How to pronounce my name: "Your ghost", without the "t".
This page is not being updated anymore.
Using a dataset of every product page display, purchase and review submission that happened for a large online retailer throughout the entire year of 2015, we estimate the value of reviews, i.e., how much did the conversion rate for a product change as it accumulated more reviews. We control for exogenous trends, not related to the existence of reviews, by using a dataset of sessions in which users did not viewed or read reviews.
Note: This work is in partnership with PowerReviews, a review platform provider, and it was the focus of one the company's whitepapers
You can find the whitepaper here or download it directly. You can also read their blog post here.
Spiegel Research Center also has a post about this work. I also wrote a blog post discussing a different part of the insights from the work.
Mobile apps is a highly competitive market and promotions is a way for apps to gain word-of-mouth. Our work tries to quantify the benefits and risks of various promotions and suggest ways to design them in ways that amplify the benefits and mitigate the risks. We study four promotions offered on Apple’s App Store that vary in scale, price discount and redemption procedure. We find that each of these characteristics has a unique effect on the sales and ratings.
Data: I have a growing amount of data that includes detailed information for tens of thousands of apps in the iOS appstore, as well as panel data for millions of reviews of these apps. If you are interested in gaining access to the data or have ideas of how to use it, please don't hesitate to contact me.
We conducted a large-scale field experiment on Reevoo, a product review platform, to test the effect of social influence on consumer reviews. While reviews exhibit no evidence of social influence prior to our experiment, we find that adding a subtle social signal to communications with potential reviewers caused a rise in product ratings.
Given a snapshot of a diffusion process, we seek to understand if the snapshot is feasible for a given dynamic, i.e., whether there is a limited number of nodes whose initial adoption can result in the snapshot in finite time. While similar questions have been considered for epidemic dynamics, here, we consider this problem for variations of the deterministic Linear Threshold Model, which is more appropriate for modeling strategic agents. Specifically, we consider both sequential and simultaneous dynamics when deactivations are allowed and when they are not. Even though we show hardness results for all variations we consider, we show that the case of sequential dynamics with deactivations allowed is significantly harder than all others. In contrast, sequential dynamics make the problem trivial on cliques even though it's complexity for simultaneous dynamics is unknown. We complement our hardness results with structural insights that can help better understand diffusions on social networks under various dynamics.
We studied a variant of stable matching motivated by the fact that, in most centralized markets, many agents do not have direct communication with each other. Hence even if some blocking pairs exist, the agents involved in those pairs may not be able to coordinate a deviation. We modeled communication channels with a bipartite graph between the two sets of agents which we call the social graph, and we studied socially stable matchings. A matching is socially stable if there are no blocking pairs that are connected by an edge in the social graph. Socially stable matchings vary in size and so we looked for a maximum socially stable matching. We proved that this problem is NP-hard and, assuming the unique games conjecture, hard to approximate within a factor of 3/2-$\epsilon$, for any constant $epsilon>0$. We complemented the hardness results with a 3/2-approximation algorithm.
We studied crowd-curation mechanisms that rank articles according to a score which is a function of user- feedback. We precisely quantified the dynamics of which articles become popular in these systems. While crowd-curation can be relatively effective for cardinal objectives like discovering and promoting content of high quality, they do not perform well for ordinal objectives such as finding the best articles. Our analysis suggests that user preferences and behavior are a far greater determinant of curation quality than the actual details of the curation mechanism. Finally, we showed that certain shifts in user voting behavior can have positive impacts on these systems, suggesting that active moderation of user behavior is important for high quality curation in crowd-sourced systems.
One of my responsibilities in the Theory Group was recording the talks given in our weekly seminar. You can visit their YouTube channel for many interesting talks