When it comes to information, connections have meaning. The value of the connections can be made explicit, as in ontology; but even implicitly, meaning exists, including in how it’s built.
The primary benefit of hierarchies is that they chunk information into digestible segments that are less overwhelming. It’s a wonderful quality.
But that simplicity comes at a cost: information density. Even when a hierarchical structure has so much information that it can be overwhelming, as in vast information stores like medicine or science, no hierarchy is going to include all the potential connectome — the pattern of the information connections.
Take this diagram for instance. You can mentally invest this with whatever information categories you want to help you understand: clothing, pet supplies, science, elementary school subjects. I’m going to use kitchen cupboards as my example going forward.
Chances are pretty good that you found it really easy to fill in and even expand. Part of this is because our culture uses hierarchical structures everywhere. It’s simple to build, helps people navigate and orient, and because it reads unidirectional it increases the likelihood of everyone staying on the same page during conversations. It’s also nice when you are thinking on your own: it chunks information. The chunking keeps the information to a finite, easily digestible set, with just enough categorical context to heighten meaning.
The dirty little secret (ok, it’s not dirty, but complex; it’s not little, but universal; and it’s not secret, just ignored) is that information is intrinsically a network. All data has multiple, undefined quantity of connections to an uncapped quantity of other data, until they are architected for a particular use.
That’s why when you go to one person’s kitchen cupboards, they might put food/dishes in chunks; all equipment and dishes in one half, all food in the other half (nearest the door). Another person might choose to put things away broadly by size first; the lobster pot and 20-pound bag of flour in a bottom cupboard (safe lifting). Another kitchen, shared amongst multiple people, might first divvy space up by exclusive cupboard use per person (community peace). But in none of these scenarios, in none of these minds, does food stop being food if it’s put away outside of those chunks. It’s less findable. Navigating the kitchen as an outsider, without understanding implicit information connections makes it harder to troubleshoot instances like Pat having been in a hurry and putting the rice bin in with the pots the last time they used it. All the information is still coherent, there’s just gaps to the connections and data incorporated.
So instead of simply labeling that top box with a mental “Pat’s kitchen”, let’s pop in what it’s really drawing from: a network. Let’s see the start of what happens as we form a hierarchical structure out of networked information, assuming we’re approaching it from a top-down perspective.
Five nodes were chosen (black). They, in more robustly designed structures, will have been well researched and gone through testing, with bets understood so they can be more easily pivoted from. Or they might have used the same organization they’ve used since their great-grandmother passed down “the way it’s done”. Or somewhere in between, such as using the best practice suggestions in the latest magazine or blog post, or finding a happy medium with others living in the same space.
- This is the first major decision with potential for cognitive bias.
- Value is applied: these nodes have certain prioritized characteristics.
- Taxonomy starts coming into play by naming the categories, implicitly or explicitly.
- Implicit in the categories is an underlying formulation of “what makes sense”, e.g., putting food closest to the door for easy unpacking, size/safety, community peace, a sense of authority.
Then, choices are made about what information is “relevant” from the network, taking into consideration the context of the selected parent nodes. To help you see this, the parent nodes were numbered and the selected data points were highlighted. Actually, I also faded the unselected data points, just like they are in our brains when we are considering hierarchies. If we’re building from the top down, we stop considering any data that doesn’t match our categories. This even affects the reading of hierarchical structures built from the bottom up, but that’s writing for another day.
- This is the second major decision point with potential for cognitive biases.
- Exclusions are made; in Stacey’s kitchen, batteries don’t exist; in Pat’s they do.
Then we make choices about how to chunk up the information so the interior categories of each parent node makes sense. But it gets even more fuzzy about what goes with what, how they might be labeled into categories, and when it makes sense to form levels. Even in a kitchen where the primary decision is “size/safety”, some small, light things might go low — like a saucepan that seems more in line with the lobster pot than wine glasses; this is implicit inclusion based on unstated metadata.
- This is the third major decision point with potential for cognitive biases.
- The decisions about levels tend to be filled with unstated values. There are implicit standards of importance and priority baked in as levels are segmented.
- Taxonomy starts really shining; choose derogatory terms, and it doesn’t really live at the designated level in terms of associated respect. Batteries living in a junk drawer become more “junk” than “L2”. The stated, findable label takes precedence over the abstract hierarchical formulation.
- Implicit metadata becomes a consideration, like whether the prioritized defining factor of the saucepan is size, use, material, etc.
Still staying with the same diagram and happening simultaneously with the levels and taxonomy decisions, we make choices about what of the understood connectome are going to be included. Networks are omnidirectional and don’t cap how many in or out connections can exist. Hierarchies only allow for a unidirectional movement of information, with one incoming option for each data point. To move from multiple to one, connections have to be selected, and the rest given up.
- This is the forth major decision point with potential for cognitive biases.
- The decisions about which connections to maintain are also filled with values, implicitly deciding which incoming node is more meaningful in this formulation to how the information is being categorized.
So, to create a hierarchical structure out of a network, we have:
- Multiple opportunities to apply/miss cognitive biases,
- Exclusion of both data and connections,
- Values applied by implication only, making them harder to communicate about and change,
- Taxonomy choices to sway assessment,
- Priorities implicitly impressed,
- Unspoken metadata swaying where to include items and implying prioritized meaningfulness.
Let’s be clear: this abstract representation is a simple network. If I were to run into this network, I’d consider it low resolution. Irrespective of inclusion, there are 59 nodes and 62 connections, with an average of 1.05 connections per node. However, the excluded information (25 nodes, 32 connections), are a comparably whopping 1.28 — yet still far from the possibilities inherent in a network. It’s enough to design a hierarchy, but we’re focusing on an overall simpler set of information. If I were working with this kind of network, I’d communicate it was temporary, as we consider it during construction. But once constructed, it can become entrenched by users and lose the temporary quality we assumed as we cobbled it together.
Let’s break down what happens what’s missing after this simple network has been architected into this particular hierarchical structure. So, in this abstracted version of the network, first let’s focus on the explicit decisions.
Twenty-five nodes were left out, or about 44%.
Thirty-two connections were ignored, or about 52%. There’s a higher instance of missing connections by quantity and percent compared to nodes because it’s a network. Connections between data nodes can be added to infinitely in a network.
Now forget all the decisions made, and look again at the network-to-hierarchy without all the visual orientation help. The selected parent nodes are hard to see in the network without definition (and yes, neither network nor hierarchy are on the page in the same orientation/order). No matter how much effort went into aligning them, juxtaposed to the complexity of the network, they are relatively arbitrary. The arbitrariness will become clearer as more information is discovered and connected. That’s why kitchens can be organized so many ways and still make sense and be navigable. If and where the network is maintained and understood, it makes it easier to find a place for a fancy new gadget or newly necessitated gluten-free foods while maintaining essential findability. In other words, it’s prudent to understand the fullness of the organization to better understand how and where to pivot.
Connectome is the real heart of information structure. It’s what helps us form context and meaning. It helps us balance how a data point weighs against interpretation. When it comes to something as literally life-or-death as medical information, the connectome can be the difference between understanding what’s happening and being a medical mystery. It’s the fungibility of the included connectome, translated from a network structure to a more friendly hierarchical structure, that we really need to wrap our heads around. So let’s highlight some of our negative spaces.
Thirteen connections to nodes are ghosted from immediate connection, e.g., not including connections/nodes after the first drop. These are those liminal calls, like deciding whether batteries belong in the kitchen.
Of those, 5 of the connections are to another of the selected “parent” nodes, either directly or with one step. That’s like deciding what quality/metadata is the defining factor for a small saucepan.
There are 2 other, deeper pathways comprised of 8 connections total, that would have connected selected “parent” nodes. Think of these like foods that you wouldn’t consider buying: gluten free pretzels, yogurt starter, boxed wine, $50/ounce ham. They exist, they are purchased by someone, but they never enter this particular kitchen so they aren’t included in the information structure. But if it was suddenly to enter the kitchen, they are still rather liminal, able to be put in a couple different places and be findable.
Connectome was collapsed once to force inclusion within the designed structure. It usually becomes an aesthetics decision — like pots literally needed a little more space, regardless of the primary organizational decision. Or if we consider navigation, keeping the click count down.
One node is included in two categories, with no explicit connection in the hierarchical structure. Think knives; many people keep the ones they use frequently in one spot, and backups/special-use elsewhere.
So, why think about this?
Human understanding expands. New data comes to light. They don’t always make it into our hierarchical constructs, and in fact the older and more used the constructs are, the harder it is to consider changing them.
Whenever solutions seem impossible to find, yet teasingly out of reach, look at the structure of the information. Is it in a hierarchy? A process? A matrix? If yes, thank goodness. All of these exclude connectome and data. Data is presented as less rich and complex than it is. Especially when it feels “teasingly just beyond reach”, it probably is. That would mean the information to rebalance the whole already exists: you might not have to go looking for a known unknown, or hope to stumble on a true unknown. It’s more a matter of tracing what connectome can be traced, backing into the structures used and then expanding it to include more information than what is surfaced.
By finding a different way into the data (big picture, data sea, different perspective, different lens) we can triangulate from multiple viewpoints. We can think about the cognitive biases, values, priorities, and unstated assumptions in play. All of these things help to move the information, connect it in a different way with different dimensions, potentially even with a much fuller network than the original structure considered.
People and information are inextricably bound. Any information that is architected for dissemination and use is potentially imprinted by the people architecting it, the people they turned to for information, the people designing from it, and the people using it. When it’s contained in information technology, those people have an even bigger impact. Data is much more stable in databases — it’s the primary benefit of capturing it, we are no longer dependent on the fungibility of memory. But data is so much more stable in databases that it can be hard to construct new connectome; it can become rigid and brittle.
Hierarchical information structures are extremely useful. They are a known and effective pattern in formal learning, and they speedily get a mind to information it contains. It’s not holistic, and close-enough understanding in a hierarchical architecture could still be leaps away from a more precise understanding based on network architecture. That is all the difference necessary to increase misunderstanding or form solutions that break more than they fix.
Our use of hierarchical structures is extensive. It is the simplest way to architect information with a structure. Keeping people on the same page is easier to do in a simpler structure, and simpler structures are easier to navigate and learn.
Simpler structures do not contain everything. They are selected windows aimed at a particular, sometimes unstated, goal — like not wanting to lug groceries in any further than absolutely necessary.