If an outbound flow is observed at the boundary of a protected network, destined to an IP address within a few addresses of a known malicious IP address, should it be considered a suspicious flow? Conventional blacklisting is not going to cut it in this situation, and the established fact that malicious IP addresses tend to be highly clustered in certain portions of IP address space, should indeed raise suspicions. We present a new approach for perimeter defense that addresses this concern. At the heart of our approach, we attempt to infer internal, hidden boundaries in IP address space, that lie within publicly known boundaries of registered IP netblocks. Our hypothesis is that given a known bad IP address, other IP address in the same internal contiguous block are likely to share similar security properties, and may therefore be vulnerable to being similarly hacked and used by attackers in the future. In this paper, we describe how we infer hidden internal boundaries in IPv4 netblocks, and what effect this has on being able to predict malicious IP addresses.

Ever wonder why that Kia Ad ran during Iron Chef? Traditional advertising methodology on television is a fascinating mix of marketing, branding, measurement, and predictive modeling. While still a robust business, it is at risk with the recent growth of online and time-shifted (recorded) television. A particular issue is that traditional methods for television advertising are far less efficient than their counterparts in the online world which employ highly sophisticated computational techniques. This paper formalizes an approach to eliminate some of these inefficiencies by recasting the process of television advertising media campaign generation in a computational framework. We describe efficient mathematical approaches to solve for the task of finding optimal campaigns for specific target audiences. In two case studies, our campaigns report gains in key operational metrics of up to 56% compared to campaigns generated by traditional methods.

This paper shows that the location of screen taps on modern smartphones and tablets can be identified from accelerometer and gyroscope readings. Our findings have serious implications, as we demonstrate that an attacker can launch a background process on commodity smartphones and tablets, and silently monitor the user's inputs, such as keyboard presses and icon taps. While precise tap detection is nontrivial, requiring machine learning algorithms to identify fingerprints of closely spaced keys, sensitive sensors on modern devices aid the process. We present TapPrints, a framework for inferring the location of taps on mobile device touch-screens using motion sensor data combined with machine learning analysis. By running tests on two different off-the-shelf smartphones and a tablet computer we show that identifying tap locations on the screen and inferring English letters could be done with up to 90% and 80% accuracy, respectively. By optimizing the core tap detection capability with additional information, such as contextual priors, we are able to further magnify the core threat.

Typical recommender systems use the root mean squared error (RMSE) between the predicted and actual ratings as the evaluation metric. We argue that RMSE is not an optimal choice for this task, especially when we will only recommend a few (top) items to any user. Instead, we propose using a ranking metric, namely normalized discounted cumulative gain (NDCG), as a better evaluation metric for this task. Borrowing ideas from the learning to rank community for web search, we propose novel models which approximately optimize NDCG for the recommendation task. Our models are essentially variations on matrix factorization models where we also additionally learn the features associated with the users and the items for the ranking task. Experimental results on a number of standard collaborative filtering data sets validate our claims. The results also show the accuracy and efficiency of our models and the benefits of learning features for ranking.

While latent factor models are built using ratings data, which is typically assumed static, the ability to incorporate different kinds of subsequent user feedback is an important asset. For instance, the user might want to provide additional information to the system in order to improve his personal recommendations. To this end, we examine a novel scheme for efficiently learning (or refining) user parameters from such feedback. We propose a scheme where users are presented with a sequence of pair wise preference questions: "Do you prefer item A over B?". User parameters are updated based on their response, and subsequent questions are chosen adaptively after incorporating the feedback. We operate in a Bayesian framework and the choice of questions is based on an information gain criterion. We validate the scheme on the Netflix movie ratings data set. A user study and automated experiments validate our findings.

This paper investigates the problem of generating on-demand recommendations over a dataset of items where the input is a selection of a few of the items. As an example in the context of a movie dataset, the user may wish to see a list of movies related to the three animation movies 'Finding Nemo', 'Up' and 'Spirited Away'. In this case, it would be expected that the list returned would contain other animation movies like 'Wall-E', 'Princess Mononoke' etc. Thus, this problem can be viewed as a type of "clustering on demand" problem [1]. It is the set form of input that distinguishes this problem from a standard information retrieval problem where the query is usually a single item or an abstraction of a single item in the dataset. In this paper, we present several new approaches to dealing with this problem. We also show some representative results on a movie text dataset.

State-of-the-art probabilistic models of text such as n-grams require an exponential number of examples as the size of the context grows, a problem that is due to the discrete word representation. We propose to solve this problem by learning a continuous-valued and low-dimensional mapping of words, and base our predictions for the probabilities of the target word on non-linear dynamics of the latent space representation of the words in context window. We build on neural networks-based language models; by expressing them as energy-based models, we can further enrich the models with additional inputs such as part-of-speech tags, topic information and graphs of word similarity. We demonstrate a significantly lower perplexity on different text corpora, as well as improved word accuracy rate on speech recognition tasks, as compared to Kneser-Ney back-off n-gram-based language models.

Reinforcement learning (RL) is a promising technique for creating a dialog manager. RL accepts features of the current dialog state and seeks to find the best action given those features. Although it is often easy to posit a large set of potentially useful features, in practice, it is difficult to find the subset which is large enough to contain useful information yet compact enough to reliably learn a good policy. In this paper, we propose a method for RL optimization which automatically performs feature selection. The algorithm is based on least-squares policy iteration, a state-of-the-art RL algorithm which is highly sampleefficient and can learn from a static corpus or on-line. Experiments in dialog simulation show it is more stable than a baseline RL algorithm taken from a working dialog system.

We present an extension to the Lasso [6] for binary classification problems with ordered attributes. Inspired by the Fused Lasso [5] and the Group Lasso [7, 3] models, we aim to both discover and model runs (contiguous subgroups of the variables) that are highly predictive. We call the extended model LAPS (the Lasso with Attribute Partition Search). Such problems commonly arise in financial and medical domains, where predictors are time series variables, for example. This paper outlines the formulation of the problem, an algorithm to obtain the model coefficients and experiments showing applicability to practical problems of this type.

Classification problems with functionally structured input variables arise naturally in many applications. In a clinical domain, for example, input variables could include a time series of blood pressure measurements. In a financial setting, different time series of stock returns might serve as predictors. In an archaeological application, the 2D profile of an artifact may serve as a key input variable. In such domains, accuracy of the classifier is not the only reasonable goal to strive for; classifiers that provide easily interpretable results are also of value. In this work, we present an intuitive scheme for extending decision trees to handle functional input variables. Our results show that such decision trees are both accurate and readily interpretable.