Amazon recommendationsJanuary – May 2025

Finding relevancein sparse data.

Hybrid recommendation system for the Amazon Video Games dataset

A team-built hybrid recommendation system combining collaborative filtering and content signals, evaluated on more than 99% sparse user–item interactions.

My contribution

ALS model evaluation, testing, and analysis, alongside participation in project planning and implementation.

Interactions → relevance

Sparse interactionsHybrid recommendations

Conceptual project illustration

Interaction sparsity: >99%
Collaborative filtering: ALS
Evaluation measures: 4

Dataset context and evaluation approach, not claimed improvements in recommendation quality.

The problem

Most users interact with only a small fraction of a product catalog. In the Amazon Video Games dataset used for this course project, the user–item interaction matrix was more than 99% sparse. That made limited interaction history and cold-start cases central to the recommendation problem.

System overview

Interaction dataImplicit ALSCollaborative signals

Item textSentence embeddingsCosine similarity

Hybrid recommendation layerWeighted signals → recommendations

High-level system overview, based on the project description. The two branches contribute collaborative and content-based signals.

Combining two sources of relevance

The team explored implicit ALS collaborative filtering alongside content-based recommendations. The content path used all-MiniLM-L6-v2 sentence embeddings and cosine similarity to compare item text.

The hybrid approach combined these signals through weighted scoring. The project also explored frequent-item and association patterns, co-visited items, and cold-start strategies.

My focus: evaluating the system

My main contribution centered on evaluating the recommendation system and ALS model. I tested different latent-factor configurations and top-k recommendation settings, using RMSE, Precision@k, Recall@k, and F1@k.

This work focused on comparing model behavior under sparse interactions. I also participated in project planning and implementation; the full recommendation engine was a team effort.

RMSE: error in estimated interaction values
Precision@k: relevance among the top-k recommendations
Recall@k: coverage of relevant items in the top-k set
F1@k: the balance between precision and recall

What the evaluation covered

The evaluation examined latent-factor and top-k choices rather than assuming that a larger model or a longer recommendation list would always be better. The different measures provided complementary views of the system.

Scope and limitations

This was an Intelligent Recommendation Systems course project using an offline dataset. Offline evaluation does not establish how a deployed recommendation product would perform with real users.

The project provided experience with model evaluation, sparse interaction data, and hybrid recommendation design. It was not presented as a production e-commerce deployment.

Technologies

Python
pandas
NumPy
SciPy
implicit ALS
Sentence Transformers
PostgreSQL

Next case studyWalaw Aya