Event Recommendation Pipeline

Overview

This project implements an end-to-end event recommendation pipeline designed to retrieve and rank relevant activities for users based on preferences, event metadata, and ranking signals.

The system demonstrates how real-world recommendation systems are structured, including:

• Data ingestion and preprocessing
• Feature engineering
• Candidate retrieval
• Ranking
• Offline evaluation
• API serving
• Interaction logging

The pipeline operates on structured event datasets and simulated user interactions.

---

System Architecture

The system is divided into three main stages:

1. Offline Pipeline (Model Building)

• Load events from database
• Generate text and structured features
• Build hybrid feature vectors
• Build FAISS index for retrieval
• Save artifacts for serving

Output artifacts:

• TF-IDF vectorizer
• Feature scaler
• FAISS index
• Metadata
• Hybrid feature matrix

---

2. Online Serving Layer

• REST API built using Flask
• Accepts user_id and optional filters
• Retrieves candidates using FAISS
• Applies ranking logic using:
  • Similarity score
  • Recency
  • Popularity
  • User preferences
• Returns ranked recommendations

---

3. Interaction Logging & Evaluation

• Logs user interactions (click/no-click)
• Computes offline metrics:
  • Precision@K
  • Recall@K
  • NDCG
  • MAP
  • CTR (offline simulation)

---

Database Schema

Events

Stores event metadata used for recommendation.

Fields:

• event_id
• event_name
• event_type
• start_time
• end_time
• date
• event_location
• seats_available
• description

Derived features:

• timing bucket (Morning/Afternoon/Evening/Night)
• recency score
• popularity score

---

Users

Stores user preferences.

Fields:

• user_id
• preferred_types
• preferred_timings
• preferred_duration

---

Interactions

Stores behavioral signals.

Fields:

• interaction_id
• user_id
• event_id
• clicked
• timestamp

---

Feature Engineering

The system uses a hybrid feature approach:

Text features:

• Event type
• Name
• Description
• Location

Structured features:

• Seats (normalized)
• Recency
• Timing bucket

Hybrid vector

---

Retrieval and Ranking

Candidate Retrieval

• FAISS index
• Vector similarity search

Ranking Signals

Final ranking score combines:

• Similarity
• Popularity
• Recency
• Seat availability

---

Offline Evaluation

Ranking quality is evaluated using:

• Precision@K
• Recall@K
• NDCG
• MAP
• CTR (simulated)

Evaluation is performed using:

• Ground-truth preferences
• Interaction logs

---

Example API Request

POST /recommend

{
  "user_id": 3
}
{
  "user_id": 3,
  "event_types": ["Tech Meetups"],
  "timings": ["Evening"]
}
{
  "latency_ms": 65,
  "results": [...]
}

---

Separation of Concerns

The system separates:

• Storage layer (SQLite)
• Retrieval layer (FAISS)
• Ranking layer
• API layer
• This mirrors production recommender architectures.

Limitations

• Uses simulated interactions instead of real user behavior
• Small user population
• No online learning or retraining loop
• No geographic distance ranking yet

Future Improvements

• Collaborative filtering
• User embeddings
• Location-aware ranking
• Real-time feedback loop
• Periodic model retraining
• A/B testing framework