Evgeny Frolov

Making Recommender Systems better.

[Evgeny Frolov @evfro]

Fixed our text following reviewers comments, final preprint: https://t.co/CluweDnkb8.
Interesting to note that if stability constraints are not imposed, #recsys inconsistencies increase as latent space grows, i.e., exactly when recommendations are expected to be more personalized

[Evgeny Frolov @evfro]

Turns out, even models with deterministic output like #PureSVD may generate inconsistent recommendations between two consecutive trainings. Even if changes in data are small. In our new #recsys work accepted at #UMAP2021 we show how to fix it with dynamical low-rank approximation

[Evgeny Frolov @evfro]

Turns out, even models with deterministic output like #PureSVD may generate inconsistent recommendations between two consecutive trainings. Even if changes in data are small. In our new #recsys work accepted at #UMAP2021 we show how to fix it with dynamical low-rank approximation

[Evgeny Frolov @evfro]

@karlhigley @Thanos98585056 Also, observed same effects in hybrid generalization of PureSVD https://t.co/RZQ41gBUZo. Side information didn not help with the coverage. Normalization trick did.

[Evgeny Frolov @evfro]

@karlhigley Si, for PureSVD it's easy to achieve with proper normalization of rows/cold. It was proposed by @Thanos98585056 in their Eigenrec paper. I have experimented with it as well here w.r.t. coverage (see bottom graph) https://t.co/6d5Td48Cnh

[Evgeny Frolov @evfro]

@karlhigley I wish there would be more analysis with PureSVD included, though. It's much easier to tune and it has that clear dependency on rank vs sensitivity to variations in user behavior.

[Evgeny Frolov @evfro]

@karlhigley I see it that way: initially you recommend mostly popular items, which leads to small coverage. When you make your algorithm more sensitive to variations in user behavior, you at the same time increase coverage and improve quality for those users with more subtle tastes.

[Evgeny Frolov @evfro]

@karlhigley Also, I'd argue that the ability combating mainstream bias should result in better item catalog coverage without penalizing top-n recs quality metrics. And there're much simpler ways to do that, especially with standard SVD, which is not even suitable for RMSE comparisons.

[Evgeny Frolov @evfro]

@jonathanstray @ruchowdh @geomblog @ndiakopoulos @mdekstrand @LaraRedmer I'd add that modern recsys are not just bucketing. Roughly, they represent users/items as mixtures of "buckets" (e.g., latent factors) and each bucket by itself is a mixture of various unobserved/implicit aspects that govern decision making. Lots of complexity is captured there.

[Evgeny Frolov @evfro]

Nice to see that the book we contributed a chapter in is listed by @bookauthority as the recommended reading on the #recsys topic for this year! Thanks to @slavaxx, @icantador, and @domonkostikk for making this book! https://t.co/vx4BwIkCWJ

[Evgeny Frolov @evfro]

@karlhigley Oh yes, that's true. In fact, FM can be obtained by reducing tensor representation and restricting some interactions between factors (I hypothesize it's how Rendle initially came up with FM after or in parallel with developing his PITF model)

[Evgeny Frolov @evfro]

@karlhigley This way you can have a more general representation than the one based on clustering. It may be not perfect for interpretability, but can give you a better accuracy (similarly to how NNMF gives interpretability but looses in accuracy comparing to standard MF).

[Evgeny Frolov @evfro]

@karlhigley In the PinnerSage paper they implicitly build context out of past user actions based on clustering. But you can go with a tensor formulation as well: instead of user-item coordinates (matrix) switch to user-items-items tensor, where the 3rd coordinate encodes previous items

[Evgeny Frolov @evfro]

@govindgnair You can create a rank-1 matrix with column means and take truncated SVD using "sparse matrix + rank-1 matrix" form. Note, you don't need to explicitly form a dense matrix, as you only need matrix-vector multiplication. There're also more sophisticated techniques, e.g. SoftImpute.

[Evgeny Frolov @evfro]

@vlcdn 26. если работаете с разреженными данными вместо картинок (recsys, например), размещайте их сразу все на ГПУ и не используйте встроенное итерирование по батчам (будет медленно). Для этого можно воспользоваться специальными даталоадерами, например https://t.co/iCwYwRxGZY

[Evgeny Frolov @evfro]

Will be virtually presenting our #recsys2020 work today at the poster session starting at midnight (Moscow time). It's gonna be a challenge, never had to do this so late 🌆 . Also hope my neighbours won't hear me 💤🌛

[Evgeny Frolov @evfro]

@danya_merkulov Except for the music, which was added separately.

[Evgeny Frolov @evfro]

This year #RecSys2020 features teaser videos and links them to the accepted contributions! Have a look at our teaser at https://t.co/3V3Hn7gO2G