游戏邦在:
杂志专栏:
gamerboom.com订阅到鲜果订阅到抓虾google reader订阅到有道订阅到QQ邮箱订阅到帮看

Rovio:使用机器学习和人工智能设计奖励广告

发布时间:2020-08-17 09:09:14 Tags:,

Rovio:使用机器学习和人工智能设计奖励广告

原作者:Sarah Impey 译者:Willow Wu

几个月前,我们有幸参加了谷歌的GameCamp活动,见到了许许多多手游领域中顶级内容提供者。Rovio的商业智能总监Elif Büyükcan跟我们分享了他们是如何在游戏中使用奖励广告的。我们会先叙述他们的做法,然后总结出大家可以借鉴学习的知识。

Elif想表达的主要想法是:

“设计合适的盈利方案是重中之重,这需要花费很长时间,你还得不断思考、测试。持续评估、学习和改进。所以,机器学习(ML)和人工智能(AI)能成为你最好的帮手。”

今天,我们会把他们演讲中的关键点过一遍(尤其关于是机器学习),其中包括:

-Rovio是如何改进他们的盈利模式
-《愤怒的小鸟:梦幻爆破》案例研究,奖励广告的分析
-对有志向成为游戏开发者的人提出的建议

好了,我们立即开始!

一、为什么Rovio想研究奖励广告

现如今,行业中有一些经过市场检验的盈利模式可供开发者探索。但面对这么多的选项,你可能很难确定要聚焦于哪一种,要测试什么。而Rovio选择了奖励广告。

angry birds go biz(from pocketgamer.biz)

angry birds go biz(from pocketgamer.biz)

为什么?

原因一:IAP的王冠正在滑落

IAP一直是F2P游戏盈利的王牌手段,但市场上那些热门游戏的IAP转化率却在下降。

依赖IAP,实质上就是将希望寄托于那一小群愿意掏钱买东西的玩家。现在,开发者想从绝大多数不接触IAP的玩家身上争取收益,这并不令人意外。

但与此同时,开发者们也很谨慎,不想因为添加什么东西导致这些珍贵的玩家离开你的游戏。这是一个不小的挑战。

原因二:游戏内广告正变得越来越流行

最近的研究表明,如果广告融合得好且与玩家有关联性,他们确实会对广告很有兴趣。而且随着机器学习和人工智能的进化,开发者现在对玩家有了更好的认识,能够为他们量身打造广告。

广告会破坏游戏体验的这种说法,如今自然也就不成立了。

原因三:找到适合的平衡点会给你带来收入

IAP转化率并不是一成不变的,游戏的生命周期总价值也是如此。它们俩都会时间的推进而降低,但原因不同。

简单来说就是优秀的盈利模型不会完全依赖于IAP或广告。关键是要在两者间找到平衡。解决这个问题,你的收入就有可能节节攀升了。

二、案例研究:《愤怒的小鸟:梦幻爆破》

在演讲的过程中,Elif详细讲解了他们的新游戏《愤怒的小鸟:梦幻爆破》。游戏是一年多年以前发行的,现在已经是他们最成功的产品之一了。《愤怒的小鸟:梦幻爆破》进入了iOS平台畅销榜的前100,并且一直没有掉出来。

以防你还没玩过这个游戏,我先跟你讲讲:游戏使用了基于物理学的“点击&清除”机制——通过点击五颜六色的泡泡,让一定范围内的同色泡泡消失。清除大量泡泡会给你小鸟,你可以组合它们来清除更大范围的泡泡。

那么,他们是如何利用奖励广告的呢?这些广告的奖励方式很简单:

每天,玩家可以通过看广告获得每日礼物,或者更多生命/更长的游戏时间。

广告通常都是Rovio的交叉推广,这是符合逻辑的——如果玩家喜欢《愤怒的小鸟:梦幻爆破》这个游戏,那么他们很有可能也会喜欢其它的《愤怒的小鸟》游戏。

Rovio旗下还有一些游戏,但这并不意味着每一款游戏的广告效果都是一样的。奖励广告有好几种运用方式:在开局前获得免费的增强道具、重试失败关卡的机会或者是回收失去的资源。

Elif说这些方法都有不错的效果,但它们基本上都算是“通用”模式。他们希望在不影响IAP转化率的情况下,探索个性化奖励广告的投放方式,由此来提高广告收入。

于是,《愤怒的小鸟:梦幻爆破》就成为了他们的实验目标。

他们分为了三个阶段:

第一阶段:测试奖励广告

关于奖励广告模式,Rovio团队想出了一个好点子,在测试发行期间展开了实验。根据他们在其它游戏中看到的情况,游戏中的设定都相当慷慨。

但不幸的是,结果并不是他们所期待的那样。

他们测试了什么:

如果玩家在通关失败后看了一个广告(每天至少可以看一次),就可以获得额外三个步数来继续尝试闯关。

结果:

增加这种特色对游戏的盈利产生了负面影响——IAP收入和转化率都下降了。

经验教训:

1.“通用”模式并没有效果。他们从非付费玩家身上获得的收入无法弥补付费玩家那边流失的利益。

2.在测试发行期间进行实验不是个好办法。因为样本量很小,数据的不确定性很大,小小的调整都可能造成很大的影响。

计划的下一步:

获得令人失望的结果是否意味着他们永远不应该使用奖励广告?“不,当然不是,”Elif说。“我们需要研究结果、改进优化然后重新测试。”

他们知道这些广告对付费玩家产生了消极影响——这是他们最不想看到的结果,所以他们想:“要是我们把这些奖励广告只展示给非付费玩家呢?”

但是你要怎么快速试别哪些会成为付费玩家,哪些不会呢?

阶段二:几个团队聚集在一起,提出了一个假说和检验的方法

他们需要找到一种方法来预测谁是付费玩家谁是非付费玩家,然后进行A/B测试,看看他们的假说是否成立。

要完成这个任务,他们需要广告、商业智能、技术和游戏团队保持持续合作。关键人物有:

·游戏的产品经理——负责安排测试和游戏新特色。

·数据分析员——负责运行A/B测试,提供与游戏表现相关数据。

·广告营销经理——负责提出广告投放的思路,以及如何巧妙地融合。

经过长时间的头脑风暴,他们敲定了一个理论:

“如果我们只向非付费玩家提供+3步数的奖励广告,我们将看到IAP、广告收入都会增加。”

设计一个机器学习模型

为了检验他们的假设,Rovio需要一个可分配值的系统,这个值可以预测用户成为付费玩家的可能性有多大,以及大概会消费多少钱。这样他们就能够瞄准非付费玩家,展示奖励广告了。

这是个数据科学方面的挑战。但是凭借Rovio强大的机器学习和人工智能,他们早已准备好了。事实上,他们之前就在制作一个针对玩家的LTV预测模型,可以通过调整来服务于新的测试任务。

随着新模型的建立和运行,验证马上就可以展开了。

阶段三:测试新模型

他们测试了什么:

只对模型预测为非付费玩家的用户提供+3步数奖励广告。

他们做了几个星期的测试,并在此后的几个星期里继续观察长期影响。这次他们的测试环境比较好,因为游戏在全球范围内上架了,所以样本量不是问题。

结果:

仍没有看到任何积极影响,但是消极影响减少了。他们的总收入减少了一点——额外的广告收入依然没能完全弥补IAP的微小损失。

但是要记住,这对比上次测试来说效果已经有所提升了。游戏的转化率维持不变,IAP收入只稍稍减少。

经验教训:

1.广告收入很难弥补从付费玩家那里流失的收益

2.预测模型大部分是准确的,但少数几个不准确的地方会让大家付出不小的代价

3.当+3步数奖励影响到其它奖励广告的消费时,可以将每个用户的平均广告展示量降低。

4.尽管结果仍然是不理想的,但它们正朝着正确的方向发展。

计划的下一步:

Elif说他们有三个选择:

1.做一个新的A/B测试,把付费用户的概率评分调低。这是一个低风险的选项。

2.做一个新的A/B测试,调高广告的展示频率。这有可能会提高广告收入,但是也有可能会对IAP收入造成更进一步的负面影响。

3.让测试模型变得更加精准,或者换一个不一样的模型。这是技术要求最高的选择,但可以带来一些快速的效果。

那有没有考虑过放弃奖励广告这个模式?Elif说“不,没有。在强调一下,这些实验是需要时间和耐心的。我们在不断调整、测试、学习。”

他们选择了第三个方案,重点是改进或更换预测模型,这样他们就可以更有效地针对非付费玩家,降低IAP收入受影响的风险。

三、最后来讲讲我们从Elif的演讲中学到了什么

Elif谈到了很多技巧、建议和教训,以下是我们在这次演讲中学到的重要知识:

1.合适的广告投放能够提升产品的生命周期总价值

游戏内广告,关键就在于你怎么使用。你只需要把对的广告在对的时间展现给对的玩家。

2.每一次测试都是新的教训

即使你的测试结果并不乐观,这些付出也并没有白费。至少你知道了什么是行不通的,而且你一般都能从测试中看出为什么行不通。每一次测试其实都是循环中的又一个学习&提升的阶段。

3.个性化,机器学习和人工智能是关键

玩家是一个相当多元化的大群体——经过细分的目标市场也是如此。“通用”盈利策略的效果肯定比不上为玩家量身定制的个性化策略。机器学习和人工智能对定制化能起到非常关键的作用,如果你没有使用,你的收入永远都不可能持续往高处走。

4.游戏开发需要数据驱动

在游戏这个行业,我们可以快速实验、测试、吸取教训。总是有很多的新测试、新技术可以帮助我们提升优化。为了获得效果最佳的盈利模式,我们必须要加以利用。

本文由游戏邦编译,转载请注明来源,或咨询微信zhengjintiao

Acouple of months ago, we had the pleasure of attending Google’s GameCamp, which was filled to the brim with gamedev content delivered from the best and brightest in mobile gaming. While there, Elif Büyükcan, the Business Intelligence Director at Rovio Entertainment, shared with us how they’ve been using rewarded ads in their games. We’re here to report on their story, and what you can learn from them.

Elif’s main message was this:

“Getting your monetization model right is incredibly important. It takes a lot of time, thought and testing. You need to be constantly measuring, learning and adapting. This makes machine learning (ML) and artificial intelligence (AI) your best friends.”

And today, we’ll go through everything we learned from their talk (specifically about machine learning), including:

·The process Rovio use to hone their monetization models,
·a case study of Angry Birds Dream Blast and rewarded ads,
·and what their advice is for aspiring game developers.

Without further ado, let’s dig in.

Why Rovio wanted to work on rewarded ads

There are a few, solid monetization routes developers can explore these days. And with so many balls to juggle, it’s tricky to know what to focus on and test. For Rovio, they chose rewarded ads. And these were the reasons why:

Reason one: The crown of in-app purchases (IAP) is slipping

IAP has always been king of F2P monetization. But IAP conversion rates are declining for the top-performing games in the market.

When you rely on IAP, you’re only making money from the small minority of players who are willing to make those purchases. So it makes sense that developers are looking for ways to bring in revenue from the vast majority of players who don’t use IAP.

But at the same time, it’s only natural to want to avoid adding anything that might discourage these precious payers from playing your game. This is a big challenge.

Reason two: In-game ads are becoming more popular

Recent studies have shown that players do value ads, if they’re integrated well and relevant to them. And as ML and AI evolve, developers are getting better at understanding their players and personalizing ads.

So the myth that ads hurt games is quickly dissolving.

Reason three: Striking the right balance brings rewards
IAP conversion rates aren’t infinite, but neither are lifetime values (LTV) from ads. They both tend to decline over time, but for different reasons.

In short: the perfect monetization model wouldn’t rely entirely on IAP or ads. The challenge is to find the right balance between the two. Crack that problem, and you potentially stand to give your revenue a bump.

Case study: Angry Birds Dream Blast

During her talk, Elif went into detail about one of their latest games, Angry Birds Dream Blast. Having launched a year ago, it’s already one of their most successful games. It made it into the 100 top-grossing iOS games and has stayed there.

And in case you haven’t played it, the game uses a physics-based ‘tap to clear’ mechanism – you pop colorful bubbles by tapping on them. Clearing big amounts of bubbles gives you birds, and you can combine birds to clear wider areas.

So, how do they use rewarded ads in one of their performing games? Here’s what they shared.

How Rovio were using ads in Dream Blast

They already had rewarded ads, but they were only doing one thing: Each day, the player can watch an ad in return for a daily reward or a longer playtime that day.

These ads often cross-promote Rovio’s other games, when it makes sense – as players who enjoy Dream Blast are likely to enjoy the other Angry Birds games too.

How Rovio use rewarded ads in other games

Rovio has a few games under their sleeve, but that doesn’t mean ads work the same in each one of them. A few different ways they do use rewarded ads is: getting a free booster at the beginning of a level, retrying a failed level, or even reclaiming lost rewards.

Elif says these models were all working well, but these were mostly ‘one size fits all’ models. They wanted to find ways of personalizing their rewarded ads to up their ad revenue without hurting their IAP conversion rate.

So they set about finding the best way to use rewarded ads in Dream Blast. They did this in three stages.

Stage one: They started testing rewarded ads

The team at Rovio had come up with a good idea for a rewarded ad model that they expected would increase their total revenue. So they tested it during soft launch. They used fairly generous configurations, based on what they’d seen in their other games.

Unfortunately, the results weren’t what they had hoped for. Here’s what she shared.

What they tested:

If a player watches an ad (watchable at least once a day) after failing a level, they get three extra moves to complete the level.

The results:

Adding this feature had a negative impact on the game’s monetization. IAP revenue and conversion rate both dropped.

What they learned:

1.Using a ‘one size fits all’ method doesn’t work. The revenue they lost from paying players outweighed the revenue they gained from non-paying players.
2.They also realised that testing during a soft launch is tricky. Because the sample size is so small, the figures can be inconclusive and minor changes can have a big effect.

What they planned to do next:

Did these negative results mean they should never show rewarded ads ever? “No, of course not,” says Elif. “We needed to study the results, adapt, optimize and re-test.”

They knew these ads were discouraging paying players – and that’s the last thing they wanted to do. So they thought: ‘what if we could show the rewarded ads only to non-paying players?’

But how do you quickly identify what players will become payers and which won’t?

Stage two: Several teams came together to create a hypothesis and a way to test it

They needed to find a way of predicting who the payers and non-payers would be, and then run an A/B test to see if their hypothesis worked.

To do this, they needed their advertising, business intelligence, technology and game teams to be constantly collaborating. The key players were:

·The game’s product manager – responsible for planning tests and new features.
·Data analysts – responsible for running the A/B tests and providing the data on the game’s performance.
·Ad monetization manager – responsible for proposing ideas about ad placements and how to integrate them smartly.

After lengthy brainstorming, they settled on a hypothesis:

‘If we offer the +3 moves rewarded ad only to predicted non-spenders, we’ll see an increase in our total revenue from both IAP and ads.’

Creating a machine learning model

To test their hypothesis, Rovio needed a system that could assign a value to each user – a value that would predict how likely they were to become a spender and how much they were likely to spend. This would allow them to target only the non-spenders with rewarded ads.

This is a data science challenge. So with Rovio’s ML and AI capabilities, they were well prepared for it. In fact, they already had a player-level LTV prediction model in production that they could tweak to serve their new purpose.

With their new model up and running, they were ready to test their hypothesis.

Stage three: Testing a new model

What they tested:

They offered the +3 moves rewarded ads only to the players the model predicted would be non-spenders.

They ran the test for a few weeks, and kept an eye on the long-term impact for more weeks after that. Their setup was better this time as they’d launched the game globally – so small sample sizes weren’t an issue.

The results:

Rovio saw no positive impact, but less negative impact. Their total revenue was slightly lower – the extra ad revenue didn’t recover the minor loss in IAP revenue.

Keep in mind, this was still an improvement on the last test though. The game’s conversion rate stayed stable and IAP revenue was almost stable.

What they learned:

1.The cost of losing a spender is simply too high for ad revenue to compensate
2.The prediction model was mostly accurate, but the few inaccuracies were costly
3.When the +3 moves reward affects consumption of other video rewards, ad impressions per user can be lower.
4.Although the results were still negative, they were moving in the right direction

What they planned to do next:

Elif said they had three options:

1.Run a new A/B test with a lower spender probability score. This was the low-risk option.
2.Run a new A/B test with ads appearing more often. This could increase ad revenue, but would risk lowering IAP revenue again.
3.Try to make the prediction model more accurate, or try a different model. This was the most technical option, but could bring some fast results.

Did they consider giving up on making rewarded ads work? Elif said, “No, we didn’t. Again – these things take time and patience. We keep adapting, testing and learning.”

They decided to go with the third option. They focused on improving or replacing the prediction model, so they could target non-paying players more effectively, with less risk to IAP revenue.

Finally, what we learned from Elif’s talk

That was all that Elif had time to share, but boy did they cover a lot. There were a lot of tips, advice, and lessons in there, but here are our biggest takeaways from this talk:

1. Well-placed ads can boost total LTV

When it comes to in-game ads, it’s all about how you use them. You just need to show the right ads to the right players in the right ways.

2. Every test is a lesson

Even if your test has negative results, you haven’t wasted your time. Negative tests tell you what doesn’t work and they often tell you why. Every test is another stage in a constant cycle of learning and improving.

3. Personalization, MI and AI are key

Players are a hugely diverse group – even among a carefully targeted audience. A ‘one size fits all’ approach to monetization will never be more effective than tailoring your game to individual players. ML and AI are the ideal tools for personalization, so if you’re not using them, your revenue is never going to be as high as it could be.

4. Game development needs to be data-driven

In our industry, we can experiment, test and learn very quickly. There are always new tests and techniques we can use to improve. And to get monetization models right, we have to take advantage of that.

(source: gameanalytics )


下一篇: