Point-of-Interest (POI) recommendation has become an important means to help people discover attractive and interesting places, especially when users travel out of town. However, extreme sparsity of user-POI matrix creates a severe challenge. To cope with this challenge, we propose a unified probabilistic generative model, Topic-Region Model (TRM), to simultaneously discover the semantic, temporal and spatial patterns of users' check-in activities, and to model their joint effect on users' decision-making for selection of POIs to visit. To demonstrate the applicability and flexibility of TRM, we investigate how it supports two recommendation scenarios in a unified way, i.e., home-town recommendation and out-of-town recommendation. To support real-time POI recommendation, we further extend the TRM model to an online learning model TRM-Online to track changing user interests and speed up the model training. Besides, based on the learnt model, we propose a clustering-based branch and bound algorithm (CBB) to prune the POI search space and facilitate fast retrieval of the top-k recommendations. We conduct extensive experiments to evaluate the performance of our proposals on two real-world datasets including recommendation effectiveness, overcoming cold-start problem, recommendation efficiency and model training efficiency. The experimental results demonstrate the superiority of our proposals.
This paper presents a study of the question of which baseline to use when testing a new retrieval technique. We show that measuring a statistically significant improvement over a weak baseline is not necessarily a good predictor of whether the improvement will be measured on a strong baseline. Indeed, sometimes strong baselines are often made worse. We show there is little past research outlining how strong a baseline must be in order for an experimental comparison to be valid. This paper describes a series of experiments examining the testing of new techniques against multiple baselines. We show that the only way to be truly confident that a new technique is a real contribution is to compare it against, nothing less than the state of the art. However, showing consistent significant improvement over a basket of lower performing baselines can give researchers a level of confidence that a new technique will at the very least not degrade a state of the art system.
The prediction of the content evolution of news events can support or improve many online news services. For example, news websites can attract people by ranking the news events on the frontpage according their potentials of the content evolution. However, existing prediction methods are mainly based on time series historical data, which are not suitable for the news events with limited historical data and busrty property. In this paper, we propose a purely content-based method to predict the evolution of the news events. A news event at a given time point is considered as a system composed of different keywords, and the uncertainty of this system is defined and measured as the semantic uncertainty of this news event. At the same time, we have defined an uncertainty space limited by two extreme states: the most uncertain state and the most uncertain states. We believe that the semantic uncertainty has correlation with the content evolution of the news events, so it can be used to predict the evolution of the news events. The experimental results show that the correlation does exist and is stronger than the correlation of predicted value from the time-series based method with the content change.
The long tail theory for consumer demand implies the need for more accurate personalization technologies to target items to the users who most desire them. A key tenet of personalization is the capacity to model user preferences. Most of the previous work on recommendation and personalization has focused primarily on individual preferences. While some focus on shared preferences between pairs of users, they assume that the same similarity value applies to all items. Here we investigate the notion of "context", hypothesizing that while two users may agree on their preferences on some items, they may also disagree on other items. To model this, we design probabilistic models for the generation of rating differences between pairs of users across different items. Since this model also involves the estimation of rating differences on unseen items for the purpose of prediction, we further conduct a systematic analysis of matrix factorization and tensor factorization methods in this estimation, and propose a factorization model with a novel objective function of minimizing error in rating differences. Experiments on several real-life rating data sets show that our proposed models consistently yields context-specific similarity values that perform better on a prediction task than models relying on shared preferences.
Current random forest (RF) based learning-to-rank (LtR) algorithms use a classification or regression framework to solve the ranking problem in a pointwise manner. The success of this simple yet effective approach coupled with the inherent parallelizability of the learning algorithm makes it a strong candidate for widespread adoption. In this paper we aim to better understand the effectiveness of these algorithms by comparing between pointwise and listwise objective functions. We introduce what we believe to be the first listwise version of an RF-based LtR algorithm. The algorithm directly optimizes an information retrieval metric of choice in a greedy manner. Computational complexity of the listwise approach is higher than the pointwise counterpart, hence for larger datasets, we design a hybrid algorithm which combines listwise and pointwise objective in each tree. Experimental results reveal that the listwise/hybrid algorithm outperforms the pointwise approach on the majority (but not all) of LtR datasets available. We then investigate several aspects of the two algorithms to better understand the inevitable performance trade-offs. This investigation suggests that for large LtR datasets the splitting criterion (implied by the objective function) is important particularly if small trees are learnt, but as the trees are grown large the advantage of the listwise approach diminishes relative to its pointwise counterpart.
Query auto-completion (QAC) assists web search users in formulating queries with a few keystrokes. Previous work on QAC mainly centers around returning a list of completions to users, aiming to push queries that are most likely intended by the user to the top positions but ignoring the redundancy among the candidates in the list. This may push valuable suggestions out of the list, given that only a limited number of candidates can be shown to the user. In this paper, we consider the task of diversifying query auto-completion, aiming to return correct completions early in a list of candidates and reduce the redundancy among QAC candidates. We develop a greedy query selection approach that predicts query completions based on the current search popularity of candidates and on the intents of previous queries in session. The popularity of candidates can be aggregated from query logs. However, search intents are implicitly expressed by previous clicked documents in session, which are categorized using the open directory project. Bayesian probabilistic matrix factorization is applied to derive the query intent distribution over aspects. We quantify the improvement of our model against a state-of-the-art baseline using two real-world query logs and show it beats the baseline in terms of metrics used in QAC and diversification.