# The Datastream Marketplace

## A Reverse Auction Mechanism for Algorithms

Last updated on December 29, 2014

## Decision Automation

We are currently seeing how data processing needs are exploding. Consumers are leaving a growing crumb trail of events as they use the Internet. Sensor-data telemetry is growing into the trillions of machine-generated events per day. DNA sequencing is generating terascale genome databases, per human being. On top of generating and collecting more data, the expectations as of how the data is used are also growing.

Social data platforms, such as Foursquare or Twitter, have educated consumers about seamless data manipulation, in exploratory, unstructured and close to real-time ways. Businesses owners now expect similar flexibility in the analysis of inventory, catalogue and procurement data, joining their proprietary data with open data sets. The objective is to make better and faster business decisions. Just like businesses, individuals are also using data mashup capabilities to make better informed decisions when buying a house, selecting utility providers, changing insurance, etc.

Advertising exchanges, stock exchanges, energy exchanges, etc. see billions of daily automated trades, which computers make by processing large amounts of data using advanced mathematic models, e.g. segmentation, targeting, pricing, risk, sizing, etc. But outside these exchanes, only a few data technology powerhouses have today the infrastructure, skills, distribution and algorithmic bench to implement analytics, prediction and forecasting at terascale.

The basic technology infrastructure required to compute and solve big data problems, especially when real-time answers are needed, is very nascent. A few successful big data technology infrastructure pioneers like Splunk are opening the way. But we are still very early in the cycle and the skills necessary to solve big data problems are hard to find. Because of the lack of viable and easy to use technology solutions, research in scaling machine learning to big data problems has been very limited.

So far, we have seen big data solutions put most emphasis on analytics, i.e. answering questions for humans to make decisions. Little effort has been spent in using big data to dynamically change business processes through predictions and decision automation.

We need to create a business intelligence decision automation platform where the next Industrial Internet Revolution1 can happen. In this platform, machine- and human-generated data will be captured, collected and distributed at terascale, and immediately processed for analysis, estimation, prediction, forecasting and anomaly detection using highly scalable A.I. techniques. This will enable a revolution around decision automation, across agriculture, industry and services, and a new data marketplace pivoted on the intelligence / decisions inferred out of data.

## Decision Platform Capabilities

The Business Intelligence Decision Automation Platform (BIDAP) is two things: a technology platform and a marketplace. Technically, BIDAP provides a distribution and computation framework; distributed online learning; and provenance and leakage control. In addition, BIDAP is also a big data marketplace where application developers sell (publish) or buy (consume) events from datastreams in an auction. The bid is truthful to the value derived out of the event.

Out of the components of BIDAP, the marketplace is the business keystone. In other words, this is not a venture to create and sell a technology platform offering distributed learning on datastreams. Although the technology is a large part of the innovation required, we are not selling technology. In fact, the expectation is to build as little core infrastructure as possible, and to reuse existing technology and distributed machine learning frameworks as much as possible.

### Datastreams on the Cloud

We are at the right point on time when this platform can be created. Pretty much up to now datasets were physically split and stored on-premise. Data owners wanted to keep their proprietary assets physically controlled. But cloud infrastructure and Hadoop have opened the way for startups to keep all their data on the cloud. Most importantly, sensor-devices are completely shifting data to the cloud and away from on-premise. Many datasets now just flow continuously into Amazon Web Services (albeit, they are still not joined, but at least they are “physically close”).

Marketplaces around data are not new. Several types of data marketplaces exist already:

• Static dataset licensing, e.g. Infochimps, twitter’s firehose, Bloomberg, etc. These are not efficient, since there is limited transparency and almost no liquidity as one must buy a time-bound license to the datastream up-front.
• Data Management Platforms (DMP) in online advertising, e.g. BlueKai, and data clearing houses, e.g. Acxiom. Both are highly succesful businesses that operate on batch windows, providing qualified audience segment scoring lists. Their revenues depend on very difficult attribution calculations based on how well those audience segments perform compared to others not using the data. To solve the complexity of revenue attribution, these companies operate by taking a revenue share on the data consuming business.
• Ad exchanges, which operate on a per impression basis. Search, contextual and display ad exchanges run an auction, where the opportunity is the impression. They are all highly efficient and profitable.

An efficient data marketplace must allow spot trades at the lowest possible granularity, as this increases the number of trading opportunities (just like in ad exchanges). In a datastreams marketplace, application developers buy the ability to consume and compute individual events datastreams.

Trading on events in a datastream is not the only possible model. Microsoft is trying to create a marketplace where databases on Azure’s SQLServer cloud are the trading instrument. The database as trading instrument has less liquidity potential than the datastream, and therefore is less likely to become efficient in the long run.

## e-Sourcing Continuous Decisions

Deploy a fairly standard technology stack in a number of nodes, on-premise, that can consume the sensor-device datastream that is proprietary to the customer, and host custom logic to process events and produce “intelligence”. The idea is to then offer the events and a problem (as a component in a supply chain) to external teams. Each external team would use their algorithmic bench to infer a solution to the problem. The producer of the datastream feeds events into the multiple teams competing, they produce some intelligence and that intelligence gets fed consumed back by the producer of the datastream. In reality, it could be that the role of the producer and consumer are clearly defined and separate. What matters most is that the problem is clearly defined, with specified input and outputs.

Imagine then two customers that have deployed this platform on-premise, and that there is an opportunity to join their data to provide value-add. The data could flow direct between customers, or through a clearing house cluster on the cloud.

Each solution to the problem will have a different “quality score”. We want the payout to be conditioned by the quality score and the bid that the intelligence team is willing to sell for. This is an auction with multiple sellers and one buyer (reverse auction or procurement auction). Each intelligence offer gets a rank $$r = bid / quality$$, and the offer with the lowest rank wins. The buyer can set a maximum reserve price (auction ceiling). We will have $$r_1 = b_1 / q_1$$ and $$r_2 = b_2 / q_2$$, and $$r_1 < r_2$$, in an auction where bidder 1 wins. Then the payout should be such that the seller collects as much as possible:

$p / q_1 \le b_2/q_2$

which means that the final payout is:

$p = b_2 \cdot \frac{q_1}{q_2} - \epsilon$

where $$\epsilon$$ is the smallest unit of currency allowed in the trade, e.g. USD 0.01 or USD 1.00 or USD 1,000, …

The issue around pricing is then determining quality in a transparent and truthful way. One possibility is for the marketplace operator being the one determining the quality, but that makes it harder to sell the solution. The problem is that the quality is measured after taking the decision based on the intelligence provided, comparing the observed outcome with the predicted outcome.

The question then is who should take the risk for the estimation bias. The operator? The buyer? The seller?

Imagine in MB that multiple labs teams would compete, per impression. They would each be providing an ad with a predicted CTR. The team with the higest predicted CTR would win the auction and get impression. If the impression translates into a click, all parties are happy. If the actual CTR is lower than the estimated CTR (over-estimation), then the seller lab team got too many (unfair) opportunities, and the buyer incurs a loss. If the actual CTR is higher than the estimated CTR (under-estimation), then the seller lab team sold too low.

1. The Industrial Internet is a term coined by General Electric, somewhat close to what Kevin Ashton called the Internet of Things. I prefer to use GE’s term, as it speaks more to the intelligence and the revolutionary decision automation implications that machine learning and big data will have on the industry, unlike IoT which talks more about connecting physical things to the Internet and the resulting sensor-data generation, but speaks less to the decision automation factors. In fact, one could argue that there is also an Internet of Humans, an Agricultural Internet, … so perhaps we are missing a new and more encompassing term.