Exploring Deep Space: Learning Personalized Ranking in a Semantic Space

Jeroen B. P. Vuurens, Martha A. Larson, Arjen P. de Vries

Abstract

Recommender systems often struggle to balance predictive accuracy with the ability to surface novel or niche items. This paper introduces "Deep Space" (DS), a framework that reimagines recommendation as a two-step geometric problem. First, items are mapped into a high-dimensional semantic space where proximity represents "intersubstitutability." Second, the system learns a personalized linear transformation (a hyperplane) for each user to rank these items. Evaluated on the MovieLens 1M dataset, the Deep Space approach significantly outperforms traditional collaborative filtering baselines like BPRMF and UserKNN.

Introduction

Standard recommender systems generally fall into two categories: Collaborative Filtering (CF), which relies on user-item interaction history, and Content-Based (CB), which analyzes item descriptions. While CF is typically more accurate, it requires massive amounts of data and often fails to recommend novel items.

The authors identify a critical gap in item representation. They argue that for a recommendation to be truly effective, items should be positioned in a space where proximity represents intersubstitutability—the idea that if a user liked item A, they would likely accept item B as a substitute. The challenge lies in creating a unified framework that can build this space using either interaction data or raw text while allowing for fast, highly personalized ranking.

The Deep Space Framework

The proposed Deep Space (DS) framework consists of two distinct phases:

1. Learning the Semantic Space

The system utilizes a shallow neural network architecture based on Paragraph Vector (PV-DBOW). In this setup, an item ID acts as a "document," and the "words" are either the specific users who rated it (for the CF variant) or the words found in its IMDB reviews (for the CB variant). This process generates a high-dimensional vector for every item, positioning them based on the context of their ratings or descriptions.

The neural architecture used to learn item representations in the semantic space.

2. User-Specific Ranking

To personalize the experience, the system learns a unique hyperplane for each user, representing their "preference direction" in the high-dimensional space. By projecting all item vectors onto a line orthogonal to this hyperplane, the system creates a 1D ranked list. This is trained using a pairwise ranking loss: the model is presented with pairs of items (one liked, one unrated) and learns to rotate the hyperplane until the liked items consistently rank higher.

Experimental Results

The authors evaluated the framework using the MovieLens 1M dataset (1 million ratings) and IMDB Reviews for the content-based variant (DS-CB).

Key Insights

  • The Necessity of High Dimensionality: While high dimensionality often leads to the "curse of dimensionality," the authors found that a space of approximately 1,000 dimensions is necessary to provide enough "room" for overlapping user preferences to be linearly separable. Performance improved steadily as dimensionality increased, peaking at the 1,000-dimension mark.
  • Recency and Noise: The study found that using only a user's most recent ratings ($\phi_t$) to train the ranking function improved results. This suggests that user tastes evolve, and older data can act as noise that distracts the model from current preferences.
  • Efficiency: Despite the high dimensionality, the system is remarkably efficient. It takes approximately 0.3 seconds to generate a full ranked list for a user from a pool of 10,000 items.
Impact of semantic space dimensionality on Recall@10 performance for both validation and test sets.

Performance Comparison

The collaborative filtering variant (DS-CF) achieved a Recall@10 of 0.151, nearly doubling the performance of the popular Bayesian Personalized Ranking (BPRMF) baseline (0.079). Below 300 dimensions, the linear hyperplane struggled to accurately capture complex user tastes, but performance stabilized at higher dimensions.

Conclusion and Future Work

Deep Space offers a unified architecture that works effectively for both ratings and text data. However, the authors acknowledge limitations, such as the constraints of linear transformations. Future work may explore non-linear transformation functions (like deeper neural layers) for lower-dimensional spaces and methods for jointly learning item representations and user preferences simultaneously. Additionally, finding a way to better fuse text and rating data into a single "hybrid" space remains an open challenge for addressing the cold-start problem.

Summary updated

Abstract

Recommender systems leverage both content and user interactions to generate recommendations that fit users' preferences. The recent surge of interest in deep learning presents new opportunities for exploiting these two sources of information. To recommend items we propose to first learn a user-independent high-dimensional semantic space in which items are positioned according to their substitutability, and then learn a user-specific transformation function to transform this space into a ranking according to the user's past preferences. An advantage of the proposed architecture is that it can be used to effectively recommend items using either content that describes the items or user-item ratings. We show that this approach significantly outperforms state-of-the-art recommender systems on the MovieLens 1M dataset.
BibTeX
@inproceedings{vuurens2016exploring,
  author = {Jeroen B. P. Vuurens and Martha A. Larson and Arjen P. de Vries},
  title = {{Exploring Deep Space: Learning Personalized Ranking in a Semantic Space}},
  booktitle = {DLRS@RecSys},
  year = {2016},
  doi = {10.1145/2988450.2988457},
}
Figure