[Paper Review] Modeling Musical Taste Evolution with Recurrent Neural Networks

Massimo Quadrana, Marta Režnáková+3 more|arXiv (Cornell University)|Jun 18, 2018

Music and Audio Processing12 references3 citations

TL;DR

This paper proposes a recurrent neural network (RNN) model to predict the evolution of user musical taste through sequential internet radio station creation, demonstrating superior performance in long-tail music personalization and achieving a 16.7% improvement in MRR@5 when integrated into a learning-to-rank system, outperforming baselines in capturing long-term preference dependencies.

ABSTRACT

Finding the music of the moment can often be a challenging problem, even for well-versed music listeners. Musical tastes are constantly in flux, and the problem of developing computational models for musical taste dynamics presents a rich and nebulous problem space. A variety of factors all play some role in determining preferences (e.g., popularity, musicological, social, geographical, generational), and these factors vary across different listeners and contexts. In this paper, we leverage a massive dataset on internet radio station creation from a large music streaming company in order to develop computational models of listener taste evolution. We delve deep into the complexities of this domain, identifying some of the unique challenges that it presents, and develop a model utilizing recurrent neural networks. We apply our model to the problem of next station prediction and show that it not only outperforms several baselines, but excels at long tail music personalization, particularly by learning the long-term dependency structure of listener music preference evolution.

Motivation & Objective

• To model the dynamic evolution of individual musical preferences over time using sequential data from internet radio station creation.
• To address the challenge of long-tail music personalization, where users explore niche or less popular artists over time.
• To evaluate the effectiveness of RNNs in capturing long-term dependency structures in listener behavior compared to popularity-based and similarity-based baselines.
• To integrate the RNN predictions into a cascaded learning-to-rank framework to enhance recommendation quality.
• To analyze model performance across different listener cohorts based on historical station counts and popularity levels.

Proposed method

• The model uses a gated recurrent unit (GRU) architecture to encode sequences of user-created radio stations, learning temporal dependencies in listening behavior.
• The RNN is trained to predict the next station a user will create based on their historical sequence of station creations.
• A cascaded learning-to-rank (L2R) system combines RNN predictions with features from KNN, popularity (POP), and word2vec (W2V) models to re-rank candidate stations.
• The L2R model uses LambdaMart with 500 trees and is optimized on MRR@1 and MRR@5 to improve top-k recommendation accuracy.
• Features include station popularity, station typology, listener age, listening diversity, song popularity, and station count, with RNN output added as an additional feature.
• The model is evaluated on a dataset of over 11 billion radio stations from Pandora, with listeners segmented by historical station count and future behavior.

Experimental results

Research questions

• RQ1 Can RNNs effectively model the long-term evolution of musical taste from sequential station creation data?
• RQ2 How does the RNN's performance compare to popularity-based and similarity-based baselines in predicting next station creation?
• RQ3 In what user cohorts (e.g., by station count or popularity) does the RNN show the greatest advantage?
• RQ4 To what extent does integrating RNN predictions improve ranking performance in a cascaded learning-to-rank system?
• RQ5 How well does the RNN perform in recommending less popular, long-tail music items compared to other models?

Key findings

• The RNN model achieves an 8.2% improvement in recall@20 over baseline models, with performance increasing significantly as historical data grows.
• The RNN outperforms baselines most substantially for users with between 6 and 40 created stations, indicating strong learning of long-term dependencies.
• The RNN demonstrates a marked advantage in recommending less popular, long-tail music items, correctly identifying them more often than other models.
• Integrating the RNN as an additional feature in a cascaded learning-to-rank system yields a 16.7% relative improvement in MRR@5 when optimized on MRR@5.
• The model shows a 10.3% improvement in MRR@1 and 13.8% in MRR@3 when the L2R system is optimized on MRR@1, confirming strong top-1 and top-3 performance.
• The results indicate that RNN predictions are complementary to other models, with the RNN capturing unique information about preference evolution.