Embedding based models have been the state of art in collaborative filtering for over a decade. Traditionally, dot product or higher order equivalents used to combine two more embeddings, e.g., most notably matrix factorization. In recent years, it was suggested replace with learned similarity e.g. using multilayer perceptron (MLP). This approach is often referred as neural (NCF). this work, we revisit experiments NCF paper that popularized similarities MLPs. First, show proper...

The task of item recommendation requires ranking a large catalogue items given context. Item algorithms are evaluated using metrics that depend on the positions relevant items. To speed up computation metrics, recent work often uses sampled where only smaller set random and ranked. This paper investigates in more detail shows they inconsistent with their exact version, sense do not persist relative statements, e.g., recommender A is better than B, even expectation. Moreover, sampling size,...

Recommender systems personalize content by recommending items to users. Item recommendation algorithms are evaluated metrics that compare the positions of truly relevant among recommended items. To speed up computation metrics, recent work often uses sampled where only a smaller set random and ranked. This paper investigates such in more detail shows they inconsistent with their exact counterpart, sense do not persist relative statements, for example, recommender A is better than B , even...

Matrix factorization learned by implicit alternating least squares (iALS) is a popular baseline in recommender system research publications. iALS known to be one of the most computationally efficient and scalable collaborative filtering methods. However, recent studies suggest that its prediction quality not competitive with current state art, particular autoencoders other item-based In this work, we revisit four well-studied benchmarks where was reported perform poorly show proper tuning,...

We study the repeated, nonatomic congestion game, in which multiple populations of players share resources and make, at each iteration, a decentralized decision on to utilize. investigate following question: given model how individual update their strategies, does resulting dynamics strategy profiles converge set Nash equilibria one-shot game? consider particular strategies using algorithms with sublinear discounted regret. show that sequence converges sense Cesàro means. However,...

This paper presents a game theoretic framework for studying Stackelberg routing games on parallel networks with horizontal queues, such as transportation networks. First, we introduce new class of latency functions that models congestion due to the formation physical queues. For this class, some results from classical literature (in which is assumed be non-decreasing function flow) do not hold. In particular, find there may exist multiple Nash equilibria have different total costs. We...

We consider the Lighthill--Whitham--Richards traffic flow model on a junction composed by one mainline, an onramp, and offramp, which are connected node. The onramp dynamics is modeled using ordinary differential equation describing evolution of queue length. definition solution Riemann problem at based optimization use right-of-way parameter. numerical approximation carried out Godunov scheme, modified to take into account effects buffer. present result some simulations numerically check...

We study convergence properties of distributed learning dynamics in repeated stochastic routing games. The game is that each player observes a vector, the conditional expectation which equal to true loss (almost surely). In particular, we propose model every m follows mirror descent with Bregman divergence D <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ψm</sub> and rates η xmlns:xlink="http://www.w3.org/1999/xlink">t</sub> <sup...

Most optimal routing problems focus on minimizing travel time or distance traveled. Oftentimes, a more useful objective is to maximize the probability of on-time arrival, which requires statistical distributions times, rather than just mean values. We propose method estimate large-scale road networks, using probe vehicle data collected from GPS. present framework that works with large input data, and scales linearly size network. Leveraging planar topology graph, computes efficiently...

The routing game models congestion in transportation networks, communication and other cyber physical systems which agents compete for shared resources. We consider an online learning model of player dynamics: at each iteration, every chooses a route (or probability distribution over routes, corresponds to flow allocation the network), then joint decision all players determines costs path, are revealed players. pose following estimation problem: given sequence decisions corresponding costs,...

We consider a repeated routing game over finite horizon with partial control under selfish response, in which central authority can fraction of the flow and seeks to improve network-wide objective, while remaining applies an online learning algorithm. This problem is inspired from one-shot Stackelberg game. Our setting different that we do not assume players play Nash equilibrium; instead, they apply results optimal dynamics. propose methods for approximately solving this problem: A greedy...

We consider the System Optimal Dynamic Traffic Assignment (SO-DTA) problem with Partial Control for general networks physical queuing. Our goal is to optimally control any subset of agents minimize total congestion all in network. adopt a flow dynamics model that Godunov discretization Lighthill–Williams–Richards partial differential equation triangular flux function and corresponding multicommodity junction solver. The formulation generalizes SO-DTA cases where only fraction can be...

As our ground transportation infrastructure modernizes, the large amount of data being measured, transmitted, and stored motivates an analysis privacy aspect these emerging cyber-physical technologies. In this paper, we consider in routing game, where origins destinations drivers are considered private. This is motivated by fact that spatiotemporal information can easily be used as basis for inferences a person's activities. More specifically, differential mapping from flow each...

Recommender System research suffers currently from a disconnect between the size of academic data sets and scale industrial production systems. In order to bridge that gap we propose generate more massive user/item interaction by expanding pre-existing public sets. User/item incidence matrices record interactions users items on given platform as large sparse matrix whose rows correspond columns items. Our technique expands such larger numbers (users), (items) non zero values (interactions)...

We study the problem of learning similarity functions over very large corpora using neural network embedding models. These models are typically trained SGD with sampling random observed and unobserved pairs, a number samples that grows quadratically corpus size, making it expensive to scale corpora. propose new efficient methods train these without having sample pairs. Inspired by matrix factorization, our approach relies on adding global quadratic penalty all pairs examples expressing this...

We consider a routing game played on graph, in which different populations of drivers (or packet routers) iteratively make decisions and seek to minimize their delays. The Nash equilibria the are known be minimizers convex potential function, over product simplexes represent strategy spaces populations. class population dynamics only uses local loss information, can interpreted as mirror descent potential. show that for vanishing, non-summable learning rates, guaranteed converge set...

We consider the problem of projecting a vector onto simplex Δ = {x ∈ ℝ+d : Σi=1d xi 1}, using Bregman projection. This is common in first-order methods for convex optimization and online-learning algorithms, such as mirror descent. derive KKT conditions projection problem, show that divergences induced by ω-potentials, one can efficiently compute solution bisection method. More precisely, an ω-approximate be obtained O(d log 1/ω). also class exponential potentials which exact computed...

The routing game models congestion in transportation networks, communication and other cyber physical systems which agents compete for shared resources. We consider an online learning model of player dynamics: at each iteration, every chooses a route (or probability distribution over routes, corresponds to flow allocation the network), then joint decision all players determines costs path, are revealed players. pose following estimation problem: given sequence decisions corresponding costs,...

We study inefficiencies in parallel networks with horizontal queues due to the selfish behavior of players, by comparing social optima Nash equilibria. The article expands studies on routing games which traditionally model congestion latency functions that increase flow a particular link. This type function cannot capture effects queues. Latencies as density, and can decrease increasing latencies. class arises transportation networks. For static analysis parallel-link networks, we show there...