User-Side Realization

TL;DR

This work articulates user-side realization as a principled framework for enabling end-users to realize tailored, fair, and privacy-preserving functionalities without modifying service providers. It introduces wrapper and reverse approaches, then develops concrete algorithms across chapters: PrivateRank and PrivateWalk for fair recommender systems, Consul for principled on-device state extraction, Tiara and Seafaring for on-demand image retrieval and Web-scale active learning, CLEAR for a fully client-side image search, and PRISM for privacy-preserving translation. Together, these contributions demonstrate that user-owned implementations can achieve competitive performance, strong fairness/privacy guarantees, and practical deployment in real-world services like IMDb, Twitter, Flickr, and translation platforms. The research advances both theoretical foundations (e.g., metric recovery, locality, and differential privacy guarantees) and practical demonstrations, highlighting significant potential for user-centric customization in web services. The work points to broad implications for privacy, fairness, and user autonomy in large-scale online ecosystems, with concrete pathways for future tooling, libraries, and cross-domain applications.

Abstract

Users are dissatisfied with services. Since the service is not tailor-made for a user, it is natural for dissatisfaction to arise. The problem is, that even if users are dissatisfied, they often do not have the means to resolve their dissatisfaction. The user cannot alter the source code of the service, nor can they force the service provider to change. The user has no choice but to remain dissatisfied or quit the service. User-side realization offers proactive solutions to this problem by providing general algorithms to deal with common problems on the user's side. These algorithms run on the user's side and solve the problems without having the service provider change the service itself.

Figures (30)

• Figure 1: Approaches of user-side realization.
• Figure 2: Trade-off between fairness and performance: PR (blue curve) represents PrivateRank, and PW (green curve) represents PrivateWalk. The score reported in a parenthesis is the performance of each method when the recommendation is completely fair. Even though PrivateRank does not use log data, its performance is comparable with the oracle method, which adopts prohibitive information. Although PrivateWalk performs worse than PrivateRank in exchange for fast evaluation, it remains significantly better than random guesses.
• Figure 3: Sensitivity of hyperparameters: Our methods are insensitive to hyperparameters.
• Figure 4: CanAdd$(\mathcal{R}, i)$
• Figure 5: PrivateRank

Theorems & Definitions (11)

• Theorem 1
• Theorem 2
• proof
• proof
• Proposition 3
• proof
• Theorem 4
• proof
• Definition 5: Differential Privacy
• Theorem 6