Transformer-based Cross-Encoders achieve state-of-the-art effectiveness in text retrieval. However, Cross-Encoders based on large transformer models (such as BERT or T5) are computationally expensive and allow for scoring only a small number of documents within a reasonably small latency window. However, keeping search latencies low is important for user satisfaction and energy usage. In this paper, we show that weaker shallow transformer models (i.e., transformers with a limited number of layers) actually perform better than full-scale models when constrained to these practical low-latency settings since they can estimate the relevance of more documents in the same time budget. We further show that shallow transformers may benefit from the generalized Binary Cross-Entropy (gBCE) training scheme, which has recently demonstrated success for recommendation tasks. Our experiments with TREC Deep Learning passage ranking query sets demonstrate significant improvements in shallow and full-scale models in low-latency scenarios. For example, when the latency limit is 25ms per query, MonoBERT-Large (a cross-encoder based on a full-scale BERT model) is only able to achieve NDCG@10 of 0.431 on TREC DL 2019, while TinyBERT-gBCE (a cross-encoder based on TinyBERT trained with gBCE) reaches NDCG@10 of 0.652, a +51% gain over MonoBERT-Large. We also show that shallow Cross-Encoders are effective even when used without a GPU (e.g., with CPU inference, NDCG@10 decreases only by 3% compared to GPU inference with 50ms latency), which makes Cross-Encoders practical to run even without specialized hardware acceleration.
Adaptations of Transformer models, such as BERT4Rec and SASRec, achieve state-of-the-art performance in the sequential recommendation task according to accuracy-based metrics, such as NDCG. These models treat items as tokens and then utilise a score-and-rank approach (Top-K strategy), where the model first computes item scores and then ranks them according to this score. While this approach works well for accuracy-based metrics, it is hard to use it for optimising more complex beyond-accuracy metrics such as diversity. Recently, the GPTRec model, which uses a different Next-K strategy, has been proposed as an alternative to the Top-K models. In contrast with traditional Top-K recommendations, Next-K generates recommendations item-by-item and, therefore, can account for complex item-to-item interdependencies important for the beyond-accuracy measures. However, the original GPTRec paper focused only on accuracy in experiments and needed to address how to optimise the model for complex beyond-accuracy metrics. Indeed, training GPTRec for beyond-accuracy goals is challenging because the interaction training data available for training recommender systems typically needs to be aligned with beyond-accuracy recommendation goals. To solve the misalignment problem, we train GPTRec using a 2-stage approach: in the first stage, we use a teacher-student approach to train GPTRec, mimicking the behaviour of traditional Top-K models; in the second stage, we use Reinforcement Learning to align the model for beyond-accuracy goals. In particular, we experiment with increasing recommendation diversity and reducing popularity bias. Our experiments on two datasets show that in 3 out of 4 cases, GPTRec's Next-K generation approach offers a better tradeoff between accuracy and secondary metrics than classic greedy re-ranking techniques.
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.
Sequential Recommendation is a popular recommendation task that uses the order of user-item interaction to model evolving users' interests and sequential patterns in their behaviour. Current state-of-the-art Transformer-based models for sequential recommendation, such as BERT4Rec and SASRec, generate sequence embeddings and compute scores for catalogue items, but the increasing catalogue size makes training these models costly. The Joint Product Quantisation (JPQ) method, originally proposed for passage retrieval, markedly reduces the size of the retrieval index with minimal effect on model effectiveness, by replacing passage embeddings with a limited number of shared sub-embeddings. This paper introduces RecJPQ, a novel adaptation of JPQ for sequential recommendations, which takes the place of item embeddings tensor and replaces item embeddings with a concatenation of a limited number of shared sub-embeddings and, therefore, limits the number of learnable model parameters. The main idea of RecJPQ is to split items into sub-item entities before training the main recommendation model, which is inspired by splitting words into tokens and training tokenisers in language models. We apply RecJPQ to SASRec, BERT4Rec, and GRU4rec models on three large-scale sequential datasets. Our results showed that RecJPQ could notably reduce the model size (e.g., 48% reduction for the Gowalla dataset with no effectiveness degradation). RecJPQ can also improve model performance through a regularisation effect (e.g. +0.96% NDCG@10 improvement on the Booking.com dataset). Overall, RecJPQ allows the training of state-of-the-art transformer recommenders in industrial applications, where datasets with millions of items are common.
The use of pre-training is an emerging technique to enhance a neural model's performance, which has been shown to be effective for many neural language models such as BERT. This technique has also been used to enhance the performance of recommender systems. In such recommender systems, pre-training models are used to learn a better initialisation for both users and items. However, recent existing pre-trained recommender systems tend to only incorporate the user interaction data at the pre-training stage, making it difficult to deliver good recommendations, especially when the interaction data is sparse. To alleviate this common data sparsity issue, we propose to pre-train the recommendation model not only with the interaction data but also with other available information such as the social relations among users, thereby providing the recommender system with a better initialisation compared with solely relying on the user interaction data. We propose a novel recommendation model, the Social-aware Gaussian Pre-trained model (SGP), which encodes the user social relations and interaction data at the pre-training stage in a Graph Neural Network (GNN). Afterwards, in the subsequent fine-tuning stage, our SGP model adopts a Gaussian Mixture Model (GMM) to factorise these pre-trained embeddings for further training, thereby benefiting the cold-start users from these pre-built social relations. Our extensive experiments on three public datasets show that, in comparison to 16 competitive baselines, our SGP model significantly outperforms the best baseline by upto 7.7% in terms of NDCG@10. In addition, we show that SGP permits to effectively alleviate the cold-start problem, especially when users newly register to the system through their friends' suggestions.
In recent years, the rapid growth of online multimedia services, such as e-commerce platforms, has necessitated the development of personalised recommendation approaches that can encode diverse content about each item. Indeed, modern multi-modal recommender systems exploit diverse features obtained from raw images and item descriptions to enhance the recommendation performance. However, the existing multi-modal recommenders primarily depend on the features extracted individually from different media through pre-trained modality-specific encoders, and exhibit only shallow alignments between different modalities - limiting these systems' ability to capture the underlying relationships between the modalities. In this paper, we investigate the usage of large multi-modal encoders within the specific context of recommender systems, as these have previously demonstrated state-of-the-art effectiveness when ranking items across various domains. Specifically, we tailor two state-of-the-art multi-modal encoders (CLIP and VLMo) for recommendation tasks using a range of strategies, including the exploration of pre-trained and fine-tuned encoders, as well as the assessment of the end-to-end training of these encoders. We demonstrate that pre-trained large multi-modal encoders can generate more aligned and effective user/item representations compared to existing modality-specific encoders across three multi-modal recommendation datasets. Furthermore, we show that fine-tuning these large multi-modal encoders with recommendation datasets leads to an enhanced recommendation performance. In terms of different training paradigms, our experiments highlight the essential role of the end-to-end training of large multi-modal encoders in multi-modal recommendation systems.
Query Performance Prediction (QPP) estimates the effectiveness of a search engine's results in response to a query without relevance judgments. Traditionally, post-retrieval predictors have focused upon either the distribution of the retrieval scores, or the coherence of the top-ranked documents using traditional bag-of-words index representations. More recently, BERT-based models using dense embedded document representations have been used to create new predictors, but mostly applied to predict the performance of rankings created by BM25. Instead, we aim to predict the effectiveness of rankings created by single-representation dense retrieval models (ANCE & TCT-ColBERT). Therefore, we propose a number of variants of existing unsupervised coherence-based predictors that employ neural embedding representations. In our experiments on the TREC Deep Learning Track datasets, we demonstrate improved accuracy upon dense retrieval (up to 92% compared to sparse variants for TCT-ColBERT and 188% for ANCE). Going deeper, we select the most representative and best performing predictors to study the importance of differences among predictors and query types on query performance. Using existing distribution-based evaluation QPP measures and a particular type of linear mixed models, we find that query types further significantly influence query performance (and are up to 35% responsible for the unstable performance of QPP predictors), and that this sensitivity is unique to dense retrieval models. Our approach introduces a new setting for obtaining richer information on query differences in dense QPP that can explain potential unstable performance of existing predictors and outlines the unique characteristics of different query types on dense retrieval models.
Recommender systems are frequently challenged by the data sparsity problem. One approach to mitigate this issue is through cross-domain recommendation techniques. In a cross-domain context, sharing knowledge between domains can enhance the effectiveness in the target domain. Recent cross-domain methods have employed a pre-training approach, but we argue that these methods often result in suboptimal fine-tuning, especially with large neural models. Modern language models utilize prompts for efficient model tuning. Such prompts act as a tunable latent vector, allowing for the freezing of the main model parameters. In our research, we introduce the Personalised Graph Prompt-based Recommendation (PGPRec) framework. This leverages the advantages of prompt-tuning. Within this framework, we formulate personalized graph prompts item-wise, rooted in items that a user has previously engaged with. Specifically, we employ Contrastive Learning (CL) to produce pre-trained embeddings that offer greater generalizability in the pre-training phase, ensuring robust training during the tuning phase. Our evaluation of PGPRec in cross-domain scenarios involves comprehensive testing with the top-k recommendation tasks and a cold-start analysis. Our empirical findings, based on four Amazon Review datasets, reveal that the PGPRec framework can decrease the tuned parameters by as much as 74%, maintaining competitive performance. Remarkably, there's an 11.41% enhancement in performance against the leading baseline in cold-start situations.
A large catalogue size is one of the central challenges in training recommendation models: a large number of items makes them memory and computationally inefficient to compute scores for all items during training, forcing these models to deploy negative sampling. However, negative sampling increases the proportion of positive interactions in the training data, and therefore models trained with negative sampling tend to overestimate the probabilities of positive interactions a phenomenon we call overconfidence. While the absolute values of the predicted scores or probabilities are not important for the ranking of retrieved recommendations, overconfident models may fail to estimate nuanced differences in the top-ranked items, resulting in degraded performance. In this paper, we show that overconfidence explains why the popular SASRec model underperforms when compared to BERT4Rec. This is contrary to the BERT4Rec authors explanation that the difference in performance is due to the bi-directional attention mechanism. To mitigate overconfidence, we propose a novel Generalised Binary Cross-Entropy Loss function (gBCE) and theoretically prove that it can mitigate overconfidence. We further propose the gSASRec model, an improvement over SASRec that deploys an increased number of negatives and the gBCE loss. We show through detailed experiments on three datasets that gSASRec does not exhibit the overconfidence problem. As a result, gSASRec can outperform BERT4Rec (e.g. +9.47% NDCG on the MovieLens-1M dataset), while requiring less training time (e.g. -73% training time on MovieLens-1M). Moreover, in contrast to BERT4Rec, gSASRec is suitable for large datasets that contain more than 1 million items.
Performing automatic reformulations of a user's query is a popular paradigm used in information retrieval (IR) for improving effectiveness -- as exemplified by the pseudo-relevance feedback approaches, which expand the query in order to alleviate the vocabulary mismatch problem. Recent advancements in generative language models have demonstrated their ability in generating responses that are relevant to a given prompt. In light of this success, we seek to study the capacity of such models to perform query reformulation and how they compare with long-standing query reformulation methods that use pseudo-relevance feedback. In particular, we investigate two representative query reformulation frameworks, GenQR and GenPRF. GenQR directly reformulates the user's input query, while GenPRF provides additional context for the query by making use of pseudo-relevance feedback information. For each reformulation method, we leverage different techniques, including fine-tuning and direct prompting, to harness the knowledge of language models. The reformulated queries produced by the generative models are demonstrated to markedly benefit the effectiveness of a state-of-the-art retrieval pipeline on four TREC test collections (varying from TREC 2004 Robust to the TREC 2019 Deep Learning). Furthermore, our results indicate that our studied generative models can outperform various statistical query expansion approaches while remaining comparable to other existing complex neural query reformulation models, with the added benefit of being simpler to implement.