Algorithm Nudging: Down the YouTube Rabbit Hole
The unprecedented role of social media and fake news in nudging voting behavior in the recent US elections has been in the spotlight of everything from investigative articles to academic papers.
Recent Criticism
In a recent thought-provoking piece (Fiction is outperforming reality) The Guardian focused on the role of YouTube (alongside Facebook and Twitter sock puppets) in consistently exposing online users to substantial amounts of dubious content. According to their investigation, algorithms can be tricked to recommend increasingly sensational, manipulative and often fake content designed to nudge users towards ever more extreme views.
Guillaume Chaslot, ex-Google engineer, supports this view with his personal project AlgoTransparency.org. Mimicking an account with no viewing history, his project simulates how by starting with an unbiased search for information, viewers end up exposed to an ever more biased content via each subsequent step of YouTube recommendations.
According to The Guardian, “By keeping the algorithm and its results under wraps, YouTubeensures that any patterns that indicate unintended biases or distortions associated with its algorithm are concealed from public view”.
Harvard (techno)sociologist Zeynep Tufekci (author of the book Twitter and Tear Gas: The power and fragility of networked protest) shares this bleak view of online content pushing users down an ever darker rabbit hole. After spending a year analyzing the impact of technology on modern society, Tufekci acknowledges its positive contributions but also warns against the ability of algorithms to predict far more than your taste in entertainment.
By ceding information about our habits to advertisers, we allow them to cross the line of privacy in unexpected ways, just to be able to fine-tune their ad targeting efforts. Just to give you an idea of how serious this is: there are now ways to target messages to people with bipolar disorder as they enter the manic phase and are more susceptible to certain types of messages.
The chicken or the egg? Are algorithms really the ones to blame?
Exposure to certain types of content can, indeed, predispose online users to make certain purchases or leaning towards specific partisan choices. But is this really provoked by malicious algorithms or deliberate abuse of the technology behind this platform? Or is it a phenomenon emerging from the psychology of searching and interacting with online content, fueled by something much harder to police and control? In other words: is behavior being changed by the algorithms or do algorithms simply reflect (and amplify) already existing predisposition for engaging with such content?
Psychological biases and behavioral tendencies which are part of social cognition are now increasingly intertwined with the effects of technology we do not fully understand yet.
Funnelling from broad to specific
As Google engineers point out, they tweak the recommendation engine based on behavioral insights like the fact that “users often discover artists in a genre beginning with the most broadly popular before focusing on smaller niches”.
Funnelling recommendations to lead from broader categories to ever more specific and niche suggestions is an excellent approach if a person is interested in rock music and is lead to discover a selection of blues rock subgenres. This approach, however, gets a sinister twist if it drives a viewer from broader mainstream conservative content to ever more radical footage of violence and extreme right views.
Furthermore, content recommendations are based on viewing history and predictions based on collaborative filtering but w]hen content charge exists on a spectrum.
Confirmation bias bubbles
Recommendations implicitly and explicitly fuel confirmation bias bubbles. In real life, people are more likely to interact with others who sustain and confirm your preexisting beliefs. Online recommendation algorithms deepen the problem by exposing users exclusively to content which further reinforces and radicalizes their views.
Unless you consciously seek out different points of view, online interactions spiral you into seeing increasingly leftist or increasingly right-wing content, all the while alienating you from the positions expressed by the other side.
This process makes it easier for you to get enmeshed and indoctrinated and abolishes the many nuances and legitimate views which actually exist in the complicated space in the middle.
Extremes are easier to represent than complex nuances
There is another caveat: breaking down the spectrum of human beliefs and positions on economic, social and political issues is very complicated. The simplified left-right divide works ok for generalizations on a group level, but individuals often hold complex, highly nuanced and sometimes contradictory beliefs.
Tweaking algorithms to recommend highly specific suggestions within each user’s moderate leanings across a variety of topics is a tough task. So it is far easier to operate via vast generalizations (and simplifications) and nudge people towards content which is representative of the extremes of the spectrum, instead of the subtle nuances.
As journalist Tim Pool pointed out once, if you hold legitimate mainstream conservative views algorithms lump you with people across the political right spectrum whether that is a prices match or not. Let’s say you support free trade, less government intervention and smaller taxes (nothing radical about these views): by having a taste for news on these topics, altgorithms would sooner or later pair you with other users with conservative views and after enough iterations of the recommendation system, you would end up viewing extreme right content well beyond your right-of-center position on economics.
Sooner or later the information bubbles we inhabit online transfer to offline interactions as well. Access to information is probably one of the greatest features of the Internet. Unfortunately, this access is rarely used for critical fact-checking and more often than not, it only supports cherry-picking and confirmation bias.
Instead of building bridges, the abundance of information ends up creating invisible fences around our real and imagined preferences.
Algorithms do not produce biases, neither do they have an agenda to endorse fake content and conspiracy theories. The phenomenon we are observing is emergent, and it is the result of a vicious circle in content consumption.
Even if there is no conscious demand for sensational clickbait content, a cocktail of human curiosity and gullibility provides enough viewer interest to demand to rise to the occasion of content supply.
Clicking on a video is the outcome, but the reasons for doing it can vary dramatically. From actual susceptability to conspiracy theories through sheer curiosity and the urge to laugh at outrageous content: there are many reasons to watch something. As a media analyst I routinely watch content that is not personally appealing to me, but it helps me gain insight into different points of view, think critically and try to understand both the media landscape and what consumers likes and think.
Quick insights from Google Trends
The Guardian article made me think a lot about this issue of pushing forward specific types of content. Although I believe psychological biases and the appeal of emotionally laden conspiratory content is what made this content popular (and algorithms simply recognized the trend and amplified it), I cannot deny some worrisome tendencies.
I decided to run a quick Google Trends search on two ubiquitous topics: flat earth theories and the anti-vaccine movement, just to see whether any trends would pop up from the data.
Keeping in mind that correlation is not causation, I must admit that it is interesting that YouTube searches for “flat earth” (as a search term) and vaccine controversies (as a broader topic) peaked either alongside or at least a month before dramatic increases in Google web searches for them. So, it does look plausible that interest and consumption of video content on a given topic could fuel a shift in understanding or at least an openness to see more of the same.
This is a very simple insight which needs a lot of further data corroboration and when I have more time I would love to dig further into it.
What is publically known about YouTube’s recommendation system?
The use of deep neural networks and the algorithms behind YouTube’s recommender system are proprietary so to a large extent they are kept deliberately opaque to protect the company’s unique advantages and copyright over the technology. According to the publically available information, the YouTube recommendation system consists of two neural networks representing the different stages of information retrieval (See Figure 1):.
1.Deep candidate generation model
Candidate generation narrows down the massive database to a subset of videos which may be relevant to the user and uses implicit feedback (as opposed to explicit thumbs up/down) to provide a classification based on user history, collaborative filtering, etc. Search and watch history are tokenized (turned into strings with assigned meaning) and demographic information about the user is embedded.
Layers of depth (watches, searches, age of training example) enhance the model, allowing it to use additional features by modelling their interaction and outperforming older approaches.
2. Deep ranking model
The model uses impression data to personalize and calibrate prediction. Since several hundred videos are filtered out for the candidate generation model, ranking is limited to this feasible subset rather than the entire massive YouTube database.
A lot of A/B testing is used for tweaking the ranking and details like watchtime are used as a better predictor than click-through rate (which could be boosted by clickbait videos). There are also a number of categorical and continuous/ordinal features determined by a variety of factors, including properties of the content/item like impressions and properties of the user (history of search queries or number of recently watched videos).
Specialized features focusing on pas user behavior and items provide rich data for recommender systems. Having layers of depth also helps with modelling non-linear interactions between the vast number of features.
Constraints which the algorithms is trying to work around
The three most challenging constraints for the recommender system (according to Google engineers Paul Covington, Jay Adams and Emre Sargin)
- Scale: With 1.5 billion logged in users every month who spend an hour per day (on average) on the platform, YouTube recommendations need highly specialized distributed learning algorithms and efficient serving systems to function
With such an immense body of content, algorithms favour content which appears to be not only relevant to the user (and similar groups of users) but also engaging. So even if they are negative, comments and interactions with a video make it more visible. The algorithm cannot discriminate for quality and depth of information as a human reviewer would, but it can recognize and reward viral content.
2. Freshness: In 2017 1 billion hours of YouTube content were consumed daily and users uploaded hundreds of hours of new content evert minute. This creates a challenge for the recommendation system to incorporate new content and the latest user activities
3. Noise: User history is affected by external factors which do not always reflect preferences which need to be taken into account for future recommendations.
Watching more nuanced content or videos which diverge from the established viewing pattern could be rooted out as noise, thus contributing to simplification and generalization of interests towards the more extreme ends of a spectrum, instead of complex content catering to views which are harder to define.
The emerging science of click bait
There is also something interesting to take into account, which research is only recently catching up to it: the psychology of how sensational clickbait content attracts our attention and affects our viewing habits.
The Valence-Arousal-Dominance model
A few years ago MIT Technology Review covered a massive cross-lingual study by Marco Guerini (University of Trento) and Jacopo Staiano (Sorbonne Université, Paris) which researched the Relation between Emotions and Virality in 65000 stories.
They used the Valence-Arousal-Dominance model to investigate why some stories become viral, and others do not. This model focuses on three core dimensions:
- Each emotion has a Valence (positive or negative)
- It envokes a certain level of Arousal (high for emotions like anger; and low for emotions like sadness).
- Each person has a level of Dominance (or control) over the emotion (e.g emotions like fear are overwhelming and we generally have low dominance over them).
What Guerini and Staiano discovered is that the content which generates more comments or shares is associated with emotions of high arousal (like happiness and anger) and emotions where people feel less in control (like fear and sadness).
Content which generates more social votes (likes) is associated with emotions people feel more in control of: for example inspiration. Interestingly, valence did not influence virality at all. In other words, it did not matter whether content triggered positive or negative emotion.
This gives us some insight as to why sensational clickbait content (even if it contains extreme violence or borderline negative themes) is so visible: its description is specifically designed to provoke strong emotions and overwhelm viewers.
____
So, what do you think? Is there a mass effort to trick and manipulate algorithms or are we just observing how biases and tendencies inherent to human nature become distroted and amplified in the artificial social setting of the online environment?
