Posts by soudle_noop

Will-the-game-guy

I've found the bot wont let you do a manual report if you already reported someone in game.

Am I doing it wrong?

I double checked with some folks and apparently what you are observing is correct! I think my previous suggestion was in fact, wrong.

Sincere apologies.

Will-the-game-guy

Anyone that runs into accounts like this, dont report them in game.

Screenshot the name (if you can) and manually report them after the game via the website.

Itll get reviewed by a human and theyll get the boot.

Feel free to submit player tickets if you feel strongly about the name.

That being said, you SHOULD also report them in game if you think it's inappropriate.

I'm very curious to see the prediction accuracy of human agents vs a machine learning model.

The use of additional statistics to compliment the model is a pretty smart idea, especially when dealing with this kind of problem which has very few number of games available.

On top of my head, I think another interesting thing to try is to use a model that can propagate through time. Say this one (warning: I've never personally used it), or any other kinds of TTT (Trueskill Through Time) implementations that are out there. It might lead to more accurate estimates.

Bird-The-Word

Seraphine and Skarner...nsfw?

I assume you are referring to https://www.reddit.com/r/leagueoflegends/comments/nwb3gd/seraphine\_is\_so\_pretty\_which\_inspired\_me\_to/

canPaykek

if u want ill post this every monday night <--

Thanks for posting this :)

I'm only got to Platinum last year while working full time. Is my job holding me back from my TRUE POTENTIAL (tm)?

stopfeedingplz

Wait i thought every1 had this. Am i part of a small minority? Why me though?

Yes, only a small proportion of players from selected regions got this new store.

You got it because you were randomly chosen.

Mafros99

On another note, can you tell Blau to drop Yasuo? It'd greatly enhance his mmr

He is actually elevating Yasuo. https://twitter.com/blaustoise/status/1230926909640339456?s=20

Blood_Raptor

Maybe if you had the skin, you'd play him more often

I haven't played Amumu for a very long time, but after seeing Inferno Amumu from Your Shop I decided to pull him out... and now I look like an Amumu one trick.

JAZEYEN

Again, I really hate how they've made a habbit of releasing teasers after skins are hitting PBE.

Like these are going to be live in a really soon and we've already seen them fully on PBE. Why do they show the teasers now?

This builds no hype because we already know, it's wasted resources.

I love Riot's hype team!(If that even is a thing, I just assume a portion of their marketing goes to a secret hype team lol) They've just had some bad habbits recently.

Perhaps you are a highly engaged player who actively seeks out new information from the PBE. Maybe you have even played her using the skin on PBE. For you, this teaser probably feels redundant.

That doesn't automatically mean it's wasted resources. Marketing initiatives work, and we have evidence to back it up.

alanliang

For my (fellow) nerds, I wanted to address two main things.

1. Defining Plat+ (Classification Accuracy Shrinking Sample Size)

TL;DR 1: Outlined below, we currently use a classification system (A) which we believe prioritizes accuracy, but does lower the sample size pool (ex: Plat+) by a small amount. The old system (B), which most sites use, inflates win rates and sample sizes - plain and simple, it's not acceptable to use a system that says a class (Plat+) has a 55% winrate on average. That would incorrectly mean that most builds have a positive win rate in Plat+. That skews the data way more than sample size. I'd try to remember that size of sample is not the be-all-end-all of accuracy.

Let's use this sample game:

Team 1 (W): Plat 4, Plat 4, Gold 1, Gold 1, Gold 1 | Team 2 (L): Gold 1, Gold 1, Gold 1, Gold 1, Gold 1

As a simple example, here are 2 classification strategies (among many). We currently use Class A.

Classification A (Average):

Approach: average the MMRs and classify the results as the averaged MMR (here: Gold).

Result: This game counts as 5 Gold Wins and 5 Gold Losses.

Classification B (Individualized):

Approach: Each player's W/L counts towards their actual rank.

Result: This game counts as 2 Plat Wins and 3 Gold Wins and 5 Gold Losses.

Class A Pros:

+Will have 50% wins and 50% losses

+The numbers for Gold will add up to 100%

Class A Cons:

-Fewer champions analyzed. We lose 2 Plat games which are classified as Gold games in the above example.

-Less data for higher ranks. All Challenger games are rare, so mixes of Challenger + Masters will be classified as Masters games.

Class B Pros:

+More champions analyzed

+More data for higher ranks

Class B Cons:

-As MMR increases, win rates will be inflated as Plat players are more likely to win than Gold players, etc.

-Conclusion: Matches in a given MMR will be extremely unlikely to add up to 100% System B results in Plat+ players, as a class, having an average of 55%+ win rate. If this were the only game in our database, our site would say that Plat players would have a 100% winrate and Gold players have a negative winrate. You can understand how this is problematic and can inflate results across millions of games. A sample size needs to be significantly large, but size is not the absolute determinant in an accurate interpretation of the results.

2. Outages In Last Few Months Causing Small Data

Here are some of the issues that caused U.GG data outages in the past few months:

Patch 10.1: Season ID from Riot API is 13, which was the same season id from season 9. We made an assumption and expected this number to be 14 and this caused our systems a few months ago to crash.

Patch 10.4: Rewrote our match crawler to more efficiently support profiles and tier list. Previously we were seeding with a group of players and searching through their matches to add other players to our queue. Now we crawl through the leaderboard to grab every match. We had issues while transitioning this process.

Patch 10.8: Adjusted our code base because one of our databases, Elastic Search, changed many defaults with a recent update.

The shard default for the index went from 10 to 2. This caused I/O issues with our database.

I would like to reiterate that these issues are not an excuse. Because of our position in the League of Legends eco system, our users deserve better. We were too reckless in deploying our changes and have caused headache for a lot of you. We're working on refining our process to avoid these types of outages. We will make U.GG accuracy and reliability a top priority.

Kudos for using Class A instead of Class B. IMO it's the right decision. I've always worried about third party websites not doing their homework and use Class B instead.

Cool visualizations using R and ggplot!

One thing I'd suggest is a URL library such as this so you don't need piece things together manually.

mundus108

Out of curiosity, is R used at Riot? Or is it more of a "whichever tool prefered" culture?

We do use R sometimes. For example we want to run a statistical model but the model only exists in R.

_Trixrforkids_

Rioters are asleep! Quick play Aurelion Sol!

:eyes: