Recently, large language models (LLMs) have advanced recommendation systems (RSs), and recent works have begun to explore how to integrate LLMs into industrial RSs. While most approaches deploy LLMs offline to generate and pre-cache augmented representations for RSs, high-dimensional representations from LLMs introduce substantial storage and computational costs. Thus, it is crucial to compress LLM representations effectively. However, we identify a counterintuitive phenomenon during representation compression: Mid-layer Representation Advantage (MRA), where representations from middle layers of LLMs outperform those from final layers in recommendation tasks. This degraded final layer renders existing compression methods, which typically compress on the final layer, suboptimal. We interpret this based on modularity theory that LLMs develop spontaneous internal functional modularity and force the final layer to specialize in the proxy training task. Thus, we propose Modular Representation Compression (MARC) to explicitly control the modularity of LLMs. First, Modular Adjustment explicitly introduces compression and task adaptation modules, enabling the LLM to operate strictly as a representation-learning module. Next, to ground each module to its specific task, Modular Task Decoupling uses information constraints and different network structures to decouple tasks. Extensive experiments validate that MARC addresses MRA and produces efficient representations. Notably, MARC achieved a 2.82% eCPM lift in an online A/B test within a large-scale commercial search advertising scenario.

Recommender system plays a pervasive role in today’s online services, yet its closed-loop nature, i.e., training and deploying within a specific closed domain, constrains its access to open-world knowledge. Recently, the emergence of large language models (LLMs) has shown promise in bridging this gap by encoding extensive world knowledge and demonstrating advanced reasoning capabilities. However, previous attempts to directly implement LLMs as recommenders fall short in meeting the demanding requirements of industrial recommender systems, particularly in terms of online inference latency and offline resource efficiency. In this work, we propose an Open-World R ecommendation Framework with E fficient and Deployable K nowledge I nfusion from Large Language Models, dubbed REKI , to acquire two types of external knowledge about users and items from LLMs. Specifically, we introduce factorization prompting to elicit accurate knowledge reasoning on user preferences and items. With factorization prompting, we develop individual knowledge extraction and collective knowledge extraction tailored for different scales of recommendation scenarios, effectively reducing offline resource consumption. Subsequently, the generated user and item knowledge undergoes efficient transformation and condensation into augmented vectors through a hybridized expert-integrated network , ensuring its compatibility with the recommendation task. The obtained vectors can then be directly used to enhance the performance of any conventional recommendation model. We also ensure efficient inference by preprocessing and prestoring the knowledge from the LLM. Extensive experiments demonstrate that REKI significantly outperforms the state-of-the-art baselines and is compatible with a diverse array of recommendation algorithms and tasks. Now, REKI has been deployed to Huawei’s news and music recommendation platforms and gained a 7% and 1.99% improvement during the online A/B test.

Recommender systems (RSs) often suffer from the feedback loop phenomenon, i.e., RSs are trained on data biased by their recommendations. This leads to the filter bubble effect that reinforces homogeneous content and reduces user satisfaction. To this end, serendipity recommendations, which offer unexpected yet relevant items, are proposed. Recently, large language models (LLMs) have shown potential in serendipity prediction due to their extensive world knowledge and reasoning capabilities. However, they still face challenges in aligning serendipity judgments with human assessments, handling long user behavior sequences, and meeting the latency requirements of industrial RSs. To address these issues, we propose SERAL (Serendipity Recommendations with Aligned Large Language Models), a framework comprising three stages: (1) Cognition Profile Generation to compress user behavior into multi-level profiles; (2) SerenGPT Alignment to align serendipity judgments with human preferences using enriched training data; and (3) Nearline Adaptation to integrate SerenGPT into industrial RSs pipelines efficiently. Online experiments demonstrate that SERAL improves exposure ratio (PVR), clicks, and transactions of serendipitous items by 5.7%, 29.56%, and 27.6%, enhancing user experience without much impact on overall revenue. Now, it has been fully deployed in ''Guess What You Like'' on the Taobao App homepage.

Interactive recommender systems (IRSs) aim to meet user needs through natural language dialogues, optimizing recommendations with minimal interactions. Typically, IRSs are based on large language models (LLMs). Existing methods generally consist of two stages: decision-making (deciding whether to recommend or ask clarification questions) and action execution (generating recommendations or clarification questions). These methods usually follow a decision-first paradigm, where the model first decides on the action based on past conversations, and then executes the corresponding action. Since LLMs struggle to process a large number of candidate items, the recommendation process is often carried out in collaboration with external recommendation tools, which provide a small candidate set for LLMs to refine.

The past few years have witnessed a growing interest in LLM-based recommender systems (RSs), although their industrial deployment remains in a preliminary stage. Most existing deployments leverage LLMs offline as feature enhancers, generating augmented knowledge for downstream tasks. However, in recommendation scenarios with numerous users and items, even offline knowledge generation with LLMs demands significant time and computational resources. This inefficiency arises from the autoregressive nature of LLMs. A promising solution is speculative decoding, a Draft-Then-Verify approach that increases the number of tokens generated per decoding step. In this work, we first identify recommendation knowledge generation as a highly fitting use case for retrieval-based speculative decoding. Then, we discern its two characteristics: (1) the vast number of items and users in RSs leads to retrieval inefficiency, and (2) RSs exhibit high diversity tolerance for LLM-generated text. Building on these insights, we introduce Lossless Acceleration via Speculative Decoding for LLM-based Recommender Systems (LASER), which features a Customized Retrieval Pool to enhance retrieval efficiency and Relaxed Verification to improve the acceptance rate of draft tokens. LASER achieves a 3-5x speedup on public datasets and saves about 67% of computational resources during the online A/B test on a large-scale advertising scenario with lossless downstream recommendation performance. Our code is available at https://github.com/YunjiaXi/LASER

With the rapid development of online services and web applications, recommender systems (RS) have become increasingly indispensable for mitigating information overload and matching users’ information needs by providing personalized suggestions over items. Although the RS research community has made remarkable progress over the past decades, conventional recommendation models (CRM) still have some limitations, e.g., lacking open-domain world knowledge, and difficulties in comprehending users’ underlying preferences and motivations. Meanwhile, large language models (LLM) have shown impressive general intelligence and human-like capabilities for various natural language processing (NLP) tasks, which mainly stem from their extensive open-world knowledge, logical and commonsense reasoning abilities, as well as their comprehension of human culture and society. Consequently, the emergence of LLM is inspiring the design of RS and pointing out a promising research direction, i.e., whether we can incorporate LLM and benefit from their common knowledge and capabilities to compensate for the limitations of CRM. In this article, we conduct a comprehensive survey on this research direction, and draw a bird’s-eye view from the perspective of the whole pipeline in real-world RS. Specifically, we summarize existing research works from two orthogonal aspects: where and how to adapt LLM to RS. For the “ WHERE ” question, we discuss the roles that LLM could play in different stages of the recommendation pipeline, i.e., feature engineering, feature encoder, scoring/ranking function, user interaction, and pipeline controller. For the “ HOW ” question, we investigate the training and inference strategies, resulting in two fine-grained taxonomy criteria, i.e., whether to tune LLM or not during training, and whether to involve CRM for inference. Detailed analysis and general development paths are provided for both “WHERE” and “HOW” questions, respectively. Then, we highlight the key challenges in adapting LLM to RS from three aspects, i.e., efficiency, effectiveness, and ethics. Finally, we summarize the survey and discuss the future prospects.

Conversational recommender systems (CRSs) aim to capture user preferences and provide personalized recommendations through multi-round natural language dialogues. However, most existing CRS models mainly focus on dialogue comprehension and preferences mining from the current dialogue session, overlooking user preferences in historical dialogue sessions. The preferences embedded in historical sessions and the current session exhibit continuity and sequentiality, and we refer to such CRSs as sequential CRSs. In this work, we leverage memory-enhanced LLMs to model the preference continuity, addressing two key issues: (1) redundancy and noise in historical dialogue sessions, and (2) the cold-start users problem. Thus, we propose a Memory-enhanced Conversational Recommender System Framework with Large Language Models (dubbed MemoCRS), consisting of user-specific memory and general memory. User-specific memory is tailored to each user's interests and uses an entity-based memory bank to refine preferences and retrieve relevant memory, thereby reducing the redundancy and noise of historical sessions. The general memory, encapsulating collaborative knowledge and reasoning guidelines, can provide shared knowledge for users, especially cold-start users. With the above memory, LLMs are empowered to deliver more precise and tailored recommendations for each user. Extensive experiments on Chinese and English datasets demonstrate MemoCRS's effectiveness.

Recommender systems play important roles in various applications such as e-commerce, social media, etc. Conventional recommendation methods usually model the collaborative signals within the tabular representation space. Despite the personalization modeling and the efficiency, the latent semantic dependencies are omitted. Methods that introduce semantics into recommendation then emerge, injecting knowledge from the semantic representation space where the general language understanding are compressed. However, existing semantic-enhanced recommendation methods focus on aligning the two spaces, during which the representations of the two spaces tend to get close while the unique patterns are discarded and not well explored. In this paper, we propose DisCo to Disentangle the unique patterns from the two representation spaces and Collaborate the two spaces for recommendation enhancement, where both the specificity and the consistency of the two spaces are captured. Concretely, we propose 1) a dual-side attentive network to capture the intra-domain patterns and the inter-domain patterns, 2) a sufficiency constraint to preserve the task-relevant information of each representation space and filter out the noise, and 3) a disentanglement constraint to avoid the model from discarding the unique information. These modules strike a balance between disentanglement and collaboration of the two representation spaces to produce informative pattern vectors, which could serve as extra features and be appended to arbitrary recommendation backbones for enhancement. Experiment results validate the superiority of our method against different models and the compatibility of DisCo over different backbones. Various ablation studies and efficiency analysis are also conducted to justify each model component.

Recommender system plays a vital role in various online services. However, its insulated nature of training and deploying separately within a specific closed domain limits its access to open-world knowledge. Recently, the emergence of large language models (LLMs) has shown promise in bridging this gap by encoding extensive world knowledge and demonstrating reasoning capabilities. Nevertheless, previous attempts to directly use LLMs as recommenders cannot meet the inference latency demand of industrial recommender systems. In this work, we propose an Open-World Knowledge Augmented Recommendation Framework with Large Language Models, dubbed KAR, to acquire two types of external knowledge from LLMs — the reasoning knowledge on user preferences and the factual knowledge on items. We introduce factorization prompting to elicit accurate reasoning on user preferences. The generated reasoning and factual knowledge are effectively transformed and condensed into augmented vectors by a hybrid-expert adaptor in order to be compatible with the recommendation task. The obtained vectors can then be directly used to enhance the performance of any recommendation model. We also ensure efficient inference by preprocessing and prestoring the knowledge from the LLM. Extensive experiments show that KAR significantly outperforms the state-of-the-art baselines and is compatible with a wide range of recommendation algorithms. We deploy KAR to Huawei’s news and music recommendation platforms and gain a 7% and 1.7% improvement in the online A/B test, respectively.

Click-through rate (CTR) prediction has become increasingly indispensable for various Internet applications. Traditional CTR models convert the multi-field categorical data into ID features via one-hot encoding, and extract the collaborative signals among features. Such a paradigm suffers from the problem of semantic information loss. Another line of research explores the potential of pretrained language models (PLMs) for CTR prediction by converting input data into textual sentences through hard prompt templates. Although semantic signals are preserved, they generally fail to capture the collaborative information (e.g., feature interactions, pure ID features), not to mention the unacceptable inference overhead brought by the huge model size. In this paper, we aim to model both the semantic knowledge and collaborative knowledge for accurate CTR estimation, and meanwhile address the inference inefficiency issue. To benefit from both worlds and close their gaps, we propose a novel model-agnostic framework (i.e., ClickPrompt), where we incorporate CTR models to generate interaction-aware soft prompts for PLMs. We design a prompt-augmented masked language modeling (PA-MLM) pretraining task, where PLM has to recover the masked tokens based on the language context, as well as the soft prompts generated by CTR model. The collaborative and semantic knowledge from ID and textual features would be explicitly aligned and interacted via the prompt interface. Then, we can either tune the CTR model with PLM for superior performance, or solely tune the CTR model without PLM for inference efficiency. Experiments on four real-world datasets validate the effectiveness of ClickPrompt compared with existing baselines.

As a critical task for large-scale commercial recommender systems, reranking rearranges items in the initial ranking lists from the previous ranking stage to better meet users’ demands. Foundational work in reranking has shown the potential of improving recommendation results by uncovering mutual influence among items. However, rather than considering the context of initial lists as most existing methods do, an ideal reranking algorithm should consider the counterfactual context— the position and the alignment of the items in the reranked lists . In this work, we propose a novel pairwise reranking framework, Utility-oriented Reranking with Counterfactual Context (URCC), which maximizes the overall utility after reranking efficiently. Specifically, we first design a utility-oriented evaluator, which applies Bi-LSTM and graph attention mechanism to estimate the listwise utility via the counterfactual context modeling. Then, under the guidance of the evaluator, we propose a pairwise reranker model to find the most suitable position for each item by swapping misplaced item pairs. Extensive experiments on two benchmark datasets and a proprietary real-world dataset demonstrate that URCC significantly outperforms the state-of-the-art models in terms of both relevance-based metrics and utility-based metrics.

Re-ranking, as the final stage of the multi-stage recommender systems (MRS), aims at modeling the listwise context and the cross-item interactions between the candidate items. The objective is usually the overall utility (e.g., total clicks or revenue) of the re-ranked list, which is determined not only by the relevance, but also by the diversity of the list. However, existing methods equally promote diversity for all users and often compromise the relevance ranking. In reality, users have different diversity preferences and we should diversify the list tailored to individual users’ interests and needs. Users’ behavior history contains rich information which may be used for inferring their diversity preferences, but has rarely been explored in existing work. In this work, we propose a novel neural re-ranking with personalized diversification method (dubbed RAPID) to address the above challenge. RAPID explicitly models each user’s preference distribution over different topics by exploiting the intra- and inter-topic interactions from the user’s behavior history. The personalized diversity gain brought by each candidate item is then measured by the item’s marginal diversity and the learned personalized preference. The relevance and the personalized diversity are jointly optimized in an end-to-end manner to automatically manage the relevance-diversity tradeoff. Experimental results on two public datasets and a proprietary dataset show that RAPID outperforms the state-of-the-art with the highest utility and the best relevance-diversity tradeoff. We further prove that RAPID has a regret bound of $\tilde O(\sqrt n )$ on utility, which provides theoretical guarantee that its performance is near-optimal.

As an emerging field driven by industrial applications, integrated re-ranking combines lists from upstream sources into a single list, and presents it to the user. The quality of integrated re-ranking is especially sensitive to real-time user behaviors and preferences. However, existing methods are all built on the cloud-to-edge framework, where mixed lists are generated by the cloud model and then sent to the devices. Despite its effectiveness, such a framework fails to capture users' real-time preferences due to the network bandwidth and latency. Hence, we propose to place the integrated re-ranking model on devices, allowing for the full exploitation of real-time behaviors. To achieve this, we need to address two key issues: first, how to extract users' preferences for different sources from heterogeneous and imbalanced user behaviors; second, how to explore the correlation between the extracted personalized preferences and the candidate items. In this work, we present the first on-Device Integrated Re-ranking framework, DIR, to avoid delays in processing real-time user behaviors. DIR includes a multi-sequence behavior modeling module to extract the user's source-level preferences, and a preference-adaptive re-ranking module to incorporate personalized source-level preferences into the re-ranking of candidate items. Besides, we design exposure loss and utility loss to jointly optimize exposure fairness and overall utility. Extensive experiments on three datasets show that DIR significantly outperforms the state-of-the-art baselines in utility-based and fairness-based metrics.

Reranking, as the final stage of multi-stage recommender systems, refines the initial lists to maximize the total utility. With the development of multimedia and user interface design, the recommendation page has evolved to a multi-list style. Separately employing traditional list-level reranking methods for different lists overlooks the inter-list interactions and the effect of different page formats, thus yielding suboptimal reranking performance. Moreover, simply applying a shared network for all the lists fails to capture the commonalities and distinctions in user behaviors on different lists. To this end, we propose to draw a bird's-eye view of page-level reranking and design a novel Page-level Attentional Reranking (PAR) model. We introduce a hierarchical dual-side attention module to extract personalized intra- and inter-list interactions. A spatial-scaled attention network is devised to integrate the spatial relationship into pairwise item influences, which explicitly models the page format. The multi-gated mixture-of-experts module is further applied to capture the commonalities and differences of user behaviors between different lists. Extensive experiments on a public dataset and a proprietary dataset show that PAR significantly outperforms existing baseline models.

As the final stage of the multi-stage recommender system (MRS), reranking directly affects users' experience and satisfaction, thus playing a critical role in MRS. Despite the improvement achieved in the existing work, three issues are yet to be solved. First, users' historical behaviors contain rich preference information, such as users' long and short-term interests, but are not fully exploited in reranking. Previous work typically treats items in history equally important, neglecting the dynamic interaction between the history and candidate items. Second, existing reranking models focus on learning interactions at the item level while ignoring the fine-grained feature-level interactions. Lastly, estimating the reranking score on the ordered initial list before reranking may lead to the early scoring problem, thereby yielding suboptimal reranking performance. To address the above issues, we propose a framework named Multi-level Interaction Reranking (MIR). MIR combines low-level cross-item interaction and high-level set-to-list interaction, where we view the candidate items to be reranked as a set and the users' behavior history in chronological order as a list. We design a novel SLAttention structure for modeling the set-to-list interactions with personalized long-short term interests. Moreover, feature-level interactions are incorporated to capture the fine-grained influence among items. We design MIR in such a way that any permutation of the input items would not change the output ranking, and we theoretically prove it. Extensive experiments on three public and proprietary datasets show that MIR significantly outperforms the state-of-the-art models using various ranking and utility metrics.

Learning to rank from logged user feedback, such as clicks or purchases, is a central component of many real-world information systems. Different from human-annotated relevance labels, the user feedback is always noisy and biased. Many existing learning to rank methods infer the underlying relevance of query–item pairs based on different assumptions of examination, and still optimize a relevance based objective. Such methods rely heavily on the correct estimation of examination, which is often difficult to achieve in practice. In this work, we propose a general framework U-rank+ for learning to rank with logged user feedback from the perspective of graph matching. We systematically analyze the biases in user feedback, including examination bias and selection bias. Then, we take both biases into consideration for unbiased utility estimation that directly based on user feedback, instead of relevance. In order to maximize the estimated utility in an efficient manner, we design two different solvers based on Sinkhorn and LambdaLoss for U-rank+ . The former is based on a standard graph matching algorithm, and the latter is inspired by the traditional method of learning to rank. Both of the algorithms have good theoretical properties to optimize the unbiased utility objective while the latter is proved to be empirically more effective and efficient in practice. Our framework U-rank+ can deal with a general utility function and can be used in a widespread of applications including web search, recommendation, and online advertising. Semi-synthetic experiments on three benchmark learning to rank datasets demonstrate the effectiveness of U-rank+ . Furthermore, our proposed framework has been deployed on two different scenarios of a mainstream App store, where the online A/B testing shows that U-rank+ achieves an average improvement of 19.2% on click-through rate and 20.8% improvement on conversion rate in recommendation scenario, and 5.12% on platform revenue in online advertising scenario over the production baselines.

Learning to rank with implicit feedback is one of the most important tasks in many real-world information systems where the objective is some specific utility, e.g., clicks and revenue. However, we point out that existing methods based on probabilistic ranking principle do not necessarily achieve the highest utility. To this end, we propose a novel ranking framework called U-rank that directly optimizes the expected utility of the ranking list. With a position-aware deep click-through rate prediction model, we address the attention bias considering both query-level and item-level features. Due to the item-specific attention bias modeling, the optimization for expected utility corresponds to a maximum weight matching on the item-position bipartite graph. We base the optimization of this objective in an efficient Lambdaloss framework, which is supported by both theoretical and empirical analysis. We conduct extensive experiments for both web search and recommender systems over three benchmark datasets and two proprietary datasets, where the performance gain of U-rank over state-of-the-arts is demonstrated. Moreover, our proposed U-rank has been deployed on a large-scale commercial recommender and a large improvement over the production baseline has been observed in an online A/B testing.