What Else Would I Like? A User Simulator using Alternatives for Improved Evaluation of Fashion Conversational Recommendation Systems

TL;DR

This work addresses the limited realism of offline fashion CRS evaluation by introducing a meta-user simulator that leverages alternative target items and enriches two fashion datasets with alternative judgments. By evaluating three CRS models with an alternatives-aware feedback mechanism, the study shows that considering alternatives can significantly shift performance metrics and even model rankings, suggesting that single-target evaluations underestimate true effectiveness. The proposed approach improves evaluation completeness and realism, enabling faster satisfaction for simulated users and guiding more robust CRS development. Overall, the combination of dataset enrichment and an alternatives-aware simulator provides a more faithful framework for assessing multi-turn, image-based fashion recommendations.

Abstract

In Conversational Recommendation Systems (CRS), a user can provide feedback on recommended items at each interaction turn, leading the CRS towards more desirable recommendations. Currently, different types of CRS offer various possibilities for feedback, i.e., natural language feedback, or answering clarifying questions. In most cases, a user simulator is employed for training as well as evaluating the CRS. Such user simulators typically critique the current retrieved items based on knowledge of a single target item. Still, evaluating systems in offline settings with simulators suffers from problems, such as focusing entirely on a single target item (not addressing the exploratory nature of a recommender system), and exhibiting extreme patience (consistent feedback over a large number of turns). To overcome these limitations, we obtain extra judgements for a selection of alternative items in common CRS datasets, namely Shoes and Fashion IQ Dresses. Going further, we propose improved user simulators that allow simulated users not only to express their preferences about alternative items to their original target, but also to change their mind and level of patience. In our experiments using the relative image captioning CRS setting and different CRS models, we find that using the knowledge of alternatives by the simulator can have a considerable impact on the evaluation of existing CRS models, specifically that the existing single-target evaluation underestimates their effectiveness, and when simulated users are allowed to instead consider alternatives, the system can rapidly respond to more quickly satisfy the user.

Figures (4)

• Figure 1: Example of a fashion Conversational Image Recommendation scenario. At each turn, the user provides natural language feedback on a candidate item. In existing systems, users are assumed to have a specific target in mind (green). Instead, the presence of a single alternative (orange) or multiple alternative (blue) items can guide the system to find a target of a certain type.
• Figure 2: nDCG@10 for the various tolerance levels (denoted tol.) before selecting an alternative for the Shoes dataset.
• Figure 3: nDCG@10 for the various tolerance levels (denoted tol.) before selecting an alternative for the Dresses dataset.
• Figure 4: Number of target images for which the simulator selects an alternative over the target for the three CRS models for tolerance 1 and 3.