Abstract Context Software design in e-commerce can be improved with user data through controlled experiments (i.e. A/B tests) to better meet user needs. Machine learning-based algorithmic optimization techniques extends the approach to large number of variables to personalize software to different user needs. So far the optimization techniques has only been applied to optimize software of low complexity, such as colors and wordings of text. Objective In this paper, we introduce the COMBO toolkit with capability to model optimization variables and their relationship constraints specified through an embedded domain-specific language. The toolkit generates personalized software configurations for users as they arrive in the system, and the configurations improve over time in in relation to some given metric. COMBO has several implementations of machine learning algorithms and constraint solvers to optimize the model with user data by software developers without deep optimization knowledge. Method The toolkit was validated in a proof-of-concept by implementing two features that are relevant to Apptus, an e-commerce company that develops algorithms for web shops. The algorithmic performance was evaluated in simulations with realistic historic user data. Results The validation shows that the toolkit approach can model and improve relatively complex features with many types of variables and constraints, without causing noticeable delays for users. Conclusions We show that modeling software hierarchies in a formal model facilitates algorithmic optimization of more complex software. In this way, using COMBO, developers can make data-driven and personalized software products.

This work presents an extension of Thompson Sampling bandit policy for orchestrating the collection of base recommendation algorithms for e-commerce. We focus on the problem of item-to-item recommendations, for which multiple behavioral and attribute-based predictors are provided to an ensemble learner. In addition, we detail the construction of a personalized predictor based on k -Nearest Neighbors ( k NN), with temporal decay capabilities and event weighting. We show how to adapt Thompson Sampling to realistic situations when neither action availability nor reward stationarity is guaranteed. Furthermore, we investigate the effects of priming the sampler with pre-set parameters of reward probability distributions by utilizing the product catalog and/or event history, when such information is available. We report our experimental results based on the analysis of three real-world e-commerce datasets.

This work presents an extension of Thompson Sampling bandit policy for orchestrating the collection of base recommendation algorithms for e-commerce. We focus on the problem of item-to-item recommendations, for which multiple behavioral and attribute-based predictors are provided to an ensemble learner. We show how to adapt Thompson Sampling to realistic situations when neither action availability nor reward stationarity is guaranteed. Furthermore, we investigate the effects of priming the sampler with pre-set parameters of reward probability distributions by utilizing the product catalog and/or event history, when such information is available. We report our experimental results based on the analysis of three real-world e-commerce datasets.

We study bandit algorithms for e-commerce recommender systems. The question we pose is whether it is necessary to consider reinforcement learning effects in recommender systems. A key reason to introduce a recommender system for a product page on an e-commerce site is to increase the order value by improving the chance of making an upsale. If the recommender system merely predicts the next purchase, there might be no positive effect at all on the order value, since the recommender system predicts sales that would have happened independent of the recommender system. What we really are looking for are the false negatives, i.e., purchases that happen as a consequence of the recommender system. These purchases entail the entire uplift and should be present as reinforcement learning effects. This effect cannot be displayed in a simulation of the site, since there are no reinforcement learning effects present in a simulation. The attribution model must capture the uplift to guarantee an increased order value. However, such an attribution model is not practical, due to data sparsity. Given this starting point, we study some standard attribution models for e-commerce recommender systems, and describe how these fare when applied in a reinforcement learning algorithm, both in a simulation and on live sites.

We study the problem of optimizing recommendation systems for e-commerce sites. We consider in particular a combinatorial solution to this optimization based on the well known Maximum Coverage problem that asks for the k sets (products) that cover the most elements from a ground set (consumers). This formulation provides an abstract model for what k products should be recommended to maximize the probability of consumer purchase. Unfortunately, Maximum Coverage is NP-complete but an efficient approximation algorithm exists based on the Greedy methodology.

Abstract We propose a family of novel Chord‐based P2P schemes retaining all positive aspects that made Chord a popular topology for routing in P2P networks. The schemes, based on the Fibonacci number system, allow to simultaneously improve on the maximum/average number of hops for lookups and the routing table size per node. © 2008 Wiley Periodicals, Inc. NETWORKS, 2008

We propose routing schemes that optimize the average number of hops for lookup requests in peer-to-peer (P2P) systems without adding any overhead to the system. Our work is inspired by the recently introduced variation of greedy routing, called neighbor-of-neighbor (NoN), which allows to get optimal average path length with respect to the degree. Our proposal has the advantage of first "limiting" and then "eliminating" the use of randomization. As a consequence, the NoN technique can be implemented with our schemes without adding any overhead. Analyzed networks include several popular topologies: chord, hypercube based networks, symphony, skip-graphs. Theoretical results and extensive simulations show that the proposed simplifications (while maintaining the original node degree) do not increase the average path length of the networks, which is often improved in practice. The improvement is obtained with no harm to the operational efficiency (e.g. stability, ease of programming, scalability, fault-tolerance) of the considered systems.

Summary form only given. We study on-demand source initiated protocols for mobile wireless networks. In particular, we study the flooding procedure commonly used by these protocols to set up temporary communication paths. The benefit of the flooding technique is its generosity regarding changes in network structure. On the other hand, each time a message is sent, the entire network will be involved to set up the communication path from the source node, to the target node. We propose a new approach, which we call limited broadcasting. It is aimed to reduce the overhead by localizing the search for the target node both in terms of the time the process needs to globally stop after the target has been reached and/or in terms of the region which is affected by the search. It works in unknown networks and does not need any kind of additional information.

Greedy routing has been used in most of the proposed P2P networks because of several reasons. One of the main advantages is that greedy routing is very simple to implement and has some “implicit” fault-tolerance capabilities. It was however noticed that greedy routing usually produces paths of length larger than what would be required in a network of the given node degree. As an example some popular topologies like Chord have degree O(log n) and the greedy routing produces an average path length O(log n) whereas the lower bound is Ω(log n/log log n). The use of randomization allowed to show networks with optimal average path length. Recently a novel approach for routing in DHTs which improves on greedy routing has been proposed [4]. This approach, called NoN (Neighbors–of–Neighbors), substantially consists in making the greedy choice by looking not only at the neighbors of a node but at all the nodes at distance at most 2 from the node itself. The NoN approach together with the use of randomization in establishing the neighbors of the nodes which are present in the network, can optimally reduce the latency in several well known topologies. Hence the use of randomization, inspired by the Small-world idea introduced by Kleinberg [2], together with the NoN routing allows to maintain, to some extent, the advantages of greedy routing while optimizing the latency. Our goal is to retain the improvements given by the NoN routing over randomized networks, while eliminating the drawback in system overhead implied by this technique. In fact, randomization and NoN routing require the transmission to a node of its neighbors’s neighbors. While the authors in [4] argue that this can be done without extra cost by using keep-alive TCP messages, we eliminate the extracommunication at all and, similarly, eliminate the need of storing in each node its neighbors’s neighbors. To this aim we need to eliminate the random factor in establishing each neighbor of a node. In fact, determinism allows each node to calculate locally the neighbors of its neighbors. ∗ Work partially supported by EU RTN project ARACNE and by Italian FIRB WEBMINDS project

A box-tree is a \ifasci so-called \emph{bounding-volume hierarchy} \else bounding-volume hierarchy \fi that uses axis-aligned boxes as bounding volumes. The query complexity of a box-tree with respect to a given type of query is the maximum number of nodes visited when answering such a query. We describe several new algorithms for constructing box-trees with small worst-case query complexity with respect to queries with axis-parallel boxes and with points. We also prove lower bounds on the worst-case query complexity for box-trees, which show that our results are optimal or close to optimal. Finally, we present algorithms to convert box-trees to R-trees, resulting in R-trees with (almost) optimal query complexity.