Curation: How to Beat Negative Network Effects

There are two types of negative network effects in the software realm — networks can choose between five curation mechanisms to overcome them

Published in

Breadcrumb.vc

10 min readJan 11, 2021

So far, I have explained various characteristics of network effects and their impact on scalability, defensibility, liquidity, and monetization. The implicit assumption here is that interactions between participants are positive for everyone involved. This is true most of the time, but not all of the time. Interactions between network participants can also be negative. As a result, successful networks need to put curation mechanisms in place to encourage positive interactions; and dissuade or prevent negative ones. Let’s take a deeper look at what these negative network effects look like and the curation mechanisms to mitigate their impact.

Negative Network Effects

The definition of positive network effects is that the addition of a user makes the network more valuable for all users. So negative network effects exist when the addition of a user makes the network less valuable for all users. In software networks, negative network effects come in two variants:

1. Network Pollution

This type of negative network effect commonly occurs on social networks like Facebook as they scale. When a user’s social graph grows beyond a point, posts from distant acquaintances or family members can drown out those from close friends. This reduces the “signal-to-noise ratio” on the social feed, leading to a decline in engagement and network value. This pattern can also be seen in data networks that actively crowdsource information from users — for example, Tripadvisor, Yelp, or Stack Overflow. Here, as the number of contributors increases, lower-quality content (posts or reviews) can overwhelm the most relevant or in-depth ones. Keep in mind that users intend to make positive contributions in each of these cases. However, “uncurated” user growth can still cause relevant information to be drowned out. This type of negative network effect is called network pollution.

Notably, network pollution can only occur in networks reliant on user-generated content, i.e. interaction networks (e.g. Facebook) and data networks with active crowdsourcing (e.g. Stack Overflow). The equivalent problem for marketplaces or platforms would be well-intentioned suppliers who are unable to find demand or vice versa. This supply-demand mismatch is better described as a liquidity problem caused by insufficient supply density (for a region, a category, or both), rather than network pollution.

2. Bad Actors

Another problem is the presence of bad actors, i.e. participants who aim to create negative interactions to further their own interests. This includes spammers, trolls, or bots on social networks like Twitter or data networks like Yelp, scammers on marketplaces like Airbnb, malware applications on platforms like Windows, etc. The addition of a bad actor reduces the value of the network for all users and can repel positive participants. These bad actors need to be excluded from the network to ensure that all (or at least most) participants have positive interactions.

Bad actors drive away positive participants

Unlike network pollution, bad actors can affect nearly any type of network. Interaction networks, marketplaces, and platforms are at most risk because they allow participants to interact with each other directly. Data networks that actively crowdsource data from participants (e.g. StackOverflow or Yelp) can also be affected — for example, bad actors can spam the network with irrelevant information. However, data networks that passively crowdsource data (e.g. Mapbox) are largely immune from bad actors — this is because the network controls what type of information is shared and how it is shared, not the contributor.

Curation Mechanisms

The goal of curation is to minimize the negative network effects mentioned above and maximize positive ones. In other words, it aims to help users find the most relevant information or participants, and avoid irrelevant or malicious ones. This can be accomplished via five possible curation mechanisms:

1. Embedded Curation

The most basic way to curate network interactions is by limiting what the product allows participants to do. In other words, this type of curation is embedded within the constraints of the product itself. This makes it incredibly flexible and can be used by any type of network. However, its implementation and impact on negative and positive network effects tend to be very specific to each product.

Nextdoor is a great example of embedded curation. It only allows users located in nearby neighborhoods to interact which limits network pollution relative to a “horizontal” network like Facebook. This is one of the reasons why vertical or specialized networks provide a better customer experience than horizontal networks. Embedded curation is also the reason why data networks with passive crowdsourcing (e.g. Mapbox) are least susceptible to both forms of negative network effects described above — the product controls what participants share.

2. User-Controlled Curation

Another simple way to curate interactions is by giving users filters to control possible interactions — this includes controlling what information they see and who can consume the information they create. Implementations like search filters and prioritized lists of close connections can help users find relevant information and connections — this enhances positive network effects. It can also limit network pollution to an extent. Similarly, allowing users to block others can help them keep bad actors at bay. However, a network can only provide a limited number of filters before the cognitive load on users exceeds a threshold. This makes it difficult to rely on these filters exclusively as the network scales.

Search filters on marketplaces like Deliveroo, Uber Eats, Airbnb, and platforms like iOS or Android are obvious examples of user-controlled curation. Facebook’s friend lists are another example as they allow users to control who they share with.

3. Manual Curation

This is an interventional approach to managing participants and interactions. It puts the onus on the product owner to police interactions between participants. Networks can accomplish this by putting in place guidelines and removing content, supply, or participants that break them. Repeat offenders can be suspended or banned from the network. This approach has two primary advantages — its impact is instant and it is effective against bad actors. However, it is not effective against network pollution and it cannot reinforce positive interactions. It is also difficult to apply manual curation at scale (as Facebook’s struggles with misinformation have demonstrated). By itself, it is most effective for acute needs (e.g. removing bad actors) and at the earliest stages of a startup’s journey.

Crucially, manual curation has no technical or product-level requirements. This makes it universally applicable to all types of networks, including social networks like Facebook, marketplaces like Airbnb, platforms like iOS, and data networks like Tripadvisor.

4. Community Curation

Community curation, on the other hand, mirrors a “free-market” approach. It gives users the power to promote or demote content or supply units without any involvement from the product owner — this includes ratings, likes, shares, etc. depending on the network in question. The primary advantage of community curation is that it is scalable, can promote positive network effects, and also limit network pollution to an extent. However, it requires an active network to be impactful and it is less effective against bad actors. This makes it better suited for networks that have already reached a desired level of liquidity and have manual curation in place already. The last point is especially important because community curation can be problematic if used in isolation, especially early on in the adoption cycle. For example, without an app store review process, the developer of a malware app can drown out real reviews with fake positive reviews, perpetuating the very behavior that curation aims to filter out.

Community curation is also complex because its applicability varies based on the nature of each network and the number of participant types. On one-sided networks, like data networks and the majority of interaction networks, community curation can only be used when users have an opportunity to provide feedback and this feedback can be collected at scale. This restricts its use to dense, 1:many interaction networks (e.g. Twitter, Facebook, etc.) and data networks with active crowdsourcing (e.g. Waze, Tripadvisor, etc.).

Take Twitter, for example. Users can promote a tweet they see on their timeline by retweeting or liking it. Similarly, users can also provide negative feedback by reporting tweets that break guidelines — this community-triggered curation mechanism can also make manual curation more scalable, to an extent. Facebook likes, Medium claps, Github stars, “helpful” buttons on Tripadvisor, and the “thumbs up” and “not there” buttons on Waze are all different implementations of community curation.

Two-sided or multi-sided networks like marketplaces and platforms, on the other hand, can always use community curation. However, the implementation depends on the number of concurrent transactions that a single supply unit or listing can accommodate. If there is no limitation on the number of demand-side participants a supply unit can support, community curation can be implemented as a simple 1-way review system — demand rates supply after every transaction. We see this on “low-touch” marketplaces like Udemy or Deliveroo, and on platforms like Android, Shopify, or Salesforce.

However, on “high-touch” marketplaces like Airbnb and Uber, each supply unit can accommodate a limited number of demand-side participants at a time (often just one). Here, community curation requires a 2-way review system to develop counterparty trust — both supply and demand-side participants need to rate their experience with each other.

On Airbnb, for example, a property listing can only accommodate one booking at a time. Here, both the guest and the host rate their experience with each other after the transaction is completed and fulfilled.

5. Automated Curation

Automated curation attempts to find a balance between an “interventional” and a “free-market” approach. The basic idea behind automated curation is that the network applies a tag, or trains a ranking algorithm based on the behavior of participants, i.e. data. This data could include user engagement (e.g. clicks, time spent), performance metrics (e.g. delivery time, conversion rate), signals from community curation (e.g. likes, ratings), or a combination of the above. Automated curation is very effective at enhancing positive network effects and somewhat effective at countering network pollution. However, it is not as effective in dealing with bad actors. And like community curation, it is best implemented after a minimum level of liquidity has been reached — this gives the network sufficient activity and data to inform curation.

Facebook’s news feed algorithm is the most well-known example of automated curation — it ranks content based on engagement patterns and what users are most likely to interact with. Airbnb’s Superhost program is another example — it automatically applies the tag to hosts who have a rating of 4.8+, with a 90+% response rate, <1% cancellation rate, and have completed a minimum number of transactions. Uber Eats’ algorithm for ranking restaurants is yet another example as it takes a number of factors into account, including delivery time and rating.

As we can see, user feedback from community curation can be one of the datasets feeding into automated curation, but this isn’t a requirement. For example, Facebook, Uber Eats, and Airbnb’s ranking algorithms take user feedback into account. However, Snapchat Discover does not have any mechanism to collect user feedback and its algorithm is purely based on engagement.

Finally, the requirements for automated curation are less stringent than they were for community curation but more stringent than prior curation mechanisms. It can be used on any network where users actively seek out and broadcast information. This automatically includes all two-sided and multi-sided marketplaces and platforms. Among one-sided networks, this can be applied to all data networks and dense, 1:many interaction networks like Facebook, YouTube, TikTok, Snapchat, Medium, etc.

Keep in mind that interaction networks tend to be very sensitive to automated curation, with a high risk of negative side effects. Promoting content that users are more likely to engage with can create filter bubbles. This can also happen with community curation but to a lesser extent. Network dilution is another risk — excessive reliance on an algorithm can lead to content from direct connections being deprioritized in favor of “engaging content” (e.g. TikTok). This can dilute the importance of user identity and weaken the value of connections between participants, making the product more vulnerable to competitors. In other words, in the pursuit of limiting negative network effects like network pollution, startups can go too far and weaken positive network effects.

To recap, each of the curation mechanisms I have explained has different strengths and weaknesses. This is summarized in the table below:

As a result, most networks need to use a combination of applicable mechanisms to manage positive and negative network effects. Keep in mind that these curation mechanisms need to be re-evaluated constantly —the mechanisms that help a network get off the ground could become ineffective or even detrimental at scale.

What you should do next…

1. Pitch me: No warm intros required

I invest in pre-seed/seed-stage startups as a Venture Partner at Speedinvest and also via the Atomico Angel Program. If your startup is built on network effects and based in Europe, pitch me here.