Existing example-based prediction explanation methods often bridge test and training data points through the model’s parameters or latent representations. While these methods offer clues to the causes of model predictions, they often exhibit innate shortcomings, such as incurring significant computational overhead or producing coarse-grained explanations. This paper presents a Highly-precise and Data-centric Explanation (HD-Explain) prediction explanation method that exploits properties of Kernelized Stein Discrepancy (KSD). Specifically, the KSD uniquely defines a parameterized kernel function for a trained model that encodes model-dependent data correlation. By leveraging the kernel function, one can identify training samples that provide the best predictive support to a test point efficiently. We conducted thorough analyses and experiments across multiple classification domains, where we show that HD-Explain outperforms existing methods from various aspects, including 1) preciseness (fine-grained explanation), 2) consistency, and 3) computation efficiency, leading to a surprisingly simple, effective, and robust prediction explanation solution.

Model editing aims to modify the outputs of large language models after they are trained. Previous approaches have often involved direct alterations to model weights, which can result in model degradation. Recent techniques avoid making modifications to the model’s weights by using an adapter that applies edits to the model when triggered by semantic similarity in the representation space. We demonstrate that current adapter methods are critically vulnerable to strong lexical biases, leading to issues such as applying edits to irrelevant prompts with overlapping words. This paper presents a principled approach to learning a disentangled representation space that facilitates precise localization of edits by maintaining distance between irrelevant prompts while preserving proximity among paraphrases. In our empirical study, we show that our method (Projector Editor Networks for Model Editing - PENME) achieves state-of-the-art model editing results while being more computationally efficient during inference than previous methods and adaptable across different architectures.

Existing research on Retrieval-Augmented Generation (RAG) primarily focuses on improving overall question-answering accuracy, often overlooking the quality of sub-claims within generated responses.Recent methods that attempt to improve RAG trustworthiness, such as through auto-evaluation metrics, lack probabilistic guarantees or require ground truth answers.To address these limitations, we propose Conformal-RAG, a novel framework inspired by recent applications of conformal prediction (CP) on large language models (LLMs).Conformal-RAG leverages CP and internal information from the RAG mechanism to offer statistical guarantees on response quality.It ensures group-conditional coverage spanning multiple sub-domains without requiring manual labelling of conformal sets, making it suitable for complex RAG applications.Compared to existing RAG auto-evaluation methods, Conformal-RAG offers statistical guarantees on the quality of refined sub-claims, ensuring response reliability without the need for ground truth answers.Additionally, our experiments demonstrate that by leveraging information from the RAG system, Conformal-RAG retains up to 60% more high-quality sub-claims from the response compared to direct applications of CP to LLMs, while maintaining the same reliability guarantee.

Self-supervised representation learning (SSRL) has advanced considerably by exploiting the transformation invariance assumption under artificially designed data augmentations. While augmentation-based SSRL algorithms push the boundaries of performance in computer vision and natural language processing, they are often not directly applicable to other data modalities, and can conflict with application-specific data augmentation constraints. This paper presents an SSRL approach that can be applied to any data modality and network architecture because it does not rely on augmentations or masking. Specifically, we show that high-quality data representations can be learned by reconstructing random data projections. We evaluate the proposed approach on a wide range of representation learning tasks that span diverse modalities and real-world applications. We show that it outperforms multiple state-of-the-art SSRL baselines. Due to its wide applicability and strong empirical results, we argue that learning from randomness is a fruitful research direction worthy of attention and further study.

For content recommendation systems on social media platforms, timely, efficient, and accurate estimation of user preferences can effectively improve their performance and enhance the platforms’ activeness. However, efficiency and accuracy are often a pair of trade-offs; Accurate user preference estimation often requires tracking dynamic preference shifting with complex sequential modelling, whereas efficient systems may fail to follow the shift because of a lack of modelling capacity. This paper presents Asynchronous and Asymmetric User Preference Updating System AAUPU, a distributed collaborative filtering system, that can track hundreds of millions of users’ preferences in real time by processing streaming data. The AAUPU system finds a good balance between efficiency and accuracy, making it well-suited for large-scale personalization service needs on social media platforms. We implemented the system on the Google Cloud Platform and successfully tracked user preference for a population of 400 million active Twitter accounts. To evaluate the estimation quality, we conducted massive A/B tests on the two most important service surfaces of Twitter, involving more than ten million users. Our experimental results show that the recommender system based on the AAUPU system significantly improves the overall recommendation performance on the platform.

Social media platforms such as Twitter or StockTwits are widely used for sharing stock market opinions between investors, traders, and entrepreneurs. Empirically, previous work has shown that the content posted on these social media platforms can be leveraged to predict various aspects of stock market performance. Nonetheless, actors on these social media platforms may not always have altruistic motivations and may instead seek to influence stock trading behavior through the (potentially misleading) information they post. While a lot of previous work has sought to analyze how social media can be used to predict the stock market, there remain many questions regarding the quality of the predictions and the behavior of active users on these platforms. To this end, this article seeks to address a number of open research questions: Which social media platform is more predictive of stock performance? What posted content is actually predictive, and over what time horizon? How does stock market posting behavior vary among different users? Are all users trustworthy or do some user’s predictions consistently mislead about the true stock movement? To answer these questions, we analyzed data from Twitter and StockTwits covering almost 5 years of posted messages spanning 2015 to 2019. The results of this large-scale study provide a number of important insights among which we present the following: (i) StockTwits is a more predictive source of information than Twitter, leading us to focus our analysis on StockTwits; (ii) on StockTwits, users’ self-labeled sentiments are correlated with the stock market but are only slightly predictive in aggregate over the short-term; (iii) there are at least three clear types of temporal predictive behavior for users over a 144 days horizon: short, medium, and long term; and (iv) consistently incorrect users who are reliably wrong tend to exhibit what we conjecture to be “botlike” post content and their removal from the data tends to improve stock market predictions from self-labeled content.

Preserving the performance of a trained model while removing unique characteristics of marked training data points is challenging. Recent research usually suggests retraining a model from scratch with remaining training data or refining the model by reverting the model optimization on the marked data points. Unfortunately, aside from their computational inefficiency, those approaches inevitably hurt the resulting model's generalization ability since they remove not only unique characteristics but also discard shared (and possibly contributive) information. To address the performance degradation problem, this paper presents a novel approach called Performance Unchanged Model Augmentation (PUMA). The proposed PUMA framework explicitly models the influence of each training data point on the model's generalization ability with respect to various performance criteria. It then complements the negative impact of removing marked data by reweighting the remaining data optimally. To demonstrate the effectiveness of the PUMA framework, we compared it with multiple state-of-the-art data removal techniques in the experiments, where we show the PUMA can effectively and efficiently remove the unique characteristics of marked training data without retraining the model that can 1) fool a membership attack, and 2) resist performance degradation. In addition, as PUMA estimates the data importance during its operation, we show it could serve to debug mislabelled data points more efficiently than existing approaches.

Managing uncertainty in preferences is core to creating the next generation of conversational recommender systems (CRS). However, an often-overlooked element of conversational interaction is the role of clarification. Users are notoriously noisy at revealing their preferences, and a common error is being unnecessarily specific, e.g., suggesting ”chicken fingers” when a restaurant with a ”kids menu” was the intended preference. Correcting such errors requires reasoning about the level of generality and specificity of preferences and verifying that the user has expressed the correct level of generality. To this end, we propose a novel clarification-based conversational critiquing framework that allows the system to clarify user preferences as it accepts critiques. To support clarification, we propose the use of distributional embeddings that can capture the specificity and generality of concepts through distributional coverage while facilitating state-of-the-art embedding-based recommendation methods. Specifically, we incorporate Distributional Contrastive Embeddings of critiqueable keyphrases with user preference embeddings in a Variational Autoencoder recommendation framework that we term DCE-VAE. Our experiments show that our proposed DCE-VAE is (1) competitive in terms of general performance in comparison to state-of-the-art recommenders and (2) supports effective clarification-based critiquing in comparison to alternative clarification baselines. In summary, this work adds a new dimension of clarification to enhance the well-known critiquing framework along with a novel data-driven distributional embedding for clarification suggestions that significantly improves the efficacy of user interaction with critiquing-based CRSs.

Explaining the influence of training data on deep neural network predictions is a critical tool for debugging models through data curation. A recent tractable and appealing approach for this task was provided via the concept of Representer Point Selection (RPS), i.e. a method the leverages the dual form of $l_2$ regularized optimization in the last layer of the neural network to identify the contribution of training points to the prediction. However, two key drawbacks of RPS are that they (i) lead to disagreement between the originally trained network and the RP regularized network modification and (ii) often yield a static ranking of training data for the same class, independent of the data being classified. Inspired by the RPS approach, we propose an alternative method based on a local Jacobian Taylor expansion (LJE) of the Jacobian.We empirically compared RPS-LJE with the original RPS-$l_2$ on image classification (with ResNet), text classification recurrent neural networks (with Bi-LSTM), and tabular classification (with XGBoost) tasks.Quantitatively, we show that RPS-LJE slightly outperforms RPS-$l_2$ and other state-of-the-art data explanation methods by up to 3\% on a data debugging task. Qualitatively, we observe that RPS-LJE provides individualized explanations for each test data point rather than the class-specific static ranking of points in the original approach. Overall, RPS-LJE represents a novel approach to RPS that provides a powerful tool for data-oriented explanation and debugging.

Interactive (a.k.a. conversational) recommendation systems provide the potential capability to personalize interactions with increasingly prevalent dialog-based AI assistants. In the conversational recommendation setting, a user often has long-term preferences inferred from previous interactions along with ephemeral session-based preferences that need to be efficiently elicited through minimal interaction. Historically, Bayesian preference elicitation methods have proved effective for (i) leveraging prior information to incrementally estimate uncertainty in user preferences as new information is observed, and for (ii) supporting active elicitation of preference feedback to quickly zero in on the best recommendations in a session. Previous work typically focused on eliciting preferences in the space of items or a small set of attributes; in the dialog-based setting, however, we are faced with the task of eliciting preferences in the space of natural language while using this feedback to determine a user’s preferences in item space. To address this task in the era of modern, latent embedding-based recommender systems, we propose a method for coembedding user-item preferences with keyphrase descriptions (i.e., not explicitly known attributes, but rather subjective judgments mined from user reviews or tags) along with a closed-form Bayesian methodology for incrementally estimating uncertainty in user preferences based on elicited keyphrase feedback. We then combine this framework with well-known preference elicitation techniques that can leverage Bayesian posteriors such as Upper Confidence Bounds, Thompson Sampling, and a variety of other methods. Our empirical evaluation on real-world datasets shows that the proposed query selection strategies effectively update user beliefs, leading to high-quality recommendations with a minimal number of keyphrase queries.

Critiquing is a method for conversational recommendation that incrementally adapts recommendations in response to user preference feedback. Specifically, a user is iteratively provided with item recommendations and attribute descriptions for those items; the user may then either accept the recommendation or choose to critique an attribute to generate a new recommendation. A recent direction known as latent linear critiquing (LLC) takes a modern embedding-based approach that seeks to optimize the combination of user preference embeddings with embeddings of critiques based on subjective item descriptions (i.e., keyphrases from user reviews); LLC does so by exploiting the linear structure of the embeddings to efficiently optimize their weights in a linear programming (LP) formulation. In this paper, we revisit LLC and note that it’s score-based optimization approach inherently encourages extreme weightings in order to maximize predicted score gaps between preferred and non-preferred items. Noting that the overall end task objective in critiquing is to re-rank rather than re-score, in this paper we take a ranking optimization approach that seeks to optimize embedding weights based on observed rank violations from earlier critiquing iterations. We evaluate the proposed framework on two recommendation datasets containing user reviews. Empirical results demonstrate that ranking-based LLC generally outperforms scoring-based LLC and other baselines across a variety of datasets, critiquing styles, and both satisfaction and session-length performance metrics.

Providing explanations for recommended items not only allows users to understand the reason for receiving recommendations but also provides users with an opportunity to refine recommendations by critiquing undesired parts of the explanation. While much research focuses on improving the explanation of recommendations, less effort has focused on interactive recommendation by allowing a user to critique explanations. Aside from traditional constraint- and utility-based critiquing systems, the only end-to-end deep learning based critiquing approach in the literature so far, CE-VNCF, suffers from unstable and inefficient training performance. In this paper, we propose a Variational Autoencoder (VAE) based critiquing system to mitigate these issues and improve overall performance. The proposed model generates keyphrase-based explanations of recommendations and allows users to critique the generated explanations to refine their personalized recommendations. Our experiments show promising results: (1) The proposed model is competitive in terms of general performance in comparison to state-of-the-art recommenders, despite having an augmented loss function to support explanation and critiquing. (2) The proposed model can generate high-quality explanations compared to user or item keyphrase popularity baselines. (3) The proposed model is more effective in refining recommendations based on critiquing than CE-VNCF, where the rank of critiquing-affected items drops while general recommendation performance remains stable. In summary, this paper presents a significantly improved method for multi-step deep critiquing based recommender systems based on the VAE framework.

Critiquing is a method for conversational recommendation that iteratively adapts recommendations in response to user preference feedback. In this setting, a user is iteratively provided with an item recommendation and attribute description for that item; a user may either accept the recommendation, or critique the attributes in the item description to generate a new recommendation. Historical critiquing methods were largely based on explicit constraint- and utility-based methods for modifying recommendations w.r.t. critiqued item attributes. In this paper, we revisit the critiquing approach in the era of recommendation methods based on latent embeddings with subjective item descriptions (i.e., keyphrases from user reviews). Two critical research problems arise: (1) how to co-embed keyphrase critiques with user preference embeddings to update recommendations, and (2) how to modulate the strength of multi-step critiquing feedback, where critiques are not necessarily independent, nor of equal importance. To address (1), we build on an existing state-of-the-art linear embedding recommendation algorithm to align review-based keyphrase attributes with user preference embeddings. To address (2), we exploit the linear structure of the embeddings and recommendation prediction to formulate a linear program (LP) based optimization problem to determine optimal weights for incorporating critique feedback. We evaluate the proposed framework on two recommendation datasets containing user reviews with simulated users. Empirical results compared to a standard approach of averaging critique feedback show that our approach reduces the number of interactions required to find a satisfactory item and increases the overall success rate.

In many complex planning problems with factored, continuous state and action spaces such as Reservoir Control, Heating Ventilation and Air Conditioning (HVAC), and Navigation domains, it is difficult to obtain a model of the complex nonlinear dynamics that govern state evolution. However, the ubiquity of modern sensors allows us to collect large quantities of data from each of these complex systems and build accurate, nonlinear deep neural network models of their state transitions. But there remains one major problem for the task of control – how can we plan with deep network learned transition models without resorting to Monte Carlo Tree Search and other black-box transition model techniques that ignore model structure and do not easily extend to continuous domains? In this paper, we introduce two types of planning methods that can leverage deep neural network learned transition models: Hybrid Deep MILP Planner (HD-MILP-Plan) and Tensorflow Planner (TF-Plan). In HD-MILP-Plan, we make the critical observation that the Rectified Linear Unit (ReLU) transfer function for deep networks not only allows faster convergence of model learning, but also permits a direct compilation of the deep network transition model to a Mixed-Integer Linear Program (MILP) encoding. Further, we identify deep network specific optimizations for HD-MILP-Plan that improve performance over a base encoding and show that we can plan optimally with respect to the learned deep networks. In TF-Plan, we take advantage of the efficiency of auto-differentiation tools and GPU-based computation where we encode a subclass of purely continuous planning problems as Recurrent Neural Networks and directly optimize the actions through backpropagation. We compare both planners and show that TF-Plan is able to approximate the optimal plans found by HD-MILP-Plan in less computation time. Hence this article offers two novel planners for continuous state and action domains with learned deep neural net transition models: one optimal method (HD-MILP-Plan) and a scalable alternative for large-scale problems (TF-Plan).

Critiquing is a method for conversational recommendation that adapts recommendations in response to user preference feedback regarding item attributes. Historical critiquing methods were largely based on constraint- and utility-based methods for modifying recommendations w.r.t. these critiqued attributes. In this paper, we revisit the critiquing approach from the lens of deep learning based recommendation methods and language-based interaction. Concretely, we propose an end-to-end deep learning framework with two variants that extend the Neural Collaborative Filtering architecture with explanation and critiquing components. These architectures not only predict personalized keyphrases for a user and item but also embed language-based feedback in the latent space that in turn modulates subsequent critiqued recommendations. We evaluate the proposed framework on two recommendation datasets containing user reviews. Empirical results show that our modified NCF approach not only provides a strong baseline recommender and high-quality personalized item keyphrase suggestions, but that it also properly suppresses items predicted to have a critiqued keyphrase. In summary, this paper provides a first step to unify deep recommendation and language-based feedback in what we hope to be a rich space for future research in deep critiquing for conversational recommendation.

Supervised topic classifiers for Twitter and other media sources are important in a variety of long-term topic tracking tasks. Unfortunately, over long periods of time, features that are predictive during the training period may prove ephemeral and fail to generalize to prediction at future times. For example, if we trained a classifier to identify tweets concerning the topic of “Celebrity Death”, individual celebrity names and terms associated with these celebrities such as “Nelson Mandela” or “South Africa” would prove to be temporally unstable since they would not generalize over long periods of time; in contrast, terms like “RIP” (rest in peace) would prove to be temporally stable predictors of this topic over long periods of time. In this paper, we aim to design supervised learning methods for Twitter topic classifiers that are capable of automatically downweighting temporally unstable features to improve future generalization. To do this, we first begin with an oracular approach that chooses temporally stable features based on knowledge of both train and test data labels. We then search for feature metrics evaluated on only the training data that are capable of recovering the temporally stable features identified by our oracular definition. We next embed the top-performing metric as a temporal stability regularizer in logistic regression with the important property that the overall training objective retains convexity, hence enabling a globally optimal solution. Finally, we train our topic classifiers on 6 Twitter topics over roughly one year of data and evaluate on the following year of data, showing that logistic regression with our temporal stability regularizer generally outperforms logistic regression without such regularization across the full precision-recall continuum. Overall, these results establish a novel regularizer for training long-term temporally stable topic classifiers for Twitter and beyond.

Variational Autoencoder (VAE) based methods for Collaborative Filtering (CF) demonstrate remarkable performance for one-class (implicit negative) recommendation tasks by extending autoencoders with relaxed but tractable latent distributions. Explicitly modeling a latent distribution over user preferences allows VAEs to learn user and item representations that not only reproduce observed interactions, but also generalize them by leveraging learning from similar users and items. Unfortunately, VAE-CF can exhibit suboptimal learning properties; e.g., VAE-CFs will increase their prediction confidence as they receive more preferences per user, even when those preferences may vary widely and create ambiguity in the user representation. To address this issue, we propose a novel Queryable Variational Autoencoder (Q-VAE) variant of the VAE that explicitly models arbitrary conditional relationships between observations. The proposed model appropriately increases uncertainty (rather than reduces it) in cases where a large number of user preferences may lead to an ambiguous user representation. Our experiments on two benchmark datasets show that the Q-VAE generally performs comparably or outperforms VAE-based recommenders as well as other state-of-the-art approaches and is generally competitive across the user preference density spectrum, where other methods peak for certain preference density levels.

Previous highly scalable One-Class Collaborative Filtering (OC-CF) methods such as Projected Linear Recommendation (PLRec) have advocated using fast randomized SVD to embed items into a latent space, followed by linear regression methods to learn personalized recommendation models per user. However, naive SVD embedding methods often exhibit a strong popularity bias that prevents them from accurately embedding less popular items, which is exacerbated by the extreme sparsity of implicit feedback matrices in the OC-CF setting. To address this deficiency, we leverage insights from Noise Contrastive Estimation (NCE) to derive a closed-form, efficiently computable "depopularized" embedding. We show that NCE item embeddings combined with a personalized user model from PLRec produces superior recommendations that adequately account for popularity bias. Further analysis of the popularity distribution of recommended items demonstrates that NCE-PLRec uniformly distributes recommendations over the popularity spectrum while other methods exhibit distinct biases towards specific popularity subranges. Empirically, NCE-PLRec produces highly competitive performance with run-times an order of magnitude faster than existing state-of-the-art approaches for OC-CF.

In many real-world hybrid (mixed discrete continuous) planning problems such as Reservoir Control, Heating, Ventilation and Air Conditioning (HVAC), and Navigation, it is difficult to obtain a model of the complex nonlinear dynamics that govern state evolution. However, the ubiquity of modern sensors allow us to collect large quantities of data from each of these complex systems and build accurate, nonlinear deep network models of their state transitions. But there remains one major problem for the task of control -- how can we plan with deep network learned transition models without resorting to Monte Carlo Tree Search and other black-box transition model techniques that ignore model structure and do not easily extend to mixed discrete and continuous domains? In this paper, we make the critical observation that the popular Rectified Linear Unit (ReLU) transfer function for deep networks not only allows accurate nonlinear deep net model learning, but also permits a direct compilation of the deep network transition model to a Mixed-Integer Linear Program (MILP) encoding in a planner we call Hybrid Deep MILP Planning (HD-MILP-PLAN). We identify deep net specific optimizations and a simple sparsification method for HD-MILP-PLAN that improve performance over a naive encoding, and show that we are able to plan optimally with respect to the learned deep network.

Given recent deep learning results that demonstrate the ability to effectively optimize high-dimensional non-convex functions with gradient descent optimization on GPUs, we ask in this paper whether symbolic gradient optimization tools such as Tensorflow can be effective for planning in hybrid (mixed discrete and continuous) nonlinear domains with high dimensional state and action spaces? To this end, we demonstrate that hybrid planning with Tensorflow and RMSProp gradient descent is competitive with mixed integer linear program (MILP) based optimization on piecewise linear planning domains (where we can compute optimal solutions) and substantially outperforms state-of-the-art interior point methods for nonlinear planning domains. Furthermore, we remark that Tensorflow is highly scalable, converging to a strong plan on a large-scale concurrent domain with a total of 576,000 continuous action parameters distributed over a horizon of 96 time steps and 100 parallel instances in only 4 minutes. We provide a number of insights that clarify such strong performance including observations that despite long horizons, RMSProp avoids both the vanishing and exploding gradient problems. Together these results suggest a new frontier for highly scalable planning in nonlinear hybrid domains by leveraging GPUs and the power of recent advances in gradient descent with highly optimized toolkits like Tensorflow.