Deep Learning for Latent Events Forecasting in Twitter Aided Caching Networks

Zhong Yang, Yuanwei Liu, Yue Chen, Joey Tianyi ZhouarXiv 2021 · DOI · arXiv

Abstract

A novel Twitter context aided content caching (TAC) framework is proposed for enhancing the caching efficiency by taking advantage of the legibility and massive volume of Twitter data. For the purpose of promoting the caching efficiency, three machine learning models are proposed to predict latent events and events popularity, utilizing collect Twitter data with geo-tags and geographic information of the adjacent base stations (BSs). Firstly, we propose a latent Dirichlet allocation (LDA) model for latent events forecasting taking advantage of the superiority of the LDA model in natural language processing (NLP). Then, we conceive long short-term memory (LSTM) with skip-gram embedding approach and LSTM with continuous skip-gram-Geo-aware embedding approach for the events popularity forecasting. Lastly, we associate the predicted latent events and the popularity of the events with the caching strategy. Extensive practical experiments demonstrate that: (1) The proposed TAC framework outperforms the conventional caching framework and is capable of being employed in practical applications thanks to the associating ability with public interests. (2) The proposed LDA approach conserves superiority for natural language processing (NLP) in Twitter data. (3) The perplexity of the proposed skip-gram-based LSTM is lower compared with the conventional LDA approach. (4) Evaluation of the model demonstrates that the hit rates of tweets of the model vary from 50% to 65% and the hit rate of the caching contents is up to approximately 75% with smaller caching space compared to conventional algorithms.

Introduction

As global mobile data traffic continues to skyrocket—projected to reach 77 exabytes per month—network providers are struggling with backhaul congestion. Edge caching, which involves storing popular content at base stations closer to users, is a primary solution. However, its effectiveness is limited by the difficulty of predicting what users actually want to consume in real-time.

Traditional caching strategies often rely on the Zipf distribution, which assumes content popularity follows a static power law. These models are often insufficient because they ignore the semantic concurrency of words and the hyper-local, fast-changing nature of public interest. This paper introduces the Twitter Context Aided Content Caching (TAC) framework, which uses Twitter as a real-time sensor to understand local trends, allowing base stations to proactively cache relevant media before users even request them.

Methodology

The TAC framework utilizes a multi-stage machine learning pipeline to bridge the gap between social media trends and network caching.

The proposed Twitter Context Aided Content Caching (TAC) framework architecture.

1. Latent Event Extraction (LDA)

Latent Dirichlet Allocation (LDA) is used to identify "latent events" or topics from a corpus of tweets. Unlike simple word-counting, LDA identifies clusters of words that frequently appear together, providing a nuanced understanding of local trends, such as a specific sports match or festival.

2. Popularity Forecasting (LSTM)

To predict which events will trend next, the framework employs Long Short-Term Memory (LSTM) networks. Two embedding approaches were compared:

  • Skip-gram Embedding: A standard NLP technique associating words based on context.
  • Continuous Skip-gram-Geo-aware Embedding: A novel approach where latitude and longitude tokens are appended to tweet vectors, allowing the model to learn the relationship between what is being said and where it is being said.

3. Caching Strategy

The framework associates predicted topics with actual media URLs (images/videos) found in tweets. It maintains a "Prior List" (PL) for each base station, dynamically updating the cache based on the predicted probability of an event's occurrence in that specific region.

Experimental Setup and Datasets

The study utilized a custom-collected dataset from London, UK, spanning January to February 2018. Approximately 70,000 geo-tagged tweets were processed. The London area was divided into nine sub-regions based on base station density to test localized caching performance. Data cleaning involved removing non-English characters and punctuation, while URLs were extracted to link text topics to actual media files.

Key Results

  • High Hit Rates: The TAC framework achieved a tweet hit rate of 50% to 65%. Most significantly, the hit rate for actual cached media contents reached approximately 75%.
  • Efficiency and Storage Optimization: The proposed machine learning approach outperformed conventional Least Frequently Used (LFU) and Least Recently Used (LRU) algorithms. While LFU often requires caching up to 90% of content to achieve results, the TAC framework maintained high performance with a "medium" cache portion of about 75%, proving it is more selective and efficient with limited edge storage.
  • Perplexity: The Geo-aware LSTM showed significantly lower perplexity compared to standard models, indicating higher predictive accuracy.

Conclusion and Future Work

The TAC framework demonstrates that local social media data is a powerful leading indicator of local data demand. By treating geographic coordinates as tokens within a word embedding space, the authors successfully transformed a standard NLP problem into a spatial-temporal forecasting tool.

Future research will address current limitations, such as "monolithic" predictions where the model tends to predict a single outcome for an area rather than multiple concurrent events. Additionally, expanding the framework to include broader web traffic and addressing data sparsity in regions with fewer geo-tagged tweets remain key areas for development.

Summary updated

Abstract

A novel Twitter context aided content caching (TAC) framework is proposed for enhancing the caching efficiency by taking advantage of the legibility and massive volume of Twitter data. For the purpose of promoting the caching efficiency, three machine learning models are proposed to predict latent events and events popularity, utilizing collect Twitter data with geo-tags and geographic information of the adjacent base stations (BSs). Firstly, we propose a latent Dirichlet allocation (LDA) model for latent events forecasting taking advantage of the superiority of the LDA model in natural language processing (NLP). Then, we conceive long short-term memory (LSTM) with skip-gram embedding approach and LSTM with continuous skip-gram-Geo-aware embedding approach for the events popularity forecasting. Lastly, we associate the predicted latent events and the popularity of the events with the caching strategy. Extensive practical experiments demonstrate that: (1) The proposed TAC framework outperforms the conventional caching framework and is capable of being employed in practical applications thanks to the associating ability with public interests. (2) The proposed LDA approach conserves superiority for natural language processing (NLP) in Twitter data. (3) The perplexity of the proposed skip-gram-based LSTM is lower compared with the conventional LDA approach. (4) Evaluation of the model demonstrates that the hit rates of tweets of the model vary from 50% to 65% and the hit rate of the caching contents is up to approximately 75\% with smaller caching space compared to conventional algorithms.
BibTeX
@misc{yang2021deep,
  author = {Zhong Yang and Yuanwei Liu and Yue Chen and Joey Tianyi Zhou},
  title = {{Deep Learning for Latent Events Forecasting in Twitter Aided Caching Networks}},
  year = {2021},
  eprint = {2101.01149},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2101.01149},
}
Figure