October 25, 2023 Published by Enrico Palumbo, Andreas Damianou, Francesco Fabbri, Hugues Bouchard, Mounia Lalmas
Spotify’s search system performs an important function in serving to customers discover the catalog and uncover new content material. Exploratory searches corresponding to “new funk soul releases” or “guitar solos to learn” present a chance for customers to seek out their subsequent favourite music and for the platform to floor under-served content material [1]. Query suggestion companies corresponding to Related Searches assist customers formulate efficient queries and are extraordinarily helpful in exploratory search, the place the invention course of typically requires a number of iterations of question formulation. For instance, think about that the person is on the lookout for music to stream throughout a yoga session. The person initially varieties a generic ‘yoga’ question, however after seeing the question strategies, the person opts for ‘vinyasa’ and at last refines the intent with ‘vinyasa flow’.
Related Searches assist customers navigate the catalog by recommending associated queries that facilitate the exploration course of, resulting in discovery and simpler searches.
Graph Learning for Query Suggestions
To facilitate exploratory searches by way of question strategies, we developed a graph studying strategy. Graphs are very best buildings for modeling connections and associations, and so they can simply embody heterogeneous information corresponding to queries, objects, matters, musical genres. Graph studying strategies study patterns in these buildings and encode them into vectors that may successfully be used for advice by way of nearest neighbor search. In this work, we use node2vec [2] to study the graph construction and the vectorial representations.
Graph Learning algorithms (e.g. node2vec) study patterns in graph buildings and encode them in a vector house. Then, question suggestions will be obtained by discovering the closest neighbors to the person question.
We noticed in offline experiments that the graph studying answer node2vec produces correct question strategies (+22% with respect to the most effective baseline, i.e. a transformer mannequin based mostly on semantic similarity), whereas maintaining a excessive degree of variety, which is important to permit customers to examine completely different exploration paths.
Hence, we ran an A/B check experiment with this mannequin on thousands and thousands of customers the place the node2vec mannequin was added as a brand new supply of question strategies. The outcomes confirmed that this mannequin results in vital enhancements in protection (+1.42%), and variety of clicks on question strategies (+1.21%), with out affecting the general latency of the system. We observe an particularly marked enhance for clicks on exploratory queries (+9.37%), proving our speculation that the graph studying answer can assist exploration by way of search.
Conclusions
Graphs can join heterogeneous information (e.g. queries, songs, artists, podcasts, matters, musical genres), facilitating associations that assist exploratory searches. Graph Learning permits one to simply create graph-based recommender techniques by mapping graph patterns right into a vector house. We developed a question suggestion mannequin based mostly on graph studying and show that it helps exploratory searches in a large-scale on-line experiment. We consider that these outcomes show the potential of graph-based strategies for search and lots of different core merchandise at Spotify and elsewhere.
For extra info, please confer with our paper:
Graph Learning for Exploratory Query Suggestions in an Instant Search System
Enrico Palumbo, Andreas Damianou, Alice Wang, Alva Liu, Ghazal Fazelnia, Francesco Fabbri, Rui Ferreira, Fabrizio Silvestri, Hugues Bouchard, Claudia Hauff, Mounia Lalmas, Ben Carterette, Praveen Chandar, David Nyhan
CIKM23
References
[1] Tomasi, Federico, et al. “Query Understanding for Surfacing Under-served Music Content.” CIKM 2020.
[2]: Grover, Aditya, and Leskovec, Jure. “node2vec: Scalable feature learning for networks.” Proceedings of the twenty second ACM SIGKDD worldwide convention on Knowledge discovery and information mining. 2016.