Deep Structured Semantic Model (DSSM)

Dual-tower DSSM for user-item matching on MovieLens: 81.74% test AUC and 74.56% test accuracy.

Status: Research
Period: Apr 2025
Area: Recommender systems
Code: Code public
Stack: Python
PyTorch Lightning
MovieLens

Test AUC: 81.74%
Test accuracy: 74.56%

User features
User tower
Item tower
Item features
Similarity score

Fig. 2 Data flow, left to right (top to bottom on small screens).

Problem

Match users to items with a model that can embed both sides independently, the building block for two-tower retrieval.

Approach

I built a dual-tower DSSM for user-item matching on MovieLens and improved it through architecture and training optimization.

Results

81.74% test AUC.
74.56% test accuracy.

DSSM is the same retrieval family behind the production news recommender.