The Dropbox machine learning team shared in a recent blog post how the model behind their content suggestions feature allows it to deal with different types of content, incorporates folder suggestions into the existing file suggestions model, and handles cloud-based third-party content.
The content suggestions feature, which began rollout in April, added a suggested files section at the top of the Dropbox homepage, meant to facilitate quick access to the files users look for when using Dropbox, much like the Quick Access feature of Google Drive.
The feature started out supporting only file recommendations, and was based on a relatively shallow neural-network model that uses various salient signals together with task-agnostic Dropbox entity embeddings - which include deeply semantic vector representations of Dropbox users, files and folders - to predict the CTR (clickthrough rate) of the files in a heuristics-based candidate set. This approach is similar to the one used for the machine learning model behind Google’s Quick Access.
The recent improvements to the content suggestion model include several new components meant to address different challenges:
To improve predictions of the existing file suggestions model, the model incorporates more information on files, using both the name and the extension of the file. File extensions are incorporated into the feature vector not by using the naive one-hot encoding scheme, but by learning an embedding space in which semantically similar file extensions (e.g. .jpg and .png) are close, supposedly enabling the model to learn from unevenly distributed data. Additionally, an embedding for filenames was created by taking the state vector of a char-RNN, ingesting one character of the filename at a time, to help with detecting temporary file names.
Supporting suggestions for third-party cloud-based files resulting from recent partnerships, such as Google Docs and Microsoft Office 365 files, which suffer from a small amount of training data, was done by developing separate models. The same approach was also used to support suggestions for Dropbox Paper documents.
Support for folder suggestions was added by using frequently and recently used files as signals to suggest their parent folders might be a good candidate. Like with third-party files, a separate model was trained for file suggestions, with training data coming from both online experiments run with heuristic models and an unsupervised dataset of user events in Dropbox, both a prominent part of Dropbox's machine learning infrastructure.
Finally, the various models were composed into a single model by learning, for each model, a mapping from its generated score to the expected CTR for a suggested item of the given score. Since the CTR is comparable across all submodels, suggestions by all models can be ranked by their expected CTR.
This product enhancement effort can be seen as another aspect of the Dropbox intelligence initiative (DBXi), the company's effort to incorporate artificial intelligence into all of its products and services. This effort also includes automatic image text recognition feature, capable of optical character detection on photos of documents, including receipts and field reports, to enable users to search through more types of content, and integration of machine learning models into Dropbox's search capabilities.
The added ability to handle various types of content also complements the company’s efforts to position their Spaces product as a central employee collaboration tool, competing with prominent products such as Slack, Google Drive and Microsoft Teams.