very interesting. relevant to my work on creating uber models of advertising effectiveness where the signals come from different walled gardens and where the A/B test results can vary for the same publisher (e.g. Facebook) for the same brand across camigpans. True finding or statistical variation that can be 'shrunk' by prior distributions on the parameters in the uber model?
Hi Great Video 😊 I work on my thesis in Electro Engineering and your Movie helped me a lot. Can you please provide me with your sources so I can use them too? Thank you very much!
It means you you „learn“ the distribution of beta from the data. You can then sample that distribution(e.g. with MCMC). If the variance of the samples is high this means the model is uncertain about its prediction. That’s my understanding at least.
@@hristovassilev7812 I thought that after maximum likelihood estimation, you get a single set of weights and bias (B0 and B1) that maximize the probability of observing your data. So how does that turn into a "distribution"; when did you generate that distribution where you can sample a "new beta vector" from? Thanks!
@@Darkev77 What you describe is correct for maximum likelihood estimation. But since the video refers to a Bayesian method of estimating the parameters the idea is a bit different: you treat B0 and B1 themselves as a random variable. You generate the distribution of B0 and B1 using Bayes theorem: p(B | data) = p(data | B) * p(B) / p(data)
Would the diversity not only matter when the likelihood function has multiple local minima? Otherwise, why "hedge our bets" if we are likelihood-maximising values of beta are all very similar?
Ritvikmath: Building on your nice ideas.... Q1 - Is there a jupyter notebook of this Bayes sampling model yet? It's nice concept. Q2 - It was found that recency of purchases (which you explicitly incorporated) and frequency of purchases too are both predictive of churn, in Univ of Virgina's Marketing Analytics course. Let's make a ML model using Bayes sampling, but now incorporating both these factors. Yet another AI research direction stares at us in the face....