TechPrufer
Get started
Dropbox
Dropbox logoDropbox · Feature trail

Content Suggestion ML Infrastructure

1 post · Oct 2019

What this trail covers

Dropbox is continuously improving its content suggestion capabilities by enhancing the underlying machine learning infrastructure and algorithms. This includes developing more sophisticated models to handle diverse content types (files, folders, cloud-based documents, Paper docs), incorporating richer signals like file extensions and filenames through embeddings and sequential models, and building improved tooling like dbxlearn for efficient hyperparameter tuning and experimentation. The goal is to provide users with more relevant and timely access to their important content.

All Dropbox trails
  1. Using machine learning to predict what file you need next, Part 2This post details the improvements made to Dropbox's content suggestion system, focusing on handling diverse content types and enhancing ML models. Key contributions include: 1. Developing file extension embeddings trained via a weakly-supervised task to handle heavy-tailed distributions and semantically similar extensions. 2. Incorporating filenames as a signal using char-RNNs for better semantic understanding. 3. Training separate models for cloud-based documents and Paper docs due to data limitations, merging results with a combiner model. 4. Decoupling file and folder suggestions, reusing pipelines but adapting candidate generation, signal fetching, and training data definitions for folders. 5. Introducing dbxlearn, an ML infrastructure tool for elastic computing and advanced hyperparameter tuning (Bayesian Optimization), enabling faster iteration and experimentation.