Title: P2: I was at Tinkoff.AI RecSys Meetup #3 [2024-03-07 Thu]
Others: https://scikit-learn.org/stable/modules/model_evaluation.html

Python example:

from math import log
from numpy import mean

def cross_entropy(p, q):
return -sum([p[i]*log(q[i]) for i in range(len(p))])

y_true = [[0, 1, 0], [0, 0, 1]]
y_pred = [[0.05, 0.95, 0.00001], [0.1, 0.8, 0.1]]

print('for each:')
[print('-', cross_entropy(x,y)) for x,y in zip(y_true, y_pred)] #3 #dailyreport #crossentropy #logloss #loss #3

Title: P1: I was at Tinkoff.AI RecSys Meetup #3 [2024-03-07 Thu]
Used to calc entropy to protect from 'skewed distribution'.

I still don't understand Why this have minimum when p is equal
to q.

cross-entropy or log_loss used because of properties:
Gradient-Friendly, best compatible with Softmax Activation,
sensitive to small differences between the predicted and true
distributions, penalizes larger deviations more severely
than smaller ones. #3 #dailyreport #crossentropy #logloss #loss #3

Title: P0: I was at Tinkoff.AI RecSys Meetup #3 [2024-03-07 Thu]
- Yandex uses Transformer architecture in their library for recsys.
over 900f code is writen in C++ (proprietary library)
- Sber has their own open source library for recsys (RePlay?)

I have been reading about cross-entropy. It is Asymetric to
(p,q). It is a averaged probabilities of p outcomes by the
weights of amount of information in q outcomes. Averaged sum #3 #dailyreport #crossentropy #logloss #loss #3