The video game industry has adopted recommendation systems to boost users interest with a focus on game sales. Other exciting applications within video games are those that help the player make decisions that would maximize their playing experience, which is a desirable feature in real-time strategy video games such as Multiplayer Online Battle Arena (MOBA) like as DotA and LoL. Among these tasks, the recommendation of items is challenging, given both the contextual nature of the game and how it exposes the dependence on the formation of each team. Existing works on this topic do not take advantage of all the available contextual match data and dismiss potentially valuable information. To address this problem we develop TTIR, a contextual recommender model derived from the Transformer neural architecture that suggests a set of items to every team member, based on the contexts of teams and roles that describe the match. TTIR outperforms several approaches and provides interpretable recommendations through visualization of attention weights. Our evaluation indicates that both the Transformer architecture and the contextual information are essential to get the best results for this item recommendation task. Furthermore, a preliminary user survey indicates the usefulness of attention weights for explaining recommendations as well as ideas for future work. The code and dataset are available at: https://github.com/ojedaf/IC-TIR-Lol.

This paper introduces a new methodology for detecting anomalies in time series data, with a primary application to monitoring the health of (micro-) services and cloud resources. The main novelty in our approach is that instead of modeling time series consisting of real values or vectors of real values, we model time series of probability distributions over real values (or vectors). This extension to time series of probability distributions allows the technique to be applied to the common scenario where the data is generated by requests coming in to a service, which is then aggregated at a fixed temporal frequency. Our method is amenable to streaming anomaly detection and scales to monitoring for anomalies on millions of time series. We show the superior accuracy of our method on synthetic and public real-world data. On the Yahoo Webscope data set, we outperform the state of the art in 3 out of 4 data sets and we show that we outperform popular open-source anomaly detection tools by up to 17% average improvement for a real-world data set.

When the training data are maliciously tampered, the predictions of the acquired deep neural network (DNN) can be manipulated by an adversary known as the Trojan attack (or poisoning backdoor attack). The lack of robustness of DNNs against Trojan attacks could significantly harm real-life machine learning (ML) systems in downstream applications, therefore posing widespread concern to their trustworthiness. In this paper, we study the problem of the Trojan network (TrojanNet) detection in the data-scarce regime, where only the weights of a trained DNN are accessed by the detector. We first propose a data-limited TrojanNet detector (TND), when only a few data samples are available for TrojanNet detection. We show that an effective data-limited TND can be established by exploring connections between Trojan attack and prediction-evasion adversarial attacks including per-sample attack as well as all-sample universal attack. In addition, we propose a data-free TND, which can detect a TrojanNet without accessing any data samples. We show that such a TND can be built by leveraging the internal response of hidden neurons, which exhibits the Trojan behavior even at random noise inputs. The effectiveness of our proposals is evaluated by extensive experiments under different model architectures and datasets including CIFAR-10, GTSRB, and ImageNet.

The learning speed of feed-forward neural networks is notoriously slow and has presented a bottleneck in deep learning applications for several decades. For instance, gradient-based learning algorithms, which are used extensively to train neural networks, tend to work slowly when all of the network parameters must be iteratively tuned. To counter this, both researchers and practitioners have tried introducing randomness to reduce the learning requirement. Based on the original construction of Igelnik and Pao, single layer neural-networks with random input-to-hidden layer weights and biases have seen success in practice, but the necessary theoretical justification is lacking. In this paper, we begin to fill this theoretical gap. We provide a (corrected) rigorous proof that the Igelnik and Pao construction is a universal approximator for continuous functions on compact domains, with approximation error decaying asymptotically like $O(1/\sqrt{n})$ for the number $n$ of network nodes. We then extend this result to the non-asymptotic setting, proving that one can achieve any desired approximation error with high probability provided $n$ is sufficiently large. We further adapt this randomized neural network architecture to approximate functions on smooth, compact submanifolds of Euclidean space, providing theoretical guarantees in both the asymptotic and non-asymptotic forms. Finally, we illustrate our results on manifolds with numerical experiments.

The learning rate (LR) is one of the most important hyper-parameters in stochastic gradient descent (SGD) for deep neural networks (DNNs) training and generalization. However, current hand-designed LR schedules need to manually pre-specify schedule as well as its extra hyper-parameters, which limits its ability to adapt non-convex optimization problems due to the significant variation of training dynamic. To address this issue, we propose a model capable of adaptively learning LR schedule from data. We specifically design a meta-learner with explicit mapping formulation to parameterize LR schedules, which can adjust LR adaptively to comply with current training dynamic by leveraging the information from past training histories. Image and text classification benchmark experiments substantiate the capability of our method for achieving proper LR schedules compared with baseline methods. Moreover, we transfer the learned LR schedule to other various tasks, like different training batch sizes, epochs, datasets, network architectures, especially large scale ImageNet dataset, showing its stronger generalization capability than related methods. Finally, guided by a small set of clean validation set, we show our method can achieve better generalization error when training data is biased with corrupted noise than baseline methods.

We compare the discretize-optimize (Disc-Opt) and optimize-discretize (Opt-Disc) approaches for time-series regression and continuous normalizing flows (CNFs) using neural ODEs. Neural ODEs are ordinary differential equations (ODEs) with neural network components. Training a neural ODE is an optimal control problem where the weights are the controls and the hidden features are the states. Every training iteration involves solving an ODE forward and another backward in time, which can require large amounts of computation, time, and memory. Comparing the Opt-Disc and Disc-Opt approaches in image classification tasks, Gholami et al. (2019) suggest that Disc-Opt is preferable due to the guaranteed accuracy of gradients. In this paper, we extend the comparison to neural ODEs for time-series regression and CNFs. Unlike in classification, meaningful models in these tasks must also satisfy additional requirements beyond accurate final-time output, e.g., the invertibility of the CNF. Through our numerical experiments, we demonstrate that with careful numerical treatment, Disc-Opt methods can achieve similar performance as Opt-Disc at inference with drastically reduced training costs. Disc-Opt reduced costs in six out of seven separate problems with training time reduction ranging from 39% to 97%, and in one case, Disc-Opt reduced training from nine days to less than one day.

We develop Banach spaces for ReLU neural networks of finite depth $L$ and infinite width. The spaces contain all finite fully connected $L$-layer networks and their $L^2$-limiting objects under bounds on the natural path-norm. Under this norm, the unit ball in the space for $L$-layer networks has low Rademacher complexity and thus favorable generalization properties. Functions in these spaces can be approximated by multi-layer neural networks with dimension-independent convergence rates. The key to this work is a new way of representing functions in some form of expectations, motivated by multi-layer neural networks. This representation allows us to define a new class of continuous models for machine learning. We show that the gradient flow defined this way is the natural continuous analog of the gradient descent dynamics for the associated multi-layer neural networks. We show that the path-norm increases at most polynomially under this continuous gradient flow dynamics.

Forecasting of multivariate time-series is an important problem that has applications in many domains, including traffic management, cellular network configuration, and quantitative finance. In recent years, researchers have demonstrated the value of applying deep learning architectures for these problems. A special case of the problem arises when there is a graph available that captures the relationships between the time-series. In this paper we propose a novel learning architecture that achieves performance competitive with or better than the best existing algorithms, without requiring knowledge of the graph. The key elements of our proposed architecture are (i) jointly performing backcasting and forecasting with a deep fully-connected architecture; (ii) stacking multiple prediction modules that target successive residuals; and (iii) learning a separate causal relationship graph for each layer of the stack. We can view each layer as predicting a component of the time-series; the differing nature of the causal graphs at different layers can be interpreted as indicating that the multivariate predictive relationships differ for different components. Experimental results for two public traffic network datasets illustrate the value of our approach, and ablation studies confirm the importance of each element of the architecture.

A context-aware recommender system (CARS) applies sensing and analysis of user context to provide personalized services. The contextual information can be driven from sensors in order to improve the accuracy of the recommendations. Yet, generating accurate recommendations is not enough to constitute a useful system from the users' perspective, since certain contextual information may cause different issues, such as draining the user's battery, privacy issues, and more. Adding high-dimensional contextual information may increase both the dimensionality and sparsity of the model. Previous studies suggest reducing the amount of contextual information by selecting the most suitable contextual information using a domain knowledge. Another solution is compressing it into a denser latent space, thus disrupting the ability to explain the recommendation item to the user, and damaging users' trust. In this paper we present an approach for selecting low-dimensional subsets of the contextual information and incorporating them explicitly within CARS. Specifically, we present a novel feature-selection algorithm, based on genetic algorithms (GA), that outperforms SOTA dimensional-reduction CARS algorithms, improves the accuracy and the explainability of the recommendations, and allows for controlling user aspects, such as privacy and battery consumption. Furthermore, we exploit the top subsets that are generated along the evolutionary process, by learning multiple deep context-aware models and applying a stacking technique on them, thus improving the accuracy while remaining at the explicit space. We evaluated our approach on two high-dimensional context-aware datasets driven from smartphones. An empirical analysis of our results validates that our proposed approach outperforms SOTA CARS models while improving transparency and explainability to the user.

Rainfall prediction helps planners anticipate potential social and economic impacts produced by too much or too little rain. This research investigates a class-based approach to rainfall prediction from 1-30 days in advance. The study made regional predictions based on sequences of daily rainfall maps of the continental US, with rainfall quantized at 3 levels: light or no rain; moderate; and heavy rain. Three regions were selected, corresponding to three squares from a $5\times5$ grid covering the map area. Rainfall predictions up to 30 days ahead for these three regions were based on a support vector machine (SVM) applied to consecutive sequences of prior daily rainfall map images. The results show that predictions for corner squares in the grid were less accurate than predictions obtained by a simple untrained classifier. However, SVM predictions for a central region outperformed the other two regions, as well as the untrained classifier. We conclude that there is some evidence that SVMs applied to large-scale precipitation maps can under some conditions give useful information for predicting regional rainfall, but care must be taken to avoid pitfall