The Onion Rater and other mitigations for platform negative externalities
Next: Regional, supplier and customer platforms
In the previous post (part 1) on centralised platforms we introduced some of their negative aspects.
TL;DR One approach to solve the gaming attempts on large platforms (will occur both in centralized and decentralized ones) is to build for each person their own view into the content. This can be done by placing more weight into the opinions of people that I personally know (i.e., highest score is my own judgement, then my dearest and nearest, then their nearest, mixing in at some point scores from other sources like public reviews). This eliminates the “flat-earth” model of current Internet services, where the fraudulent users’ automated bots have equal weight to myself. This approach we call the Onion Rater.
Onion is role specific. My skills in one topic are not equal to any other unless by chance. Ratings also need to degrade over time unless I do transactions regularly.
Mitigations for negative effects
Unbundling Algorithms from Data
Let’s discuss in the next two chapters possible evolution paths for platforms.
In Europe there is a payment services directive (PSD2) which opens up the application programming interfaces (APIs) for financial institutions like banks. This allows 3rd parties to create services on top of information held by banks. Potential new services can be for example aggregation services with bring consumers finances from all accounts together and allow them to understand and manage own financial situation better. Or for 3rd parties to launch new ways to make payments online without credit or debit cards.
Same principle can be applied to new platforms like social media, content sites or commerce. When I own my content and can grant 3rd party apps access via an open API, new services tailored to my preferences or needs can spring up.
For example:
Content reaper. Auto delete my postings after a while
Only show content from verified users
Show how many people have muted this person or how many of people that I know have blocked this person
Being able to delete direct messages from my side and the recipient. Allowing both sender and receiver to do this. Same for group messages.
Commerce filter. Examples: “do not show low-ball offers to my items on sale”
Personal feed. Content platforms try to do this but with them a single algorithm “rules them all”. Open APIs allow free innovation to meet the needs of various groups to surface just stuff you want or use very simple logic like show newest always. Or order based on different and sometimes personalized criteria like a combination how many followers this person has, how recent it is, who I follow on other services, how many times it has been liked, pick my close friends individuals to top of feed always etc. See Onion Rater below for an example.
Mood feeds. Change my filter preferences based on what I intend to do or am interested at the moment. Since people cannot be bothered to change any filters, the automations would have to figure this out somehow.
Filter Recommender. A service that has many customers using it, can start building aggregate services. When I am in the “car mood”, filter my feed based on other people in same city who share my interest in car related topics. Or on commerce platforms filter out buyers who promise to buy but then vanish.
Own algorithm also prevents large platforms from shadow banning prestigious medical journals or other fact-based science journals by their opaque fact checking partners.
You can get a feel of the power of filtering if you view the existing Internet companies as opt-in filters (no one is forced to use any particular service on the Internet). Search engines are filters for content or service pages, market place searches filter for what they think you are most likely to buy (or what gives them best margin if it is their own brand item), app stores filter for applications etc. Current services lock their data as it is the most valuable service they have.
Unbundling is about giving users’ data back to them and allowing users to collaborate and bring in 3rd party filtering.
For commerce platforms there is an additional option. Large retail companies lose a lot of products to professional criminal organisations who sell them via the online platforms. This raises the prices to everyone while online commerce platforms reap the benefits. Some of them make it as hard as possible to investigate albeit their public relations messages beg to differ. If anonymised transaction data were available via an interface or in a sandbox
law enforcement officials could apply machine learning to automatically flag potential abnormal activity for further study.
Private Chat Groups and Feature Changes
Second alternative is for discussions to vanish from the public sphere into private chat rooms in various messaging applications where all participants know each other having common hobbies or interests, working together or so. A variant of this is private discussion services where you join discussion by invite and the host manages participants blocking trolls. In this model the inviters’ reputation rests on the people they have invited.
This development is currently in full swing.
Third alternative is different type of social media or changes to current user interfaces– i.e., not having some features that cause bad side effects. For example, real time video meeting rooms bringing back real time interaction instead of browsing video catalogs and leaving snarky comments.
Overlay Services as Mitigation for social media
Today’s platforms are closed gardens and may be excellent to surface best results within their walls but many interesting people or offers are scattered over several of them. Open APIs allow building aggregate overlay services that fish their results from a much bigger pool. Larger choice leads to improved service. Say you are engaged in a discussion on some platform but very few people participate in it. An aggregator can pinpoint you to an alternate platform where just now an active discussion on that is taking place. Or if you have inherited an old watch or other item, you might want to know all relevant places where people discuss and help each other or where old, pre-owned items are sold. Aggregator has that view. The latter case also served by search engines but they do not have the real time view that a specialized aggregator is likely to have.
Overlays can also implement their own database and start capturing data there. Storing all interactions first to own database and then publishing it to one or more underlying platforms. Over time own database grows and at some point, the overlay platform can think about omitting the underlying platforms if it makes sense.
To combat trolls and fake news, overlays can also take benefit from the generative adversial network (GAN) method discussed in earlier albeit in a somewhat unorthodox way.
Today there are numerous bots on the social networks that are teaching people antisocial behavior and trying to get different user groups to feel alienated and hate each other. Reasons for this are political as discussed. Best ones are using machine learning models to improve performance. If I ban or engage with them in meaningful ways, they can use this as a signal in learning how to work worse.
A better method it to do the filtering is outside of the platforms and derive any learning possibility from these computer programs. This can be done with GAN models. Generative adversial networks are a concept where computers are made to play against each other. One model trying to post antisocial messages and another trying to block and filter them. Over time they learn each other’s tricks and get better and better. In our case we would use only the filtering model in production.
To maximize its performance, it is better to not only passively filter out content but to be a proactive member of the underlying platforms and teach the anti-social bots to degrade in performance. This is done by generating confusing signals towards the bot networks – posting incoherent pseudo-babble, liking and blocking same type of content and so on.
One effect of this is naturally, that to remaining user on the native underlying platform, the behavior generated by the overlay appears unconventional. But as weird is the new normal, this should be fine. Just proves that one is in sync with the times.
Over time if overlay network becomes popular, they can think about changing the underlying backend platform to its own – unbundling itself from its genesis network. This leaves the bot networks and their star students to have fun with each other on the old one.
The Onion Rater
One variant of the mechanisms possible with open APIs is so called Onion Rater. This is a model where spheres of decreasing influence are built around each user.
At the most important layer is content, products and vendors I have liked or purchased and given good reviews. Next important is what my immediate family and closest friends have done. This takes into consideration the fact that success and quality is highly subjective and this is especially true of services. It is easiest to understand when I have kind of common reference with the reviewers.
Next is what the friends of friends have said followed by either what friends of those friends have reviewed/liked or what reputable big brand news medias are posting and so on. This works like social circles in life - when I know someone personally over longer time, I know what “flavor” their opinions have and immediately know what matches mine and what not.
One of the problems with platforms is that they put all users on the same level. If a lot of people I do not know like or buy something and they appear to be similar to me, the platforms recommend this type of results to me. These other users can however be synthetic people - computer programs - created by a company selling “reviews” to not-so-honest companies or an adversary state with bad intentions. Platforms operate based on the “flat earth” model.
Responses also need to age with time. If someone has good ratings as a carpenter, but has not had any interactions or work for the past 5 years, one can question how relevant the reputation is. Different skills are usually independent of each other. For a person who does accounting and has excellent reputation, that reputation reflects only slightly their skills in carpentry (perhaps indicating commitment to good results).
The image below shows how the onion rater concept works.
Onion rater concept is based on a paper by a Brazilian researcher João Marcos Barguil.
The onion works very well for products or experiences that divide people. These are called multimodal products. Many films have divisive effect – some folks giving ten starts on rating services, others one. Presenting rating averages for these contains no real value as they end up somewhere in the middle mediocre category. Nor does surfacing the movies with most likes as a substantial part of people might find it poor. Personalized lens lets each group of people discover their own flavor.
The onion can also improve matching process on platforms. A recent ILO study on micro-tasking platforms found that on average, workers spent 20 minutes on unpaid activities for every hour of paid work - searching for tasks, taking unpaid qualification tests, researching clients to mitigate fraud and writing reviews. That’s a transaction cost of 1/3 from their productive time. The Onion let’s people find trusted parties through the digital “word of mouth” eliminating unnecessary search and re-qualification activities.
Such an Onion based filter takes into account two factors – identity and the service being offered. Many people on tasking platforms offer different types of work and obviously their skill levels differ.
In the next post we talk about regional and expert platforms that slightly differ from centralised, global ones.