Reading reflections for “Recommender Systems”

Resnick and Varian talk about how “it is often necessary to make choices without personal experience of the alternatives.” In their framing, recommender systems are primarily a decision-making tool. On the other hand, DellaPosta, Shi and Macy suggest that the critical insight behind item-based recommenders (“people who like one thing in common also tend to like other things”) results in lifestyle politics and culture wars, even when the commonalities have no underlying logic (horoscopes and liberalism?)

In your reflections, consider writing about the implications of these ideas: if correlations are not reflective of deeper shared desires and values, what does this mean for using recommenders to help make choices when we don’t know the alternatives well? What other implications does reading the two papers together suggest?

4 Comments

  1. Article 1 more focuses on the functionality of the recommender system. They consider it as a tool to help people make decisions without having experience with the alternatives. Most of real life scenarios do have to make decisions without experience. For instance, if you are shopping on amazon, people will not buy lots of similar items with different brands. They will directly pick one of them. The recommendation system helps the consumer to pick the most suitable one. On the other hand, article 2 authors believe the correlations are not very effective. The demographic or peer influence will have very subtle effects on the decision of individuals. Due to the network autocorrection, the choices of individuals may differ from the prediction, so that the prediction won’t be accurate. In my opinion, even the prediction might be completely wrong, the recommendation system still has its reason to exist. Some of the recommendation system won’t take the correlation into considerations at all. If someone has ordered bedsheets, pillows, and bedspreads. The shopping website will recommend a lot of bedroom appliances to the consumer. Under this circumstance, the correlations of the users are not very important. As long as the website correctly estimates the users’ potential interest and the following action, it would provide a valid recommendation to the user.

  2. Resnick and Varian talk about how “it is often necessary to make choices without personal experience of the alternatives.” In their framing, recommender systems are primarily a decision-making tool. On the other hand, DellaPosta, Shi and Macy suggest that the critical insight behind item-based recommenders (“people who like one thing in common also tend to like other things”) results in lifestyle politics and culture wars, even when the commonalities have no underlying logic (horoscopes and liberalism?)

    In your reflections, consider writing about the implications of these ideas: if correlations are not reflective of deeper shared desires and values, what does this mean for using recommenders to help make choices when we don’t know the alternatives well? What other implications does reading the two papers together suggest?

    In Resnick and Varian’s article, recommender systems are studied primarily in the light of technicality, like the dimensions of the technical design or domain space. One of the few socio political discussions emerging from this is when the business model comes in. In designing the domain space, evaluations are structured, gathered, distributed and displayed based on priority & cost structure, thus the presence of a for-profit component often leads to corrupted systems in terms of credibility. In this case, the interests of advertisers or people who are willing to pay for advertising outweigh the independence and integrity of recommendations themselves; or a privacy concern of to whom and how users’ data is sold by the recommender platforms. Still this very much means they frame recommender systems primarily as a tool that aids decision-making, which doesn’t have any intrinsic political attributions and of which good or bad impacts depend on a man made physical arrangements.

    The research of DellaPosta, Shi and Macy, however, accentuates on a critical aspect of lifestyle politics: An agent’s opinion is influenced by demographic background, but the effect is dramatically amplified by the reinforcing dynamics of homophily and social influence – being surrounded by birds from the same flock. This kind of influence is not sole to carving lifestyle, but is central to expanding and deepening the cultural enclaves – be it normative or aesthetic – that symbolizes particular kinds of characters or statuses.
    This idea suggests a very interesting implication that the lifestyle – political profile correlation do not necessarily reflect one’s true desire and value, but is rather formulated through a self-reinforcing dynamic that is embedded in most recommender systems.
    The more people have in common, the more likely they are to interact, the more they interact, they more likely they are to influence one another, and the more they influence one another, the more they are likely to have in common – as the authors put it to summarize their research on the LSE US Centre’s daily blog.
    When looking at this along with how social media recommender algorithms such as Facebook’s work, news feed are so compelling because it is customized to allow the most possible interaction with people of similar interests and tastes. As this interaction is further amplified, the smallest affinities between lifestyle and ideology are amplified, thus creating distinctive groups with their lifestyle tied to identifiable political profile.

    What’s worth noting is that we as users are situated at a very low level of consciousness throughout this whole social grouping – it almost feels so natural and harmless that we are having this current news feed. But what we essentially lose is a large degree of autonomy in creating our social and ideological networks – yes you can be a little more conscious and ignore unreliable things, but if you’re exposed to its derivatives everyday it’s likely that you’ll lose such consciousness and naturally take in those ideas.

    Not only on a personal level but largely, this self-reinforcing effect of homophily and social influence will make the lifestyle distinctions more symbolic of political profiles. Consequently, the cultural common ground is eroded, leading to intense polarization and cultural wars without much opportunity for reconciliation as we witness today.

  3. If correlations aren’t reflective of shared values, recommender systems still provide us with viable options. Recommender systems help narrow down our choices and give us the opportunity to learn more about them on a deeper level. For example, if someone likes a mystery novel and the recommender system gives recommendations of other novels people who read that book also bought, then even if those recommendations don’t represent a deeper connection it still provides a smaller subset of possible options for a person to look into compared to if they were to look for any book in general. This will allow a customer to spend more time evaluating the recommended options and then deciding if they want to buy them.

    Recommender systems need to be designed properly so that they can filter out biased reviews which can skew a person’s interests towards possible hidden agendas. The second article showed how our lifestyles can be influenced by political agendas and different candidates may take advantage of poorly structured recommender systems to help forward their cause.

  4. Resnick and Varian (1997) mainly discuss the technical aspects of recommendation systems- with some emphasis on their social implications and long-term impact. They describe recommendation systems as a ‘decision making tool’ but can this decision making be affected/altered by some external factors? DellaPosta, Shi and Macy (2015) discuss an aspect of aforementioned question in their paper by calculating the extent our lifestyle and decision making in everyday life is correlated with our political and ideological alignment. They find that our lifestyle preferences (choice of coffee, music, art, moral judgements) are correlated with our political alignments (while factoring demographics and socio-economic backgrounds as a 3rd variable).
    One implication of the ideas discussed by both papers is that our recommendation systems (which essentially is a decision making tool) might behave just like human beings. They might introduce biases or preferences certain groups or topics. Recommendation system is built on data trained and provided by human beings, and if human beings provide their “biased” recommendations influenced by certain political or ideological beliefs, then the recommendation system will result in a biased system – with its recommendations not being accurate. Even though result might not be accurate or valuable, recommendation system can still help individuals in their decision-making tasks and help them narrow down their options (and make fast decisions). In other words, we can still rely on recommendation systems to an extend even if correlations do not reflect deeper shared values and desires.

Leave a Comment

Your email address will not be published. Required fields are marked *