Abstract
In this work we describe a hybrid recommendation approach for recommending sounds to users by exploiting and semantically enriching textual information such as tags and sounds descriptions. As a case study we used Freesound, a popular site for sharing sound samples which counts more than 4 million registered users. Tags and textual sound descriptions are exploited to extract and link entities to external ontologies such as WordNet and DBpedia. The enriched data are eventually merged with a domain specific tagging ontology to form a knowledge graph. Based on this latter, recommendations are then computed using a semantic version of the feature combination hybrid approach. An evaluation on historical data shows improvements with respect to state of the art collaborative algorithms.
Introduction
Recommending audio samples, such as those found on platforms like Freesound, presents unique challenges compared to more structured media like movies or books. Audio samples often lack rich, structured metadata, forcing many systems to rely on Collaborative Filtering (CF). While CF is effective in many domains, it suffers from "cold start" problems for new items and fails to capture the underlying conceptual relevance of a sound.
User-generated tags offer a potential solution but are frequently noisy, ambiguous, or overly specific. This research addresses the challenge of "semantic enrichment"—transforming raw, ambiguous tags into structured knowledge to improve recommendation accuracy.
Methodology
The proposed system utilizes a hybrid approach that integrates collaborative data with a semantically enriched Knowledge Graph (KG).
Entity Linking and Knowledge Graph Expansion
The process begins with Entity Linking and Disambiguation. Using Babelfy, raw tags and descriptions are mapped to unique identifiers (URIs) in WordNet and DBpedia. This step resolves linguistic ambiguities, ensuring that a tag like "bark" is correctly identified as either a canine sound or a tree texture.
Once entities are identified, the system performs Knowledge Graph Expansion. It crawls the graph to find related concepts, such as WordNet hypernyms (broader terms) and DBpedia categories. To prevent "semantic drift," where a sound becomes related to irrelevant concepts through too many links, the system defines an $h$-hop neighborhood (typically 4 hops) and applies a logarithmic decay factor: $1 + \log(l)$, where $l$ is the distance from the original tag.
Feature Combination
The recommendation engine employs a Feature Combination strategy. It concatenates three distinct data types into a single high-dimensional vector:
- Graph-based features: Representing the semantic context.
- Collaborative features: Representing user download history.
- Textual features: Representing raw tags and keywords.
These vectors are then processed by a content-based recommender framework using graph kernels to determine item similarity.
Evaluation and Results
The model was evaluated using a dataset from Freesound spanning 2005 to 2013, encompassing 20,000 users and over 12 million downloads. The semantic hybrid approach was compared against several baselines: Popularity-based (MostPop), Matrix Factorization (BPR-MF), and Sparse Linear Methods (SLIM and BPR-SLIM).
As shown in the Precision-Recall curves, the proposed semantic hybrid approach significantly outperforms all baseline algorithms. The improvements were found to be statistically significant ($p < 0.01$). While BPR-SLIM emerged as the strongest baseline, it consistently lagged behind the hybrid model across the entire curve.
Conclusion
By shifting from "strings" to "things," this semantic approach allows for more nuanced recommendations. For instance, the system can recommend a "howl" to a user interested in "wolves" even if the specific keyword is missing.
Future work will address the computational overhead of building multi-hop graphs and explore metrics beyond accuracy, such as novelty and diversity, which are critical in creative audio production environments.
Summary updated