Abstract
Traditional recommender systems are designed to help consumers navigate existing catalogs. However, item producers—such as movie studios or product designers—face a different challenge: deciding what to create next to appeal to a broad audience. This paper introduces the Maximum Coverage Novel Itemset Problem (MCNIP), shifting the focus from recommendation to generation. By employing Variational Autoencoders (VAEs) to learn a joint latent space of item features and user preferences, the authors propose a method to generate a small set of novel items that maximize total user satisfaction. Evaluations on synthetic, artistic, and cinematic datasets demonstrate the model's ability to "discover" high-potential product concepts and archetypes.
Introduction: The Producer's Dilemma
Standard recommender systems solve the "discovery" problem for users but ignore the "creation" problem for producers. Designers and manufacturers often suffer from a "cold-start" problem: how can one design a product that does not yet exist but is guaranteed to be successful? Furthermore, because a single item rarely satisfies an entire market, producers must curate a diverse portfolio (a "set") of items to cover various user preference groups.
The authors formalize this as the Maximum Coverage Novel Itemset Problem (MCNIP). The goal is to identify a set of $K$ non-existent items that maximize the total predicted ratings while ensuring that as many users as possible are "covered"—meaning they find at least one item in the set highly appealing.
Methodology: Generative Recommendation
The proposed solution involves a three-stage pipeline centered on a shared latent representation of items and users.
1. Learning the Latent Design Space
The core of the system is a model that maps item features (e.g., movie tags, image pixels) into a low-dimensional latent space $\mathcal{Z}$. Unlike standard VAEs that only focus on reconstruction, this model uses a joint loss function to ensure the latent space captures both item characteristics and user preferences (ratings). While a linear model (LM) serves as a baseline, a deep Variational Autoencoder (VAE) is used to capture the complex, non-linear manifold of the design space.
2. Greedy Latent Search
Finding the optimal set of $K$ items is NP-hard. To approximate the solution, the authors discretize the latent space by sampling thousands of candidate "imaginary" items. A greedy algorithm then iteratively selects items that provide the highest marginal increase in coverage for previously unsatisfied user groups.
3. Decoding into Reality
Once the optimal latent points (represented as stars in the latent space) are identified, they are passed through the VAE’s decoder. This component acts as the "designer," translating abstract mathematical coordinates back into human-interpretable features, such as a set of movie tags or a generated image.
Experimental Evaluation
The framework was tested across three diverse domains:
- Synthetic Data: The model successfully "rediscovered" hidden ideal items from noisy rating data, outperforming linear baselines in reconstruction fidelity.
- MART Abstract Art: Using a dataset of 500 paintings and 10,000 ratings, the model generated visual features. A set of just five generated images was able to "cover" 82% of the user base.
- MovieLens Tag Genome: The model generated intuitive movie archetypes from a genome of 1,128 tags. For instance, it proposed a "sentimental drama" archetype and a "big-budget action sci-fi" archetype, both of which aligned with successful real-world movie patterns despite being novel generated concepts.
Key Insights
- The Decoder as a Creative Tool: While most research focuses on the encoder (understanding the user), this work highlights the decoder's role in product synthesis.
- Emergent Audiences: In this paradigm, groups are not pre-defined; instead, they emerge based on the items generated. The item effectively "creates" its own audience.
- VAE Superiority: Deep generative models are essential for this task. Linear models often produce "blurry" or "average" items that lack the distinct features necessary to satisfy specific user niches.
Conclusion and Future Directions
This research moves beyond traditional recommendation by providing a "brainstorming" tool for creators. Future work aims to incorporate feasibility constraints—ensuring generated items are physically or economically viable—and exploring human-in-the-loop interfaces where designers can interactively refine the model's proposals.
Summary updated