Lecture 43 - Collaborative Filtering | Stanford University

🔔 Stay Connected! Get the latest insights on Artificial Intelligence (AI) 🧠, Natural Language Processing (NLP) 📝, and Large Language Models (LLMs) 🤖. Follow ( / mtnayeem ) on Twitter 🐦 for real-time updates, news, and discussions in the field.
Check out the following interesting papers. Happy learning!
Paper Title: "On the Role of Reviewer Expertise in Temporal Review Helpfulness Prediction"
Paper: aclanthology.org/2023.finding...
Dataset: huggingface.co/datasets/tafse...
Paper Title: "Abstractive Unsupervised Multi-Document Summarization using Paraphrastic Sentence Fusion"
Paper: aclanthology.org/C18-1102/
Paper Title: "Extract with Order for Coherent Multi-Document Summarization"
Paper: aclanthology.org/W17-2407.pdf
Paper Title: "Paraphrastic Fusion for Abstractive Multi-Sentence Compression Generation"
Paper: dl.acm.org/doi/abs/10.1145/31...
Paper Title: "Neural Diverse Abstractive Sentence Compression Generation"
Paper: link.springer.com/chapter/10....

[@ahishchinnu1843]

This is hands down the best video I have ever seen that instills the intuition for the concept. Absolutely brilliant!!!

[@SpaceXevent6]

Subscribe To get Ethiopian Education Tutorials የኢትዮጲያን ትምህርቶች ቲቶርያል ከፈለጉ ሰብስክራይብ በማድረግ ይመልከቱ

[@23karthikb]

Finally, a good concise explanation of these concepts!

[@prabhacar]

Your explanation of why item based CF is better than user-based CF in the last 60 seconds was very intuitive and logical. thanks for bringing so much clarity through easy-to-understand movie examples.

[@mathsLaw]

How simple and how elegant , i guess this is the Stanford way , thank you !

[@subratpati9166]

Thanks a lot......Finally found some tutorial which is easy to learn !! Great work

[@PotatoMan1491]

Prof is excellent in delivering essential pieces, I find this very informative

[@parthgupta6604]

One of the best videos on collaborative filtering!

[@yakhoubsoumare6944]

Sir, you are a genius!

[@marigam]

Thank you so much! I’ve heard it tried to be explained before but it was too difficult! This actually made sense! 💜

[@tanmaythaker2905]

Best explanation ever! Simple yet comprehensive.

[@dewinmoonl]

what an excellent video. clear with examples, explain the reasonings. stanford is no joke :)

[@shubhamtalks9718]

Finally, someone explained it properly.❤

[@Girisuable]

Awesome explanation. Very clear with the scenarios. Thank you so much.

[@curryfavours]

Very insightful comparison of item-item vs. user-user approaches towards the end.

[@jackshaak]

Thank you for putting this together in a fairly easy way to understand. You're a very good teacher.

[@SanjarAhmadov]

Very insightful. Thank you.

[@manisthashrestha250]

thank you! that was great :D

[@luis-benavides]

Great way to explain it! Thank you.

[@jakman85]

Excellent explanation over neighborhood-based CF. Anyway you could explain Factor-Analysis based CF? Thanks!

[@akshaysrivastava4304]

great video, very informative, thanks professor!

[@gourabnath9789]

Awesome explanation! Thank you

[@connorphoenix1107]

Thank you this is exactly what I was looking for

[@lucaecari2928]

Thank you, very nicely explained, both the intuition behind the concepts and the theory.

[@johng5295]

Thanks in a million. Awesome. Where have you been all these years.

[@karthikrajeshwaran1997]

Helped me totally grok! Thanks so much prof.

[@lukamandic5700]

At Option 3 it is mentioned that cosine distance treats zero value of movies as negative. Actually cosine zero values do not have impact on cosine distance. Cosine distance: 1 - (A · B) / (||A|| * ||B||) (A · B) represents the dot product of vectors A and B. ||A|| represents the Euclidean norm or magnitude of vector A. ||B|| represents the Euclidean norm or magnitude of vector B. If one component of vector A or Vector is 0 then it won't have impact on cosine similarity.

[@yuzhao5268]

great lecture- wondering what are some deep learning based approach to solve this problem today

[@shreyaraut9270]

thank you... explanation is simple and understandable

[@saqisyed5299]

Wow. Now this is what I was looking for.

[@josewesley2384]

Thank you, it's the best explanation that i saw by now. The idea is very simple, how i didn't think on it ?

[@ernestdesouza8888]

great video !!!!

[@ericarnaud5062]

Amazing explanation

[@mr68clubshorts]

Great work I dont know how to thank you really

[@saurabhsingh-nm8zj]

Excellent way of teaching

[@sparshnagpal1509]

Best Description.

[@ramiroarizpe3135]

My Sat Guru came to save me, thanks, really usefull information.

[@shauvikpujari9522]

Thank u sir your explanation is awesome

[@Krimson5pride]

great explanation. So you know both baroque and acid rock? I will make sure this guy gets more exposure.

[@HazemAzim]

Excellent .. Very well structured

[@Shivamagrawal50]

Hi Sir, like you said Pearson's correlation similarly takes out the individual users harshness to the rating scale while finding out user to user collaborative filtering, but why use same method for item to item collaborative filtering?

[@dmitrizakharov9068]

Hi, thanks for the great explanation. Is there a reason to use centered cosine distance over normalized cosine distance? Would dividing by the standard deviation of the row after subtracting the mean not better account for different levels of variance within user ratings? Is there a downside to normalizing instead of centering?

[@rc_matic]

What you're describing is standardization, not normalization. In general, that would be an idea worth trying. However in this case, an issue with standardizing here MIGHT be the fact that a user's ratings are not normally distributed. Also, since standard deviation relies on the sample/population size, you could get vastly different results comparing a user who has seen 10 movies and a user who has seen 1000 movies.

[@whiteF0x9091]

Excellent ! Thank you.

[@qihanguan8105]

Thank you! Super helpful!

[@danielclark739]

I get what you're saying that 0 is not a great assumption for a rating if a user did not watch a movie because they might have liked it if they watch it. However, it ignores the notion that a user had the opportunity to watch and wasn't interested enough to do so. Which is data that reflects their interest in a movie.

[@rpcroft]

Does one need to restrict the similarity set to positive values? I calculated the entire matrix (item-item) using python and pandas. I had values outside the 1 to 5 range. For example the inferred value for user2 / film2 was 20.07 and it was -1.27 for User4 / film 1. Negative centered cosine similarities seem to throw the whole thing out of whack. It is possible to have a small denominator in the 18.43 calculation magnifying values and also negative results.

[@junaidahmad-oe4oh]

Very Informative

[@thasin9671]

very very very very ... helpful.

[@rayane4975]

great explanation!

[@binhnguyenan6549]

Thanks a lot!

[@albb762]

Centered cosine similarity is a very bad algorithm for this problem, because a person rated all the movies with 5 star will be similar to the person who rated all the movies with 1 stars.

[@aashudwivedi]

This example is the case where one of the person is a tough rater and the other one is an easy rater.

[@himanshuladia9099]

what if the ratings are equal and therefore absolute vector is equal to zero ??? we will can't divide them on zero @@aashudwivedi

[@harshsandesara9470]

I always wondered why we needed to center everything around 0 instead of just filling in the missing values with the average ratings... basically the same thing but the calculations are done in the original frame instead of a shifted frame, and the problem that you mentioned most likely wouldn't arise then

[@abhinaygupta]

Great Explanation.

[@Robin-zc2iw]

Very helpful, thanks man

[@Flaviaestat]

Great explanation

[@abhisheksahu6032]

can anyone explain how do we pick the neighbor hood size?

[@alvaropatricioiribarren3791]

Thank you!

[@darkeddarksecret6279]

good day sir, how to compute the cosine similarity between the row? like 1.00, -0.18, 0.41, 0.10, -0.31 and 0.59. please replay asap i needed it in my thesis project, this explaination helps me a lot sir.

[@Skandawin78]

I like this video. But I'm confused as we are trying to solve 2 problems here, 1. Find the missing ratings 2. Find the item (IBCF) or user (in case of UBCF) which/who are similar to each other we use centering+ cosine similarity and KNN to address problem 1 or 2 or both? We use rating to identify similar user/items or do we find similar items/users and then fill the missing rating?

[@aydin3852]

thank you, now I get it

[@codelover847]

In item-item , shouldn't we be computing cosine similarities between colums?? Because Si,j denotes the similarity of items (i.e movies).

[@edwinchong37]

Better then my Prof.!!!

[@foodie2841]

Sir I have a query if I want to find ratings for movie 3 by user 4 then what should I do?

[@nginfrared]

Awesome explanation

[@ArtificialIntelligenceAllinOne]

Thanks!! Keep Learning and Sharing :)

[@aramun7614]

Very good explanation.

[@abdelrhmanwagdy3383]

what if the rates are equal and there absolute vector is equal to zero ??? we will can't divide them on zero

[@ramans11]

very nice..

[@PumarkoHD]

When predicting the weighted avg at around 18:40, is it a coincidence that .41 and .59 add up to 1? Or should that always be the case?

[@jrafaelvalle]

It's a coincidence. They represent the similarity, sim, between 1 and 3 and 1 and 6, sim(1, 3), sim(3,6). Consider k values different from 2, for example 1 or 3...

[@mamatizm817]

i need an explanation for propagation affinity algorithme with example please

[@mohammadhegazy1285]

Thank you Very much sir

[@celinexu6598]

Tried the centralized consine, have 2 quetions here: 1. if we have the example as D (in the vedio) which have all the record rating as 0, the consine distence will any other member A,B,C will be NaN. becuase, there is O at the denominator. any treatment? 2. if i have 1000 customer then I have to comuter 499500 number for any 2 of the customer Combinations. If i have 1 Million customer , the computation is really large, not sure nomal server can handel. Any suggestion or short cut? Thank you!

[@aymoon8080]

Nice work

[@eti-inyeneibia6442]

i couldnt get 1.00 , 0.18 0.41 with the cosine similiarity formula..pls explain

[@RahulPatel-hr4qe]

give me your email-id

[@megabites3649]

@@RahulPatel-hr4qe can you expalain it to me? [e-mail]

[@philipjaykho114]

@@RahulPatel-hr4qe please help me getting those answer too Sir. this is my email address [e-mail].. thanks a lot sir

[@almasazimkhan7170]

In the formula(option2) of user-user collaborative filtering we have "Ryi" which is rating. But what we should do if Ryi in original dataset is 0. Should we ignore it?

[@rhealisa9268]

It is not important. The formula uses the *sum* of Ryi ratings, so absence of rating (in this case, considered zero) doesn't make a difference.

[@abderahmanehm8117]

Thank you

[@philipjaykho114]

Hi Artificial Intelligence, Im so confuse because I tried to solve the similarity (those value in green-font) but it seems so hard for me. please help me. thank you so much sir.

[@aishi99]

very helpful. thanks

[@nishthavarshney9548]

hw to calcutate the cosine value?

[@AhmedRaza-kp8io]

Centered Cosine similiarity???

[@K-mk6pc]

Sir never addressed about user - user similarity metrics.

[@akshayjagtap7834]

Where to find code for the same

[@lishannavishka1783]

I have a Problem I can not say it proper way but I will try to explain it. think there are two points A and B, A = [1,0], B=[0,1] cosine similarity between A and B is Cos90 = 0 right? like wise in this example how did you find Cosine similarity between A and B sir? how can we graph A and B points

[@iiTE4L]

17:12 in green 0.41 is wrong. I get 0.55 and I double checked it. Everything else is correct.

[@megabites3649]

how to solve that 1.00, -0.18, 0.41, 0.10, -0.31 and 0.59?

[@PrateekAgrawal]

How many fans of "Acid Rock" here ;)

[@Sudharsan-dd1qi]

Sir i have one problem In Collaborative filterations can u help to solve?

[@hasanalikhan3543]

How they rated overall 5 movies its 6 not 5

[@maganaluis92]

This is horrible

[@voiceofwomen2420]

you are just reading slides, not explaning well. disappointed

[@ibrahimiltifat7045]

Thank you