Deep Learning for Latent Events Forecasting in Content Caching Networks

Zhong Yang, Yuanwei Liu, Yue Chen, Joey Tianyi Zhou

Abstract

Traditional content caching strategies in mobile networks often rely on static popularity distributions, failing to account for real-time local context. This paper introduces the Twitter context Aided Content caching (TAC) framework, which leverages geo-tagged social media data to predict local content demand. By integrating Latent Dirichlet Allocation (LDA) for topic modeling and Long Short-Term Memory (LSTM) networks for location-aware forecasting, the TAC framework proactively caches media associated with emerging "latent events." Experimental results using a dataset of 2 million tweets from London demonstrate that TAC significantly improves cache hit rates and reduces delivery latency compared to traditional frequency-based methods.

Introduction

The exponential growth of mobile data traffic necessitates efficient edge caching at base stations to alleviate backhaul congestion and minimize user latency. Conventional strategies typically follow Zipf’s Law, assuming a static power-law distribution of content popularity. However, these methods are often "spatially blind" and lack the contextual awareness needed to anticipate hyper-local surges in demand driven by real-world events.

Social media platforms like Twitter serve as real-time sensors of public interest. By analyzing the textual content and geographic metadata of tweets, network operators can identify emerging trends—"latent events"—before they translate into massive content requests. The challenge lies in accurately mapping these colloquial, unstructured data points to specific network locations for proactive caching.

The TAC Framework

The proposed Twitter context Aided Content caching (TAC) framework consists of a multi-stage pipeline designed to extract, forecast, and deliver content based on social media signals.

Figure 1: Overview of the TAC framework architecture, illustrating the flow from Twitter data extraction to caching decisions.

Latent Event Extraction

The framework utilizes Latent Dirichlet Allocation (LDA) to perform unsupervised topic modeling. Unlike frequency-based models, LDA identifies clusters of co-occurring words, allowing the system to recognize holistic "events" (e.g., a local festival or sports match) rather than isolated keywords.

Popularity and Location Forecasting

To handle the temporal and spatial dynamics of these events, the framework employs a deep learning approach:

  1. Skip-gram Embedding: Tweets are converted into vector representations that capture semantic context.
  2. Geo-aware LSTM: A multi-layer LSTM network processes these sequences. Crucially, latitude and longitude coordinates are treated as tokens within the sequence, enabling the model to learn spatial-temporal correlations directly.
  3. Content Prediction: The model predicts the probability of specific topics trending at particular base stations, which is then mapped to associated media files (images/videos) via embedded URLs.
Figure 2: The Multi-layer LSTM Network architecture used for popularity forecasting, incorporating skip-gram pre-training.

Experimental Evaluation

The framework was evaluated using a dataset of approximately 2 million tweets collected in London over a 31-day period. Filtering for geo-tagged data yielded 70,000 tweets, which were mapped to nine geographic regions corresponding to base station clusters.

Key Results

  • Cache Performance: The TAC framework achieved a media caching hit rate of approximately 75%, outperforming Least Frequently Used (LFU) and Least Recently Used (LRU) schemes by 12% and 35%, respectively.
  • Latency Reduction: By integrating the predictive model with Non-Orthogonal Multiple Access (NOMA) for content delivery, the system reduced transmission delays by 25% compared to conventional OMA-based LFU schemes.
  • Scalability: Model performance, measured by perplexity, improved as the dataset size increased, indicating that the system benefits from higher volumes of social media data.

Conclusion and Future Work

The TAC framework demonstrates the viability of using social media as a proactive demand predictor for edge caching. By moving from reactive, frequency-based models to context-aware deep learning, network efficiency can be significantly enhanced. Future research directions include addressing data sparsity in geo-tagged content, integrating multi-platform data (e.g., Instagram, TikTok), and exploring cooperative caching strategies between adjacent base stations to further optimize resource allocation.

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 collected 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 because of the superiority of 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. Furthermore, we associate the predict latent events and the popularity of the events with the caching strategy. Lastly, we propose a non-orthogonal multiple access (NOMA) based content transmission scheme. Extensive practical experiments demonstrate that: 1) the proposed TAC framework outperforms 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 conventional LDA approach; and 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. Simulation results also shows that the proposed NOMA-enabled caching scheme outperforms conventional least frequently used (LFU) scheme by 25%.
BibTeX
@article{yang2022deep,
  author = {Zhong Yang and Yuanwei Liu and Yue Chen and Joey Tianyi Zhou},
  title = {{Deep Learning for Latent Events Forecasting in Content Caching Networks}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2022},
  doi = {10.1109/twc.2021.3096747},
}
Figure