Building a conversation platform — Part 11

Why graph?

Published in

The Graph

6 min readSep 5, 2016

In the previous part I proposed another five additions to our graph model: tags, products, places, revisions or versions, and impressions as nodes. We saw how this augmented model created an interesting network of people, places, products and topics, interconnected by stories and discussions.

But we never really discussed what is the actual benefit of a graph approach, in general — why such a network is interesting. After all, a graph is the most generic model, that can be. Everything is a graph, that consists of things that are related to each other. And almost anything can be described in terms of interconnected constituents. So a graph model of conversations — or any content in general — is hardly revelatory. What does it really give us to think about something as a graph?

One answer to this question is more intuition. Using a graph as a mental model of online conversations makes it more intuitive to find patterns of how content connects people together, or how content connects people to tags or products. Obviously any possible model lets us compute what topics a particular user tends to comment, or what other users a particular user tends to converse with, and what topics they tend to discuss, etc. But a graph approach makes things like the notion of distance between users so much more intuitive, more visual, more natural.

Distance of two users, in the sense that how many next arrows separate blocks authored by them, on average, is actually a simple but prominent example of this, as a generally useful metric. Of course, average may not necessarily be the relevant or meaningful measure, it just rather stands for any possible function, that takes the distribution of block distances to a single number. Such a measure can be used both to identify clusters of users — users who tend to converse together — , and both to actually guide the reader when navigating the graph of content.

So far, we have mentioned only one way of how to prioritize and highlight routes for a reader, in Part 8, when we described a method of attaching metadata to audience arrows, that encode a preferred route in the graph. But that was a very specific feature, while filtering and prioritizing options for readers at every block is a crucial point, that can make the difference between a pleasant user experience of systematic exploration of knowledge and engaging opinions, or a baffling labyrinth of noise. Using distance in the graph can provide a more generic approach to this.

So in addition to that idea of sharing routes by encoding them in audience arrows, we can also take the distance of users, and what content connects them and how, and use this data to indicate directions to readers at every junction. And by indication I really mean sign posts like “this way are people you tend to discuss cars with”, “read forward this way to read from people you always discuss science with” etc. Factoring in tags, and distance from tags, or even places, products, groups etc., can truly enrich the repertoire of guiding readers through.

Figure 1: routes highlighted for a particular reader at Block 3

This is nothing more than applying the same idea of how recommender systems and graphs go hand in hand, to our problem too. There’s vast literature on this topic out there. We argue that for the same reasons why representing the data of a platform as a graph is the natural choice for a recommender system, it is also the most useful way of modeling such a platform in general. It makes the description of all the elements of the platform and their relations, much easier to reason about. And distance was just one example of this.

Of course, the actual model we have built throughout this series of posts, may feel ad-hoc here and there. The choice of terminology, or the direction of arrows, or many other details for that matter, might come off as being arbitrary in some cases. One might even argue that our model should not even be a directed graph, as the direction of arrows — probably with the single exception of next arrows — does not add any extra meaning to the model, that isn’t already there. The fact that most arrow types could be flipped without changing the meaning of the model, is a clue indeed. And without arrow directions, how is an undirected graph as a model any different from a relation model? — one might ask rightly.

But the intention behind this series was more to introduce an approach, and less to iron out every detail of building a model of online conversations and collaborative browsing, reading and storytelling. I believe that the concepts and relationships that we described along the way, are all relevant and key to these applications, no matter how we name them, and whether we model them as arrows pointing in a particular direction, or rather nodes, or whatever else.

The approach we have introduced, was to take inventory of common features and applications, and then generalize and unify these seemingly unrelated concepts, like chatrooms and blogs, bookmarks and ratings etc. And thus we arrived at a graph model. This is another answer to our original question about the benefits of thinking of online conversations and commenting as a graph: the graph appears to be the common denominator.

But is it really meaningful and useful to make all these generalizations, to unify these concepts, and blur these distinctions? Is it really practical to say that a blog and a chatroom are fundamentally the same thing, instead of actually distinguishing them, as their mechanics and rules seem pretty different?

There is a principle, that the weakest possible language should always be used to describe a model. And I use the term description in the sense of building it up from building blocks, like sentences from words. What this principle is really about, is that the building blocks of a model, and the rules of combining these building blocks (and this is what I mean by language), should be such that only certain combinations are allowed. Specifically, only ones that have meaning. A good model does not allow erroneous or meaningless combinations of its basic concepts.

Do we violate this principle, when we make the generalizations and abstractions, that reduce a blog and a chatroom to the same concept of a group? Isn’t it a weaker and therefore stricter model, with fewer possible combinations, that rather distinguishes messages, posts and comments, or links from replies, or chatrooms from blogs?

Our model certainly opens up the door to recombining these concepts and features in novel and unexpected ways — as we pointed it out multiple times — , but in my opinion this is rather a fertile ground for innovation, than room for error. I believe that our model is still restricting enough to not allow meaningless combinations, but generic and flexible enough to leave room for new configurations of existing conversation and storytelling features. I also believe that serendipity is the true source of innovation, and our graph model was built in the spirit of making this possible.

Finally, I feel like we need to answer one more question, that may naturally arise at this point. Is it even worth to think about these problems at all? Do we really need more innovation in the field of online conversations? Aren’t online conversations fatally toxic, and aren’t these platforms just playgrounds for bullies, and the ultimate arenas of polarization, simplified interpretations and pathological narcissism? Isn’t social media causing enough damage already, by ruining public discourse, and removing subtlety, rationality and depth from exchanges?

While I’m obviously not suggesting that our toying around with a graph model might have anything to do with addressing such issues, but in general I believe that thinking about online conversations is still important. This unprecedented interconnectedness of people, that bridges geographic distances and ignores social and cultural barriers, is also a historic opportunity. The fact that everybody can have a voice, and how these voices can coalesce, is not only a danger, and a source of all kind of harm and malice — although this may seem less and less obvious in 2016. But it also holds the potential of both cooperation, and control of any form of power, on levels never seen before. Also, let’s face it, social media is not going away, and neither does commenting. Our only chance is to innovate in this field, and find the right way, so we can reduce harm, while exploiting the full potential.

Building a conversation platform — Part 11

Why graph?

Written by Mikolaj Szabó