MIND: multi-interest network with dynamic routing

MIND multi-interest retrieval with capsule-style dynamic routing and label-aware attention: 83.92% validation accuracy, HR@50 of 0.104 on MovieLens 1M.

Status: Research
Period: 2025
Area: Recommender systems
Code: Code public
Stack: Python
PyTorch Lightning
MovieLens 1M

Validation accuracy: 83.92%
HR@50: 0.104

User behavior
Dynamic routing (capsules)
Several interest vectors
Label-aware attention
Candidate retrieval

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

Problem

One embedding per user cannot represent someone with several unrelated interests.

Approach

I implemented the MIND multi-interest retrieval model, with capsule-style dynamic routing and label-aware attention, to represent each user with several interest vectors.

Results

83.92% validation accuracy.
HR@50 of 0.104.
Dataset: MovieLens 1M.