White paper March 2019
Landing Valuable App Users How Deep Learning Enhances Acquisition Campaigns
EXECUITVE SUMMARY App installation metrics are no longer the numbers that matter the most to marketers. To counter app abandonment and fraudulent users, marketers need to maximize the lifetime value (LTV) of a customer, which requires a comprehensive understanding of user engagement and behavior. Using evolving deep learning – an advanced artificial intelligence (AI) technique, marketers can make better sense of data to target the highest performing users in their user acquisition campaigns while avoiding app install fraud. This requires a more complex understanding of data that leads to a lower cost per acquisition and captures more valuable customers – those who are more likely to be engaged and less likely to churn – eventually enhances the LTV of the user base. In this manner, deep learning makes the customer journey more predictable and marketing to them more versatile and holistic. Marketers can use deep learning to build campaigns that are time-efficient and cost-effective to boost revenue and profitability.
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
1
INTRODUCTION: SEEKING QUALITY APP USERS App downloads are on the rise, but app abandonment is still sitting at a significant 21%š. This tells us that it’s no longer enough to invest in acquiring users who install apps onto their mobile devices if one in five of them never open it again after one use.
App abandonment
Instead, the new measure in the app economy is in-app spending, and the best metric for that is customer engagement and loyalty. Artificial intelligence can identify the traits and behavior of high-value app users and help marketers acquire them, as well as support their retention and reengagement strategies. AI can also counter fraudulent app installations and fake in-app actions, a growing challenge in which bots steal from app marketers via methods such as false referral attribution, pushing up the cost per acquisition and wasting marketers’ budgets on fake users.
1
http://info.localytics.com/blog/21-percent-of-users-abandon-apps-after-one-use
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
2
OPPORTUNITIES AND CHALLENGES IN APP MARKETING Global mobile app downloads are set to rise 45% to 258 billion from 2017 to 2022, according to a recent forecast by App Annie². In that same period, predictions project that consumer in-app spending will reach nearly US$157 billion, a 92% increase from 2017. Thanks to the continuing strength of the app markets in China and India, and the emergence of maturing markets in Indonesia and Vietnam, the Asia Pacific region will see the fastest acceleration in both app downloads and spend³. This boom has both advantages and disadvantages for app marketers. While users in the market have shown a willingness to download and use apps, competition from app developers and marketers is fierce. How can app marketers maximize their spend by targeting and acquiring the highest value users?
Two key challenges in app marketing
Research shows that retention is the key marker for customer LTV4, but it is also a major pain point, with 55% of marketers nominating user engagement and retention the biggest challenge they face5. While retention is improving overall, the app abandonment rate remains quite high at 21%,
User retention
Fraudulent installs
which means more than one in five users never open an app again after initial use6.
2&3 https://s3.amazonaws.com/files.appannie.com/reports/1805_Report_2022_Forecast_EN.pdf 4 https://www.appsflyer.com/resources/2018-retention-benchmarks-2/ 5 https://www.statista.com/statistics/753889/leading-marketing-challenges-app-marketers/ 6 http://info.localytics.com/blog/21-percent-of-users-abandon-apps-after-one-use
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
3
This suggests app marketers need to engage users from the moment they install and open the app to increase the chances the user will complete an in-app event, whether that is agreeing to receive push notifications, starting a subscription or making a purchase. Lifting engagement reduces churn, increases revenue and therefore ensures a higher return on investment. Marketers also need to be wary of paying to acquire fake users, so these acquisition and engagement strategies must also filter out fraudulent installs and false in-app actions. Unfortunately, according to a report by a mobile attribution firm, the number of fraudulent app installs rose significantly in 20187, with more than 11% of all global downloads fraudulently installed8. The fraud extends to software that mimics use behavior when bad actors are paid per impression for in-app actions. All this comes at a price: about US$19 billion9 for app marketers.
7 https://www.emarketer.com/content/is-app-install-fraud-on-the-rise 8 https://hub.appsflyer.com/hubfs/State%20of%20Mobile%20Fraud%20Q1%202018%20AppsFlyer.pdf 9 https://www.juniperresearch.com/press/press-releases/ad-fraud-to-cost-advertisers-$19-billion-in-2018
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
4
HOW DEEP LEARNING IMPROVES ANALYSIS App marketers need to avoid fraudsters by identifying real and valuable users before bad actors have a chance to hijack the install. This starts at the user acquisition (UA) phase: identifying the right audience greatly reduces the chances of app fraud because you are looking for traits and behaviors prior to download, which is much harder for scammers to fake. Deep learning is also one weapon marketers are deploying to find valuable users and purchasing patterns10. To understand deep learning, you first have to understand how humans analyze data. A human approach consists of heuristic rules that are highly dependent on each person’s capabilities and experience, which means a marketing analyst who has a decade of experience is faster and has a better chance of understanding audience data than someone who just started in the role last week. Experience not only teaches this senior analyst which patterns are useful, but also which anomalies to take notice of, and which to ignore. Advance to machine learning, one of the popular applications of AI, which starts to become scalable and available in real time, with multi-dimensional predictions. However, humans still need to guide the program and often select the data to input, but the actions are not as straightforward as ‘if this, then that’. Deep learning, on the other hand, sees AI process large volumes of scattered, abstract data by itself to make predictions with higher accuracy.
10 https://theconversation.com/when-ai-meets-your-shopping-experience-it-knows-what-you-buy-and-what-you-ought-to-buy-101737
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
5
How is DL different from other approaches to ad bidding in app campaigns Human Approach
Machine Learming
Deep Learming
Approach
Human heuristic rules
Scalable analysis
Multi-task learning
Analysis of dimensions
1-3 dimensions
Multiple dimensions
Multiple dimensions
Processing Time
Days to weeks
Real-time
Real-time
Human efforts
Huge
Medium
Less
Data process
Only process data that makes sense to human
Data needs to be preprocessed by human
Process large volumes of scattered, abstract data by itself
Predictie ability
Rule-based
Data-driven prediction
Data-driven prediction with higher accuracy
For example, an app marketer for a retailer who relies on human analysis might notice that a certain segment of their audience, females aged 25-40, are more likely to purchase a handbag in the evening than any other time of day. That marketer might then use that information to run a campaign to promote handbags to that segment of users in the evening hours. Machine learning would reveal other layers of data. For instance, it is not just females aged 25-40, but women in that segment who are tertiary educated, who are more likely to make the purchase and who generate the most revenue see the product at least four times prior to purchase. Once deep learning is involved, the predictions start to take into account aspects of the purchase a human may never have considered. This might include how the weather that day affects which colours and styles sell best, how stock market movements in the morning might influence spending patterns in the evening and what other category items could be promoted successfully alongside handbags. This information, evaluated in tandem, makes it a powerful tool. Deep learning not only amalgamates all the patterns from previous purchases to make a prediction, it has also run a number of possible scenarios through its system to figure out which is the most likely to work. One big advantage is that it can do it at scale, resulting in highly targeted marketing campaigns. A deep learning-powered campaign might decide that Soo Lin in rainy Singapore is likely to purchase a green handbag on the train home, whereas Trudy in Hong Kong, after turbulence at the stock exchange that morning, just wants to browse luggage late that night as she is not yet ready to buy. Trudy does not need to own any stocks or even check the market for the deep learning algorithm to know this. The data indicates the likelihood and deep learning has already sized up Trudy’s ability to be influenced by various factors.
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
6
Based on Appier’s data, deep learning-based prediction is proven 17% more accurate than a human statistic predictive method, a smarter way to verify traffic that is good enough to bid on.
In-app events prediction accuracy: Human statistic approach vs. Deep learning
14%
Human statistic approach Deep learning
17%
17%
22%
13%
Total
Open
35% 4%
Tutorial
Register
Login
Search
Conversion
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
7
OPTIMIZING CAMPAIGNS WITH DEEP LEARNING Deep learning helps marketers improve campaign performance through AI return on advertising spend (ROAS) optimization. It learns from similar past campaigns using historical data and makes predictions based on results. This not only provides marketers with an in-depth understanding of their audience, right down to likely LTV, it can also indicate whether the campaign will achieve its conversion aims. AI tools can incorporate audience sampling from new traffic sources as well. Marketers can access these predictions before or during the campaign, so if it looks like the campaign will fail to reach its KPIs, AI will recommend that campaigns be tweaked or stopped so marketers can make adjustments that ensure cost effectiveness. More importantly, it tracks how users are engaging with the app. This is essential for behavioral analytics as research shows those users who opt-in to push notifications are more likely to use the app several times a day than those who don’t11 and that those who are engaged are less likely to churn. Actions such as signups and purchases also help AI characterize users as valuable so that UA campaigns can target a similar segment. This helps marketers target quality traffic, which in
11
http://info.localytics.com/blog/21-percent-of-users-abandon-apps-after-one-use
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
8
turn leads to more installs by valuable users and well-performing inapp events. Deep learning can therefore support re-engagement and retargeting campaigns aimed at marketers’ existing user base through user segmentation and customized marketing campaigns.
Case study Indonesia’s leading ride-hailing app wanted to attract as many users as possible in a competitive market. It used AI tools to recruit valuable users – those who were likely to make more bookings – to lower its customer acquisition cost (CAC). The company deployed Appier’s multi-layered
The best use of deep learning in UA campaigns is its ability to create valuable lookalike audiences based on historical data and a holistic view of LTV. Deep learning helps these tools evolve in a dynamic market to follow and predict shifts, be proactive instead of reactive. Based on Appier’s analysis of 244 app campaigns with 0.5 billion datapoints in 13 Asian markets between March and October 2018, deep learning-based prediction is more accurate than a machine learning-based approach without deep learning, when it comes to optimize the most challenging KPIs, such as retention rates and ROAS, which are 7% and 10% more accurate respectively.
deep learning technology, which predicted and optimized future events, including retention and
Deep learning optimizes the most challenging KPIs
purchase, in the conversion
Machine Learning Applied
journey by analyzing early user
Machine Learning + Deep learning Applied
patterns, such as clicks and
35%
installs.
36% 9%
7%
10%
Paired with the ad fraud predictor, which blocked suspicious traffic, the predictor helped the company increase its install rate by 119% and its booking rate by 63%, while reducing the CAC by 45%.
Retention Rate
Registration Rate
Booking Rate
Purchase Rate
ROAS
Data from 244 Appier app campaigns, Mar to Oct 2018
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
9
DEEP LEARNING VERSUS FRAUD Deep learning can also employ multi-stage fraud detection strategies, where the algorithm learns to identify new and evolving fraud patterns to develop new rules to respond. In this way, deep learning not only recognizes patterns from previous fraudulent events, it can also create and apply fresh parameters as fraudsters change tactics.
Outlier patterns detected by deep learning Publisher
Suspicious Patterns
Normal Pattern
CTIT Browser
Install Time
Device Location
Publisher Time
12
https://www.fico.com/blogs/analytics-optimization/5-keys-to-using-ai-and-machine-learning-in-fraud-detection/
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
10
Marketers use several datapoints to identify behavior that indicates whether the user is likely to be engaged and therefore likely to conduct valuable in-app actions. Deep learning takes this knowledge and compares new data to distinguish outliers12. Rather than discard this information as irrelevant to the campaign, it will take it into account when trying to detect fraud. This way it teaches itself how to discern the difference between a high-value user’s behavior and suspicious activity – and everything in between. Numbers related to downloads and installs are no longer the end goal for app marketers. Today, app use – from engagement to in-app purchases – are the metrics by which marketers measure success. Deep learning can make sense of complexity and, as a tool, support marketers to optimize campaigns that capture and retain real, valuable users, and negate fraud. This reduces the need for trial and error, which lowers the cost per acquisition and increases profitability.
Want to find out more about how you can leverage deep learning to enhance your app acquisition campaigns? Contact us and explore our CrossX Advertising Solutions today.
12
https://www.fico.com/blogs/analytics-optimization/5-keys-to-using-ai-and-machine-learning-in-fraud-detection/
Landing Valuable App Users: How Deep Learning Enhances Acquisition Campaigns
11