features, and the number of items that need to be returned.
The ranking service specific for cold start algorithms generates context feature vectors based on the feature generation configuration file. During this process, item features may need to be obtained from the feature storage system.
The ranking service specific for cold start algorithms synchronizes model parameters from a model parameter storage system such as a Hologres database and caches the obtained model parameters.
The ranking service specific for cold start algorithms calls the contextual bandit model to predict a score for each candidate item. This process needs to implement the exploration and exploitation policy to balance exploration and exploitation.
The ranking service specific for cold start algorithms selects the Top N candidate items based on the scores predicted in the previous step.
The ranking service specific for cold start algorithms writes the feature vectors of the Top N candidate items into the feature storage system in real time. This provides materials for subsequent model training.
The ranking service specific for cold start algorithms returns the Top N candidate items to the recommendation service.
The recommendation service combines the obtained Top N cold start items with traditional recommendations and returns them to the app for presentation.
The app uploads behavioral logs such as presentation and click logs to the log server in real time.
A Flink job reads the real-time behavioral logs collected on the log server from message-oriented middleware such as Kafka and DataHub.
The Flink job parses the behavioral logs and filters the behavioral logs of the target scenario based on configurations to build samples and train the model.
After a short period of time, the Flink job builds sample labels based on behavior types and obtains the feature vectors of samples from the feature storage system.
The Flink job performs online learning for model training, and synchronizes model parameters to the model parameter storage system in real time so that the ranking service specific for cold start algorithms can obtain the model parameters.
Data transmission view
