August 2016 Summaries

The Multi-Armed Bandit (MAB) problem is a decision-making challenge that involves selecting amongst a set of options, or "arms," to maximize expected rewards based on limited feedback. It is analogous to choosing the most rewarding biased coin from a set without sufficient trials or when penalties are incurred for poor choices. This problem is relevant in real-life scenarios like clinical trials and ad placements where only the outcomes of chosen actions are observable. In machine learning, MAB algorithms, such as ε-greedy and UCB1, are used to balance exploration and exploitation, making intelligent decisions in dynamic environments. These algorithms adapt to changes in reward distributions, as demonstrated in experiments where they adjust to optimal strategies over time. Contextual Bandits, an extension of MABs, utilize additional information to make informed decisions, and tools like Vowpal Wabbit offer pre-built algorithms for large datasets. The text also discusses how bandit algorithms can be applied to personalize content feeds and improve user engagement, with examples illustrating their adaptability and effectiveness compared to traditional full-information models.

Feeds play a crucial role in enhancing user engagement and retention by allowing users to explore, share, and interact with content tailored to their interests, as illustrated by Google Music's use of activity streams and personalization. The platform's "Recent Activity" section enriches the user experience by facilitating seamless music discovery and resumption, while its optional personalization process—requiring minimal input like selecting a single music genre—ensures quick and efficient onboarding. Leveraging user activity and alternative data sources, such as YouTube and geolocation, Google Music offers a curated "Recommended For You" section that highlights the importance of integrating data-driven feeds into apps for improved user engagement. Additionally, the article emphasizes the potential benefits for developers in incorporating such feeds and activity streams to create a more personalized and sticky user experience, suggesting that services like Stream can aid in the efficient implementation and scaling of these features.

Apache Cassandra, originally developed by Facebook in 2009, is a free and open-source distributed database designed to manage large volumes of data across numerous servers. At Stream, Cassandra is utilized as the primary data store for handling feeds due to its ability to automatically shard data, manage partial outages without data loss or downtime, and scale almost linearly. Typically, Cassandra clusters are configured to withstand the loss of one or two nodes, but Stream has contributed a guest post on HighScalability detailing how to configure Cassandra to survive a complete availability zone outage.