Solving the Cold Start Problem on One's Own as an End User via Preference Transfer

TL;DR

This work tackles the cold-start problem from the end-user perspective by proposing Pretender, an algorithm that enables users to transfer preferences from a source service to a target service without provider support. It formulates the objective as minimizing an integral probability metric distance between the source distribution $oldsymbol{ u}_S$ and the target distribution $oldsymbol{ u}_T^{oldsymbol{w}}$ of labeled items, then solves a continuous relaxation over item weights and uses randomized rounding with postprocessing to output exactly $K$ items. Theoretical guarantees are established for two IPM instantiations: Maximum Mean Discrepancy (MMD) and 1-Wasserstein distance, including convergence rates $O(L^{-1/2})$ for MMD and $O(K^{-1/(d+2)})$ for Wasserstein, under mild kernel and density Assumptions; results extend to general target-data usage via a quadrature perspective. Empirically, Pretender achieves near-continuous-optimal transfer across MovieLens, Last.fm, and Amazon domains, outperforming baselines and illustrating the practicality and robustness of user-side preference transfer. The work introduces a novel problem setting and demonstrates that users can meaningfully influence recommendations without changes from service providers, potentially broadening the applicability of personalized recommender systems.

Abstract

We propose a new approach that enables end users to directly solve the cold start problem by themselves. The cold start problem is a common issue in recommender systems, and many methods have been proposed to address the problem on the service provider's side. However, when the service provider does not take action, users are left with poor recommendations and no means to improve their experience. We propose an algorithm, Pretender, that allows end users to proactively solve the cold start problem on their own. Pretender does not require any special support from the service provider and can be deployed independently by users. We formulate the problem as minimizing the distance between the source and target distributions and optimize item selection from the target service accordingly. Furthermore, we establish theoretical guarantees for Pretender based on a discrete quadrature problem. We conduct experiments on real-world datasets to demonstrate the effectiveness of Pretender.

Figures (2)

• Figure 1: MMD as a function of the number of selected items $K$. The optimal value of the combinatorial optimization problem is intractable but is guaranteed to lie somewhere between the red and blue lines. We can see that the difference between the proposed method and the optimal value decreases as $K$ increases. As we discussed in Section \ref{['sec: monotone']}, the optimal value is not monotonic in $K$, and the sweet spots lie around $32$ to $64$ in both cases. We can also see that the optimal values get worse more quickly in (a) because the source data contain fewer items, the relevant items in the target set are exhausted more quickly, and we are forced to select less relevant items when we increase $K$.
• Figure 2: Table 1. Performance Comparison. Each value represents the average MMD across all users. Lower is better. The standard deviation is computed across users. The proposed method is much better than the baseline methods and is close to the optimal continuous solution.

Theorems & Definitions (32)

• Proposition 3.1
• proof
• Proposition 3.3
• Proposition 3.4
• Theorem 3.5
• Corollary 3.6
• proof
• Theorem 3.8
• Corollary 3.9
• Corollary 3.10