Events Add an event Speakers Talks Collections
 
Request Q&A
Request Q&A
Video
Engaging new and diverse audiences with ML
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
596
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About the talk

Recent events have drastically accelerated player growth in the gaming industry, and have brought new audiences to titles. By promoting positive experiences and understanding these new gamers and what they enjoy, studios can not only help retain these new players, but promote further diversity and inclusion in their games. Learn how Google used ML to help studios understand nuanced and contextualized interactions to detect and remediate toxicity and disruptive behavior, and how to provide personalized experiences for every type of player.

About speakers

Patrick Smith
Machine Learning Specialist at Google Cloud for Games
Vishu Cheruku
Customer Engineer at Google Cloud for Games

Patrick Smith is a Machine Learning Specialist and the global ML/AI Solutions Lead for Google Cloud for Games. With a background in data science and data engineering, Patrick has experience in leading consulting and technical customer engagements for Fortune 500 companies and public-sector agencies. He is the author of the book "Hands-On Artificial Intelligence."

View the profile

Skilled in managing diverse teams and resources to scale impact across enterprise organizations. Expertise in streamlining operations, and driving deadlines to ensure maximum customer satisfaction and business revenue. Outstanding critical thinking and communication skills. Able to successfully build strong working relationships with coworkers and clientele, while ensuring cooperation among company departments.

View the profile
Share

HarborOne. My name is bisha trucu, and I'm a customer engineer was in Google Cloud for games today. What we talkin about engaging, you and diverse audiences with machine, learning specialist with Google cardboard games for getting into how you can use machine learning to engage audiences. Let's first, talk about the current gaming landscape. The covid-19 pandemic has had a significant impact on gaming 38% of video. Game players reported increase in their play time for socializing at over 65% of their play sessions

are longer than ever. The overall landscape has really started to shift and the need to account for diversity. And the types of players has become a parent. As the number of players that are in gaming increases. Trust is at the core for a positive social experience. While psychological safety social cohesion, and camaraderie. Also play a critical part in the shirt interactive experience of what kind of social experiences eclair needs and enabling them to make the most of their time during gameplay, is critical to your satisfaction and retention. Let's take a look at some

examples of the types of players were saying today. Where is a female player in her? Forties, she spends a significant amount of time. Playing multiplayer games and works hard to nourish. The relationships with the other players. As soon as the new game releases clairee, spent the entirety of our weekend playing the game on her console in the weeks following. She figured out of War controller pro server. Gameplay, navigate areas and having to make decisions on how to clear. A, so, extremely accomplished gaming helps. Her feel more personally, satisfied and increases, her feeling a

competent. Where is a male player in his mid-twenties with strangers and does not enjoy multiplayer games for this reason? He prefers playing mobile games in single player. Console games in terms of time cuz it might spare be used as his commute time, in between appointments and classes and his free time in the evenings for game play. Enjoy being able to expand his thinking through these games by getting to explore and being able to build things. He loves the autonomy of the gameplay. On the other hand, where see is a female player in her thirties,

to primarily uses gaming, as a means for social interaction. Play receipt has both single and multiplayer games on her mobile device and her console. She plays throughout the day whenever she has some down time, including during breaks in between her meetings, loves to play games to connect with her friends. Although she has moved away from them. Now. Jimmy play allows players need to connect and interact with her friends on a regular basis. What are the feeling of accomplishment in player? A space to explore and be adventurers League player be or using gameplay. As a means to take connected

like players to eat Compton economy, and social connection, help improve the well-being of players by helping various players achieve, their distinct goals has added to the overall solution accounting for the diversity and players helps create a more positive and trusted experienced players to come with social interaction and particularly what that means for new players and play retention. Thanks, miss you throughout the whole covid-19 pandemic in continuing to this day.

New players are coming into video games that Studios and Publishers have not seen before. It's really a nude first audience of players from across the world, across all spectrums and backgrounds with that though, come Z. The increased probability friction. Between those players comes toxicity, either from dealing with it on the developer side or experiencing it ourselves. But the problem is what is toxicity MPG with charm is colloquial. It's loaded. We don't necessarily know what

defines it. We just know that it's bad. That makes it hard to solve, but we don't know exactly how to solve things, or what the salt or it's hard to build Wheel Solutions, and that's why we call it disruptive behavior. That you may have seen that the Fairplay Lions, who many of the studios and Google cloud in Google belong, to has really pushed the term disruptive behavior, and disruptive behavior shows. And says what it is that's happening here. Is that players games are being disrupted their ability to enjoy their game is being disrupted. Their individual play snails are being

disrupted. If it's really just, that players cannot enjoy the game in the way that they want to takes a lot of different faces. It could be lately bad like racism or criminal Behavior, but it could also be inappropriate sharing or it could be anti-social behaviour unattended. Disruption. Maybe you're playing music loud, music in the background while you're playing a game, and that's preventing somebody else from being able to concentrate on the game. This is all disruptive behavior, and we want to solve for that. And also be sensitive to a different players

play Styles at the same time. There, many different causes for disruptive, behavior. They're all overlapping to an extent. What I'd really like to emphasize here is it's usually not bad players. It's bad. Moments players are generally not out to ruin another player's game. But if they do have to do something at another player finds disruptive, usually letting them know what it is. Explicitly telling them what they did, a result of Behavior. Now, it's only about 1 to 2% of players that are really those true trolls, really the true problem players and we do want to find

those players and we do want to take action before everybody else. We want to just make the game has less friction and make it more enjoyable overall. Don't, let's talk about the taxes. If you stirred 5 main types of abuse that we see in games and they all have different presents levels offense of speech, which is what we see most often, which is fleurs hate speech things that have directed that we know are bad, that's in about 25%. Load out personal attacks. So these could start as a joke and get very personal, but it's an attack on an individual discrimination threats and sexual

harassment come below that at very similar percentages. In all these together and create what we all know, as a toxic video game department have been reported to avoid games because of the sort of abusive and destructive Behavior. So, where is this abuse happening? Most often split on whether the sort of abusive communication happens in voice chat or text chat, but they do feel worse when it happens in voice chat. It's important to understand the toxicity and destructive behavior is caused by bad moments. Not that players. The solution to this problem is

to create a healthy ecosystem that has built-in proactive and reactive protections that create a more positive experience for players. No, tackling disruptive behavior is so challenging just because it's so nuanced and there's so many layers. That when we talked about destructive Behavior, right? We have toxic speech, talk chat griefing cheating in the other disruptive actions, that could be in your game. These could all happen at the same time or they could be isolated in detecting the relative levels of them and when they are happening and why they're having makes this an

extremely hard problem solve, one that we can solve with machine learning and it's the humans in an ultimately adjudicating on these issues. But it is a tough one. Effectively monitoring removing censoring, providing Downstream actions against players because of disruptive behavior. All of those can be tricky as well. Again, we don't want to prematurely remove a player from a game or or mute a player. If they honestly, we're just having friendly banter with their friends. And that goes into balance and free speech and enforcement. We want to make sure that we're encouraging bantering, these

environments encouraging players to talk in the way that the game environments have really to have the community support themselves and understand what is wrong and what is not right and balance. Those two can be tricky to the city with matchmaking. We definitely want to make sure that we're keeping players evenly matched. We find that touches me happens. When there are moments of friction, which can be players that are in this match, my skill or other types of mismatches, and we want, it kind of is effective to bring those two together over there because he

is poised to tackle the whole problem together. To take a look at what I'll call traditional rule-based methods, where we have a bad word that you might want to fill it around or something else that we want to catch that great in. That works for some circumstances, but not all circumstances. Now to be trying to detect toxicity in usernames. So that's Beach and usernames. That might work. But if you're trying to text disruptive behavior overall, in interactions between two players, it's going to be more new. It's going to depend if they're in a certain group. Their backgrounds.

Are it is vishu alluded to, in the beginning of this. That's what was making this even more difficult. And even more pressing issue. Now, with the increased player bases that were seen it. So it's important to take that context in which leads to natural language processing methods. And what I mean by that are machine learning and artificial intelligence driven methods to understand natural language and what the intent is behind, those are great in and of themselves, but they also take good data to make sure that they're affected and they need to be monitored so that they don't

necessarily skew and end up providing more harm than help. Anomaly detection, methods are also useful for detecting Bangla cheating, you briefing. But they don't handle banal, peace sign. And when it comes to manual review you only have so many people. And so many hours. We find that a very little amount of tickets actually get results. If you only have a manual review process. And so all of these things need to be combined in one way or another to bring defected disruptive behavior in education system. Don't really want to emphasize here, that it is important to get your machine learning

algorithms, right? Having a bad machine learning algorithm, a bad machine Learning System can do as much harm as not having any at all. If not worse. You can eat unfairly Target players. If the data is bad, if the model starts drifting or even worse at the model, was prone to abuse. We all know that data is the food for algorithms. And if that food is not good. You're just going to have that quintessential garbage in garbage out. We need to make sure that our data is treated for appropriate categories. As fairness measures in, it is representative of the diverse group of

players or no singing gaming. So that any model train is going to be representative of the overall population. We especially want to be sensitive to a particular language that might be harmful of one circumstance but not in another. And that's why we need to ensure these algorithms are fair and unbiased. That users can trust them that data is kept private and that they're reliable and robust Beach emphasis on a IX in ensuring that fairness is built into our what is to land some of the bills, intense full

capabilities for model fairness and model monitoring that you can use in our AI platform, but I do want to emphasize that this is a critical thing that you should all be thinking about. The successful strategies include both proactive and reactive measures. No proactive is encouraging. Those matches, encouraging positive experiences, reducing friction, but if it does lead to disruptive behavior to be Dulce. Disruptive, behavior and games we want to respond, effectively. We want to respond in a way to contextualize that can either automatically detect what's happening or sister player

support groups in detective. What's happen? Sir, Patrick what you're saying is that it's just as important to create a positive environment for players, but also thinking about how to mitigate destructive Behavior. Once it does happen. Can you help me understand how a particular as helping solve these issues and I really saw this one strong and when we work with individual Studios and individual providers on this problem, we are bringing or MLP methods,

training them up on individual game data, whether that be Bar-B-Que an hour out of a male models, when we're detecting grieving, and cheating, or when we're trying to text, talk to Steven speech in Chad, or natural language, processing methods, which come through our products or or bring them out altogether, wrapping that in our Google infrastructure and bring that to the customer, so that I can run at Planet scale for millions of concurrent users. Also, bringing the best models that we have to offer these LP models that I just mentioned, are the models that underlie a lot of are Google

products. And also, as I mentioned a little bit earlier, bringing the nation's, it's an integrated, a lot of our capabilities book from our internal capabilities in for pets are flow of model fairness and AI ethics into Rai platforms. That are that comes in the form of the water to where you can monitor model, skew, you can monitor how an individual change in model input, might affect output of classification, things of that nature, which as I am just emphasizes is pretty important. In an overall. We're trying to be sustainable here and so altogether or bring the Forefront of AI technology

to the space. So let's talk about some of the proactive measures that we can use in-game specifically fixing flat experiences in for matchmaking, reduce friction between players and better overall experience in. These are two methods in two ways that we can help us improve for the player and reduced toxicity and games, based matchmaking. What say you have a nice towards title or first person shooter and it's just watched chances are your players come to that title at play. Other titles in the same category. They're going to have skills Weatherby. They're really good

sniper in the first-person shooter. Are there really good shot, and he's smart style. Those skills are going to determine their natural play level or their natural play capabilities. Do you need to make sure that those players are ranked appropriately and in their place style, is being taken into account. When we drank the moon when we match them, their traditional methods might walk through players will have to grind up individual levels and individual rights, but if we focus more on their innate played skill and their innate play style. We can better understand how good they are where

they might improve and then have to match them with players that were they'll get a good experience. You could do something like this, using a simple clustering algorithm was you can get in rbq ml private. So, let's say, you have individual players, we can cluster them into playstyles, understand those clusters, and then predict based off of those clusters or you could do something a little more advanced, like a problem, same idea, but it's understanding those individual play style so that we can better match those players. And now offering Dynamic experiences players get bored when they

have to grind through a game. It's just a fact and 40% of players are likely to leave the game due to that frustration and boredom. So why change the game adaptively to the individual players preference? So that they don't leave the game. They don't turn it reduces friction again it. So if you think about getting rid of something like reinforcement learning methods, or even just a classification, or regression methods to proactively suggest things in game that a player could do what you just really help them along or dab, the game varmint to their skill level. That

helps make the game more enjoyable, more enjoyment creates less frustration over all. It's good for the player base. At what you have those models is very easy to deploy them in. Came with our new Vortex IAAI platform, you have access to Auto amount. You can you use dtml on top of a query or you can use the individual. Always have a one-click to Florida to a robust scalable and point that you can use for your game. And we seen you support millions of concurrent players to make it a lot easier, where your data scientist in. Ml Engineers to focus more on making those quality models, unless around

the OPP's also focusing on, in on the OPP's. You can use capabilities like our pipelines offering, which utilizes tf-x underneath to really make robust AI ecosystems. Village books in on some of those reactive measures specifically on detecting toxicity in speech and chat. Now, many of you are probably familiar with the concept of word embeddings where we can take a sentence or a piece of natural language and turn it into two. The problem is not all word beddings are necessarily

text Elias. And as I mentioned earlier, contextualisation in this space is key. We want to understand what a player is saying, within the context of their friends, their group, their play style, everything, so that we're understand was actually toxic in. What's not important. Then to use contextual representation me contextual representations are bidirectional and take into account the entire sentence. The entire paragraph in this is where you want to use something like a bird model,

that has more contextual representations. Now, focusing in on those unidirectional versus bi-directional models. Unidirectional models, build representations, incrementally in bi-directional models have context again, focusing specifically on it's a great method for fine-tuning to your games, dataset in a way that it understands the contextualisation in your game and the new ones looking more geriatric protections. Let's look at analysis to cheat and briefing detection. Really what we want to do here is detective normal

patterns. So whether that be in a game stream, or whether that the in other game data, we want to detect when something is a little anomalous in this is pretty similar to traditional, anomaly detection. If you think about griefing, it's somewhat of a normal action if we Define what a normal gameplay tickets, if we think about cheating, having too many bullets are making a shot that you never would have been able to make it first person. Shooter is anomalous and so we want to reduce that risk by Is detecting, what is within the bounds of normal or decent possible behavior in your games, physics

environment? And then flagging. When it subsides that and for this, you can use ottoman covers to use like I said, with a traditional time series where they can use other forms of data, but was really great. Here is detecting an anomalous Behavior. Now what an auto motor does is it takes an individual input and reconstructs the output, which might seem simple, but then it gets really, really good. At understanding what that output and what the date is supposed to look like anything is flag anything. That isn't almost doesn't look like an output that is expected. Do I what colors are

available right out of the box? Envy? Qml, both for time-series data and other data. And so if you have your data easily in in big Prairie, you can deploy one of these not trained. One of these models into play, one of these models of the few lines of sequel. You can start detecting cheating in in greeting Behavior very easily, lay off the bat together. So we have an individual for a model fine-tuned set around toxicity and we have an anomaly detection model inaudible. Kotor trains around detecting breathing in cheating. How do you put this together? What's

soon for the sake of this example, that we have individual scores? It's a float between dr21. Those scores are going to come out of the individual models, and we need to aggravate them in some sense to Monitor and keep track of the overall level of disruptive behavior that's happening within a squad with in a match, with a particular person individual scores with Google Cloud Pub sub. Is a message cute in that can individually receive the, the message with a pub sub topic, which would be the toxicity score that's coming out of the

model. We can then send that over the data flow which handles data streaming in all of your individual score aggregation live in stream can be done, via data flow out of that. We have an individual score. That's horrible. Measure the overall level of disruptive behavior on. How to wrap this. All up with the Google Cloud platform. You can manage this entire ecosystem of disruptive behavior. My talking scale at the planets Gilliland support, millions of concurrent players focusing in on our platform. You can train your models and deploy them with ease and not have to worry about the office

in the auto scaling around that, we handle the oddest feeling for you. Depending on when your muscles are needed, or how many players are playing in La, I pipelines ecosystems through our pipelines offering and we bring in the best of tensorflow around tfx around model fairness and wrap it up into the hole. System to really bring a holistic view of how to create and manage and deploy your models. Some of the discussion from today, in order to engage New and diverse audiences. Using machine learning. It's important to

understand the types of players that are now gay men. Understanding the diversity of players and helping them achieve their distinct goals during gameplay. It's critical to player attention, especially given the increase in the number of new players. In the pen, demick toxicity and destructive Behavior has a major risk that sex player attention, especially for these newer players. Finally, in order to tackle disruptive behavior. Google, Cloud offers various machine, learning solution to Taco Bus proactive measures that are focused on creating positive Community,

address problematic Behavior. Thank you so much.

Cackle comments for the website

Buy this talk

Access to the talk “Engaging new and diverse audiences with ML”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “Google for Games Developer Summit 2021”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “IT & Technology”?

You might be interested in videos from this event

November 9 - 17, 2020
Online
50
95
future of ux, behavioral science, design engineering, design systems, design thinking process, new product, partnership, product design, the global experience summit 2020, ux research

Similar talks

Mike Nadal
Customer Engineer at Google Cloud for Games
+ 1 speaker
Emma Haruka Iwao
Developer Advocate at Google Cloud for Games
+ 1 speaker
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Adrian Graham
Senior Cloud Solutions Architect at Google Cloud for Games
+ 1 speaker
Travis Williams
Enterprise Cloud Architect at Google Cloud for Games
+ 1 speaker
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Sebastian Weigand
Systems Architect at Google Cloud for Games
+ 1 speaker
Susan Romera
Customer Engineer at Google Cloud for Games
+ 1 speaker
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video
Access to the talk “Engaging new and diverse audiences with ML”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
949 conferences
37757 speakers
14408 hours of content
Patrick Smith
Vishu Cheruku