Learning to Cluster Documents into Workspaces Using Large Scale Activity Logs

Google Drive is widely used for managing personal and work-related documents in the cloud. To help users organize their documents in Google Drive, we develop a new feature to allow users to create a set of working files for ongoing easy access, called workspace. A workspace is a cluster of documents, but unlike a typical document cluster, it contains documents that are not only topically coherent, but are also useful in the ongoing user tasks.

To alleviate the burden of creating workspaces manually, we automatically cluster documents into suggested workspaces. We go beyond the textual similarity-based unsupervised clustering paradigm and instead directly learn from users' activity for document clustering. More specifically, we extract co-access signals (i.e., whether a user accessed two documents around the same time) to measure document relatedness. We then use a neural document similarity model that incorporates text, metadata, as well as co-access features. Since human labels are often difficult or expensive to collect, we extract weak labels based on co-access data at large scale for model training. Our offline and online experiments based on Google Drive show that (a) co-access features are very effective for document clustering; (b) our weakly supervised clustering achieves comparable or even better performance compared to the models trained with human labels; and (c) the weakly supervised method leads to better workspace suggestions that the users accept more often in the production system than baseline approaches.