Conventional multi-modal sequential recommenders usually employ sequence models (e.g., SASRec and GRU4Rec) as the recommender framework for sequential dependency modeling and multi-modal information as add-on knowledge for fine-grained preference learning. However, the existing methods ignore the time interval information in the interaction sequences, which reflects the user’s preference transition. Consequently, they acquire biased preference representations, leading to suboptimal performance. Along these lines, we concentrate on incorporating the time interval into the multi-modal sequential recommendation and investigating the internal influences across multiple modalities. Firstly, we treat the time interval as an independent modality and exploit a Time Interval Encoder (TIE) to quantify the time interval in the sequences. Secondly, we consider the heterogeneity between multiple modalities and design a two-stage fusion strategy. The first stage injects time intervals into each modality (ID, image, and text) and the second stage aggregates the user preferences from each modality to get the user’s sequence preferences. To this end, we build a model-agnostic framework for the multi-modal sequential recommendation, namely the Time-interval-aware Multi-Modal Sequential Recommender (TMMSRec) . The empirical validations on four real-world datasets prove the effectiveness of our method, where the maximum performance improvement against several state-of-the-art baselines achieves up to 16.05%. Our code is available at https://github.com/wangpp0602/TMMSRec .

Invariant user preference learning is a core task in recommender systems. Accurately modeling user preferences is crucial, as it directly impacts the quality of recommendations. However, heterogeneous user preferences in feedback data often exhibit mixture distributions, which obscure invariant preferences and introduce bias. Existing methods typically address this issue by partitioning feedback data into multiple environments and learning invariant preferences across them. Nonetheless, these approaches often lack theoretical guarantees for environment construction and fail to capture the dynamic nature of user preferences across different environments. Along these lines, we propose a novel framework, Invariant and Environment-specific Preferences for unbiased recommendation (IEPref). IEPref leverages auxiliary information as a reliable signal to guide the environment classifier in partitioning the environment, thereby enabling the learning of more stable and generalizable user preference with theoretical guarantees. Additionally, we design environment-specific proxy modules to capture context-dependent preference patterns unique to each environment. The environment classifier assigns each user-item interaction to its corresponding latent environment, and both invariant and environment-specific preferences are integrated for recommendations. Extensive experiments on five real-world datasets demonstrate that IEPref achieves superior performance over existing baselines, effectively mitigating recommendation bias while preserving personalized modeling capabilities.

In this paper, we focus on the often-overlooked issue of embedding collapse in existing diffusion-based sequential recommendation models and propose ADRec, an innovative framework designed to mitigate this problem. Diverging from previous diffusion-based methods, ADRec applies an independent noise process to each token and performs diffusion across the entire target sequence during training. ADRec captures token interdependency through auto-regression while modeling per-token distributions through token-level diffusion. This dual approach enables the model to effectively capture both sequence dynamics and item representations, overcoming the limitations of existing methods. To further mitigate embedding collapse, we propose a three-stage training strategy: (1) pre-training the embedding weights(2) aligning these weights with the ADRec backbone, and (3) fine-tuning the model. During inference, ADRec applies the denoising process only to the last token, ensuring that the meaningful patterns in historical interactions are preserved. Our comprehensive empirical evaluation across six datasets underscores the effectiveness of ADRec in enhancing both the accuracy and efficiency of diffusion-based sequential recommendation systems.

In recommender systems, it is frequently presumed that missing ratings adhere to a missing at random (MAR) mechanism, implying the absence of ratings is independent of their potential values. However, this assumption fails to hold in real-world scenarios, where users are inclined to rate items they either strongly favor or disfavor, introducing a missing not at random (MNAR) scenario. To tackle this issue, prior researchers have utilized explicit MAR feedbacks to infer the propensities of unobserved, implicit MNAR feedbacks. Nonetheless, acquiring explicit MAR feedbacks is resource-intensive and time-consuming and may not reflect users’ true preferences. Furthermore, most methods have only been tested on synthetic or small-scale datasets, thus their applicability and effectiveness in real-world settings without MAR feedbacks remain unclear. Along these lines, we aim to predict MNAR ratings without MAR prior propensities by exploring the consistency between MAR and MNAR feedbacks and narrowing the gap between them. From the empirical study and preliminary experiment, we hypothesize thatuser preferencescan be treated as the common prior propensity for both MAR and MNAR generative processes. In this way, we extend this hypothesis to a more general MNAR scenario: user preferences learned from MNAR can partially substitute for the prior propensities derived from MAR feedbacks for MNAR recommendation tasks. To validate our hypothesis and approach, we develop a lightweight iterative probabilistic matrix factorization framework (lightIPMF) as a practical method of our methodology, utilizing user preferences extracted from MNAR, not MAR, to estimate MNAR feedbacks. Finally, the experimental results show that modeling user preferences can effectively improve MNAR feedback estimation without MAR feedback, and our proposed lightIPMF outperforms the state-of-the-art MNAR methods in predicting MNAR feedbacks.

Inferring user preferences from users’ historical feedback is a valuable problem in recommender systems. Conventional approaches often rely on the assumption that user preferences in the feedback data are equivalent to the real user preferences without additional noise, which simplifies the problem modeling. However, there are various confounders during user–item interactions, such as weather and even the recommendation system itself. Therefore, neglecting the influence of confounders will result in inaccurate user preferences and suboptimal performance of the model. Furthermore, the unobservability of confounders poses a challenge in further addressing the problem. Along these lines, we refine the problem and propose a more rational solution to mitigate the influence of unobserved confounders. Specifically, we consider the influence of unobserved confounders, disentangle them from user preferences in the latent space, and employ causal graphs to model their interdependencies without specific labels. By ingeniously combining local and global causal graphs, we capture the user-specific effects of confounders on user preferences. Finally, we propose our model based on Variational Autoencoders, named Causal Structure Aware Variational Autoencoders (CSA-VAE) and theoretically demonstrate the identifiability of the obtained causal graph. We conducted extensive experiments on one synthetic dataset and nine real-world datasets with different scales, including three unbiased datasets and six normal datasets, where the average performance boost against several state-of-the-art baselines achieves up to 9.55%, demonstrating the superiority of our model. Furthermore, users can control their recommendation list by manipulating the learned causal representations of confounders, generating potentially more diverse recommendation results. Our code is available at Code-link ( https://github.com/MICLab-Rec/CSA ).

The past flourishing years of sequential location-based services began with the introduction of the Self-Attention Network (SAN), which quickly superseded CNN or RNN as the state-of-the-art backbone. Recent works utilize modified attention mechanisms or neural network layers to process spatial-temporal factors to realize fine-grained individual behavior pattern modeling. However, we argue these methods can be further improved due to the significant increase in the model's parameter scale or computational burden. In this paper, we first exploit two lightweight approaches, Rotary Time Aware Position Encoder (RoTAPE) and multi-head Interval Aware Attention Block (IAAB), to impel SAN by efficiently and effectively capturing spatial-temporal intervals among the user's visited locations, which require neither extra parameters nor a high computational cost. On the one hand, RoTAPE encodes the day- and hour-level timestamps into sequence representation simultaneously via a sinusoidal encoding matrix, and the corresponding time intervals can be explicitly captured by SAN. Specifically, the multi-level temporal differences are mutually independent to reflect the periodical pattern and jointly complete to measure the absolute time interval. On the other hand, IAAB, point- wise injecting the historical spatial-temporal intervals into the attention map, can promote SAN attaching importance to the spatial relations under the constraints of time conditions. Then, we design a novel MLP-based module, Spatial-Temporal Relation Memory (STR Memory), implemented with fully connected linear layers and matrix transpose operations. STR Memory, endowing the interactions inside historical intervals along different directions, can convert the historical intervals into spatial-temporal relations in future trajectories for accurate predictions. To this end, we propose an end-to-end mobility trajectory prediction framework, namely STiSAN$^+$, employing RoTAPE, stacking multiple layers of IAAB-based encoder-decoder architecture, and coupling with STR Memory. We conducted numerous experiments on six public LBSN datasets to evaluate our proposed algorithm. From Next Location Recommendation to Multi-location Future Trajectory Prediction, our STiSAN$^+$gains average 15.05% and 18.35% improvements against several state-of-the-art sequential models, respectively. Ablation studies demonstrate the effectiveness of RoTAPE, IAAB, and STR Memory under our framework. Moreover, we separately validate the extensibility and interpretability of RoTAPE and IAAB through non-sampled metric evaluation and visualization.

Recommender Systems (RSs) typically face the cold-start problem and the filter-bubble problem when users suffer the familiar, repeated, and even predictable recommendations, making them bored and unsatisfied. The key to solving these issues is learning users' fine-grained preferences and recommending appealing and unexplored items deviating from users' historical items. However, existing models consider cold-start or filter bubble problems separately and ignore that they can reinforce mutually and damage the models' performance accuracy. To this end, we devise a novel serendipity-oriented recommender system (GenerativeSelf-constrainedSerendipitousRecommenderSystem,GS$^{2}$-RS) that generates users' fine-grained preferences to enhance the recommendation performance. Specifically, GS$^{2}$-RS extracts users' interest and satisfaction preferences and generates virtual but convincible neighbors' preferences from themselves with a twin Conditional Generative Adversarial Nets (not from real neighbors). Then we introduce the serendipity item, which is low-interest but high-satisfaction among candidate items. We use the serendipity item to improve the diversity of recommended items, which relieves the filter-bubble problem. Along with this line, a gated mechanism is applied to their fine-grained preferences (interests, satisfactions) to obtain their serendipity items. Finally, these serendipity items are inversely injected into the original user-item rating matrix and build a relatively dense matrix as the input for backbone RS models. Note that GS$^{2}$-RS tackles cold-start and filter-bubble problems in a unified framework without any additional side information and enriches the interpretability of recommendation models. We comprehensively validate GS$^{2}$-RS for solving cold-start and filter bubble problems on four real-world benchmark datasets. Extensive experiments illustrate GS$^{2}$-RS's superiority in accuracy, serendipity, and interpretability over state-of-the-art models. Also, we can plug our model into existing recommender systems as a preprocessing procedure to enhance their performance.

In this work, we present TriMLP as a foundational MLP-like architecture for the sequential recommendation, simultaneously achieving computational efficiency and promising performance. First, we empirically study the incompatibility between existing purely MLP-based models and sequential recommendation, that the inherent fully-connective structure endows historical user–item interactions (referred as tokens) with unrestricted communications and overlooks the essential chronological order in sequences. Then, we propose the MLP-based Triangular Mixer to establish ordered contact among tokens and excavate the primary sequential modeling capability under the standard auto-regressive training fashion. It contains (1) a global mixing layer that drops the lower-triangle neurons in MLP to block the anti-chronological connections from future tokens and (2) a local mixing layer that further disables specific upper-triangle neurons to split the sequence as multiple independent sessions. The mixer serially alternates these two layers to support fine-grained preferences modeling, where the global one focuses on the long-range dependency in the whole sequence, and the local one calls for the short-term patterns in sessions. Experimental results on 12 datasets of different scales from 4 benchmarks elucidate that TriMLP consistently attains favorable accuracy/efficiency tradeoff over all validated datasets, where the average performance boost against several state-of-the-art baselines achieves up to 14.88%, and the maximum reduction of inference time reaches 23.73%. The intriguing properties render TriMLP a strong contender to the well-established RNN-, CNN-, and Transformer-based sequential recommenders. Code is available at https://github.com/jiangyiheng1/TriMLP .

Data analysis usually suffers from the Missing Not At Random (MNAR) problem, where the cause of the value missing is not fully observed. Compared to the naive Missing Completely At Random (MCAR) problem, it is more in line with the realistic scenario whereas more complex and challenging. Existing statistical methods model the MNAR mechanism by different decomposition of the joint distribution of the complete data and the missing mask. But we empirically find that directly incorporating these statistical methods into deep generative models is sub-optimal. Specifically, it would neglect the confidence of the reconstructed mask during the MNAR imputation process, which leads to insufficient information extraction and less-guaranteed imputation quality. In this paper, we revisit the MNAR problem from a novel perspective that the complete data and missing mask are two modalities of incomplete data on an equal footing. Along with this line, we put forward a generative-model-specific joint probability decomposition method, conjunction model, to represent the distributions of two modalities in parallel and extract sufficient information from both complete data and missing mask. Taking a step further, we exploit a deep generative imputation model, namely GNR, to process the real-world missing mechanism in the latent space and concurrently impute the incomplete data and reconstruct the missing mask. The experimental results show that our GNR surpasses state-of-the-art MNAR baselines with significant margins (averagely improved from 9.9% to 18.8% in RMSE) and always gives a better mask reconstruction accuracy which makes the imputation more principle.

Urban prediction tasks that aim to model the complicated spatial and temporal patterns of urban indicators (such as weather, vehicle charging demand, etc.) for accurate prediction, have been increasingly important in constructing smart cities and accelerating the urbanization process in the modern era. However, most existing works of urban prediction have only concentrated on spatial and temporal correlations, but ignored the effect of distribution shift from spatial and temporal perspectives; this could largely hinder the performance of urban prediction tasks. In order to solve this problem, in this paper, we propose a Shift-Aware Urban Prediction (SAUP) framework to eliminate the inherent shift effect among spatial-temporal urban time series data. Specifically, SAUP starts with a Shift Elimination Module, built upon our proposed Spatial-Temporal Attention Flows (STAF) composed of invertible attentions and coupling layers of normalizing flows in order to transform the raw shifted data into a unified distribution to remove the spatiotemporal shift. After the shift effect is eliminated, the Correlation Processing Module of SAUP further captures the core correlations to learn spatiotemporal dependencies, in which topological correlations and geographic correlations are jointly learned by GCN and CNN based on pre-defined graphs and extracted POI information. In addition, SAUP includes a model-agnostic Forecasting Module, which can be employed as any forecasting architecture to accomplish the predictions. To recover the raw distribution information, the output of the Forecasting Module is further taken for the inverse transformation of the Shift Elimination Module to produce the final forecasts. We have conducted extensive experiments in the SAUP framework, coupled with six state-of-the-art spatiotemporal forecasting models on two real-world datasets. Experimental results have demonstrated the consistent improvements of SAUP over the baseline algorithms.

Points-of-interest (POIs) recommendation plays a vital role in location-based social networks (LBSNs) by introducing unexplored POIs to consumers and has drawn extensive attention from academia and industry. Existing POI recommender systems usually learn fixed latent vectors to represent both consumers and POIs from historical check-ins and make recommendations under the spatio-temporal constraints. However, we argue that the existing works still suffer from the challenges of explaining consumers’ complicated check-in actions. To this end, we first explore the interpretability of recommendations from the POI aspect, i.e., for a specific POI, its function usually changes over time, so representing a POI with a single fixed latent vector is not sufficient to describe the dynamic nature of POIs. Besides, check-in actions to a POI are also affected by the zone where it is located. In other words, the zone's embedding learned from POI distributions, road segments, and historical check-ins could be jointly utilized to enhance POI embeddings. Along this line, we propose aTime-zone-spacePOI embedding model (ToP), which integrates multi-knowledge graphs and topic model to introduce not only spatio-temporal effects but also sentiment constraints into POI embeddings for strengthening interpretability of recommendation. Specifically, ToP learns multiple latent vectors for a POI in a different period with spatial constraints via knowledge graph learning. To add sentiment constraints, ToP jointly combines these vectors with the zone's representations learned by topic models to make explainable recommendations. ToP considers the time, space, and sentiment of POI in a unified embedding framework, which benefits the POI recommendations. Extensive experiments on real-world Changchun city datasets demonstrate that ToP achieves state-of-the-art performance in terms of common metrics and provides more insights for consumers’ POI check-in actions.

The past flourishing years of sequential point-of-interest (POI) recommendation began with the introduction of Self-Attention Network (SAN), which quickly superseded CNN or RNN as the state-of-the-art backbone. To realize the fine-grained users' behavior patterns modeling, recent works utilize modified attention mechanisms or neural network layers to process spatial-temporal factors. However, due to the significant increase on either model's parameter scale or computational burden, we argue that these methods can be further improved. In this paper, we exploit two lightweight approaches, Time Aware Position Encoder (TAPE) and Interval Aware Attention Block (IAAB), to impel SAN by considering the spatial-temporal intervals among POIs separately, where requiring neither extra parameters nor high computational cost. On the one hand, TAPE, adjusting the positions in sequences based on the timestamps dynamically and generating positional representations with sinusoidal transformation, can enhance sequence representations to reflect both the absolute order and relative temporal proximity among all POIs. On the other hand, IAAB, point-wise adding the scaled spatial-temporal intervals to the attention map, can promote the attention mechanism attaching importance to the spatial relation among all POIs under the constraints of time conditions and providing more explainable recommendation. We integrate these two modules into SAN and propose a Spatial-Temporal Interval-Aware sequential POI recommender, namely STiSAN, as an end-to-end deployment. Experimental results based on three public LBSN datasets and one real-world city transportation dataset demonstrate STiSAN's superior performance (average 13.01% improvement against the strongest baseline). Moreover, we validate the extensibility and interpretability of TAPE and IAAB through metric evaluation and visualization separately.

With the boosting of neural networks, recommendation methods become significantly improved by their powerful ability of prediction and inference. Existing neural-network based recommender systems (NN-RSs) usually first employ matrix embedding (ME) as a pre-process to learn users’ and items’ representations (latent vectors), then input these representations to a specific modified neural network framework to make accurate Top-k recommendations. Obviously, the performance of ME has a significant effect on RS models. However, most NN-RSs focus on accuracy by building representations from the direct user-item interactions (e.g., user-item rating matrix), while ignoring the underlying relatedness between users and items (e.g., users who rate the same ratings for the same items should be embedded into similar representations), which is an ideological disadvantage. On the other hand, ME models directly employ inner products as a default loss function metric that cannot project users and items into a proper latent space, which is a methodological disadvantage. In this paper, we propose a supervised collaborative representation learning model - Magnetic Metric Learning (MML) - to map users and items into a unified latent vector space, enhancing the representation learning for NN-RSs. First, MML utilizes dual triplets to model not only the observed relationships between users and items, but also the underlying relationships between users as well as items to overcome the ideological disadvantage. Specifically, a modified metric-based dual loss function is proposed in MML to gather similar entities and disperse the dissimilar ones. With MML, we can easily compare all the relationships (user to user, item to item, user to item) according to the weighted metric, which overcomes the methodological disadvantage. We conduct extensive experiments on four real-world datasets with large item space. The results demonstrate that MML can learn a proper unified latent space for representations from the user-item matrix with high accuracy and effectiveness, and lead to a performance gain over the state-of-the-art RS models by an average of 17 percent.

In e-commerce, online retailers are usually suffering from professional malicious users (PMUs), who utilize negative reviews and low ratings to their consumed products on purpose to threaten the retailers for illegal profits. PMUs are difficult to be detected because they utilize masking strategies to disguise themselves as normal users. Specifically, there are three challenges for PMU detection: 1) professional malicious users do not conduct any abnormal or illegal interactions (they never concurrently leave too many negative reviews and low ratings at the same time), and they conduct masking strategies to disguise themselves. Therefore, conventional outlier detection methods are confused by their masking strategies. 2) the PMU detection model should take both ratings and reviews into consideration, which makes PMU detection a multi-modal problem. 3) there are no datasets with labels for professional malicious users in public, which makes PMU detection an unsupervised learning problem. To this end, we propose an unsupervised multi-modal learning model: MMD, which employs Metric learning for professional Malicious users Detection with both ratings and reviews. MMD first utilizes a modified RNN to project the informational review into a sentiment score, which jointly considers the ratings and reviews. Then professional malicious user profiling (MUP) is proposed to catch the sentiment gap between sentiment scores and ratings. MUP filters the users and builds a candidate PMU set. We apply a metric learning-based clustering to learn a proper metric matrix for PMU detection. Finally, we can utilize this metric and labeled users to detect PMUs. Specifically, we apply the attention mechanism in metric learning to improve the model’s performance. The extensive experiments in four datasets demonstrate that our proposed method can solve this unsupervised detection problem. Moreover, the performance of the state-of-the-art recommender models is enhanced by taking MMD as a preprocessing stage.

Identifying the dynamic functions of different urban zones enables a variety of smart city applications, such as intelligent urban planning, real-time traffic scheduling, and community precision management. Traditional urban function research using government administrative zoning systems is often conducted in a coarse resolution with fixed split, and ignore the reshaping of zones by city growth. To solve this problem, we propose a two-stage framework in order to represent the high-definition distribution of urban function across the city, by analyzing continuous human traces extracted from the dense, widespread, and full-time cellular data. At the representation stage, we embed the locations of base stations by modeling the user movements with staying and transfer events, along with the consideration of dynamic trip purposes in continuous human traces. At the annotation stage, we first divide the city into the finest unit zones and each covers at least one base station. By clustering the base stations, we further group the unit zones into functional zones. Last, we annotate functional zones based on the local point-of-interest (POI) information. In experiments, we evaluate the proposed high-definition function study in two tasks: (i) in-zone crowd flow prediction, and (ii) zone-enhanced POI recommendation. The results demonstrate the advantage of the proposed method with both the effectiveness of city split and the high-quality function annotation.

Recommender systems have been playing an important role in providing personalized information to users. However, there is always a trade-off between accuracy and novelty in recommender systems. Usually, many users are suffering from redundant or inaccurate recommendation results. To this end, in this article, we put efforts into exploring the hidden knowledge of observed ratings to alleviate this recommendation dilemma. Specifically, we utilize some basic concepts to define a concept, Serendipity , which is characterized by high-satisfaction and low-initial-interest. Based on this concept, we propose a two-phase recommendation problem which aims to strike a balance between accuracy and novelty achieved by serendipity prediction and personalized recommendation. Along this line, a Neural Serendipity Recommendation (NSR) method is first developed by combining Muti-Layer Percetron and Matrix Factorization for serendipity prediction. Then, a weighted candidate filtering method is designed for personalized recommendation. Finally, extensive experiments on real-world data demonstrate that NSR can achieve a superior serendipity by a 12% improvement in average while maintaining stable accuracy compared with state-of-the-art methods.

In real-world recommendation scenarios, there are two common phenomena: 1) users only provide ratings but there is no review comment. As a result, the historical transaction data available for recommender system are usually unbalanced and sparse; 2) Users' opinions can be better grasped in their reviews than ratings. This indicates that there is always a bias between ratings and reviews. Therefore, it is important that users' ratings and reviews should be mutually reinforced to grasp the users' true opinions. To this end, in this paper, we develop an opinion mining model based on convolutional neural networks for enhancing recommendation (NeuO). Specifically, we exploit a two-step training neural networks, which utilize both reviews and ratings to grasp users' true opinions in unbalanced data. Moreover, we propose a Sentiment Classification scoring method (SC), which employs dual attention vectors to predict the users' sentiment scores of their reviews. A combination function is designed to use the results of SC and user-item rating matrix to catch the opinion bias. Finally, a Multilayer perceptron based Matrix Factorization (MMF) method is proposed to make recommendations with the enhanced user-item matrix. Extensive experiments on real-world data demonstrate that our approach can achieve a superior performance over state-of-the-art baselines on real-world datasets.