A latent space is a learned lower-dimensional representation space where users and items are positioned according to characteristics inferred from interaction data. In a matrix-factorization recommender, user and item factors form these representations: their distributions can reveal changing preferences or item popularity, while their proximity can support recommendation explanations. section 11.2.2 section 11.3.2
Use latent representations in a production recommender by:
- Train a matrix-factorization model on a user-item rating matrix. section 11.2.2
- Extract user and item embeddings directly from the trained model. section 11.2.2
- Compare factor distributions across retrained models to detect drift in preferences or item popularity. section 11.2.2
- Find similar users or items in the space and use neighborhood ratings to explain recommendations. section 11.3.2
Explainable movie recommendations
An explainable matrix-factorization recommender places similar users and items near one another. For one recommendation, an explanation can report that eight of nine similar users rated the movie four or higher, giving the prediction an interpretable neighborhood-based rationale. section 11.3.2
from deepchecks.tabular.checks import FeatureDrift
drift_check_item = FeatureDrift(
columns=["item_factor_2"]
).run(dataset_item_factors_t1, dataset_item_factors_t2)
drift_check_item.save_as_html()
Run a Deepchecks FeatureDrift check on an item-factor column to compare latent-factor distributions between datasets. section 11.2.2
