# The User Data Marketplace

Last updated on December 29, 2014

There is a nascent and potentially large opportunity to create a user data marketplace, whereby users have full control of which and how their personal data is collected and used.

## The Consumer View

Our activity on the Internet leaves behind a trail of behavioural events from which information inference can occur: what we buy, what we visit, what we read, etc. Hundreds of players on the Internet collect these behavioural events in order to apply complex algorithms and convert the raw behavioural events into valuable information. Ad networks, ad exchanges, data management platforms, demand-side platforms, personalization providers, app developers, cell networks, data clearing houses, spy agencies, etc. The derived information is used to optimize what each Internet player is trying to achieve: publishers seek to improve engagement and yield; marketplace operators seek better margins; advertisers look for optimizing the return on investment for their marketing budgets driving awareness and/or conversions; and finally, consumers are attracted by interesting experiences that keep them engaged and solve their tasks.

In order to construct such an event trail, web technology uses cookies, and users can block or remove these cookies from their browser, therefore granting and revoking access and erasing their digital footprint. But users are seldom aware that 3rd parties, with which they may have never interacted, visited and definitely not given permission to, are also collecting their behavioural data through the use of so-called 3rd party cookies. And beyond the browser, app developers collect user data by proprietary means and users have little ability to control what’s being collected nor how it’s being used, or reused, except for some network-centric advertising settings such as IDFA in Apple’s iOS.

It seems that across the board, users do not have enough granular controls to how their data is being used, nor by whom, and usually the data arrangements between the service provider and the consumer are all-or-nothing type of schemes. Given the current state of affairs in regards to user privacy on the Internet, the hypothesis we are formulating is that most consumers would like to:

• Have visibility into what events, facts, dimensions and features have been collected or inferred about me, anywhere on the Internet, i.e. in and by any service provider, across all browsers and applications.
• Have feature-level granular controls to grant, revoke and reset access to data and information about me, whether they are declared, collected or inferred.
• Avoid data and information leakage, i.e. inferred features should only be exploitable to those I have granted access to.

## The Application Developers View

Developers need to be able to collect, retrieve, and extract information out of the user’s declared and behavioural data with two objectives in mind: improving engagement and driving revenue.

In the revenue case, particularly digital advertising, developers will select technologies that make it easy for the players to adopt and therefore maximize likelihood of future revenue. For torso and tail developers the notion of pooling supply to meet the reach and frequency advertisers require is usually achieved by aggregating inventory and packaging in SSPs and ad networks, as only a handful of developers have independently the necessary volume.

In the engagement case, early stage developers want to pool behavioural data with other developers, and the concept of an external pool of user declared or activity data they can access becomes critical since independently they would not be big enough to gain sufficient insights. In fact, the biggest open pool of data today is Facebook, via it’s Share and Like actions scattered across the web. Developers give up behavioural signals in exchange for future traffic in the form of social referrals.

Another concern for developers is the one of not crossing the stream: users reject ads targeted to them via inference from information they disclosed in a non-advertising context. Even large players like Google or Yahoo are cautious not to cross the streams, and when they do, users revolt, e.g. when using mail data to target ads.

The hypothesis that we are formulating is that developers would like to:

• Store and retrieve both declared and behavioural data.
• Monetize, with the users consent and adequate controls, for the data generated on their sites, possibly advertising but also other means.
• Pool their data with other developers to create more engaging experiences, in a model whereby they both control what’s been pooled as well get additional revenue from other developers leveraging their data, hence introducing a data lineage and protection requirement.

Online advertising is a game of predicting response probabilities, whether that is in display, search, contextual or native, and across all pricing model, be it in impressions, clicks or conversions. Advertisers want to maximize the return on their marketing spent, by reaching the right audience, and/or getting the “correct” clicks that lead to conversions. But across the board, the key problem is estimating response probabilities given the historical users’ behaviour. All players (DSP, DMP, SSP, etc.) collect user’s data in order to create response prediction models.

Response prediction probabilities improve with the size of the training dataset. Although subsampling is an adequate technique, especially when all positives are preserved and negatives are subsampled, we know that the estimation bias is smaller when the training dataset is larger, more so in online learning systems. Advertising players therefore have a large incentive to get access to the largest available pool of behavioural data.

Additionally, the latency at which new data gets processed is critical for applications that have high churn, such as advertising. Advertising players have a large incentive to reduce the time it takes them to react to new demand.

The hypothesis is that advertising players would like to:

• Have access to the largest learning capabilities, for the largest possible pool of users and with as much behavioural data as possible.
• Create their own models to predict probabilities, e.g. clicks and/or conversions, to optimize when impressions happen so that campaign objectives are met.
• Have models (re-)trained in seconds, not hours.

The problem of creating a platform is figuring out where to subsidise adoption to achieve critical mass. Once critical mass is reached, network effects make the platform self-sustained (it gets adopted because it works, and it works because it gets adopted). A possible adoption tactic lies in figuring out where the biggest pain is today.

Although privacy is a problem for users, it’s not such a huge problem that will lead them to change online habits, at least statistically speaking. And without users, there are no advertisers (and creating demand is a people-intensive and sales-centric process which does not resonate with a platform business).

That only leaves developers as the initial targets of the platform. To attract developers, these are possible steps:

• Provide iOS and Android application developers user-based storage, similar to what parse.com used to be.
• Provide a minimum schema for event logging, that can then be used to do relatively simple learning tasks.
• Give a closed learning platform with simple binary classification, based on online logistic regression, to predict most popular X, where X can be articles, movies, games, currency, posts, friends, etc.

The next step into the platform is to provide users with granular controls:

• All facts (events and features) are encrypted.
• All facts have lineage to the source (for now).

Finally, the last step is to onboard advertising players:

• Advertising players can train their own online logistic regression and random forest learning models, against the available data. Models are encrypted.
• Models operating on the data are encrypted and can only access and produce the facts for which they signed for (through the equivalent of a model certificate authority).
• All facts carry lineage to players involved in the generation or aggregation of such fact. Such lineage is visible and can be controlled by the user.

There is an alternative adoption tactic, more labour intensive, by on-boarding existing ad tech players that generate events, e.g. mobile SSPs, etc. However, this feels more risky due to the longer sales cycle.

## Pricing Schemes

The key idea is to price facts as events, be it raw events or inferred events, in an open spot market for events. Facts can be published, by developers and advertisers, and consumed, again by developers or advertisers.

For example, an ad network will log a click event, which will get used by a DMP to create a click prediction model, which a DSP will use to bid on a CPC auction an ad exchange. The raw click could be priced at $0.0001, but the final bid on$0.24 max CPC could have been only possible via the intermediaries. The DSP will take perhaps a 10% commission, of $0.02 per click in this case, which might go$0.005 to the DMP, etc. Note however that the data lineage is used for control purposes, by the user and the publisher of the fact. The pricing only reflects the value of the “intelligence” added on top, e.g. a logistic regression model calculating the probability of click being the intelligence added on top of the raw impression and click events.

Consumers may decide to multi-plex events, and re-publish to allow multiple consumers upstream, e.g. a DMP might feed multiple DSPs. The assumption is that the market will be price efficient and information exclusivity will be priced higher than commoditized information, so an equilibrium will be met that maximizes the total information welfare.

NB: what this is saying is that BlueKai has a market incentive to only do business with the largest media buyers, and that by selling data to more players BlueKai is actually eroding their own pricing and total revenue.

## Swarms of Intelligence

An evolution of the marketplace, particularly interesting for users, but also to reduce computational costs, is to leverage device storage and computation power.

Whereas the learning phases require access to as much data as possible, the scoring or inference test) only requires a particular instance or record of the data and the model. This allows for keeping the user data always on-device, perhaps synchronized across devices, and having the trained models be encrypted be shipped to the phone, where the scoring happens. The device would then call the model service provider with the outcome of the scoring. Service providers wanting to access the user activity need their algorithms to be signed by a “user data certificate authority”.

## Genomics

One final extension could be to apply the technology to Genomics, allowing users to control the privacy around their sequencing, granting selective and limited access to service providers in order to avoid data leakage.