Proactive Ad Fraud Prevention With Artificial Intelligence

As marketers grapple with the problem of ad fraud and its mounting losses, artificial intelligence (AI) is proving to be an effective weapon that can reverse the tide.

Marketers in Asia Pacific continue to throw money at advertising, as ad spending is expected to increase 10.7 percent to US$210.43 billion in 2018, according to eMarketer. However, the ever-growing problem of ad fraud is skewing their reporting and standing in their way of showing better returns.

Even mobile marketers who expected more safety with app installs, faced 30 percent more fraud during the first quarter of 2018 compared to the same period last year, according to AppsFlyer’s “The State of Mobile Fraud: Q1 2018” study. Mobile app marketers were exposed to US$700-US$800 million in ad fraud losses worldwide. What makes ad fraud such a challenging problem today?

More Sophisticated Ad Fraud Methods Today

In the early days of ad fraud, the methods adopted by fraudsters were relatively simple. They used bots focused on driving large volumes of traffic to websites, bought cheap traffic through auto redirects or employed people to install apps in click farms. Once a click was made or an app was installed, their job was done. However, this was soon caught by advertisers and their focus began shifting to examining post-install quality, engagement, last-click attribution and return on investment (ROI).

Today, ad fraud has evolved into a completely different beast. For example, in a technique called ‘click injection’, they try to steal credit for app installs by triggering a click right before these apps get installed. Or, they stack up ads on top of each other to generate impressions for multiple ads.

Fraudsters also realize that to remain undetected, they not only need to drive traffic to websites or generate large amount of app installs, but also remain engaged subsequently and mimic human behavior. They even employ human workers in click-farms to imitate real human interactions, develop apps that hijack devices to generate additional clicks and create simulators that generate fake installs from bot networks.

Brands Turn to Technology for Help

Big brands and major publishers have begun to act against ad fraud. The Guardian recently collaborated with Google and MightyHive, a programmatic solution provider, to investigate programmatic fraud. Adobe Advertising Cloud has partnered with cybersecurity firm White Ops to tackle the problem in the streaming TV media.

Another technology that is gaining traction among marketers is blockchain, thanks to its ability to enforce decentralized monitoring and independent verification. However, there are still many challenges to be addressed, like handling the massive transaction volumes involved in real-time bidding and getting universal acceptance from everyone involved.

Joe Su, Chief Technology Officer and Co-founder of Appier, explains, “I think it will be a while before everyone in the supply chain – from media buyers to ad exchanges and publishers – cooperates to opt into a universal standard to make it feasible.”

Sophisticated Detection Powered by Machine Learning

Traditional approaches in fraud detection rely on simple rules created by humans through the measurement of three signals (metrics like conversion rate and click time) or dimensions at most. For example, IP blacklists to block suspicious traffic and filter out installs with low click-to-install-time (CTIT) have worked in the past.

However, the problem with these fixed rules is that they are pre-defined, as advertisers approach detection knowing what they are looking for. For example, one might decide to exclude app installs with CTIT below 10 seconds, due to the higher likelihood of bot-operated installs in these cases. But if such rules are fixed ahead of time, it’s only a matter of time before fraudsters figure out ways to circumvent them.

A more effective solution is to leverage technology that can keep pace with, and more importantly, stay ahead of today’s ad fraud techniques. That means marketers need to go beyond simple, fixed rule-based criteria, towards AI-powered solutions that are capable of learning new fraud patterns and refining the rules on their own.

As fraudsters employ new techniques that are capable of mimicking human behavior, these machine learning algorithms can help marketers look for fraudulent behavior not immediately evident to the human eye. This is especially critical in the case of app installs, where detecting fraudulent clicks and impressions before install becomes paramount.

“Looking for signals like suspiciously low time between clicks and installs or CTIT can indicate fraud, but by the time the install has occurred, it would be too late and an attribution for the install has already been counted,” said Su.

One foundational AI-approach, called the “tree-based model” works by analyzing a massive number of signals to achieve maximum coverage and accuracy in detecting outlier behavior. Consider the case of “the chameleon”, where fraudsters mimic legitimate publishers and generate installs at a later date, when the natural user retention is expected to drop.

Another scenario is that of an “inventory burst” where inventory count from a suspicious publisher spikes at a time when generally the in-app registration falls. As machine learning algorithms learn from gathered data over time, both these sophisticated ad patterns can be detected and fed back into the filters for improved detection in the future.

By detecting more cases of ad fraud, marketers can weed out poor quality traffic and measure their returns on advertising spend (ROAS) more accurately. In a study of 5.2 billion data points from mobile app campaigns in the region, Appier detected twice the number of suspicious installs using an AI-based approach and realized 4 percent more ROAS, compared to a traditional approach.

It won’t be long before fraudsters develop even more novel techniques to try and escape detection. Advertisers who are vigilant and proactive in preventing this with AI-based fraud detection will be in a better place than their competitors to reap the benefits of their investment.

How Far Are We From Explainable Artificial Intelligence?

Artificial intelligence (AI) is heralding a revolution in how we interact with technology. Its capabilities have changed how we work, travel, play and live. But this is just the beginning.

The next step is explainable AI (XAI), a form of AI whose actions are more easily understood by humans. So how does it work? Why do we need it? How will it forever change the way industries – especially in marketing – function?

The Mystery of the Black Box: The Problem With Current AI

No one would deny that artificial intelligence produces amazing results. Computers that can not only process vast amounts of data in seconds, but also learn, decide and act on their own have turned many industries on their heads – according to PricewaterhouseCoopers, the market worth of AI is around US$15 trillion. However, in its current form, AI does have one major weakness: explanation.

Namely, it can’t explain its decisions and actions to humans. This is sometimes referred to as the “black box” in machine learning – for example, the calculations and decisions are carried out behind the scenes with no rationale given as to why the AI arrived at that decision.

Why is this a problem? It doesn’t engender trust in the AI, which in turn raises doubt about its actions. Explainable AI is expected to solve that.

How XAI Works

XAI is much more transparent. The human actors interacting with the AI are informed not only of what decisions it reached and actions it will take, but how it came to those conclusions based on the available data. It aims to do this while maintaining a high level of learning performance.

Current AI takes data into its machine learning process and produces a learned function, leaving the user with a number of questions such as: Why did it do that? Why didn’t it do something else? When will it succeed? And when fail? How can I trust it? And how do I correct an error?

By contrast, XAI uses a new machine learning process to produce an explainable model with an explainable interface. This should answer all the questions above.

This carries its own risks. Any decision made by an AI is only as good as the data used to make it. While XAI increases trust in the decision made, that trust could be misplaced if the data is unreliable.

Another problem is how well the AI explains its decisions. If it is not comprehensible to the user – who could be a lay person with no technical background – the explanation will be worthless. Solving this will involve scientists working with UI experts, along with complex work on the psychology of explanation.

Risk, Trust and Regulation: Why We Need XAI

In so-called “big ticket” decisions like military, finance, safety critical systems in autonomous vehicles and diagnostic decisions in healthcare, the risk factor is high. Hence it is crucial that the AI explains its decisions in order to boost trust and confidence in its ability. However, there are a host of benefits for businesses in other industries.

XAI can address pressures like regulation, as it will enable full transparency in case of an audit. It will encourage best practice and ethics by explaining why each decision is the right one morally, socially and financially. It will also reinforce confidence in the business, which will reassure shareholders.

It will also put businesses in a stronger position to foster innovation, as the more advanced the AI, the more capable it is in terms of innovative uses and new abilities. Interacting with AIs will soon be standard business practice in many industries, including marketing. Hence it is vital that users can do so comfortably and with confidence.

Experts think this will empower marketers, effectively turning AI into a co-worker rather than a tool.

“In order to trust AI, people need to know what the AI is doing,” says Hsuan-Tien Lin, Chief Data Scientist, Appier. “Much like how AlphaGo is showing us new insights on how to play the board game Go, explainable AI could show marketers new insights on how to conduct marketing. For instance, AI can reach the right audience at the right time now, but if future XAI can explain this decision to humans, it would help marketers understand their audience more deeply and plan for better marketing strategies.”

It could also usher in a new way of working, with marketers accepting or rejecting XAI’s explainable suggestions with reasons in order to help the AI learn. “Today, it is likely that many great suggestions are rejected because they are not explained, and so humans overlook their power,” says Min Sun, Chief AI Scientist, Appier. However, these days could soon be over…

The Defense Advanced Research Projects Agency is currently running an XAI program until 2021. The program is expected to enable “third-wave AI systems”, where machines can build underlying explanatory models to describe real-world phenomena based on their understanding of the context and operating environment. Other experts also predict XAI will become a reality within three to five years.

XAI is no doubt the next step for AI, improving trust, confidence and transparency. Businesses would be wise not to overlook its potential.

Pinpoint Your Ideal Audience With Machine Learning

Consumer behavior is increasingly fragmented, and that has resulted in the tsunami of data, which can be overwhelming for marketers. Thanks to the behavior-based model of today’s artificial intelligence (AI) tools, marketers can now leverage these tools to segment the audience beyond the traditional parameters to build a more accurate portrait of them as individuals.

By employing AI, you will be able to segment your audience into more granular tiers, and to see which are more valuable to your end goals. Considering that more than half of customers say they will walk away from brands that send messages they find irrelevant, it’s vital that you understand your customers’ current and future intentions as well as possible.

Individuals vs Categories: Looking Beyond the ‘Types’

Artificial intelligence can find hidden user patterns that have a positive or negative impact towards achieving the goal that marketers want to achieve. Hence it can help marketers reach a predetermined end goal, be it finding the more-likely-to-purchase audience segment to drive sales volume, or giving an existing audience a personalized product or content catering to their interest.

This is very useful for marketers. If a website has an audience of one million people, a marketer might want to target a certain segment of that audience, one which is more valuable to them (maybe they are more likely to buy their product, or read more of their articles).

The traditional way of doing this is by considering three dimensions: their demography (age, gender, social aspects); their behavior (how recently they visited the website, for example); and their interests. Using these, you can target people who put one item in their online shopping cart within the last seven days, for example. It can be quite accurate, but a lot of it is guesswork, which can also be biased or wrong. It would also only allow marketers to reach limited users with purchasing intent, whereas AI has the added value of finding all possible valuable-to-marketer-defined goal users.

AI also eliminates the guesswork. It considers many more factors to produce countless combinations. Let’s take a simple example. If we consider degree of interest (say 200), internet browser behavior such as how active each individual is in a period of defined time duration (1 million), site action (6), site frequency (100), site action duration (180), and multiply them all together, you would have an enormous number. Add in more age groups, users accessing the website on certain devices and another couple of factors, and you can see how quickly the possibilities escalate.

“AI gives you literally countless combinations,” says Magic Tu, VP of product management at Appier. “Usually people are phased by this level of complexity, and have limited time in which to interpret it, so they simplify the parameters. However, that loses the richness of the data.”

Following the Behavior Patterns

Not only does AI give you this incredibly granular view of your audience, it also tells you which factors will have the most influence over whether they convert into paying customers or not, and will spot those behavior patterns. It learns from the historical data and continuously adjusts the predictive result by incorporating whichever new data becomes available. All you have to do is to create the model and tell it your end goal. With AI doing all the heavy lifting, marketers are freed up to explore the potential opportunities hidden in the data.

By predicting what users are likely to do next, marketers can adjust their campaigns accordingly. For example, if a company used AI to find the customers most likely to buy a new product, they could segment these users and target them with personalized offers and/or messaging to help convert their intention into a sale.

Chicken, Cakes and Lookalikes: Getting to Know Your Customers Better

AI tools can also provide greater insight into your customers’ interests, and they can do so with an incredible level of specificity – instead of a subject like ‘food’, which is too broad to be of any use, AI tools will analyze the terms users use, such as ‘tonkotsu recipes’ or ‘chicken restaurants’.

The application of AI can make sense of social media posts too. By using natural language processing (NLP), it picks out the keywords and phrases (such as ‘I love this type of cake’) to gain a representative understanding of the post, and then to see its outcome (for example, whether someone who wrote that post went on to buy that type of cake). The more of these posts you feed it, the more accurately it’s able to analyze patterns of behavior, which it uses as a basis for building a predictive model.

“More structured data – like numbers – are of more use to AI,” says Tu. “By using NLP, we can turn articles, blog posts and social media posts into more structured data, which makes use of a huge repository of internet content.”

Using machine learning, ultimately we could segment audiences not only by their static attributes such as gender, age, interest and so on, but also predictive behavior in the near future which is the key marketers always want to know. Combining both static attributes and predictive behaviors together creates the most efficient audience segments that marketers could ever build to achieve their goals.

Changing the Conversation

Such innovations are changing how marketers reach out to customers. For instance, Appier’s Aixon solution recently integrated with Line, one of the biggest messaging platforms in Asia. By cross-mapping Line user IDs with Appier’s CrossX database of over two billion device profiles, it can segment them into keyword and interest areas. The customer organization then uses Line Business Connect to send personalized messages to the users most likely to respond positively.

If a Line user puts a product in their online shopping basket without completing the purchase, for example, the site owner can send them a personalized message through Line with a special offer to convince them to buy the item.

Thanks to AI tools, beauty brand Lancôme discovered that skincare shoppers and make-up shoppers are two very different segments. This lets it show a set of products, videos or articles for eye creams but not mascaras to the skincare shoppers, to avoid serving them with content they will find irrelevant. It also suggests foundation in shades closer to the user’s skin tone, instead of bombarding them with the full 185 shades, most of which will not be of interest. This has resulted in a conversion rate three times greater than previously.

By segmenting their audience using machine learning, marketers can build a more dynamic picture of who is using their service. By going beyond the static traits and targeting those customers who are more valuable, they can focus their marketing campaigns accordingly and meet their customers’ needs more effectively.

Are Data Scientists Evolving With the Rise of Artificial Intelligence?

As developments in machine learning (ML) are expected to progress at a phenomenal pace, it is set to become one of the most powerful tools for businesses to enhance productivity and drive innovation. While ML, one of the most popular artificial intelligence (AI) applications, holds a lot of promise for businesses, is the role of data scientist today already evolving in order to keep up with the change?

What Is Next in AI

Continued advances in AI will see autonomous systems perceive, learn, decide, and act on their own, but to ensure the effectiveness of these systems, the machine will need to be able to explain their decisions and actions to humans. This is so called explainable AI.

“In the future, many AI systems are going to interact with people, especially those who will take responsibilities, hence the reason why AI needs to be explainable, meaning that the behavior of the system needs to be easily expected and interpreted by people,” said Min Sun, Chief AI Scientist at Appier.

Sun also pointed out that in the future, AI is going to be less supervised, which means that it will require less human inputs, and be more creative.

Data science was previously concerned with time-consuming ML tasks, such as data wrangling and feature engineering, which could take up 80 percent of data scientist’s time, but such tasks can be automated sooner or later, according to Deloitte’s Technology, Media and Telecommunications Predictions 2018 report.

Such advances in AI will give data scientists more time to execute more complex tasks. However, it brings up a problem: a majority of data scientists doesn’t possess the required advanced machine learning skills, such as deep learning (DL), a subfield of ML.

The Impact of Machine Learning on Businesses

Previously, companies might have spent a lot of time doing guesswork based on consumer data gathered online and offline, which is usually fragmented and siloed. With an AI-based approach, brands are able to unify data across different channels for a holistic view and analysis of the audience and their conversion journey.

Machine learning and deep learning allow a computer to take in huge sets of data and not only predict the outcome, but also understand what the desired output should be. It can be integrated into many aspects of digital marketing, such as predicting consumer behavior and campaign outcomes, marketing automation, sophisticated buyer segmentation and sales forecasting.

With these technologies, businesses have a more efficient and cost-effective way to build trustworthy AI systems to be used by professionals and/or to be naturally interacted with human users, according to Hsuan-Tien Lin, Appier’s Chief Data Scientist.

So, it’s no surprise to see that businesses are increasingly catching up on the adoption of AI technology. According to the International Data Corporation (IDC), AI continues to be a key spending area for companies in the near future, with worldwide spending on cognitive and AI systems increasing 54.2 percent in 2018 to US$19.1 billion. That number could go up to US$52.2 billion in 2021, IDC predicted.

Bridging the Machine Learning Skills Gap

As more businesses look to adopt AI techniques like machine learning and deep learning, data scientists are urged to upskill, in order to keep up with the current trends. Rudina Seseri, Founder and Managing Partner at Glasswing Ventures, wrote in Forbes, “Data scientists – at least the successful ones – will evolve from their current roles to becoming machine learning experts or some other new category of expertise, yet to be given a name”.

Leading tech companies such as Google and Microsoft have already been offering relevant courses aiming to help bridge the talent gap. For example, Google not only made its ‘Machine Learning Crash Course’ available to the general public earlier this year as part of the company’s ‘Learn With Google AI’ initiative, it has also launched a machine learning specialization on Coursera, an online learning platform.

Andrew Ng, one of the world’s best-known AI experts, also launched a set of courses on deep learning through Coursera in 2017, hoping to help more people get up to speed on key developments in AI.

While technical skills will be foundation of the role of data scientists, it’s crucial for them to master human-centric skills too. Data scientists will need to develop a better understanding of the overarching business strategy and business challenges in real-world scenarios, in order to create solutions that can solve real problems.

Businesses are looking for a total solution, Sun pointed out. For instance, self-driving car manufacturers need a system consisting of perception, communication, decision-making and control. In the old days, each module was designed separately, but this has been transitioning to more jointly design since the fatal self-driving Uber crash, where the perception system identified the pedestrian, but the decision-making module failed to react.

The ability for scientists to design a complete system consisting of multiple ML modules will become more and more important,” he said. “In the future, data scientists will need to have the modeling and analysis skills at the system-level to provide business people with the right total solution to the market.”

What It Means to Be a Data Scientist Today

By Yao-Nan Chen, Machine Learning Scientist, Appier

Unless you have been hibernating under a rock for a few years now, you already know that explosive growth in the volume of available data is disrupting business as we know it. This data can be a goldmine for businesses that know how to capture, analyze and use it to power artificial intelligence (AI) technology. And that’s where data science and my role come in.

IBM has predicted that demand for data scientists will increase by 28 percent by 2020. The Harvard Business Review, way back in 2012, said that being a data scientist is the sexiest job of the 21st century.

I have been working in data science since 2013 and I still come into work at Appier each day eager to solve new problems.

What Data Scientists Actually Do

Simply put, data science involves using data to generate solutions that solve practical, real-world problems. In the business world, examples revolve around AI-powered solutions, such as pushing recommendations for users based on their demographic or usage pattern, or analyzing why sales of a particular product is dropping.

Data scientists set out on solving such problems by first extracting and consolidating data, which we then analyze for patterns and trends. We use this to build predictive models, derive insights, and implement proof of concepts to test the proposed solution to the problem at hand. The problems that we work on are very specific and often have no one standard solution. Hence, data scientists are tasked with thinking out of the box to come up with a variety of possible solutions.

The impact of our solutions is known only when they are implemented; so often, if the solution fails to meet the desired outcome, we have to go back to the drawing board and start over. But this just adds to the challenge and the excitement of trying to pin down that elusive solution and make it work.

What Makes for a Good Data Scientist

Of course, every job has some less lovable bits and the burden of the data scientist is data cleaning! In most cases, the data we gather is ‘dirty’, with errors and discrepancies in it. For example, data showing that sales of a product have dropped dramatically may simply mean that malfunctioning machines have failed to capture the data accurately.

Most data scientists will agree that data cleaning is the most boring part of this job. Our inside joke is that data science is 80 percent cleaning of data and 20 percent complaining about it!

But jokes aside, data cleaning is painstaking but important work. If not done right, it can have a huge impact on the accuracy and reliability of insights.

Aside from this kind of assiduity and attention to detail, a good data scientist, no matter how good they are technically, must also have a thorough understanding of business domain and the organization’s business goals. Our solutions have to be creative, but also useful and practical.

Keeping Up with the Latest Research

In this context, keeping up with the latest research in the area of machine learning can help us stay on top of trends and monitor breakthroughs to specific problems. We don’t need to reinvent the wheel – if a particular problem has been solved before, we can always work off that.

I regularly read papers on advances in machine learning, as well as in the specific domains that I am interested in.

It’s equally important to engage in discussions with peers, keep track of their recent research and poll their opinions on machine learning trends. This will help you keep abreast of all that is happening in this area.

Growing Demand for AI Expertise

Unfortunately, there is a gap between the growing demand for data scientists and the supply of talent in the area. AI is a new track and there is a shortage of people with the required expertise. What widens the gap is that not every data scientist is a good business person. They may be stellar at solving problems in an academic or research-based environment, but often fall short when it comes to real-world business problems.

Data scientists today must constantly evolve in terms of skill set. As the adoption of AI and deep learning grows, we are automating lower-level tasks and moving onto more complex problems. We already have some mature tools that can be used to build simple models for many business cases, and these are becoming simpler to use.

In the near future, data scientists will be required to know how to leverage and use problem-specific information. As AI becomes more complex, data scientists will need to work on more abstract problems and leave simple processing and analyses to automation software.


About the author:

Yao-Nan Chen is Machine Learning Scientist at Appier. He has more than five years of experience in Machine Learning, Data Science and Data Engineering and three years of experience in practical E-commerce recommendation system. Prior to joining Appier, he worked at Yahoo Taiwan on E-commerce Recommendation System, App notification recommendation system, Model tuning for sales volume prediction, etc.

 

 

AI 101: Deep Learning

Imagine that you are a marketer looking to run a targeted marketing campaign. What if you had a tool that could easily segment your market on the basis of factors like economic status, purchasing preferences, online shopping behavior, etc. so that you could customize your approach and messaging to each segment for maximum impact and conversion?

These are the kind of insights that deep learning (DL)* can offer.   

DL refers to a family of advanced neural networks that mimic the way the brain processes information and extract goal-oriented models from scattered and abstract data. What differentiates it from traditional machine learning is the use of multiple layers of neurons to digest the information.  

A DL program trains a computer to perform human-like tasks, such as speech recognition or predicting consumer behaviors. It is fed large amounts of data and taught what the desired output should be. The more data it’s fed, the better performance.

The program then applies calculations to achieve that output, modifying calculations and repeating the cycle until the desired outcome is achieved. The ‘deep’, hence refers to the number of processing layers that the data must pass through to achieve the outcome, and how the learning algorithms are stacked in a complex, hierarchical manner. The more levels or layers there are, the ‘deeper’ the learning.

DL can analyze huge volumes of data to detect patterns and predict trends and outcomes. This is especially interesting to marketers, finding application in predicting consumer behavior and campaign outcomes, marketing automation, sophisticated buyer segmentation and sales forecasting, to name a few use cases.

*Deep learning is not magic, but it is great at finding patterns.