Putting everything in its right place with ML-powered file organization

From Manual Drudgery to ML Suggestions

Dropbox's "smart move" feature, released in November 2021, uses machine learning to analyze a user's subfolder structure and suggest destination folders for files. Users can drop a batch of unorganized files into a home directory, and smart move suggests where each file might belong, prioritizing the most likely moves. The human-in-the-loop design lets users quickly scan, accept, reject, or edit recommendations before committing, reducing the tedium of one-by-one manual moves.

The initial scope was deliberately constrained: suggestions only applied to files in a chosen folder when a move to an immediate subfolder made sense. This kept the problem tractable and aligned with the "tidying up" persona the team wanted to support.

Defining "Organized" Is the Hardest Part

The biggest obstacles weren't ML model development; they were product design and user expectations. Organization is deeply personal. User research showed that many people are wary of letting others—even colleagues—move or rename their files. Any automation had to keep the user in control, with the ability to approve, edit, or reject every suggestion.

What does an organized folder even look like? There are many valid organizational patterns, and different users approach tidiness in entirely different ways. Relying on user research focusing on Dropbox on the web, the team acknowledged early that a single model was unlikely to perform well across all scenarios.

Several constraints shaped the work:

• Data sensitivity: Filenames are sensitive, so engineers couldn't manually review production records to develop hypotheses. Instead, Dropbox employees consented to share files from the company's internal Dropbox instance. New workflows and storage solutions were developed to keep this sensitive data isolated, with access restricted to a limited team for a defined period.
• Dataset curation: Many seemingly organized folders are auto-generated by desktop applications, which weren't part of the target use case. Significant time went into selecting, filtering, and validating data, with multiple rounds of cleaning as new edge cases surfaced.
• Latency: The UX flow serves recommendations synchronously—the user triggers the workflow and waits. Longer response times degrade the experience, so performance optimization was critical.

Why a Prior Model Didn't Carry Over

From 2018 to 2019, the Dropbox user research team interviewed power users about file organization on the web. One standout persona was the "organizer for teams"—a manager or administrator who spends substantial time renaming and moving files to keep content findable for their team. Using this research, the team built a prototype by repurposing an earlier "suggested destinations" model that used recent user activity and filenames to suggest a single folder for one file at a time.

Internal testing revealed the reuse wasn't viable:

1. Non-deterministic results: Suggestions were based on the user's most recent activity, so the same file could get different recommendations depending on navigation history—confusing for testers expecting consistency.
2. Unclear rationale: The model didn't reflect how users actually reason. Testers said they relied primarily on filenames (and occasionally file contents) when organizing, and the model's outputs didn't match those expectations.
3. Confidence tiers helped: Testers appreciated having suggestions split into high- and medium-confidence groups, letting them focus on the most likely beneficial moves.

The lesson: playing with a prototype surfaces unexpected needs that interviews miss. The team recommends building low-overhead prototypes early, as testers gave feedback that fundamentally reshaped the approach.

A New Model Trained on the "End State"

How do you predict how someone will organize files they haven't organized yet? The answer: find existing folders that look organized and treat them as the labelled end state of a successful smart move action. This file structure can be broken down into a hypothetical pre-move case, yielding millions of suitable training examples from Dropbox's internal data alone.

## Existing folder structure (desired end state)
root
|
+---- folder_A/ 
|          |
|          +---- file_1.pdf
|
+---- folder_B/
|          |
|          +---- file_2.pdf
|                    
+---- folder_C/
     |
     +---- file_3.jpg
     +---- folder_D/
     +---- folder_E/
           |
           +---- file_4.pdf

The model relied on a minimal but effective set of signals:

• The name of the file being tidied (including its extension).
• The name of each candidate folder.
• The names of files and folders within each candidate folder—potential siblings. These often reveal the folder's true purpose better than its name (e.g., files like w-2.pdf, taxreturn_2020.pdf, and 2019taxes.pdf clarify the intent of a folder named simply finance).

Internally, the team tested a simple filename-similarity heuristic against a trained neural network. The eventual architecture tokenized the file being moved (the context), candidate folder names, and potentential sibling names. These tokens were passed through an in-house encoder using character-level and GloVe word-level embeddings, capturing semantic similarity and file type relationships—understanding that png, jpg, and img are all image types, for instance.

Embeddings for context, candidate folder, and siblings were converted into similarity matrices, then combined with basic feature engineering and selective weighting for certain file and folder types. The features fed into a deep neural network with dropout, testing a range of architectures before settling on one with under 20 hidden layers. Options like label smoothing and architecture changes were explored along the way.

The model scored each file/candidate-folder pair. The top-ranked candidate became the suggested destination. Scores also drove confidence tiers: roughly the top 20% of scored recommendations were high-confidence and shown most prominently; medium-confidence suggestions appeared less centrally but remained available in the review screen; the bottom tranche wasn't displayed at all.

Validation relied on usage patterns from Dropbox's internal instance. Since the dataset came only from approved folders within the company's enterprise account, small-scale manual review and data cleaning were essential—even modest review captured undocumented assumptions about what smart move should do that were trip-ups for the proof of concept.

Heuristic vs. model: what the numbers said

Offline evaluation on internal data gave the trained model a slight edge over the similarity heuristic when classifying (file, candidate folder) pairs. That outcome aligned with expectations: a model should capture more complex relationships between file and folder names — semantic links like a document mentioning health insurance belonging under “Medical Docs” — as well as patterns in folder contents and file extensions, e.g., “Vacation 2022” being image-heavy and thus a more likely home for beach.png and summer_trip.jpg. For both approaches, adding child features from files and folders already sharing the candidate folder improved results.

Internal testing largely mirrored those offline findings, with the model again outperforming the heuristic. But the picture changed once suggestions went into the end-to-end UX for a limited external alpha. There, the heuristic slightly beat the model in raw acceptance: 61% of suggestions were accepted without edits versus 59%. When restricted to the high-confidence suggestions shown prominently in the UI, the heuristic’s acceptance rate jumped above 94%, with the model at 90%.

That reversal is not necessarily a mark against the model. Internal Dropbox organization habits may have skewed training, and user priorities for how they organize their own content differ. The model was, however, reusable for other feature prototypes, where it did outpace the heuristic.

Accuracy isn’t everything

The alpha also underscored that raw prediction accuracy is a poor proxy for user satisfaction. A recommendation could be technically correct — a reasonable match for a folder’s contents — yet still appear wrong to the user. Conversely, users sometimes accepted a folder that seemed less obviously right than an alternative suggestion, based on where they ultimately relocated the files. Making smart move genuinely useful will require quantifying ease of use and interpretability in ways that capture that qualitative feedback.

Applied lessons beyond smart move

The smart move work was reframed as a reusable capability for rapid prototyping rather than a single feature. Notably, the model was redeployed to suggest folder destinations for files not yet in Dropbox — in the Save-to-Dropbox browser extension and for bulk upload in Dropbox on the web. Those cases needed quick suggestions based on minimal data from the files in question. Interestingly, the model outperformed the heuristic there: it handled sparse, missing fields better, while the heuristic faltered when many attributes were absent. That work ultimately wasn’t continued, but it confirmed that fast, low-data suggestions matter.

Two areas stand out for future investigation:

1. Alternative UX and workflow tools for organization. ML-assisted filing appears promising for reducing toil beyond the team admin/organizer persona, but extra processes must not clutter the user experience.
2. Fine-tuning an LLM to compare performance. When the smart move model was built in 2021, LLM frameworks were less accessible. It would be informative to benchmark an LLM against a model trained exclusively on internal data, particularly for non-English languages.

Acknowledgements: Thanks to Morgan Zerby and Tristan Inghelbrecht for product management support, as well as Theo Champlin, Mike Lyons, and Jiayi Zeng for their work on the human-in-the-loop content organization experience.