This episode explores disentanglement and interpretability in recommender systems, contrasting handcrafted features with representation learning. Guest Irvin Derbyschei discusses his research, which quantitatively evaluated disentanglement, finding a strong positive correlation with interpretability but no consistent link to recommendation performance, suggesting a trade-off where disentanglement acts as a regularizer.
Summarized by Podsumo
Representation learning allows models to learn their own features, leading to complex latent spaces, while handcrafted features offer direct interpretability.
The concept of disentanglement aims for independent factors within a representation (e.g., t-shirt size vs. price), making models more interpretable by allowing isolated perturbation.
A strong positive correlation was found between disentanglement and interpretability, but crucially, no consistent correlation was observed between disentanglement and recommendation performance across various models and datasets.
Enforcing disentangled representations can act as a regularizer, potentially sacrificing some performance for increased interpretability, highlighting a crucial trade-off.
Many prior studies lacked specific hyperparameters, exact data splits, or ground truth for disentanglement evaluation, making their results difficult to reproduce.
"The entire premise of learning this antenna representation is that you can't find in the representation space, you can sort of separate these different aspects of a given input. And the idea here is that if you are able to do this separation, then the representation are more interpretable."
"The correlation between the disentanglement and interpretability was quite strong, whereas the connection between the disentanglement and the effectiveness of the performance of the recommendation models, there we did not find consistent correlation."
"when we push the model to learn disentangled representations, this sort of acts like regularization... you sacrifice a bit of performance and gain a bit more in expandability in interpretability."