Using Social Media Background to Improve Cold-start Recommendation Deep Models

Yihong Zhang, Takuya Maekawa, Takahiro HaraarXiv 2021 · arXiv

TL;DR

This paper tackles the notorious cold-start problem in recommender systems by using general social media trends as a dynamic, temporal background. By extracting emerging topics from Twitter, converting them into semantic embeddings, and fusing them into a deep recommendation model via an attention mechanism, the authors significantly improve recommendation accuracy—in some cases more than doubling the hit rate.

What Problem Does This Paper Address?

Recommender systems excel when they can mine a rich history of user-item interactions. However, they run into a wall called the cold-start problem when a new user joins or a new item is introduced to the catalog. Traditional fixes rely on static contextual data, such as user demographics or item descriptions.

What these approaches miss is the fluid, real-world context of the moment. For instance, a sudden natural disaster might temporarily spike public interest in emergency supplies, or a trending food craze might drive coupon sales for a local restaurant. While previous research has tried to leverage social media to capture these trends, it usually required explicitly linking user accounts across platforms (a privacy and data-collection nightmare) or relied on crude aggregate metrics like binary sentiment scores.

This paper asks: Can we use completely unlinked, general social media chatter as a dynamic background context to improve cold-start recommendations?

Methods

The authors build upon an existing deep learning recommendation architecture (NeuMF) that combines Generalized Matrix Factorization (GMF) and a Multi-Layer Perceptron (MLP) to map user and item embeddings to a prediction score.

To inject social media context, they introduce a three-step pipeline:

  1. Time-Aware Target Modification: They segment the recommendation day into three distinct periods (morning, afternoon, and evening) to make the model sensitive to timing.
  2. Emergence Analysis: To separate background noise from actual trends, they use a foreground/background frequency comparison to detect words that are actively "surging" or trending on Twitter.
  3. Item Social Trend Encoder (ISTE): Instead of just averaging these trending words, they map them to dense semantic embeddings using Word2Vec. They then introduce an attention mechanism that dynamically aligns hourly social media trends with specific item embeddings. This allows the model to learn which real-world trends actually matter for which products.
Fig. 3. Fusing social media background with an existing model

Datasets

The study evaluates its approach using two primary datasets aligned to the same four-month period in 2017:

  • E-commerce Dataset: Provided by an industry partner, this dataset tracks 6,883 users, 18.4 million browsing records, and 85,963 purchases. The items are time-limited discount coupons ("deals") that naturally suffer from cold-start issues due to their short lifespans.
  • Social Media Dataset: A curated collection of 2,464,645 Japanese tweets from 33,443 accounts. To ensure the tweets reflected genuine domestic affairs rather than spam or international noise, the authors specifically tracked the active Japanese followers of local politicians.

Key Insights and Ideas

  • The "Steak vs. Cheese" Phenomenon: In a preliminary analysis, the authors discovered that some words have clear, lagged predictive power for e-commerce purchases, while others do not. For example, a spike in tweets containing the word "steak" statistically predicted an increase in steak coupon purchases 2 to 3 hours later. Conversely, "cheese" showed no such correlation.
  • Semantics Over Word Counts: Rather than relying on rigid, high-dimensional bag-of-words representations, projecting trending topics into a semantic embedding space allows the deep learning model to automatically discover latent relationships between real-world events and consumer preferences.
  • Attention is Key: Simply averaging social media embeddings across a time window provides a minor performance bump, but using an attention mechanism to match specific items to specific trends yields massive gains.

Key Results

The experimental results demonstrate that fusing social media background data steadily improves recommendation performance across the board:

  • Massive Top-K Gains: The improvements were most pronounced at small values of $K$ (the very top recommendations). For models using metapath-based user embeddings, adding the ISTE attention module improved the Hit Rate at 1 (HR@1) by 180% relative to the base model (jumping from 0.55% to 1.53%).
  • Better Ranking Quality: Normalized Discounted Cumulative Gain (NDCG), which measures ranking quality, increased by over 100% for $K \le 3$ when using the attention-fused model.
  • Consistency: The social-media-fused models consistently outperformed both the base models and a standard popularity-based baseline (PrevPop).

Limitations and Future Work

While the model proves that social media background data contains valuable predictive signals, it operates largely as a black box. The authors note that it is still difficult to explicitly diagnose why certain social media contexts predict specific product sales simply by looking at the model itself.

Additionally, the current system assumes that a user's purchase intent has already been detected at a given time $t$. In a production environment, the system would need an extra component to predict whether a user is in a shopping mindset before deciding to serve a trend-aware recommendation. The authors plan to tackle these explainability and intent-detection challenges in future work.

Summary updated

Abstract

In recommender systems, a cold-start problem occurs when there is no past interaction record associated with the user or item. Typical solutions to the cold-start problem make use of contextual information, such as user demographic attributes or product descriptions. A group of works have shown that social media background can help predicting temporal phenomenons such as product sales and stock price movements. In this work, our goal is to investigate whether social media background can be used as extra contextual information to improve recommendation models. Based on an existing deep neural network model, we proposed a method to represent temporal social media background as embeddings and fuse them as an extra component in the model. We conduct experimental evaluations on a real-world e-commerce dataset and a Twitter dataset. The results show that our method of fusing social media background with the existing model does generally improve recommendation performance. In some cases the recommendation accuracy measured by hit-rate@K doubles after fusing with social media background. Our findings can be beneficial for future recommender system designs that consider complex temporal information representing social interests.
BibTeX
@misc{zhang2021using,
  author = {Yihong Zhang and Takuya Maekawa and Takahiro Hara},
  title = {{Using Social Media Background to Improve Cold-start Recommendation Deep Models}},
  year = {2021},
  eprint = {2106.02256},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.02256},
}
Figure