Source-linked AI summary
Metadata Embeddings for User and Item Cold-start Recommendations
Maciej Kula
TL;DR
Cold-start recommendation suffers when sparse interactions prevent reliable latent-factor estimation. The paper presents LightFM, a hybrid model that builds user and item representations from content-feature embeddings while learning from collaborative interactions. It performs at least as well as specialised models across sparsity settings and produces semantic feature embeddings usable for tag recommendation.
Problem
Sparse interaction data makes conventional matrix factorisation unreliable for new users and items, while content-based models lack transfer learning across users.
Method
LightFM represents users and items as linear combinations of content-feature embeddings and uses a sigmoid-linked interaction model for binary recommendation data.
Results
LightFM performs at least as well as specialised models across collaborative-data sparsity scenarios and produces feature embeddings that capture semantic information for related tasks.
Takeaways & Limitations
The model is applicable in both cold- and warm-start settings, while its feature embeddings can support tasks such as tag recommendations.
Takeaways & Limitations
LightFM’s tag performance depends on high-quality metadata, and its current formulation lacks an easy way to incorporate visual or audio features directly.
Abstract
from arXiv · showhide
I present a hybrid matrix factorisation model representing users and items as linear combinations of their content features' latent factors. The model outperforms both collaborative and content-based models in cold-start or sparse interaction data scenarios (using both user and item metadata), and performs at least as well as a pure collaborative matrix factorisation model where interaction data is abundant. Additionally, feature embeddings produced by the model encode semantic information in a way reminiscent of word embedding approaches, making them useful for a range of related tasks such as tag recommendations.
1. INTRODUCTION
Cold-start recommendation is difficult because sparse interactions make latent-factor estimation unreliable, while pure content-based methods do not transfer learning across users. LightFM combines content and collaborative information, matching specialised models across sparse and dense settings.
- Sparse interactions make standard matrix factorisation poorly suited to new users and items.
- Content-based models support new-item recommendations but require substantial per-user data and do not benefit from other users’ interactions.
- LightFM represents users and items through linear combinations of latent embeddings for their content features.
- LightFM performs at least as well as pure content-based models in cold-start and low-density scenarios, with larger gains when collaborative information or user features are available.
- In warm-start settings with abundant collaborative data, LightFM performs at least as well as matrix factorisation.
- Its feature embeddings encode semantic information and support related tasks such as tag recommendations.
- Using both user and item metadata makes LightFM applicable to user and item cold-start scenarios, while one model covers dense and sparse data.
2. LIGHTFM
LightFM is designed to learn interaction-based representations while computing predictions immediately for users and items described by metadata. Its feature-based structure connects matrix factorisation with content-based modelling and adapts across sparsity levels.
- 2.1 Motivation: LightFM must learn user and item representations from interactions while supporting recommendations for new users and items.
- 2.1 Motivation: Interaction-based embeddings can transfer preference information between similar features, such as recommending ball gowns after pencil skirts.
- 2.1 Motivation: It represents users and items as sums of embeddings for their known content features, enabling immediate cold-start predictions.
- 2.2 The Model: The model predicts interactions from the dot product of user and item representations, adjusted by feature biases, with a sigmoid link for binary data.
- 2.3 Relationship to Other Models: LightFM reduces to standard matrix factorisation with only user and item indicators, while metadata features extend it beyond that model.
- 2.3 Relationship to Other Models: Fewer metadata parameters can reduce overfitting and improve generalisation when training data is limited.
- 2.3 Relationship to Other Models: Metadata-based representations make cold-start predictions possible, and indicator-only features should perform on par with standard matrix factorisation.
- 2.3 Relationship to Other Models: LightFM contains pure content-based and matrix-factorisation cases at opposite sparsity extremes, and empirical results show it matches the appropriate specialised model.
3. RELATED WORK
LightFM extends hybrid recommender approaches by jointly using user-item, item-feature, and user-feature structure. Its formulation emphasises a single collaborative objective and interpretable feature embeddings.
- Prior hybrid models jointly model content and collaborative data to address cold-start recommendation.
- Soboroff et al. use collaborative information in user-feature representations but model preferences over individual features rather than item feature sets.
- Saveski et al. jointly factorise user-item and item-feature matrices with a weight controlling their relative reconstruction accuracy.
- LightFM uses a single objective to factorise the user-item matrix, avoiding a separate balance between collaborative and content reconstruction losses.
- Shmueli et al. model item features but do not extend their approach to user features.
- LightFM jointly factorises user-item, item-feature, and user-feature matrices and can be viewed theoretically as a Factorisation Machines special case.
- By restricting interactions to user and item features, LightFM supports interpretable feature embeddings.
4. DATASETS
The evaluation uses MovieLens for dense interactions and CrossValidated for sparse question-answering recommendations. Both datasets include item tags, while CrossValidated also provides user metadata.
- The two datasets span dense interaction data, where matrix factorisation is expected to work well, and sparse data, where content-based models tend to perform better.
- MovieLens 10M is combined with the Tag Genome tag set for the first experiment.
- MovieLens contains approximately 10 million ratings from 71,567 users on 10,681 movies, with genre and tag metadata.
- After binarisation and tag filtering, the MovieLens dataset contains 69,878 users, 10,681 items, 9,996,948 interactions, and 1030 unique tags.
- CrossValidated contains 5,953 users, 44,200 questions, 188,865 answers and comments, and 1032 question tags.
- CrossValidated user metadata comes from users’ profile About Me sections.
- The recommendation task treats answered questions as positive signals and unanswered questions as negative signals, with three sampled negatives per positive training pair.
- The CrossValidated model focuses on willingness to answer rather than user expertise.
5. EXPERIMENTAL SETUP
The experiments compare matrix-factorisation, content-based, hybrid, and LightFM variants using warm-start and item cold-start evaluations. Accuracy is measured with ROC AUC across repeated train-test splits, with user and item feature configurations tested.
- Experimental design: The evaluation uses warm-start and item cold-start experiments, with 20% of interactions or all interactions for 20% of items assigned to testing.Warm-start retains every user and item in training, whereas item cold-start removes training interactions for selected items, leaving metadata available.
- Evaluation metric: Model accuracy is measured by mean ROC AUC, the average probability that a positive item ranks above a negative item across test users.The final score averages results from 10 repeated 80% training and 20% test splits.
- Compared models: The comparison includes conventional MF, two LSI-based baselines, and LightFM models using tags, tags plus item indicators, or tags plus user features.User features are available only for CrossValidated and are derived from users’ profile About Me text.
- Compared models: LightFM models use latent vectors defined from content features, while the tags and tags-plus-item-indicators variants describe users only with indicator features.The tags-plus-user-feature variant incorporates both item and user metadata when available.
- Training: Training uses stochastic gradient descent with learning rate 0.05 and 64 latent dimensions across models and experiments.The dimensionality is selected as a balance between accuracy and production computational cost, with early stopping used for regularisation.
6. EXPERIMENTAL RESULTS
Across sparse, dense, and cold-start scenarios, LightFM matches or outperforms specialised recommendation models. Its embeddings also capture feature semantics and support related recommendation tasks, although tag-based performance depends on high-quality metadata.
- Recommendation accuracy: LightFM outperforms or matches the specialised model for each evaluated scenario.This includes both sparse and dense interaction settings.
- Recommendation accuracy: In warm-start MovieLens, LightFM using tag and item indicator features slightly outperforms MF.The result suggests metadata features can remain valuable when interaction data is abundant.
- Recommendation accuracy: Across cold-start datasets, LightFM variants outperform other models, with LightFM (tags + about) providing the best performance.On MovieLens, LightFM (tags + indicators) slightly outperforms LightFM (tags), while MF performs no better than random.
- Scope boundary: LightFM performance depends on the availability of high-quality metadata, which may come from descriptions, tagging, or computer vision systems.The embeddings can also assist the tagging process by suggesting related tags.
- Parameter sensitivity: LightFM performs well with small latent dimensions, consistently outperforming other models and matching 512-dimensional LSI-LR with fewer than 32 dimensions on CrossValidated.Smaller vectors reduce memory use and query-time computation while preserving recommendation accuracy.
- Embedding applications: LightFM feature embeddings capture semantic relationships and support tag, genre or category recommendation, and recommendation justification.Tag similarity is measured using cosine similarity, while feature distances can help explain recommendations.
7. USAGE IN PRODUCTION SYSTEMS
LightFM is deployed as a production recommender at Lyst, combining online feature updates, metadata-derived representations, and scalable nearest-neighbour search. Its engineering choices support continual adaptation and predictable query performance across a catalogue of millions of products.
- Online training: LightFM is trained online, continually updating existing feature representations and creating representations for previously unseen users and items.Model state and accumulated squared-gradient information are stored in a database so training can resume as new interaction data arrives.
- Online training: New products require no retraining because their representations are computed immediately as sums of their feature representations.
- Feature engineering: Product representations combine textual features, structured metadata, moderator-derived categories, and tags detected from product images by deep convolutional neural networks.
- Nearest-neighbour search: Related-product recommendations across 8 million products use approximate nearest-neighbour search for on-demand queries and exact search for near-line computation.
- Nearest-neighbour search: Random Projection trees provide more predictable query performance than LSH by splitting points near the median distance at each internal node.This allocates approximately half the points to each leaf, avoiding highly unbalanced hash-code distributions that can make queries slow.
8. CONCLUSIONS AND FUTURE WORK
The paper concludes that LightFM is an effective hybrid recommender for both sparse and abundant collaborative data, while its feature embeddings support related tasks. It also identifies improved training objectives and direct multimodal modelling as future directions.
- Conclusions: LightFM performs at least as well as a specialised model across collaborative-data sparsity scenarios and outperforms existing content-based and hybrid models in specified cold-start settings.
- Conclusions: LightFM produces content-feature embeddings that capture important semantic information and support related tasks such as tag recommendations.
- Conclusions: These properties make LightFM applicable in both cold-start and warm-start settings.
- Future work: The model could be extended with more sophisticated training objectives, including Weighted Approximate-Rank Pairwise loss or direct mean reciprocal rank optimisation.
- Future work: The present formulation has no easy way to incorporate visual or audio features directly.The current workaround uses CNN-generated binary tags, while the paper conjectures that recommendation-loss training could improve results.