Problem
LinkedIn's JYMBII ("Jobs You May Be Interested In") recommends jobs from the member's profile. A profile misses a change in direction: during the pandemic many members began dismissing jobs like their current one and applying to remote-friendly ones. The usual fix is to sum or average sparse features (such as required skills) over the jobs a member interacted with. That feature has tens of thousands of dimensions, needs its own aggregation for each entity (skills, industry, company) and uses the simplest aggregation function.
Example from the article: a member applies to jobs 33, 11 and 54; the "activity skills" feature is the elementwise average of those jobs' skill vectors.
Solution in brief
Represent each job by a dense embedding and aggregate the member's last 28 days of actions into one activity embedding. Start with a recency-weighted average, then replace it with a CNN trained to predict the last action from the earlier ones. Random negatives and a sliding window fix the training data. Embeddings are precomputed on a schedule.
Questions this case answers
- How to represent a user's recent action history as one dense vector when summing sparse item features is too large and ignores order and recency
- How to train a sequence encoder on implicit feedback where about 97% of labels are positive and most users have few actions
- How to turn a user's recent activity history into a compact feature for a recommendation model?
- How to make recommendations follow a change in what the user wants, which the profile cannot show?
Tools and methods
- Job embeddings (Pensieve) — dense job vectors, the input of every model
- Plain and decaying average — baselines per action type
- CNN sequence model — learned aggregation over the mixed action sequence
- Random negatives, sliding window — training data fixes
- TensorFlow — packaging of every aggregation model for inference
- Feature Marketplace — LinkedIn's feature store serving the embeddings
Google or a one-time email code. No password.