MediaMath Blog

main

Data Events Trends Uncategorized

5 Questions About Dmexco 2016

September 29, 2016 — by MediaMath — 1

After two overwhelming days of nonstop networking and social events, the recovery from Dmexco is never easy, but the knowledge and lessons learned make it all worthwhile.

This year, we shared our booth space with our Audience Partners, VisualDNA and Eyeota, and hosted Listen and Learn sessions with both of them as well as with our B2B data partner, Bombora. We felt this was a great way to highlight our close industry partnerships, highlight our differentiated solutions together and collaborate closely with their teams while at Dmexco.

Now that it’s been almost two weeks since we are all back to our home offices, I decided to catch up with each of them and see how their Dmexco experience went.

• What are some of the emerging trends that you’ve seen come out of this year’s Dmexco?

VisualDNA:

We saw a marked transition, even from recent expos, that audience data for targeting and analysis has gone from a nice-to-have in previous years to a must-have. We got great feedback from clients that they are targeting significant proportions of their prospecting campaigns using data.

Eyeota:

Audience data is definitely becoming essential in programmatic campaign planning. We did a lot less explaining and educating about what we do and more about how we can work together with partners and clients. Exciting stuff!

Bombora:

While a lot of discussion from Dmexco 2016 has centered on the rise of programmatic, one of the perhaps hidden trends that is taking a foothold in driving the industry is content production. Set to be worth $313 billion in 2019 according to PQ Media*, content truly is king! And with the explosion of content across the B2B landscape, companies need to focus on delivering personalized, relevant and timely information. In addition to this, the diversification of technology and channels (e.g. augmented reality, mobile, social, etc.) means that companies need to get even more creative with the production and delivery of their content.

• Moving into 2017, do you think Dmexco has changed your positioning in the EMEA market?

VisualDNA:

We launched our Audiences With Personality on the first day of Dmexco, and the feedback was superb, it is the right time to have the best psychographic data with all of the debate about creative coming together with tech and data in programmatic offerings to improve digital advertising. An issue for planners and traders has been getting consistent high reach audiences across Europe. VisualDNA has answered that with high reach audiences in France, Germany, Italy and Spain as well as UK which we are famous for, and of course USA. As a result, I’m sure that our 2017 positioning will be greatly steered by these and more brands with EMEA and global planning will engage with our data.

Eyeota:

Not really changed our perspective — more and more US companies are showing up in full force, making Dmexco an even stronger international show for us, and we see that continuing for 2017.

Bombora:

With 47 percent of our B2B interactions captured outside of the US, Bombora’s focus has always been global. We recognize that in order to build the richest source of B2B intent data we need to surf the ‘waves of intent data’ across the world! Dmexco gives an opportunity to gauge the market, as well as expand our brand and footprint into the EMEA market, which is a strategic focus for us in 2017.

• Can you provide some key tips/hints/tricks on how to prepare for Dmexco 2017?

VisualDNA:

Clarity of proposition is becoming more key with over 1,000 exhibitors present. The puck has moved to a place where there is more clarity required about why your offering is better and different, it is no longer acceptable just to be capable. That is key to attracting new clients, for your existing ones plan ahead, book meetings where possible, people’s time is precious and their schedules are crazy. And wear comfortable shoes…

Eyeota:

Have routine prep meetings leading up to Dmexco to clearly map out goals and objectives, by team member, adding metrics to strive for. This makes your overall investment in the event measurable.

Bombora:

Getting there and beyond — book in your passes, travel and accommodation as early as you can. This just means it’s all taken care of and you can focus on all of the other fun stuff that comes with the conference sooner!

• What’s been your highlight of this year’s Dmexco?

VisualDNA:

Two stand outs:

Receiving great feedback and validation for VisualDNA’s new Audiences with Personality product, it is generation 5 of our psychographic based data and the best product / market fit we have achieved so far.

We spoke to many current clients about how our audience data frequently outperformed others. We could have candid conversations and found that people are more open when away from the office.

Eyeota:

A major highlight was being a part of the MediaMath booth and having an opportunity to present on audience data trends. Another highlight was our pre drinks event on Tuesday Sep 16^th at Fruh. It was a huge success and we had a much larger turnout than expected. We’re already planning on doing it again next year! Lastly, since Dmexco, we have been running a wildly popular 20% off deal on seasonal segments available through MediaMath’s T1 platform.

Bombora:

In addition to the announcement of a new partnership with content marketing platform, Outbrain, our presentation at the MediaMath booth was equally exciting. It was a great opportunity for our CEO, Erik Matlick, to share insights about the influence of intent data on driving content and Account-Based Marketing. We received fantastic exposure to the Dmexco attendees, including some key decision makers who expressed excitement our B2B intent data.

• What was your biggest learning/key takeaway from the event?

VisualDNA:

Location — it was great partnering on MediaMath’s booth in the heart of coveted hall 7 with high foot traffic. Just like a property agent we can verify that location is key…

Eyeota:

Let’s go bigger next year!

Bombora:

Our biggest take away was the opportunity presented through the content marketing industry and how important personalization is, as well as the dynamic growth of the industry in EMEA.

Data Uncategorized

5 Questions with Mobilewalla

September 22, 2016 — by MediaMath

We recently partnered with Mobilewalla, a mobile audience provider with a rare focus on APAC markets with full US coverage as well. I was fortunate to be introduced to their CEO and Founder, Anindya Dutta, via our Singapore office a few months ago, and we have not looked back since. From making their standard and custom audiences readily available to us via our Lotame integration to participating in our Q4 holiday seasonal promotions, Mobilewalla has rapidly come up to speed to become one of our preferred mobile audience targeting partners. I recently had a chance to chat with Anindya on some of the recent trends in mobile targeting, why the APAC market may be emerging as a leader in mobile and how he and his company are uniquely differentiated in this niche marketplace.

You say on your website that Mobilewalla has created an entirely new class of data science techniques to enable marketers to achieve 1-to-1 addressability of consumers on mobile. Can you share a bit of what has gone into these techniques and how they make your platform so powerful?

A consumer’s mobile smart device has substantially more insight and discernment into the owner’s lifestyle and behavior than any other digital device. It is privy to their movements and is co-located with the individual at his/her home, workplace, shopping/entertainment locations and when they drop their kids at school. A smartphone also knows what the owner’s favorite music and video apps are and even has valuable insights into their voting preferences. Marketers increasingly regard mobile as the medium of choice to engage consumers through contextually-served, individualized content—the holy grail of 1-to-1 marketing.

Yet, there are a number of technical impediments that prevent this goal from being realized, making mobile consumers even harder to identify and address than those on desktop. Two noteworthy issues are:

On desktops, consumers are identified and tracked based on cookies. In the industry, cookie-based addressing is well understood. In mobile environments, cookies largely don’t work when consumers use apps and websites on their smartphones and tablets. On apps, the notion of cookies doesn’t exist. On the mobile web, major browsers (such as Safari Mobile) do not support third-party cookies. The result is that consumer identification strategies need to be re-thought on mobile.
Observing and analyzing signals from consumer devices form the basis of consumer modeling. The mobile scenario is no different—a variety of signals continuously emerge from smartphones (ad requests and cell tower handshakes, for example) that embed information-rich data elements useful in understanding the consumer behind the signal. However, there are key differences between signals from mobile vs. those from fixed-line digital devices, such as desktops.
- First, the sheer volume of signal data from mobile devices dwarfs data sourced from desktops. Consider ad requests, a widely recognized source of valuable consumer data. Just in the US alone, over 12 billion ad requests are observed daily, accounting for almost 50 PB of raw data. For longitudinal analysis, simply accumulating a month’s worth of ad requests requires 1.5 exabytes of storage. Clearly, this is infeasible to do on an economical basis.
- Second, the individual data fields available in mobile signals, such as ad supply, are noisy. Take, for instance, two of the most information-rich attributes that enable understanding of the consumer: the media source of the ad request and device location. Both of these fields are messy, with the media source often being blinded and the device location rerouted often misleading.

These are just some of the issues that render the mobile consumer harder to understand and model than traditional digital users. Mobilewalla has made key technical breakthroughs that address these issues and enable mobile consumer addressability in a cookie-less environment.

You come from the world of academia. How has this past experience influenced your time at Mobilewalla?

Two clear examples come to mind. First, as a practicing computer scientist, I was motivated to solve interesting, hard problems whose solutions had practical applications. I started Mobilewalla with the same mindset—the mobile landscape offered a rich trove of hard data problems that needed to be addressed in order for the medium to achieve its potential. In many ways, Mobilewalla was more driven by my love of interesting data problems than by a clear focus to solve a specific business challenge. However, I knew that if we could address some core issues (like how to store massive amounts of mobile data), the applications would be manifold. A second way my academic past has influenced Mobilewalla is in recruitment. The initial engineers, who now represent a large part of the key technical team, are my students from my prior life. Universities offer perhaps the best selection of bright engineers in a single location, and I doubt I could have been able to put together a powerful technical team as quickly as we did without that connection. And, without the right team, Mobilewalla could not have traveled the distance it has.

It’s estimated that APAC will become the leader in mobile ad spend globally by 2017, according to eMarketer. How do marketers best take advantage of the opportunity in this region?

Because of the relatively significant population segment outside of the purview of fixed line telephony in the most populated countries in Asia (China, India and Indonesia, for example) mobile devices have become the ubiquitous form of communication, contributing to media consumption and ad spend. To take advantage of this, marketers should employ best practices from the industry, while accounting for special characteristics of consumers in their local markets. There is widespread agreement that APAC marketers should consider the following:

Deeper understanding of consumer needs
Smarter use of location technology
Programmatic techniques
Video content

It is interesting to note that a common thread running across all of the above is data. It is important that APAC marketers become attuned to data-driven marketing to address the needs of the geographically and culturally diverse continent.

According to a just-released Google report, Asian consumers are far more likely than those in the US to make quicker purchasing decisions because of online research. How do marketers in APAC capitalize on this tendency in smart, data-driven ways?

This is a great question. Perhaps the most defining aspect of online commerce in Asia is that a vast majority of it is on mobile-web, as PC-centric broadband connections are either absent or unaffordable by a majority of the populace (and smartphone penetration, while increasing, is still small in large markets like India and China). With mobile devices, all of a sudden, hundreds of millions of potential customers are within reach of online merchandisers. While on one hand this is great, on the other, it gives rise to challenges that marketers must address to exploit the opportunity. It turns out that the science of acquiring, retaining and marketing to customers online is well-understood and is dependent on the ability of merchandisers to identify customers in and across online visits. In the desktop world, these occur based on cookies, but in the mobile world, cookies don’t work. In order to fulfill the massive potential of m-commerce in Asia, marketers must look into and deploy techniques of mobile consumer identification and tracking, which are based on innovative data science techniques.

It’s often said that despite how much increasing budget is being put towards mobile advertising, marketers by and large have yet to “get mobile right.” What do you think are next steps marketers need to take to harness the power of mobile?

This is a nuanced, complex question. One way to look at it is to recognize that mobile marketers are, for the large part, desktop experts in their prior lives. Digital activities in advertisers, agencies and publishers have traditionally been desktop-driven, and desktop experts are the ones who are fashioning mobile strategies for these organizations. It follows, therefore, that the mobile marketing strategies of today are heavily influenced by the best practices of traditional digital (aka desktop) strategies. Yet, as we discussed in our answer to Question 1 above, the mobile medium, in some fundamental ways, is different from desktop—the core mechanisms to identify consumers in desktop don’t work on mobile, for example. By extension, given the importance of consumer data forms in digital marketing, “getting mobile right” will be heavily dependent upon marketers’ willingness to understand the differences of the mobile medium from traditional desktop and appreciate and address the concomitant complexities manifested in mobile consumer data. In particular, understanding mobile consumer identification strategies (and their nuances), understanding what can and cannot be done with location data available in mobile and understanding mobile consumer addressability contexts (time, place, event) will greatly aid the flourishing of mobile marketing.

Data

Top Four Approaches to Building a Data-Driven, Customer-Centric Strategy

September 13, 2016 — by MediaMath

Why is a data-centric approach important in marketing? For two main reasons:

Cross-device usage is increasing: Who is the person on your site and what devices is she using? The average US consumer owns four digital devices and consumers spend 60 hours a week consuming content across all devices.
Data drives performance: It’s hard to get the right data and use what’s effective—but it’s a must because data drives efficiency. Seventy-one percent of marketers say second-party data increases revenue. Personalized ads based on previous purchases are 92 percent more likely to lead to a conversion than static ads.

Enriched first-party data is the best data. But there’s limited scale and insights. You need to overlay it with other data from other activity from across the internet to get the right messages in front of the right consumers. Second-party data—other brands’ first-party data–is one way to do this. In the webinar “Top Four Approaches to Building a Data-Driven, Customer-Centric Strategy” with the DMA, Heather Blank, SVP of Helix, walks through how marketers can do more with their data. Watch the webinar below or here on YouTube.

Data Uncategorized

Programmatic Advertising: Where Does the Journey Go in Austria?

September 9, 2016 — by MediaMath

Where is the current opportunity and challenge for programmatic marketing in Austria? In this translated excerpt from the German publication Werbeplanung.at, Viktor Zawadzki, Region Manager, DACH, for MediaMath, says that one still untapped opportunity in the region is data management.

“A large area in which advertisers still limit themselves is data management,” Zawadzki says. He recommends a data management platform to bring more structure to all of the data that lives inside an organization. “Very few have so far managed their internal databases to connect so that they have a holistic perspective on their internal processes and their interaction with external activities.”

Zawadzki admits that it is not lack of willingness but simply a matter of the complexity of an undertaking like this. In Germany, programmatic marketing has yet to be adopted as the standard for media buying. In the rest of the world, marketers are making giant strides. “In North America, in 2016, already 63 percent of spending on display ads is made programmatically, and the UK is also positioning itself as an aggressive growth driver, spending two billion pounds,” he says. “We see in the United States that analog channels such as print, out-of-home, radio or connected TV are now able to be bought programmatically.”

For more on how marketers can better manage their data, read the following posts:

Data Uncategorized

5 Questions with Webbula

September 8, 2016 — by MediaMath

With all the different audience targeting options available, we often get asked about data hygiene and quality. We do have one specific data partner in our audience marketplace, Webbula, that has created a niche space for themselves by bringing this component to the forefront and has based their data set around it. I recently had lunch with Vince Cersosimo, CEO and Founder of Webbula, and found out more about where they are heading and some tips on how marketers can better improve the quality of their email lists to achieve the ROI they are seeking.

1. You’ve spoken a little bit on your company and how you got your start. Can you go into more detail about the origin of Webbula and where the company is headed today?

In 2009, Webbula was started to solve email sending problems for organizations. While Webbula initially started as an Email Service Provider (ESP), we quickly realized there is an even bigger problem for organizations sending email—fraudulent email data affecting brand reputation and deliverability. Because data quality is significant in the email space, Webbula developed a proprietary data technology called CloudHygiene to address fraudulent and toxic email data. Once our clients realized the results of our CloudHygiene technology, we became known in the industry for our email quality. We then considered ourselves as more of a technology company and less of an email marketing company and expanded our technology to solve the data quality problems surrounding other portions of the identity layer and demographics. Our clients grew to love our data quality not only for email, but also for data acquisition and enhancement.

As we started aggregating and compiling more of our own data, we realized our offline data enhancement and cleansing technology could be applied to the online ad targeting marketplace. Today, Webbula is able to make our quality audience data available to online advertisers by working with companies such as MediaMath and other DMPs, DSPs and data on-boarders that have build a robust ecosystem over these past few years. Now our clients and partners consider Webbula as the industry-leading data quality provider in both the email and online advertising industries, and we continue to offer the scale our customers need to launch relevant campaigns across channels.

2. What are three tips you have for marketers to improve the quality of their email lists?

A lot of marketers come to us and say, “I want to clean my list and remove all the bad email addresses,” and while that is important, our strategy is to also identify the best, high-quality email addresses for our customers. First, it’s vital to clean your list to identify active threats. While verification checks for bounces and offers some intelligence about your list, it does not address deliverable email addresses like spam traps, honeypots and moles which are the most toxic and harmful to your brand and campaign. Only email hygiene can identify deliverable email threats and many marketers make the mistake of thinking that checking for bounces is enough. When hygiene and verification are used together, marketers receive comprehensive intelligence on the quality and health of their email list and that’s exactly what we do.

Next, it is important to figure out how to reach your targeted email audience on mobile devices. This can easily be done by identifying email addresses associated with social media accounts. Why? Because email addresses associated with social media activity are likely connected to a mobile device. Social media authentication allows marketers to identify email addresses with active social media accounts so they can achieve email marketing gold.

Lastly, we always encourage marketers to learn more about their customer list and increase the insights they have on individuals – their lifestyle and interests, political leanings, online behaviors, demographics and more. By appending additional data to their customer list, marketers are able to craft precise, relevant email campaigns that drive opens, clicks and revenues.

3. So you have a cleaner list—what are the next steps to ensuring you are actually getting ROI?

While opens and clicks are important metrics to drive to ensure campaign ROI, they aren’t the best engagement metrics to measure in your email marketing campaigns. This is because spam traps can open your email messages and seeded trackers can click on them. To ensure the engagement metrics you are tracking are not fraudulent, it is essential to clean your list regularly to avoid low-quality emails and focus on high-quality emails that are most likely to engage and convert. We encourage our customers to set a baseline in their list cleansing efforts, and measure results according to that baseline overtime. Aside from cleaning their list, we also encourage marketers to grow their list through healthy list growth practices like data enhancement and social media authentication.

4. How does email hygiene affect the larger marketing ecosystem inclusive of both martech and adtech?

Webbula views the email address as the universal key identifier. Why? Everyone has an email address and uses it to log in or register for just about everything. The email has become just as important as an individual’s name and the hashed email address acts as a secure passport for identifying everything about an individual. However, dirty or fraudulent email data offers little insight and understanding about an actual individual. By using clean email data that has been filtered through rigorous email hygiene filters, marketers find confidence in their data sets and can apply these identification insights to their campaigns to better target individuals across devices.

Additionally, Webbula brings this data hygiene process to the audience data space and provides our DMP and DSP partners like MediaMath with data that has been cleaned and, in turn, improves performance for their advertising clients. While Webbula’s audience data is at scale to run large campaigns, we believe in quality over quantity and work to offer the marketplace only the best quality data.

5. How do you see email and data quality further influencing advertising in the next five years?

Today, we are already fueling identity graphs with hashed emails, and we feel identity graphs will continue to grow and create a more comprehensive picture of individuals in the future. Our customers and partners choose Webbula because we are known for our data quality, which is what we see as the main focus for the future of advertising in the next five years. Today, there is so much fraudulent, dirty data in the marketing and advertising ecosystem, and as marketers continue to identify their audiences, the demand is going to be for high-quality data that actually works. We see a lot of data providers advertising their large scale, and many marketers are initially attracted to larger audiences. However, when marketers prioritize quantity over quality, they often find their audiences are inaccurate and missing significant pieces of data, which makes their campaigns less significant or relevant to targets. Webbula’s non-modeled data for advertising is reported directly from the individual, and we apply our proprietary email and data hygiene technology to allow marketers to filter and narrow their targeting to high-quality, highly-relevant audiences.

Data Uncategorized

5 Questions with Datonics

August 24, 2016 — by MediaMath

Search data is considered especially valuable in audience targeting. If a user is actively searching for keywords, their intent to find information on these topics can be considered very strong. I recently had lunch with Michael Benedek, CEO and President of Datonics, to learn more about this niche targeting and understand their expansion into mobile audiences and coverage in Canada. We had previously worked together when I was at Interclick and he was at AlmondNet to really grow that partnership, so it was great to catch up again a few years later and see the changes that have since happened in the industry.

You offer search, behavioral purchase intent and life-stage data—what’s the importance of all of these data types to a marketer’s toolbox?

We want to be versatile in our offerings as there is often more than one method of reaching the same audience. For example, sometimes you want to target them by their behavioral or life stage profile, other times you may want to target them based on search behavior. Our segments come from the searches that a user conducts when they are online. We use search because it is a highly reliable indicator of user interest in a specific product.

What has been your strategy in how you have introduced new segments to the market?

It’s crucial for us to maintain our identity as a provider of keyword-driven segments as well as adjust to the demands of our clients. When demand for B2B data accelerated, we saw a rise in usage of our business-related segments, so we added new business audiences to enhance this vertical. In addition to keeping up with recent trends (health foods, mixed martial arts, Pokemon Go, etc.), we also saw that demo data was a staple in most advertisers’ campaigns and decided it would benefit our partners to add this important targeting attribute so that they could receive one-stop-shop fulfillment of their data needs.

You expanded your mobile data coverage in the spring. What are the still untapped opportunities with mobile data?

One of the bigger challenges for brands is understanding who is arriving on their mobile web pages. Since the Safari browser is cookie-less and more than 60 percent of US and Canadian mobile web traffic arrives through Safari, advertisers are essentially blind to these customers who have visited their digital storefront. We have a unique solution that helps reveal these visitors in a privacy-sensitive manner and we expect this to be a big shifter of industry dollars in 2017.

You currently offer data on 180+ million US and Canadian consumers, the latter of which is a rarer market for this type of data. Why have you focused your efforts there?

The decision to focus on Canada was a result of our publisher relationships. It’s easy to start chasing a new butterfly every time an advertiser asks about a region in Europe, Asia or Latin America, but we understood that our strength was here in North America, so we honed in on working with the publishers here. It is also a funny coincidence that the CEO of MediaMath is Canadian and so is the CEO of Datonics. We’re looking forward to growing our partnership together in Canada.

Can you tell us more about the Datonics Data Lab?

Our Datonics Data Lab is tasked with the mission of analyzing, classifying and optimizing the collection, categorization and distribution of incoming and outgoing non personally-identifiable information that is collected by Datonics from data provider partners and distributed onward to platform partners and their marketers. In addition to 500+ prepackaged segments, Datonics uniquely offers partners an unlimited number of custom segments powered by a keyword-driven taxonomy, allowing platform partners and their marketers to reach the exact audience of consumers they currently buy on search engines, wherever that audience goes online.

Analytics Data DIGITAL MARKETING Intelligence Media Technology Uncategorized

Machine Learning Without Tears, Part Two: Generalization

August 22, 2016 — by MediaMath

In the first post of our non-technical ML intro series we discussed some general characteristics of ML tasks. In this post we take a first baby step towards understanding how learning algorithms work. We’ll continue the dialog between an ML expert and an ML-curious person.

Ok I see that an ML program can improve its performance at some task after being trained on a sufficiently large amount of data, without explicit instructions given by a human. This sounds like magic! How does it work?

Let’s start with an extremely simple example. Say you’re running an ad campaign for a certain type of running shoe on the NYTimes web-site. Every time a user visits the web-site, an ad-serving opportunity arises, and given the features of the ad-opportunity (such as time, user-demographics, location, browser-type, etc) you want to be able to predict the chance of the user clicking on the ad. You have access to training examples: the last 3 weeks of historical logs of features of ads served, and whether or not there was a click. Can you think of a way to write code to predict the click-rate using this training data?

Let me see, I would write a program that looks at the trailing 3 weeks of historical logs, and if N is the total of ad exposures, and k is the number of those that resulted in clicks, then for any ad-opportunity it would predict a click probability of k/N.

Great, and this would be an ML program! The program ingests historical data, and given any ad-serving opportunity, it outputs a click probability. If the historical (over the trailing 3-weeks) fraction of clicked ads changes over time, your program would change its prediction as well, so it’s adapting to changes in the data.

Wow, that’s it? What’s all the fuss about Machine Learning then?

Well this would be a very rudimentary learning algorithm at best: it would be accurate in aggregate over the whole population of ad-exposures. What if you want to improve the accuracy of your predictions for individual ad-opportunities?

Why would I want to do that?

Well if your goal is to show ads that are likely to elicit clicks, and you want to figure out how much you want to pay for showing an ad, the most important thing to predict is the click probability (or CTR, the click-through-rate) for each specific ad opportunity: you’ll want to pay more for higher CTR opportunities, and less for lower CTR opps.

Say you’re running your ad campaign in two cities: San Francisco and Minneapolis, with an equal number of exposures in each city. Suppose you found that overall, 3% of your ads result in clicks, and this is what you predict as the click-probability for any ad opportunity. However when you look more closely at the historical data, you realize that all ad-opportunities are not the same: You notice an interesting pattern, i.e. 5% of the ads shown to users in San Francisco are clicked, compared to only 1% of ads shown to users logging in from Minneapolis. Since there are an equal number of ads shown in the two cities, you’re observing an average click-rate of 3% overall, and …

Oh ok, I know how to fix my program! I will put in a simple rule: if the ad opportunity is from San Francisco, predict 5%, and if it’s from Minneapolis, predict 1%. Sorry to interrupt you, I got excited…

That’s ok… in fact you walked right into a trap I set up for you: you gave a perfect example of an ad-hoc static rule: You’re hard-coding an instruction in your program that leverages a specific pattern you found by manually slicing your data, so this would not be an ML program at all!

So… what’s so bad about such a program?

Several things: (a) this is just one pattern among many possible patterns that could exist in the data, and you just happened to find this one; (b) you discovered this pattern by manually slicing the data, which requires a lot of time, effort and cost; (c) the patterns can change over time, so a hard-coded rule may cease to be accurate at some point. On the other hand, a learning algorithm can find many relevant patterns, automatically, and can adapt over time.

I thought I understood how a learning algorithm works, now I’m back to square one!

You’re pretty close though. Instead of hard-coding a rule based on a specific pattern that you find manually, you write code to slice historical data by all features. Suppose there were just 2 features: city (the name of the city) and IsWeekend (1 if the opportunity is on a weekend, 0 otherwise). Do you see a way to improve your program so that it’s more general and avoids hard-coding a specific rule?

Yes! I can write code to go through all combinations of values of these features in the historical data, and build a lookup table showing for each (city, IsWeekend) pair, what the historical click-through-rate was. Then when the program encounters a new ad-opportunity, it will know which city it’s from, and whether or not it’s a weekend, and so it can lookup the corresponding historical rate in the table, and output that as its prediction.

Great, yes you could do that, but there are a few problems with this solution. What if there were 30 different features? Even if each feature has only 2 possible values, that is already 2^30 possible combinations of values, or more than a billion (and of course, the number of possible values of many of the features, such as cities, web-sites, etc could be a lot more than just two). It would be very time-consuming to group the historical data by these billions of combinations, our look-up table would be huge, and so it would be very slow to even make a prediction. The other problem is this: what happens when an ad opportunity arises from a new city that the campaign had no prior data for? Even if we set aside these two issues, your algorithm’s click-rate predictions would in fact most likely not be very accurate at all.

Why would it not work well?

Your algorithm has essentially memorized the click-rates for all possible feature-combinations in the training data, so it would perform excellently if its performance is evaluated on the training data: the predicted click-rates would exactly match the historical rates. But predicting on new ad opportunities is a different matter; since there are 30 features, each with a multitude of possible values, it is highly likely that these new opportunities will have feature-combinations that were never seen before.

A more subtle point is that even if a feature-combination has occurred before, simply predicting the historical click-rate for that combination might be completely wrong: for example suppose there were just 3 ad-opportunities in the training data which had this feature-combination: (Browser = “safari”, IsWeekend = 1, Gender = “Male”, Age = 32, City = “San Francisco”, ISP = “Verizon”), and the ad was not clicked in all 3 cases. Now if your algorithm encounters a new opportunity with this exact feature-combination, it would predict a 0% click-rate. This would be accurate with respect to the historical data your algorithm was trained on, but if we were to test it on a realistic distribution of ad opportunities, the prediction would almost certainly not be accurate.

What went wrong here? Suppose the true click-rate for ads with the above feature-combination is 1%, then in a historical sample where just 3 such ad-opportunities are seen, it’s statistically very likely that we would see no clicks.

But what could the learning algorithm do to avoid this problem? Surely it cannot do any better given the data it has seen?

Actually it can. By examining the training data, it should be able to realize, for example, that the ISP and Browser features are not relevant to predicting clicks (for this specific campaign), and perhaps it finds that there are a 1000 training examples (i.e. ad-opportunity feature-combinations) that match the above example when ISP and Browser are ignored, and 12 of them had clicks, so it would predict a 1.2% click-rate.

So your algorithm, by memorizing the click-rates from the training data at a very low level of granularity, was “getting lost in the weeds” and was failing to generalize to new data. The ability to generalize is crucial to any useful ML algorithm, and indeed is a hallmark of intelligence, human or otherwise. For example think about how you learn to recognize cats: you don’t memorize how each cat looks and try to determine whether a new animal you encounter is a cat or not by matching it with your memory of a previously-seen cat. Instead, you learn the concept of a “cat”, and are able to generalize your ability to recognize cats beyond those that exactly match the ones you’ve seen.

In the next post we will delve into some ways to design true learning algorithms that generalize well.

Ok, looking forward to that. Today I learned that generalization is fundamental to machine-learning. And I will memorize that!

Data Uncategorized

5 Questions with Cardlytics

August 16, 2016 — by MediaMath

In our quest to provide as many interesting and innovative audience targeting solutions as possible, we have recently focused on transactional data providers, who can provide data based on credit card and bank card purchasing power. We are one of the first platforms to carry Cardlytics syndicated audiences, and also have their measurement solutions as well as custom segment capabilities. I recently had a chance to talk to Dani Cushion, CMO of Cardlytics, to understand better how our data buyers can best use this type of data.

The headline on your website says “Every purchase tells a story.” This seems especially resonant at a time when more and more consumers are blocking ads. How do you help your customers turn these purchase stories into better marketing stories?

We believe that what you’ve purchased in the past is a great indicator of what you’ll buy in the future. And, in a world that is becoming more and more about relevance and personalization, this ability to reach consumers at the right time with the right content is increasingly important. We have insight into $1.5 trillion in consumer spending across 120 million bank accounts, which provides us a unique perspective on how to best reach consumers with relevant content.

For a long time, demographics have served as a way for marketers to segment consumers. These were a nice proxy that got you close to a target, but they weren’t perfect. Demographic-based audiences assume that everyone within a particular age range or gender spends in the same way. But that’s simply not the case.

As an alternative to demographics, at Cardlytics, we live instead by “purchasegraphics”—segmenting that uses past purchases as a predictor of future purchases. These purchase intelligence-based audiences segment people based on what they buy, which helps brands tell more tailored stories and improve their campaign targeting accuracy.

What informed your decision to offer your services outside the banking channel?

We’ve used our purchase intelligence to target and measure campaigns within our native bank advertising channel for the past eight years. Our brand clients saw great success in this channel and wanted the ability to use that same purchase intelligence in campaigns across connected media. Earlier this year, we launched Platform Solutions to do just that, using purchase intelligence to make all marketing more relevant and measurable.

How do you see yourself complementing an advertiser’s first, second- and third-party data?

Our purchase intelligence gets richer as you append other data to it. For example, geolocation data gives you a good picture of foot traffic in your brick and mortar stores, but it doesn’t tell you if consumers actually made a purchase. When you layer our purchase intelligence onto geo data, you get a great picture of consumer behavior around your store. What percentage of people that come into the store make a purchase? Are people coming in and not buying? These insights help brands make smart decisions about things like merchandising and in-store customer experience.

This is just one example, but you can imagine the possibilities when you complement our purchase intelligence with other sources, such as CRM or web browsing data.

Since you launched your Platform Solutions service, what are some surprising insights clients have seen in terms of where their customers shop and how they spend?

One area where clients are often surprised is around loyalty. As marketers, we count on our loyal customers. The ones who keep coming back to express their devotion through frequent purchases. The ones who remain steadfastly immune to competitors’ marketing efforts…or so we think. We’re able to provide clients with a more complete view of consumer spending—we like to call it a “whole-wallet view”—to help brands understand where customers are spending when they’re not spending with the client’s brand. This eliminates a blind spot for brands and gives them a true understanding of who their loyalists are. Often times customers that appear to be brand loyal are actually heavy category spenders in disguise. This new view of loyalty helps brands refine targeting to capture headroom and nurture true loyalty.

In a February AdAge article, you said you were “setting up the pipes” to enable advertisers to purchase audience segments based on Cardlytics data for ad targeting via ad exchanges. We are really pleased you have chosen to partner with MediaMath on making your syndicated segments as “always on” with us. Can you tell us how advertisers can use this targeting capability best?

When we launched Platform Solutions earlier this year, a large part of our strategy was forging partnerships with the right players to make our products easy for advertisers to consume. MediaMath was one of our first choices to make syndicated audiences available. These are prepackaged audiences that help marketers reach high-intent consumers based on actual past purchases. For example, a brand could buy an audience of casual dining customers, frequent travelers or specialty grocery shoppers.

Syndicated Audiences complement Cardlytics’ custom audiences that allow brands to create unique audience segments based on tailored criteria. For example, consumers who spend with a brand’s top three competitors. Or, consumers who spend at least $100 a month in a given category.

These “purchasegraphic” audiences are powerful, helping brands reach people based on what they buy versus who they are demographically. This reduces ad waste, allowing advertisers to get more from every impression.

Data Technology Uncategorized

MediaMath and TruSignal Help ShopStyle Increase Conversions 200%

August 5, 2016 — by MediaMath

Leading online retailer ShopStyle, a division of POPSUGAR Inc., wanted to exceed their digital marketing objectives through increasing ROI on existing direct response and retargeting campaigns. With campaigns already in place on MediaMath’s TerminalOne Marketing OS™, ShopStyle sought to achieve an even higher conversion rate and lower cost per order. ShopStyle called on MediaMath and TruSignal to create and distribute a solution that would drive the ROI improvement they craved. Learn more about this engagement by downloading our full case study here.

Analytics Data Intelligence Technology Uncategorized

Machine Learning: A Guide for the Perplexed, Part One

July 21, 2016 — by MediaMath

With the increasingly vast volumes of data generated by enterprises, relying on static rule-based decision systems is no longer competitive; instead, there is an unprecedented opportunity to optimize decisions, and adapt to changing conditions, by leveraging patterns in real-time and historical data.

The very size of the data however makes it impossible for humans to find these patterns, and this has lead to an explosion of industry interest in the field of Machine Learning, which is the science and practice of designing computer algorithms that, broadly speaking, find patterns in large volumes of data. ML is particularly important in digital marketing: understanding how to leverage vast amounts of data about digital audiences and the media they consume can be the difference between success and failure for the world’s largest brands. MediaMath’s vision is for every addressable interaction between a marketer and a consumer to be driven by ML optimization against all available, relevant data at that moment, to maximize long-term marketer business outcomes.

In this series of blog posts we will present a very basic, non-technical introduction to Machine Learning. In today’s post we start with a definition of ML in the form of a dialog between you and an ML expert. When we say “you”, we have in mind someone who is not an ML expert or practitioner, but someone who has heard about Machine Learning and is curious to know more.

Can we start at the beginning? What is Machine Learning?

Machine learning is the process by which a computer program improves its performance at a certain task with experience, without being given explicit instructions or rules on what to do.

I see, so you’re saying the program is “learning” to improve its performance.

Yes, and this is why ML is a branch of Artificial Intelligence, since learning is one of the fundamental aspects of intelligence.

When you say “with experience,” what do you mean?

As the program gains “practice” with the task, it gets better over time, much like how we humans learn to get better at tasks with experience. For example an ML program can learn to recognize pictures of cats when shown a sufficiently large number of examples of pictures of “cat” and “not cat”. Or an autonomous driving system learns to navigate roads after being trained by a human on a variety of types of roads. Or a Real-Time-Bidding system can learn to predict users’ propensity to convert (i.e. make a purchase) when exposed to an ad, after observing a large number of historical examples of situations (i.e. combinations of user, contextual, geo, time, site attributes) where users converted or not.

You said “without being given explicit instructions.” Can you expand on that a bit?

Yes that is a very important distinction between an ML program and a program with human-coded rules. As you can see from the above examples, an ML system in general needs to respond to a huge variety of possible situations: e.g., respond “cat” when shown a picture of a cat, or turn the steering wheel in the right direction in respond to the visual input of the road, or compute a probability of conversion when given a combination of features of an ad impression. The sheer variety of number of possible input pictures, or road-conditions, or impression-features is enormous. If we did not have an ML algorithm for these tasks we would need to anticipate all possible inputs and program explicit rules that we hope will be appropriate responses to those inputs.

I still don’t understand why it’s hard to write explicit rules for these tasks. Humans are very good at recognizing cats, so why can’t humans write the rules to recognize a cat?

That’s a great question. It’s true that humans excel at learning certain tasks, for example recognizing cats, or recognizing handwriting, or driving a car. But here’s the paradoxical thing — while we are great at these tasks, the process by which we accomplish these tasks cannot be boiled down to a set of rules, even if we’re allowed to write a huge number of rules. So these are examples of tasks where explicit rules are impossible to write.

On the other hand there are tasks at which humans are not even good at: for example trying to predict which types of users in what contexts will convert when exposed to ads. Marketing folks might have intuition about what conditions lead to more conversions, such as “users visiting my site on Sundays when it’s raining are 10% likely to buy my product”. The problem though is that these intuition-guided rules can be wrong, and incomplete (i.e. do not cover all possible scenarios). The only way to come up with the right rules is to pore through millions of examples of users converting or not, and extract patterns from these, which is precisely what an ML system can do. Such pattern extraction is beyond the capabilities of humans, even though they are great at certain other types of pattern extraction (such as visual or auditory).

I see, so ML is useful in tasks where (a) a response is needed on a huge number of possible inputs, and (b) it’s impossible or impractical to hard-code rules that would perform reasonably well on most inputs. Are there examples where the number of possible inputs is huge, but it’s easy to write hard-coded rules?

Sure: I’ll give you a number, can you tell if it’s even or odd? Now you wouldn’t need an ML program for that!

In a future post we will discuss at a conceptual level how ML algorithms actually work.

Newer Posts Older Posts