Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Time and Place to Preference: LLM-Driven Geo-Temporal Context in Recommendations

Yejin Kim, Shaghayegh Agah, Mayur Nankani, Neeraj Sharma, Feifei Peng, Maria Peifer, Sardar Hamidian, H Howie Huang

TL;DR

The paper tackles the challenge of enriching recommendations with real-world geo-temporal context by extracting semantic signals from simple inputs (timestamp and coarse location) using large language models. It introduces a structured geo-temporal context extraction module, a lightweight informativeness test, and multiple integration strategies into sequential recommenders (notably SASRec), evaluated on MovieLens, LastFM, and a production dataset. Key findings show that geo-temporal signals can provide predictive utility, with general users benefiting from content+geo-temporal fusion and explorer users gaining from item-level or training-time GT guidance; production data particularly benefits from GT integration and flexible deployment. The work contributes a scalable, domain-agnostic approach to context-aware recommendations and provides a context-enriched MovieLens dataset to spur further research and practical deployments.

Abstract

Most recommender systems treat timestamps as numeric or cyclical values, overlooking real-world context such as holidays, events, and seasonal patterns. We propose a scalable framework that uses large language models (LLMs) to generate geo-temporal embeddings from only a timestamp and coarse location, capturing holidays, seasonal trends, and local/global events. We then introduce a geo-temporal embedding informativeness test as a lightweight diagnostic, demonstrating on MovieLens, LastFM, and a production dataset that these embeddings provide predictive signal consistent with the outcomes of full model integrations. Geo-temporal embeddings are incorporated into sequential models through (1) direct feature fusion with metadata embeddings or (2) an auxiliary loss that enforces semantic and geo-temporal alignment. Our findings highlight the need for adaptive or hybrid recommendation strategies, and we release a context-enriched MovieLens dataset to support future research.

From Time and Place to Preference: LLM-Driven Geo-Temporal Context in Recommendations

TL;DR

The paper tackles the challenge of enriching recommendations with real-world geo-temporal context by extracting semantic signals from simple inputs (timestamp and coarse location) using large language models. It introduces a structured geo-temporal context extraction module, a lightweight informativeness test, and multiple integration strategies into sequential recommenders (notably SASRec), evaluated on MovieLens, LastFM, and a production dataset. Key findings show that geo-temporal signals can provide predictive utility, with general users benefiting from content+geo-temporal fusion and explorer users gaining from item-level or training-time GT guidance; production data particularly benefits from GT integration and flexible deployment. The work contributes a scalable, domain-agnostic approach to context-aware recommendations and provides a context-enriched MovieLens dataset to spur further research and practical deployments.

Abstract

Most recommender systems treat timestamps as numeric or cyclical values, overlooking real-world context such as holidays, events, and seasonal patterns. We propose a scalable framework that uses large language models (LLMs) to generate geo-temporal embeddings from only a timestamp and coarse location, capturing holidays, seasonal trends, and local/global events. We then introduce a geo-temporal embedding informativeness test as a lightweight diagnostic, demonstrating on MovieLens, LastFM, and a production dataset that these embeddings provide predictive signal consistent with the outcomes of full model integrations. Geo-temporal embeddings are incorporated into sequential models through (1) direct feature fusion with metadata embeddings or (2) an auxiliary loss that enforces semantic and geo-temporal alignment. Our findings highlight the need for adaptive or hybrid recommendation strategies, and we release a context-enriched MovieLens dataset to support future research.

Paper Structure

This paper contains 17 sections, 7 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLMs for Contextual Enrichment
  4. Approach
  5. Prompt Design for Geo-Temporal Context
  6. Geo-Temporal Embedding Informativeness Test
  7. Model Integration
  8. Experiments
  9. Dataset and Contextual Enrichment
  10. Evaluating Geo-Temporal Embedding
  11. Evaluation Metrics
  12. Model Variants and Terminology
  13. Main Results
  14. Geo-Temporal Context Adaptation
  15. Qualitative Results
  16. ...and 2 more sections

Figures (2)

  • Figure 1: From timestamp and location to LLM-generated geo-temporal context for sequential recommendation.
  • Figure 2: (a) Baseline model (SASRec) (b) Auxiliary loss contrasting semantically and geo-temporally similar vs. dissimilar items (c) Replaced item Id embeddings with metadata embeddings as input, and added geo-temporal embeddings to metadata embeddings at the loss stage (d) Concatenated item Id and metadata embeddings as input, and added geo-temporal embeddings at the loss stage