Published at #IRRJ: "On the challenges of studying bias in Recommender Systems: The effect of data characteristics and algorithm configuration" by Savvina Daniil, Manel Slokom, Mirjam Cuper, Cynthia Liem, Jacco van Ossenbruggen, and Laura Hollink. #RecSys #Bias #DataSynthesis #Reproducibility
New at my blog: I'd been bothered for a few years now in teaching #recsys by the need for a biased lift computation for informative product-relatedness measures that aren't dominated by low-information items. So this term, I sat down to solve the problem. https://md.ekstrandom.net/blog/2025/01/biased-lift
Do you appreciate news? Do you want to help #recsys research? POPROX News is open for subscriptions!
The Platform for OPen Recommendation and Online eXpermentation is an #NSFfunded project to help scientists and journalists explore better ways to select and present news stories. Our MVP is a daily personalized newsletter — if you'd like to receive that, and periodic surveys, sign up at https://user.poprox.ai/enroll?source=ff&subsource=memd
After signup, you'll receive a confirmation email — check spam if needed.
#recsys fail: this is the second time — at least — LinkedIn only showed me the post(s) advertising an event I wanted to attend after registration closed.
might have developed a new #recsys hyperparameter search strategy in class last night.
CalvinSearch: try hyperparameter settings until The Complete Calvin and Hobbes is the top recommendation.
TikTok opensourcing recsys tooling, based on keras (from @fchollet.bsky.social on other place) github.com/bytedance/mo... #recsys
GitHub - bytedance/monolith: B...
This is just out, and looks very comprehensive: ACM Book "Information Retrieval: Advanced Topics and Techniques." Includes foundations of IR (neural search, #RecSys, etc.), Adaptations and concerns (conversational, privacy and bias), and IR for different verticals: dl.acm.org/doi/book/10....
Information Retrieval: Advance...
At the #dayofdh2024 introducing students on the basics of #textencoding (
#unicode, #xml, @TEIConsortium), #geneticediting , #criticalApparatus , managing https://dhinfra.at (#gpu clusters for the Austrian #dh community), talking about scholarly editing Ceija #Stojka notebooks https://www.ceijastojka.org/12924460-the-notebooks and chatting with Dominik Kowald from KNOW Center Graz about #recsys for DH. Now to preparing my presentation on https://gams.uni-graz.at/o:voccod the #SKOS version Denis Muzerelle's Vocabulaire de la #Codicologie .
Ik was te gast in de Filosofie in Actie Podcast, waar ik vrij mocht oreren over hoe tof aanbevelingssystemen eigenlijk zijn, ethische #AI, en de filterbubbel (sommigen geloven nog steeds in dit hypothetische construct). Beluister de podcast hier:
graus.nu/outreach/te-... #recsys
te gast in Filosofie in Actie ...
China is curbing addictive algorithms (and #recsys?).
"System providers should avoid recommendation algorithms that create “echo chambers” and induce addiction, allow manipulation of trending items, or exploit gig workers’ rights"
www.scmp.com/news/china/p...
Clear algorithms: China’s Big ...
And we have a #Dagstuhl Report!
Evaluation Perspectives of #RecSys. Edited by Christine Bauer, @evazangerle and myself. Written by a whole host of fantastic #recsys people. Too many to mention
https://drops.dagstuhl.de/entities/document/10.4230/DagRep.14.5.58
Still trying to figure out how to scratch your #recsys itch after this year's @Recsys? Have a look at the #ACMTORS Call for papers for the Special Issue on #RecommenderSystems for Good. Will you be be submitting a paper as an xmas gift to the RecSys community? // @Nava et al.
https://dl.acm.org/pb-assets/static_journal_pages/tors/pdf/TORS_SI-Recommender-Systems-for-Good-1721933817383.pdf
What if #RecSys put as much energy into explicitly anti-fascist recommendations as it does into viewpoint-diverse recommendations?
What if the goal isn’t to model tolerance for diversity, but to identify items that will increase the reader’s empathy and understanding of their neighbors and their needs?
This paper provides some thoughts applicable to that direction. https://doi.org/10.1108/JD-01-2020-0002
Part2: #dailyreport #negativesampleing #sampling #llm #recsys
target word w and the negative samples
For binary Classification: Negative sampling transforms
the problem into a series of binary classification tasks,
where the model learns to distinguish between positive
and negative samples.
Example "The dog is playing with a bone," and assume a
window size of 2 positive samples for the target word
"dog" would include:
- ("dog", "The")
- ("dog", "is")
- ("dog", "playing")
- ("dog", "with")
- ("dog", "a")
- ("dog", "bone")
Negative Samples: ("dog", "car"), ("dog", "apple"),
("dog", "house"), ("dog", "tree")
calc: logσ(vdog⋅vbone) +
logσ(−vdog⋅vcar)+logσ(−vdog⋅vapple)+logσ(−vdog⋅vhouse)
蠡
Part2: Part1: #dailyreport #negativesampleing #sampling #llm #recsys
We use:
: L = log(sigmoid(v_w * v_c)) + sum(log(sigmoid(-v_w * v_neg_i))) for i in range(k)
where:
- vw - vector representation of the target word
- vc - vector representation of context word
- v_neg_i - vector representations of the k negative
sample.
- k - number of negative samples
- log(sigmoid(v_w * v_c)) - positive term with dot product
or cosine simularity.
- sum(log(sigmoid(-v_w * v_neg_i))) for i in range(k) -
negative term - minimize the similarity between the
Part1: Part1: #dailyreport #negativesampleing #sampling #llm #recsys
Negative sampling used in NLP, RecSys,retrival and
classification tasks to address the computational
challenges associated with large vocabularies or item
sets. It modifies the training objective: Instead of
computing the softmax over the entire vocabulary, it
focuses on distinguishing the target word from a few
randomly selected "noise" or "negative" words.
Instead of loss:
: softmax(x_i) = e^(x_i) / (sum of e^(x_j) for all j from 1 to n)
: L = -log(p(w | c)) = -log(softmax(x_i))
This is your irregular reminder that “algorithm” isn’t a synonym for “model” in #RecSys. There’s a lot that goes into producing recommendations that is algorithmic but isn’t trained with gradient descent.
Parents please check your kids’ deep #RecSys models, I just found Neural Collaborative Filtering hidden inside yet another fancy architecture
we will be at #recsys2024 from monday on! with a set of papers, talks and presentations throughout the week.
Come join us on Monday, Oct. 14 for the 7th FAccTRec workshop at #RecSys2024! We've got a great program of papers, a keynote from @jvinagre, and discussion about fair and responsble #recsys. https://facctrec.github.io/facctrec2024/program/
