We conduct a 6 month field experiment on a movie-recommendation platform to identify if and how recommendation systems affect consumption. We use within-consumer randomization at the good level and elicit beliefs about unconsumed goods to disentangle exposure from informational effects. We have three experimental groups: (a) control, (b) exposed, and (c) recommended + exposed goods where only goods in (c) are recommended and we elicit beliefs about goods in (b) and (c). Comparing across these treatment arms we find recommendations increase consumption beyond its role in exposing goods to consumers. We provide support for an informational mechanism: recommendations affect consumers' beliefs, which in turn explain consumption. Recommendations reduce uncertainty about goods consumers are most uncertain about and induce information acquisition. Finally, we find evidence for spatial correlation in beliefs.
Query categorization at customer-to-customer e-commerce platforms like Facebook Marketplace is challenging due to the vagueness of search intent, noise in real-world data, and imbalanced training data across languages. Its deployment also needs to consider challenges in scalability and downstream integration in order to translate modeling advances into better search result relevance. In this paper we present HierCat, the query categorization system at Facebook Marketplace. HierCat addresses these challenges by leveraging multi-task pre-training of dual-encoder architectures with a hierarchical inference step to effectively learn from weakly supervised training data mined from searcher engagement. We show that HierCat not only outperforms popular methods in offline experiments, but also leads to 1.4% improvement in NDCG and 4.3% increase in searcher engagement at Facebook Marketplace Search in two weeks of online A/B testing.
A single digital newsletter usually contains many messages (regions). Users’ reading time spent on, and read level (skip/skim/read-in-detail) of each message is important for platforms to understand their users’ interests, personalize their contents, and make recommendations. Based on accurate but expensive-to-collect eyetracker-recorded data, we built models that predict per-region reading time based on easy-to-collect Javascript browser tracking data. With eye-tracking, we collected 200k ground-truth datapoints on participants reading news on browsers. Then we trained machine learning and deep learning models to predict message-level reading time based on user interactions like mouse position, scrolling, and clicking. We reached 27% percentage error in reading time estimation with a two-tower neural network based on user interactions only, against the eye-tracking ground truth data, while the heuristic baselines have around 46% percentage error. We also discovered the benefits of replacing per-session models with per-timestamp models, and adding user pattern features. We concluded with suggestions on developing message-level reading estimation techniques based on available data.
Large organizations use bulk email to communicate with employees about events, policies, organizational updates, and other information they feel will be useful or interesting to the employees. Such an organizational system has many stakeholders including information producers (often organizational leaders), communications professionals, recipients (employees), and management. We find this system to be inefficient - sending messages broadly appears free to senders, but shifts costs to recipients and their managers. But careful targeting is viewed as expensive by senders, as the cost is incurred locally. As a result, employees often waste time reviewing poorly targeted messages. This chapter proposes an economic model to describe a multi-stakeholder organizational communications system based on bulk email. In addition to highlighting the various costs and values perceived by stakeholders, the model captures the reputation of bulk email channels. We report on a detailed study of one organization and highlight potential interventions prompted by the model.
In this study, we partition users by rating disposition - looking first at their percentage of negative ratings, and then at the general use of the rating scale. We hypothesize that users with different rating dispositions may use the recommender system differently and therefore the agreement with their past ratings may be less predictive of the future agreement. We use data from a large movie rating website to explore whether users should be grouped by disposition, focusing on identifying their various rating distributions that may hurt recommender effectiveness. We find that such partitioning not only improves computational efficiency but also improves top-k performance and predictive accuracy. Though such effects are largest for the user-based KNN CF, smaller for item-based KNN CF, and smallest for latent factor algorithms such as SVD.