Friday, October 12, 2012

What does ‘Emergent Properties in Network Dynamics’ have to do with Shopping?

Dr. PG Madhavan was CTO Software Solutions at Symphony Teleca Corp. Previously, he was the CTO & VP Engineering for Solavei LLC and the Associate Vice President--Technical Advisory for Global Logic Inc.  PG has 20+ years of software products, platforms and framework experience in leadership roles at major corporations such as Microsoft, Lucent, AT&T and Rockwell and startups, Zaplah Corp (Founder and CEO) and Solavei. Application areas include mobile, Cloud, eCommerce, banking, retail, enterprise, consumer devices, M2M, digital ad media, medical devices and social networking in both B2B and B2C market segments.  He is an innovation leader driving invention disclosures and patents (12 issued US patents) with a Ph.D. in Electrical & Computer Engineering. More about PG at

Can you think of a framework that considers everything that humans do? Such an attempt borders on hubris! In any case, is there a “concept” (or concepts) under which we can understand all human-scale activity? Try ‘Emergent Properties in Network Dynamics’ concept on for size! We will explore how this concept holds up by thinking about a practical day-to-day “shopper scenario”.

Facebook connected us in a vast network – this is only a first step. The deep reason for the fascination with social networking can be understood from the shopper example. Shoppers are enmeshed in an ever-changing network of social interactions and preferences; desirable behavior can be made to emerge by perturbing the interactions within the network.

Let us explore the fundamentals . . .

Thinking About Network Dynamics:
We know networks – social networks, neuronal networks, societal networks, family networks and so on. They are characterized by nodes and connections among them. There is another important network characteristic – it’s “dynamics” or changes over time. Time is only one of the possible fabrics over which networks evolve (or explode); another could be “space”.

Consider the time before flight. Communicable diseases in one region of the world were not very likely to spread to far-flung regions in days. After airplanes, the space dimension “shrank”; SARS and AIDs spread much more rapidly among the human network to many parts of the world. So dynamics of the network in some cases have an important spatial dimension and the dynamics changes as new inventions and technologies proliferate. New developments can also render spatial dimension insignificant for certain “infections”. Think of the Internet and the spread of information. Before the Internet, news of Chinese leadership intrigues would have taken a long time to reach US shores. With the Internet, as far as information spread is concerned, space dimension is largely obliterated.

Thinking About Emergent Properties:
Why do we talk about “emergent” instead of “emerging”? The Joint Information Systems Committee in the UK has a convenient definition: “In systems theory emergent properties of the system are those that can’t be explained in a reductionist manner, that is to say, by the actions of the constituent parts of the system, but rather which arise from the interaction from those parts. The basic idea is well captured in the ordinary language phase ‘the whole is more than the sum of the parts’. Emergence can be observed in both the natural and the social world. A common example from nature is colour. Elementary particles have no colour, but when they are combined into atoms, they begin to absorb specific wavelengths of light.”

This definition is a good springboard to our understanding of the Emergent Properties concept. Remember your high school Physics? You know of atoms and nuclei and electrons. You can combine them to form molecules such as H2O or water. Water has some physical properties such as viscosity, surface tension, etc., that are not apparent at the atomic level. Water can go through “state transitions” under certain conditions (such as lowered temperature) to become ice. Ice has an “emergent” property of being a solid, totally unsuspected when it was water. Sometimes, you need external agents such as impurities (or intentional catalysts) to “seed” the crystallization process into ice. Water under certain other conditions (say, heat) can transition to another state such as vapor which has very different physical properties from ice. These state transitions and resulting emergent properties are the stuff of Nobel Prizes. Anderson (1977 Nobel winner) has made important contributions to concept of emergent phenomena (

The website above explores emergent phenomena in philosophy, religion, art and human sciences; there are brief discussions on topics such as emergent properties and processes in physical systems, biological systems, economics, the World Wide Web and the Internet. Cuts a very wide swath indeed!

In summary, an emergent behavior or emergent property can appear when a number of simple entities operate in an environment, forming more complex behaviors as a collective. Though not widely accepted in Complexity theory circles, I also believe that “seeds” are important in precipitating emergent behaviors (lower temperature is a necessary condition for water to freeze but not sufficient; impurities (“seeds”) are required for crystallization to occur). A group of shoppers connected by social networks can exhibit emergent behavior (of increased foot traffic to a store location) if we seed the network with the appropriate spot-discounts spread across portions of the social network.

Putting Emergent Properties & Network Dynamics Together:
Consider our brain – billions of neurons and trillions of synapses. Neuron bodies tend to cluster together into nodes or ganglia. Network connectivity among them is via long axons or short dendrites. In recent years, Neuroimaging scientists have made a living out of locating specific activity or emotion or intelligence at specific brain sites. Such images are produced by massive averaging of activities over subjects, task and time; considering how human being processes information on the go in milliseconds, these experimental conditions have little similarity!

Also, consider the recent revelations about “junk DNA”. “Coding DNA” was where all the action was supposed to be. But lo and behold, we now find that junk DNA dynamics determines when the coding happens and in what sequence and proportion, which is what gives rise to proteins that can create various disease states. Similarly, I believe that it is the dynamics of neuronal network that codes for information storage and processing. In other words, there is no physical locus of a CPU-like intelligence processor or RAM-like memory store. The pattern of neuronal activity (akin to junk DNA dynamics) over the space of the brain and over time determines the information processing that goes on in our brains. So if you look for a brain location to explain a specific human activity, just like considering Coding DNA only, you will get a partial and frustrating picture that is tantalizing but never complete in explaining how thoughts or actions happen!

Network Dynamics picture complicates life considerably. Current medical research tools are totally inadequate to measure specific activity of many millions of neuronal nodes and connections over time and space while an intact human being is performing complex tasks. Therefore, we have to resort to “concepts” instead of testable theories at this time. Let us embrace Network Dynamics as a valid concept for modeling the activity of the brain, society, family, etc.

Within the social, family or brain network and their dynamics, how do behaviors and actions arise? How can we modify them? For example, if we had a societal network dynamics model for infections, can we modify (and eventually stop) the spread of viruses such as AIDs or flu? How do pandemic patterns emerge in these networks and can we predict them so that we can position first responders at the right place ahead of time?

Putting Emergent Properties & Network Dynamics To Work:
Let us consider a down-to-earth business problem and see how we can apply the concept of Emergent Properties in Network Dynamics. We are all shoppers at one time or the other (mostly weekly for household items) – let us see how we can make shopping delightful and profitable for both the shopper and the retail shop using the concept of Emergent Properties in Network Dynamics.

Retailer wants two issues addressed: (1) how to attract shoppers and (2) how to encourage a specific shopper action such as purchasing an item. To attract shoppers in general, retailer can make the aisles wider, décor attractive, etc. To encourage me, a specific shopper, to make a purchase the retailer can recommend specific items to me for purchase (a la Amazon Recommendations). Both involve Network Dynamics since in one case, all shoppers are “connected” in some fashion or the other (by the simple fact that they shop at this store) and in the individual shopper case, I am networked to bunch of my friends and relatives whose shopping preferences affect mine. Utilizing our new-found concept of Emergent Properties in Network Dynamics, the retailer’s objective is to (a) represent and understand emergent properties such as collective behaviors and (b) design interventions in the network so that desirable actions emerge and undesirable actions are suppressed (applicable to both the individual shopper as well as to a group of shoppers).

Let us map the concepts of Emergent Properties and Network Dynamics to the shopper scenario. At this point, it will be helpful if you review the Social Network section of my earlier blog, “So-Mo-Clo Framework for CUS”,

If a social network graph of the shoppers at this particular store was created, you will see that some shoppers are well-connected (called “Hubs” denoted by “H” in the figure) but most are “Clusters” and “Cluster Groups” corresponding to our family, work and play groups. This is a one-time or “static” map of the network. Denoting its dynamics is not easy in such a graph.

If we recreated this network graph at some time intervals and considered a collection of them, we may be able to see the dynamics of the overall network graph (in practical terms, this is difficult to visualize since there are many nodes and they may change positions).

A more tractable problem is to track an individual shopper over time (or space) and give up on the much more desirable collective evolution (purely due to the limitations of current techniques). Any individual shopper can be denoted by his “attributes”. Attributes could be physical (such as height and weight), products he buys or attributes of the products he buys (cheap, sweet, fried, etc.). In abstract terms, Shopper exists as a point in an A-dimensional space where A is the number of independent attributes (this is similar to denoting a location in physical space as a point on an X-Y graph but in this case, with many more than 2 dimensions). Over time, if we repeatedly measure this shopper’s attributes (such as likes and dislikes of various products), his location (marked by “X” in the figure) in the A-dimensional graph will move around as his likes and dislikes change. If we connect these points (fuzzy blue path), we will get a “trajectory” in the state space (or attribute space in this case) of this particular shopper denoting dynamics over attributes. We will call this “Shopper Attribute Map”.

(For the initiated, this sets the stage for eigen-decomposition of shopper “attribute space”. Even though, Attributes 1 through A may be independent, they may not be the best set of orthogonal coordinates to represent the shopper – the eigenvectors are always the best (due to ‘compactness’). It is also likely that a set of axes less than A will represent the shopper better. This reduced eigen-representation forms the basis of most of the sophisticated recommendation engines).

So now we have a somewhat satisfactory pair of methods to represent the network and dynamics. How do we bring in Emergent Properties into the picture? Remember that the retailor wants two things: (1) attract shoppers and (2) a specific shopper action.

A trivial method to attract all shoppers is for the retailor to give away a crisp $100 bill to every shopper when they check out – every shopper in the neighborhood and beyond will show up! However, his business situation may not allow such an intervention. Retailor wants to create a state transition using more of a business friendly intervention such that a desirable emergent property appears (such as a step-change increase in foot traffic).

Without interventions or “seeding” to precipitate emergent properties, we are left to simply simulating/ modeling collective shopper behavior. This in itself is useful in that we can observe patterns such as Hubs and Clusters. More interestingly, just as cooling and seeding (with impurities or catalysts) engendered the emergent property of “solid” from water, the retailer can identify the conditions and interventions to attract all shoppers. Understanding of how infections spread in a network helps him in this effort. Linear Infection (or Influence) Model introduced in my blog mentioned earlier (“So-Mo-Clo Framework for CUS”) provides the methodology.

“Viral” marketing is an exact analog of the Infection Model. Using the social networks of the shoppers, the retailer can offer store-credits to a “Cluster Group” if they can reach some purchasing target as a group (such as buy 10 items of a specific brand in 2 days) – very useful in clearing the store shelf of about-to-expire merchandise. Using the “Hub” to spread this viral message, the retailor may be able attract an even larger contingent of shoppers to his store.

Retailer also wants to precipitate a specific buying behavior (treated as an emergent property here) within a specific customer. It is well-known that the most opportune time to influence a shopper and set the stage for conversion to purchase is during the purple-shaded portion (Interest-Desire-Action phase) of the Purchase Funnel in the figure. “Proximity” marketing identifies shopper’s interest via techniques such as Passive Organic Search and proffers immediate spot-discounts via shopper’s mobile phone or other in-shop displays. It may appear that influencing an individual shopper is not a “network” emergent property since an individual is identified and separately targeted for a spot-deal that is tailored to him. The point to note, however, is that for the most effective targeting of a promotion or offer to a particular shopper, we have to consider him as a node in a network that includes his social network which creates latent influences and perhaps also his “Shopper Attribute Map” made up of his likes and dislikes. Merging of these two gives the most complete and powerful network graph of the shopper to guide the Proximity marketing actions.

Lessons Learned:
·         Collection of Shoppers forms a social network.
·         An individual’s social graph and her attribute map constitute her total network.
·         Networks have temporal (and spatial) dynamics which are hard to model; an approximate approach is to recompute them repeatedly.
·         Shopper and her attributes may be modeled more compactly using their eigenvector space (in other words, distance metrics can be ortho-normalized which will improve recommendation systems).
·         Influence/ Infection model can be used to proactively study emergent properties due to interventions allowing fine-tuning of marketing strategies (Viral or Proximity) before actual deployment.

What does Emergent Properties tell us? If we want specific shopping behaviors, we need to remember that the shopper is embedded in a social graph and an attribute map. Viral and Proximity marketing are but primitive techniques to perturb the network. Embedding produces many influences and constraints on the shopper that may be varying.

For predictable success, new marketing techniques will emerge designed to create desirable emergent behaviors (beyond using one-to-one causal relationships) in a cohort of shoppers in their network-embedded state; let us call it “Emergent Marketing” as opposed to the traditional “resultant” marketing that is cause-and-effect aimed at a stand-alone individual.

Some general comments:
·         Hubs in the networks should be afforded “concierge-level” customer support because of their wide influence.
·         Products can also have “social graphs”; they are nodes of a network that are connected by links such as manufacturer, low-cal, green or organic, salty, and many more. Such a Product Social Graph will find applications in merchandising which will be of interest to CPG manufacturers and Retailors.
·         New frontiers of R&D lie in merging People and Product social graphs, improving the modeling of dynamics in social graphs and better perturbation techniques (“Emergent Marketing”) to create purposeful emergent properties in the network and its dynamics.

Within the current state of the art, we are limited in fully describing the network dynamics of shoppers. We have to rely on the shopper social graph (and the relationships along with decoding messages being passed among the nodes) and an individual’s Shopper Attribute Map. Given the graph and map as a-priori information and the Influence Model, we can design interventions via viral and proximity marketing to create certain desirable emergent buying behaviors of the shopper as an individual or shoppers as a collective.

1 comment:

  1. Hi, thanks for sharing, PG!
    I will get back with comments soon.