A CONTROLLED STUDY OF PARAMETER-EFFICIENT FINE-TUNING IN COLD-START RECOMMENDATION

Abstract

Cold-start recommendation remains one of the most persistent and practically conse quential challenges in the design and deployment of personalised information retrieval systems at scale. When users arrive on a platform with no prior interaction history, or when new items are introduced to a catalogue before any user has engaged with them, conventional collaborative filtering and matrix factorisation methods suffer se vere and often catastrophic degradation in their ranking quality, because these methods derive their predictive power entirely from the density of the observed user–item inter action matrix. The emergence of Large Language Model (LLM) based recommender systems has opened a qualitatively different path toward solving this problem: by en coding user preference signals and item characteristics as structured natural language prompts, and by leveraging the rich semantic world knowledge encoded in billions of pretrained model parameters, LLM-based recommenders can reason about relevance and preference even in the near-complete absence of interaction data. This capability has made LLM-based recommendation a topic of intense research interest, and has driven substantial investment in their deployment across commercial recommendation platforms. However, the practical utility of LLM-based recommenders in cold-start settings is fundamentally contingent upon an efficient and effective adaptation strategy. A general-domain LLM, even a very capable one, must be aligned to the specific vo cabulary of preferences, the item ontology, and the interaction patterns of the target recommendation domain. Full fine-tuning of billion-parameter models for this pur pose is computationally prohibitive, memory-intensive, and poorly suited to the dy namic, rapidly evolving user–item distributions that characterise real recommendation platforms. Parameter-Efficient Fine-Tuning (PEFT) methods have been developed to address this adaptation bottleneck, selectively updating a small fraction of model pa rameters while keeping the vast majority of the pretrained weights frozen. Despite their growing importance, however, no prior work has provided a rigorous, unified, and controlled comparison of PEFT methods specifically in the cold-start recommen dation setting. This thesis addresses that gap directly. We design and execute a comprehensive benchmark framework for the fair comparison of three major PEFT families — Low Rank Adaptation (LoRA), prompt-based tuning (soft prompt tuning and prefix tuning), and adapter-based tuning, including the specialised variant of Late Adapter Tuning (LAT) — under both the user cold-start and item cold-start regimes. The framework holds the backbone model, training optimiser, evaluation protocol, and dataset prepro cessing pipeline constant across all compared methods, ensuring that observed perfor mance differences are attributable to the adaptation mechanism alone. Six benchmark datasets are used for evaluation: the MIND news recommendation dataset, the Amazon product domains of Games, Toys, and Office, and the Netflix streaming dataset, eval uated under the all-ranking protocol using Recall@K, NDCG@K, and Hit Ratio@K metrics. i The central empirical finding of this thesis is that adaptation performance in cold start recommendation is governed not primarily by the number of parameters updated, but by where in the model those updates are applied. Late Adapter Tuning, which confines adapter modules exclusively to the upper transformer layers of the pretrained backbone, achieves performance at or above full fine-tuning on multiple benchmarks while updating only 2.37% of total model parameters, reducing GPU memory con sumption by 76.5% relative to full fine-tuning, and attaining the highest inference throughput among all compared methods. Prompt-based tuning, while offering the smallest possible parameter overhead at less than 0.01% of total parameters, proves to be expressively limited in highly personalised and long-tail recommendation set tings where surface-level input conditioning is insufficient to redirect the model’s in ternal representations toward task-specific preference patterns. LoRA occupies a well balanced intermediate position, delivering near-full-fine-tuning accuracy with 0.22% of total parameters and strong optimisation stability. A structured cold-start robust ness analysis reveals an important asymmetry between user- and item-cold-start failure modes: item cold-start is primarily a knowledge availability problem best addressed by retrieval-augmented methods, while user cold-start performance scales monotonically with interaction count and is governed by the model family and adaptation quality. Together, these findings establish a unifying principle that reframes cold-start recom mendation from a data compression problem into a representation alignment problem, in which the goal of PEFT is to identify the minimal and most task-relevant parameters in the upper model layers and update them with maximal precision.