CoupleNet: Paying Attention to Couples with Coupled Attention for Relationship Recommendation

Yi Tay, Anh Tuan Luu, Siu Cheung HuiICWSM 2018 · OpenReview

TL;DR

This paper introduces COUPLENET, a novel deep learning model for relationship recommendation that moves beyond traditional dating site recommendations. By leveraging a new dataset of real-world couples from Twitter and employing a coupled attention mechanism, COUPLENET can predict relationship compatibility based solely on social media profiles, even without explicit relationship signals.

What Problem Does This Paper Address?

The paper tackles the problem of "relationship recommendation" (RSR), which is distinct from the more common "online dating recommendation" problem. While dating recommendation systems focus on predicting reciprocal interest (e.g., likes or matches) on platforms like Tinder, this work aims to predict long-term relationship compatibility between users on regular social networks (RSNs) like Twitter. The authors argue that signals from RSNs are more indicative of genuine compatibility, as dating sites may encourage self-presentation bias. A key challenge is the lack of publicly available datasets for this specific task, and the difficulty in inferring genuine relationship status from social media.

Methods

COUPLENET employs an end-to-end deep learning architecture. It utilizes a "Siamese" network structure, meaning two identical networks process each user's profile independently, sharing parameters.

  1. User Profiling: Hierarchical Gated Recurrent Units (GRUs) are used to encode each user's tweets, learning tweet-level representations.
  2. Tweet-level Attention: An attention mechanism is applied to the GRU outputs to create a single, informative vector representation for each tweet, filtering out noise.
  3. Coupled Attention: This is the core novelty. It learns to attend to parts of one user's profile dynamically based on the other user's profile. This is achieved by computing a similarity grid between all tweets of user 1 and user 2, and then using this grid to learn weighted representations for each user. This allows the model to identify mutual interests and relevant signals between the pair.
  4. Learning to Rank: The final user representations are fed into a similarity modeling layer that uses cosine similarity, and the model is trained using a pairwise margin-based hinge loss to distinguish between true couples and negative samples.

Datasets

The paper introduces the LOVEBIRDS2M dataset, a novel benchmark for relationship recommendation. It was constructed using a distant supervision scheme from Twitter. The process involved:

  • Crawling tweets containing "heart" emojis and love-related keywords.
  • Filtering these tweets to remove non-romantic affections (e.g., family, music).
  • Further filtering to remove tweets mentioning multiple people or from highly popular accounts.
  • For the remaining ~12K "cleaned" couple tweets, the timelines of the involved users were crawled, collecting 200 latest tweets per user.
  • Explicit couple information (like "love" keywords and @mentions) was removed from these timelines to prevent direct signal leakage.

The final dataset consists of approximately 1.858 million tweets from 9290 users, forming 4645 couple pairs. This dataset is split into training (80%), testing (10%), and development (10%) sets.

Key Insights and Ideas

  • RSNs as a Richer Source: Social media profiles on RSNs, despite lacking explicit relationship status, contain valuable linguistic cues for inferring personality, interests, and ultimately, compatibility.
  • Beyond Reciprocal Signals: True relationship compatibility is more nuanced than simple reciprocal liking or messaging found on dating sites.
  • Coupled Attention for Mutual Understanding: The coupled attention mechanism is crucial for capturing how users' interests and profiles align or complement each other, going beyond individual user profiling.
  • Explainability through Attention: The attention maps provide insights into which tweets and topics contributed most to the recommendation, offering a degree of interpretability.
  • Implicit Signals Matter: COUPLENET can successfully rank couple pairs even when explicit matching signals are absent, suggesting it learns latent patterns like emotions and personality.

Key Results

  • Significant Performance Gains: COUPLENET significantly outperforms strong baselines like DeepCoNN and GRU models, achieving up to a 180-200% relative improvement in precision metrics like Hit Ratio (HR@N).
  • High Hit Ratio: The model achieves a HR@10 of approximately 64%, meaning that for a given user, the true partner appears within the top 10 recommendations 64% of the time. Accuracy (ranking the correct partner at the top position out of 100 candidates) is around 25%.
  • Impact of Tweet Count: Performance consistently improves with an increased number of tweets per user, indicating that more data leads to better user representations, though with diminishing returns.
  • Ablation Study Highlights: The pairwise ranking loss, coupled attention, and GRU attention layers are all critical components for COUPLENET's performance. Removing pairwise ranking or cosine similarity significantly degrades results.
  • Efficient Training: The model trains relatively quickly, with approximately 2 minutes per epoch on a medium-grade GPU for 200 tweets per user.

Limitations and Future Work

  • Data Scarcity: While LOVEBIRDS2M is a significant step, the authors acknowledge that more user pairs would be beneficial for a larger benchmark.
  • Interpretability Challenges: While attention provides some explainability, it can be difficult to interpret when user profiles lack explicit signals or when the matching is based on subtle emotional or personality traits.
  • Privacy Concerns: The authors note that while their dataset is publicly obtainable, using conversation history from dating sites raises privacy concerns.
  • Future Work: The authors suggest exploring other social platforms like Instagram (though noting challenges with multi-modal data and masking) and further refining the interpretability of the model. They also propose that the model's output could be used to enable more efficient analysis by humans, such as social scientists.

Summary updated

Abstract

Dating and romantic relationships not only play a huge role in our personal lives but also collectively influence and shape society. Today, many romantic partnerships originate from the Internet, signifying the importance of technology and the web in modern dating. In this paper, we present a text-based computational approach for estimating the relationship compatibility of two users on social media. Unlike many previous works that propose reciprocal recommender systems for online dating websites, we devise a distant supervision heuristic to obtain real world couples from social platforms such as Twitter. Our approach, the CoupleNet is an end-to-end deep learning based estimator that analyzes the social profiles of two users and subsequently performs a similarity match between the users. Intuitively, our approach performs both user profiling and match-making within a unified end-to-end framework. CoupleNet utilizes hierarchical recurrent neural models for learning representations of user profiles and subsequently coupled attention mechanisms to fuse information aggregated from two users. To the best of our knowledge, our approach is the first data-driven deep learning approach for our novel relationship recommendation problem. We benchmark our CoupleNet against several machine learning and deep learning baselines. Experimental results show that our approach outperforms all approaches significantly in terms of precision. Qualitative analysis shows that our model is capable of also producing explainable results to users.
BibTeX
@inproceedings{tay2018couplenet,
  author = {Yi Tay and Anh Tuan Luu and Siu Cheung Hui},
  title = {{CoupleNet: Paying Attention to Couples with Coupled Attention for Relationship Recommendation}},
  booktitle = {ICWSM},
  year = {2018},
  url = {https://openreview.net/forum?id=5r_IALpucCD},
}
Figure