Every day, we make approximately 35,000 decisions.  It is undeniable that technology has played, and continues to play, a huge role in making our lives more convenient. However, information overload and the sheer abundance of choice that the internet offers only adds to the decision fatigue epidemic. Research has shown that our cognitive resources are scarce and we tend to feel depleted by decision-making.  Machine learning (ML) is helping to limit the effects of information overload on crowdsourced review platforms, making them better decision-making aids. The machine learning megatrend is moving review platforms towards becoming recommendation platforms, as users want technology to be more decisive. One such review platform is TripAdvisor, which has become an instrumental tool for travelers looking to make decisions about what to do on their trips.
At present, TripAdvisor is working on hyper-personalized rankings and optimizing the helpfulness of reviews. By limiting the amount of information that users have to decipher to make a decision, TripAdvisor is moving closer to becoming a recommendation platform.
To help frame how ML is shaping TripAdvisor’s product direction, it is important to understand the current limitations for recommendation platforms. As per diagram 1:
- Reliance on a large community that the user can lean on for advice
- Friction for users to contribute their opinions.
Machine learning can solve for these, as it can reduce platforms’ reliance on a large user base that is incentivized to contribute reviews. All they would need is a critical mass of user contributions as “training data”. However, full reliance on AI with no additional user contributions would require a steady state for users’ preferences. That’s when the model’s learning patterns are not future-proof and the “over-fitting” phenomenon comes into play. 
Besides the current ranking algorithms, there is a slew of launches around personalized recommendations in different parts of the product. For example, TripAdvisor’s email digests include a collection of algorithmically-curated suggestions.
TripAdvisor has too many reviews for human moderators to rank which ones to display. In response to this, they built a classifier for scoring whether a review’s text is likely to be helpful to other travelers or not. However, the current performance is only sufficient for filtering reviews to route to moderators, but “even at its best precision, it has too many false positives to auto-reject any reviews.” The classifier still isn’t capable of operating without human intervention. 
Longer term, there are indications that TripAdvisor is working on more predictive tools. For example, for review helpfulness, “taking into account the user’s review writing history; if they’ve written helpful reviews in the past, perhaps we’ll give them more benefit of the doubt on a marginal review.” . TripAdvisor is feeling the pressure to become more directive. Hence, the real power of the mega-trend will be realized with TripAdvisor’s broader vision to be able to take into account user preferences, as well as various user-generated reviews, to provide users with one definitive suggestion that would best suit them and the occasion.
Training the classifier to serve one recommendation based on a weighted average of user-generated reviews (for example, weighted by the credibility of the reviewer) has major limitations.
Reflecting on how we seek recommendations in reality, we engage with one person at a time and seek their advice. With each 1:1 interaction, we can get an extra ‘datapoint’ that helps us make a reasoned decision. We also sometimes poll a large group of people to gauge the majority view on a topic. Despite taking longer, I would hypothesize that a string of 1:1 interactions are more effective at taking into account context. This dichotomy is visualized in diagram 2.
In the context of ML on TripAdvisor, the latter case holds and it is a major concern that only a limited amount of information is included in predictive models. The issue of TripAdvisor tapping into the hivemind and averaging responses to compute a recommendation in real-time is an important consideration for their product team.
In general, it is important to keep in mind that personalized recommendations are “forecasts of people’s preferences, and they are helpful even if they won’t tell you why people like the things they do, or how to change what they like.”  This raises the question of what users’ expectations of a machine-generated recommendation are?
Additionally, there is the issue of the standardization of decision-making. The call for a specific decision – which restaurant should I go to, for instance – would be met by the standardized result of social understanding, therefore flattening individual choice into social conformity, which raises the question of whether minimizing decision fatigue through machine learning is at the expense of individual agency?
(Word Count: 782)
 Sahakian, B. J.; Labuzetta, J. N. (2013). Bad moves: how decision making goes wrong, and the ethics of smart drugs. London: Oxford University Press.
 Rudiger et al. (2013). Effort reduction after self-control depletion: The role of cognitive resources in use of simple heuristics. Journal of Cognitive Psychology.
 Polman, E., & Vohs, K. D. (2016). Decision Fatigue, Choosing for Others, and Self-Construal. Social Psychological and Personality Science, 7(5), 471–478.
 Amis, G. (2015) Which of TripAdvisor’s reviews are actually helpful? Engineering & Product Operations TripAdvisor Blog. http://engineering.tripadvisor.com/which-of-tripadvisors-reviews-are-actually-helpful/
 Yeomans, M. (2015). What every manager should know about machine learning. Harvard Business Review.
 Lake. Stitch Fix’s CEO on selling personal style to the mass market. Harvard Business Review 96, no. 3 (May/June 2018): 35-40.