January 2017 Summaries

A data scientist working on feed personalization explores the use of Factorization Machines (FM) to enhance recommendation systems, particularly for addressing the cold-start problem. FM models are advanced machine learning techniques capable of generalizing methods such as Matrix/Tensor Factorization and Polynomial Kernel regression, allowing for the inclusion of additional features like item metadata in the model. These models can capture higher-order interactions, offering a rich way to improve predictions beyond basic user-item interactions. The blog post demonstrates the practical implementation of FM using TensorFlow and the RecSys 2015 challenge dataset, showcasing how historical engagement data and category metadata are utilized to train the model. Through experimentation, it is revealed that while access to comprehensive data yields more accurate predictions, the FM model still performs reasonably well in cold-start scenarios by leveraging aggregated category data. The study underscores the potential of FM models in building effective, personalized recommendation systems that enhance user engagement and conversion.

Algolia is a hosted search engine designed for developers to integrate real-time search capabilities into websites or applications without the need for managing infrastructure. It offers a 14-day free trial and a free Hacker plan, allowing users to index up to 10,000 records and perform 100,000 operations monthly. The process of setting up Algolia involves creating an index, importing data, configuring relevance settings, and implementing a search user interface using Algolia's front-end API clients. The platform supports typo tolerance, customizable ranking strategies, and real-time search results, making it suitable for various use cases. Additionally, Algolia provides libraries such as autocomplete.js and instantsearch.js to enhance the user experience. Alternatives like Solr and Elasticsearch offer similar functionalities, with Elasticsearch being available as a fully managed service through Amazon. Algolia is presented as a straightforward solution for developers seeking to build efficient search experiences.

Stream, co-founded by Tommaso and his partner over two years ago, has rapidly grown to serve over 500 companies with its feed technology reaching more than 50 million end users. The company, with offices in Boulder, Colorado, and Amsterdam, The Netherlands, is focused on further enhancing its API and performance metrics in 2017 with a team of 11 employees. Stream is committed to improving feed personalization and performance, currently boasting an average response time of ~40ms, and aims to reduce volatility and improve stability, having maintained a 99.99% uptime over the past six months. The rollout of API Version 1.1 is expected to resolve minor user issues, while full multi-region support is planned to enhance service for customers in Europe and Asia. Stream's analytics and insights will also be refined to better cater to diverse app requirements, and the company is keen on receiving feedback to shape its future developments.

Custom Ranking for activity streams and news feeds allows users to have greater control over how content is sorted, as opposed to the default chronological order on Stream feeds. By defining a scoring formula, users can prioritize activities based on specific attributes like popularity, timing, and promotions, enhancing the visibility of certain content types like events about to start or editorial content. The tutorial provides several examples, such as using logarithmic functions to limit the influence of very popular activities or pinning specific activities to the top of the feed. Furthermore, analytics data can be integrated to rank activities based on engagement metrics, offering a dynamic approach to feed management. Personalized feeds differ from ranked feeds by utilizing custom endpoints and machine learning for more tailored content delivery, enhancing features like discovery-based feeds and content suggestions.