Improving search functionality poses challenges such as data scarcity for model training, metadata enrichment for comprehensive document indexing, and the labor-intensive manual annotation for evaluation. Traditionally, iterative methods relying on human annotators and customer feedback have been used. However, recent advancements in Large Language Models (LLMs) offer new solutions. This paper focuses on applying LLMs to playlist search. Leveraging LLMs’ contextual understanding and generative capabilities automates metadata enrichment, reducing manual efforts and expediting training. LLMs also address data scarcity by generating synthetic training data and serve as scalable judges for evaluation, enhancing search performance assessment. We demonstrate how these innovations enhance playlist search, overcoming traditional limitations to improve search result accuracy and relevance.

Music Search encounters a significant challenge as users increasingly rely on catchy lines from lyrics to search for both new releases and other popular songs. Integrating lyrics into existing lexical search index or using lyrics vector index pose difficulties due to lyrics text length. While lexical scoring mechanisms like BM25 are inadequate and necessitates complex query planning and index schema for long text, text embedding similarity based techniques often retrieve noisy near-similar meaning lyrics, resulting in low precision. This paper introduces a proactive approach to extract catchy phrases from song lyrics, overcoming the limitations of conventional graph-based phrase extractors and deep learning models, which are primarily designed for extractive summarization or task-specific key phrase extraction from domain-specific corpora. Additionally, we employ a multi-step mechanism to mine search query logs for potential unresolved user queries containing catchy phrases from lyrics. This involves creation of word and character k-gram index for lyric chunks, careful query and lyrics domain-centric normalization (and expansion) and a re-ranking layer incorporating lexical and well as semantic similarity. Together these strategies helped us create a high retrieval source specifically for serving lyrics intent queries with high recall.

Music and media search spell correction is distinct as it involves named entities like artist, album and podcast names, keywords from track titles and catchy phrases from lyrics. Users often mix artist names and keywords from track title or lyrics making spell correction highly contextual. Data drift in search queries caused during calendar event days or a newly released music album, brings a unique challenge of quickly adapting to new data points. Scalability of the solution is an essential requirement as the Music catalog is extremely large. In this work, we build a multi-stage framework for spell correction solution for music, media and named entity heavy search engines. We offer contextual spelling suggestions using a generative text transformer model and a mechanism to rapidly adapt to data drift as well as different market needs by using parameter efficient based fine tuning techniques. Furthermore, using a reinforcement learning approach our spell correction system can learn from a user's implicit and explicit feedback in real-time. Some key components of this system are being used in search at Amazon Music and showing significant improvements in customer engagement rate and other relevant metrics.

This work presents the search and discovery architecture of Amazon Music, a highly efficient system designed to retrieve relevant music content for users. The architecture consists of three key stages: indexing, retrieval, and ranking. During the indexing stage, data is meticulously parsed and processed to create a comprehensive index that contains dense representations and essential information about each document (such as a music or podcast entity) in the collection, including its title, metadata, and relevant attributes. This indexing process enables fast and efficient data access during retrieval. The retrieval stage utilizes multi-faceted retrieval strategies, resulting in improved identification of candidate matches compared to traditional structured search methods. Subsequently, candidates are ranked based on their relevance to the customer’s query, taking into account document features and personalized factors. With a specific focus on the podcast use case, this paper highlights the deployment of the architecture and demonstrates its effectiveness in enhancing podcast search capabilities, providing tailored and engaging content experiences.

We introduce and formalize a novel constrained path optimization problem that is the heart of the real-time ad serving task in the Yahoo! (formerly RightMedia) Display Advertising Exchange. In the Exchange, the ad server's task for each display opportunity is to compute, with low latency, an optimal valid path through a directed graph representing the business arrangements between the hundreds of thousands of business entities that are participating in the Exchange. These entities include not only publishers and advertisers, but also intermediate entities called "ad networks" which have delegated their ad serving responsibilities to the Exchange. Path optimality is determined by the payment to the publisher, and is affected by an advertiser's bid and also by the revenue-sharing agreements between the entities in the chosen path leading back to the publisher. Path validity is determined by constraints which focus on the following three issues: 1) suitability of the opportunity's web page and its publisher 2)suitability of the user who is currently viewing that web page, and 3) suitability of a candidate ad and its advertiser. Because the Exchange's constrained path optimization task is novel, there are no published algorithms for it. This paper describes two different algorithms that have both been successfully used in the actual Yahoo! ad server. The first algorithm has the advantage of being extremely simple, while the second is more robust thanks to its polynomial worst-case running time. In both cases, meeting latency caps has required that the basic algorithms be improved by optimizations; we will describe a candidate ordering scheme and a pre-computation scheme that have both been effective in reducing latency in the real ad serving system that serves over ten billion ad calls per day.

Recommender Systems (RS), allow users to share information about items they like or dislike and obtain, in a timely fashion, recommendations based on predictions about unseen items (physical or information goods and/or services). In this process, users' preferences are considered to be the learning target functions. We study Agent-based Recommender Systems (ARS) under the scope of online learning in Multi-Agent systems (MAS). This approach models the problem as a pool of independent cooperative predictor agents, one per each user (the masters) in the system, in situations in which each agent (the learners) faces a sequence of trials, with a prediction to make in every step, eventually getting the correct value from its master. Each learner is willing to discover the degree of similarity among the target function of its master and those of other agents' masters (i.e. preference similarity). The agent uses this information for the calculation of its own prediction task, the goal being to make as few mistakes as possible. A simple, yet effective method is introduced in order to construct a compound algorithm for each agent by combining memory-based individual prediction and online weighted-majority voting. We give a theoretical mistake bound for this algorithm that is closely related to the total loss of the best predictor agent in the pool. Finally, we conduct some experiments obtaining results that empirically support these ideas and theories.