Back office ops · Production

Dropbox builds ML-powered content suggestions to surface relevant files for users

The problem

Searching through Dropbox content was tedious for users, who lacked a smart, contextual way to surface the files they needed without manual browsing.

First attempt

Initial rule-based heuristics showed unrelated files together, surfaced files accessed by background programs rather than users, and grew too complex to maintain; CTR also proved a poor and slow proxy for measuring model accuracy.

Workflow diagram · grounded in source
1
Content suggestions triggered
trigger
“we built content suggestions to make it easier to find the files you need, when you need them”
2
Fetch candidate files
integration
“we can limit to the most recent files that the user has interacted with, without a significant loss in accuracy”
3
Fetch signals per file
ai_action
“For each candidate file, we need to fetch the raw signals we're interested in related to that file. These include its history (of opens, edits, shares, etc.), which users have worked on the file, and other properties of the file such as …”
4
Encode feature vectors
ai_action
“we encode the raw signals into what is called a feature vector”
5
ML model scores and ranks
ai_action
“We pass the feature vector for each file to the ranking algorithm, get back a score per file, and sort by that score”
6
Permission check before display
validation
“The top-ranked results are then permission-checked again before being shown to the user”
7
Hit ratio feedback loop
feedback_loop
“For any given suggestion, we checked if the user accessed that file in the subsequent hour, no matter how they arrived at the file. If so, we counted it as a "hit". We could measure this both online and offline (by looking at historical …”
Reported outcome

Replacing heuristics with an iteratively improved ML model significantly boosted the hit ratio and overall click-through rate for file suggestions.

Reported metrics
Hit ratiosignificantly boost our hit ratio
click-through rate (CTR)improve overall CTR
Reported stack
StormcrowSVMneural networkLearning-To-Rank
Source
https://dropbox.tech/machine-learning/content-suggestions-machine-learning
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Replacing heuristics with an iteratively improved ML model significantly boosted the hit ratio and overall click-through rate for file suggestions.

What tools did this team use?

Stormcrow, SVM, neural network, Learning-To-Rank.

What results were reported?

Hit ratio: significantly boost our hit ratio; click-through rate (CTR): improve overall CTR (source-reported, not independently verified).

What failed first in this deployment?

Initial rule-based heuristics showed unrelated files together, surfaced files accessed by background programs rather than users, and grew too complex to maintain; CTR also proved a poor and slow proxy for measuring mo…

How is this back office ops AI workflow structured?

Content suggestions triggered → Fetch candidate files → Fetch signals per file → Encode feature vectors → ML model scores and ranks → Permission check before display → Hit ratio feedback loop.

Related back office ops cases