How Item Features Power Music Recommendations: A Hands‑On Guide

Using Item Features for Recommendations

Unlike collaborative filtering that relies on user‑generated data, this chapter shows how to recommend items by comparing their intrinsic features, using Pandora Music as a concrete example.

Based on item features classification

Pandora employs a "music‑gene" project where professional musicians annotate songs with 400+ attributes (e.g., genre, mood, instrumentation). Each song is represented as a vector of scores from 1 to 5, enabling similarity calculations.

Importance of Feature Selection

Choosing meaningful features is crucial; poor choices can lead to misleading similarity scores. The article demonstrates this with a toy example where genre and mood scores are assigned, showing how Manhattan distance identifies the most similar song.

A Simple Example

A small dataset with seven musical features (piano, vocals, driving beat, blues influence, dirty electric guitar, backup vocals, rap influence) is scored on a 5‑point scale. Manhattan distance between two songs (e.g., Dr. Dog's "Fate" and Phoenix's "Lisztomania") is computed as 9.

Displaying Recommendation Reasons

By comparing feature vectors of a target song with liked songs, the system can highlight high‑scoring matching features (e.g., strong rhythm, vocal sections, electric guitar) as the rationale for the recommendation.

Standardization Issues

When feature scales differ (e.g., BPM vs. genre scores), raw distance calculations become dominated by the larger‑scale feature. The article introduces min‑max scaling (0–1) and robust standardization (using median and absolute deviation) to mitigate this problem.

Applying the Method to Other Domains

The same feature‑based classification is applied to athlete data (height, weight) to predict sport, to the classic Iris dataset for species classification, and to an automobile fuel‑efficiency dataset for mpg prediction, illustrating the versatility of the approach.

Practical Implementation

Python code snippets (e.g., getMedian function) are discussed for computing median, absolute deviation, and normalizing columns before classification.

Key Takeaways

Feature engineering transforms items into comparable vectors.

Proper scaling ensures balanced distance calculations.

Nearest‑neighbor classification works across diverse datasets.

Standardization can dramatically improve prediction accuracy.

End of chapter.