Data-Driven Contex-Aware Pipelining for Multimodal Music Systems

Modern music platforms consume heterogeneous, multi-modal data: audio, imagery, text, and behavioral streams, yet raw signals from these sources rarely map directly onto the human-understandable concepts (mood, genre, context) that recommendation and discovery applications need. This thesis proposes a modular, input-agnostic multimedia pipeline that bridges this semantic gap through a three-layer architecture: Data Preparation, Semantic Enrichment, and a shared Semantic API, supporting both batch and streaming operation. As a concrete demonstration, a context-aware music Recommender Engine is implemented as one Layer-3 consumer of the generic Semantic API, spanning three independent use-case modes: batch/reactive, streaming/reactive, and streaming/proactive. This document reports the design of the full architecture and the implementation and empirical verification of both the batch/reactive (8.1) and streaming/reactive (8.2) modes over a 411-track licensed seed dataset, using CLAP audio embeddings, a Neo4j knowledge graph, Milvus vector search, and a Kafka/Flink/Redis streaming path, together with a feasibility-and-design analysis of the streaming/proactive mode (8.3), demonstrated through a minimal real vertical slice rather than implemented and evaluated to the same standard as the other two modes.