MyEachtraX: Semantic Event Retrieval with Time-Weighted Fusion

We present an enhanced version of MyEachtraX, a mobile-first lifelog question answering (QA) system developed for the Lifelog Search Challenge 2025.The system centers the QA process around early, high-quality retrieval of relevant lifelog events, using visual, spatial, and temporal context to guide answer generation.To address the unique challenges of lifelog data, including redundancy, ambiguity, and noisy signals, we propose a threefold set of improvements.First, we apply data quality filtering using semantic density estimation and perceptual hashing to eliminate low-information and duplicate images.Second, we integrate visual, spatial, and temporal cues into a unified scoring framework, enabling contextaware retrieval for complex, real-world queries.Third, we mitigate embedding hubness through query and gallery bank normalisation, enhancing ranking diversity and stability.Evaluations on LSC'24 tasks show significant improvements in top- retrieval accuracy, particularly for time-sensitive and semantically rich queries, demonstrating the effectiveness of retrieval-focused design in lifelog QA systems.