Engineering architecture

Architecture diagram

The following diagram shows the overall framework of the cold start pipeline:

Cold start pipeline workflow

1. The app requests a list of recommended content from the recommendation service.

2. The recommendation service calls the cold start recall module to retrieve matching candidate items.

3. The recommendation service calls the cold start algorithm scoring EAS service and passes parameters such as the candidate items to score, user/item features, and the number of items to return.

4. The cold start algorithm scoring EAS service generates context feature vectors based on the feature generation configuration file. This process may require retrieving item features from the feature storage system.

5. The cold start algorithm scoring EAS service synchronizes and caches model parameters from the model parameter storage system, such as a Hologres database.

6. The cold start algorithm scoring EAS service calls the model to calculate a prediction score for each candidate item. This step applies an E&E strategy to balance exploration and exploitation.

7. The cold start algorithm scoring EAS service selects the top N candidate items based on the scores from the previous step.

8. The cold start algorithm scoring EAS service writes the feature vectors of the top N candidate items to the feature storage system (backend log storage) in real time, providing data for subsequent model training.

9. The cold start algorithm scoring EAS service returns the top N candidate items to the recommendation service.

10. The recommendation service blends the N cold start items with the regular recommendation list and returns the combined results to the app for display.

11. The app uploads behavior logs, such as impressions and clicks, to a log server in real time.

12. A Flink task reads the real-time behavior logs from message middleware, such as Kafka or DataHub.

13. The Flink task parses the logs, filters the behavior logs for the target scenario based on its configuration, and uses these logs to build samples and train the model.

14. After a short delay, the Flink task creates sample labels based on the behavior type and retrieves the sample feature vector from the feature storage system.

15. The Flink task trains the model using online learning and synchronizes the model parameters to the model parameter storage system in real time, making the parameters available to the cold start algorithm scoring EAS service.

Data flow view