User Profiling from Reviews for Accurate Time-Based Recommendations

Oznur Alkan, Elizabeth DalyarXiv 2020 · arXiv

TL;DR

This research introduces a novel method for enhancing recommender systems by extracting "target age" preferences from product reviews. By modeling age as a linear, time-dependent variable, the system can predict a user's current needs (e.g., transitioning from baby rattles to toddler toys) and filter out-of-date recommendations, significantly improving accuracy.

The Challenge of Temporal Dynamics

Standard recommendation algorithms often fail to account for how user interests evolve over time. In specific domains like childcare, purchase history can be a misleading indicator. If a parent purchased a bassinet six months ago, traditional collaborative filtering might continue to suggest similar infant products. However, as the child grows, the parent's needs shift predictably.

The primary obstacle is that "target age"—the age of the person for whom the item is intended—is rarely explicitly stated in user profiles. This paper bridges that gap by mining latent temporal information from the text of user-generated reviews to build dynamic profiles that age alongside the user's family.

Methodology: From Text to Trajectory

The researchers proposed a four-step pipeline to integrate temporal age data into the recommendation process:

  1. Mining Age-Related Terms: Using Natural Language Processing (NLP), the system identifies age-specific noun phrases (e.g., "18 months"). It specifically targets possessive phrases like "my 5-year-old" to ensure the age refers to the intended recipient.
  2. Item Age Prediction: The system aggregates age mentions across all reviews for a specific product to establish a "target age range," using statistical tests to remove outliers.
  3. User Age Profiling: This core innovation maps the ages mentioned in a user's reviews against the dates they were posted. By applying linear regression, the system models the recipient's age as a function of time, allowing it to predict the target age at any future point.
  4. Recommendation Generation: After generating a candidate list via standard algorithms, a post-filtering (PF) step removes items whose target age range does not align with the user’s predicted trajectory.
Illustration of item age extraction from reviews (a) and user age profiling via linear regression (b).

Dataset and Evaluation

The study utilized the Amazon.com Public Dataset (Baby Category), focusing on a subset of reviews to ensure statistical relevance. The evaluation used a temporal split (80% training, 20% testing) to simulate real-world prediction scenarios.

Statistics of the Amazon Baby category dataset used for evaluation.

Key Results

The researchers compared baseline algorithms—User-Based Collaborative Filtering (UB-CF), Item-Based (IB-CF), and Matrix Factorization (MF-SVD)—against versions enhanced with their Post-Filtering (PF) method.

  • Accuracy Improvements: The post-filtered versions consistently outperformed baselines. For instance, UB-CF-PF achieved an NDCG@10 of 0.0120, a significant jump from the standard UB-CF's 0.0083.
  • Relevance: Qualitative analysis showed the system successfully replaced "stale" items (like a jumperoo for an infant) with age-appropriate gear as the predicted target age increased.

Conclusion and Future Directions

By treating reviews as latent profiles, this approach moves beyond simple ratings to understand the context of a purchase. While highly effective in the "Baby" category due to the linear nature of aging, future work will explore handling multiple target profiles per user and applying these temporal filters to more complex domains where interests shift less predictably.

Summary updated

BibTeX
@misc{alkan2020user,
  author = {Oznur Alkan and Elizabeth Daly},
  title = {{User Profiling from Reviews for Accurate Time-Based Recommendations}},
  year = {2020},
  eprint = {2006.08805},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2006.08805},
}
Figure