There's still a bit of time — shall we try to optimally rewrite the ROC-AUC solution (using sorting and ranks in O(n log n))?
Machine Learning / AI
How will you form a dataset for training such a ranker?
Moreover, perhaps the quality of the player should be somehow evaluated online — if the user started watching, how much did they watch?
What is the difference between list and tuple?
Are we leaving boosting as a model, or how?
What do you mean by depth when scrolling a page or a block?
What is the difference between self-attention and cross-attention?
Recommendations for blocks on the main page of an online cinema Write a plan to solve a product task from reviewing required data to training, testing, and deploying a model (or, if possible, without ML - then without a model). The online cinema has a main page. It consists of different blocks. There is a catalog of blocks. Blocks can contain different amounts of content. There is a ready-made movie recommender, i.e., a catalog of movies ordered by relevance. Design a personalized block recommender and its application algorithm, which satisfies the business rule - show one content on the main page once.
How difficult will it generally be to adapt to the time zone?
What do we take as positive and negative when training a double-headed network, and what loss is used?
What is Precision?
What is the difference between self-attention in the encoder and self-attention in the decoder?
You used the term 'PropNail/TopNail Optimization' — what is it?
Tell me, what is the offline metric NDCG?
What is the QPS if N movies (posters) are shown per page?
Regarding location — do you consider remote work?
How would you solve this problem — what algorithm/approach would you use?
Is it always good when the scroll depth is large?.
We have a catalog of movies ranked by a model, and separately blocks-collections of these movies with different content quantities and types. How would you approach ranking these blocks on the main page for a user? What metrics would you suggest?
Tell me more about FAISS: how is the index structured, how is it deployed, and how does the end-to-end system work?