Data selection bias has been a long-lasting challenge in the machine learning domain, especially in multi-stage recommendation systems, where the distribution of labeled items for model training is very different from that of the actual candidates during inference time. This distribution shift is even more prominent in the context of online advertising where the user base is diverse and the platform contains a wide range of contents. In this paper, we first investigate the data selection bias in the upper funnel (Ads Retrieval) of Pinterest's multi-cascade ads ranking system. We then conduct comprehensive experiments to assess the performance of various state-of-the-art methods, including transfer learning, adversarial learning, and unsupervised domain adaptation. Moreover, we further introduce some modifications into the unsupervised domain adaptation and evaluate the performance of different variants of this modified method. Our online A/B experiments show that the modified version of unsupervised domain adaptation (MUDA) could provide the largest improvements to the performance of Pinterest's advertisement ranking system compared with other methods and the one used in current production.

Knowing the perceived economic value of words is often desirable for applications such as product naming and pricing. However, there is a lack of understanding on the underlying economic worths of words, even though we have seen some breakthrough on learning the semantics of words. In this work, we bridge this gap by proposing a joint-task neural network model, Word Worth Model (WWM), to learn word embedding that captures the underlying economic worths. Through the design of WWM, we incorporate contextual factors, e.g., product's brand name and restaurant's city, that may affect the aggregated monetary value of a textual item. Via a comprehensive evaluation, we show that, compared with other baselines, WWM accurately predicts missing words when given target words. We also show that the learned embeddings of both words and contextual factors reflect well the underlying economic worths through various visualization analyses.

Although widely used in various areas, the Deep Neural Network suffers from the lack of interpretability. Existing works usually focus on one data instance and the found explanations are thus limited. We argue that a good understanding of a model should contain both systematic explanations of model behavior and effective detection of its vulnerability. Particularly we propose to use backward deduction to achieve these two goals. Given a constraint on the model output, the deduction backtracks the architecture to find corresponding data ranges in the input. In each layer, depending on the type, specific rules and/or algorithms are developed. The resulted ranges in the end can be interpreted by sampling exemplary data from them. In experiment we show that with different strategies in selecting the data ranges, the sampled fake data can either explain the model or reflect its vulnerability.

The often fierce competition on crowdfunding markets can significantly affect project success. While various factors have been considered in predicting the success of crowdfunding projects, to the best knowledge of the authors, the phenomenon of competition has not been investigated. In this paper, we study the competition on crowdfunding markets through data analysis, and propose a probabilistic generative model, Dynamic Market Competition (DMC) model, to capture the competitiveness of projects in crowdfunding. Through an empirical evaluation using the pledging history of past crowdfunding projects, our approach has shown to capture the competitiveness of projects very well, and significantly outperforms several baseline approaches in predicting the daily collected funds of crowdfunding projects, reducing errors by 31.73% to 45.14%. In addition, our analyses on the correlations between project competitiveness, project design factors, and project success indicate that highly competitive projects, while being winners under various setting of project design factors, are particularly impressive with high pledging goals and high price rewards, comparing to medium and low competitive projects. Finally, the competitiveness of projects learned by DMC is shown to be very useful in applications of predicting final success and days taken to hit pledging goal, reaching 85% accuracy and error of less than 7 days, respectively, with limited information at early pledging stage.

Offering products in the forms of menu bundles is a common practice in marketing to attract customers and maximize revenues. In crowdfunding platforms such as Kickstarter, rewards also play an important part in influencing project success. Designing rewards consisting of the appropriate items is a challenging yet crucial task for the project creators. However, prior research has not considered the strategies project creators take to offer and bundle the rewards, making it hard to study the impact of reward designs on project success. In this paper, we raise a novel research question: understanding project creators' decisions of reward designs to level their chance to succeed. We approach this by modeling the design behavior of project creators, and identifying the behaviors that lead to project success. We propose a probabilistic generative model, Menu-Offering-Bundle (MOB) model, to capture the offering and bundling decisions of project creators based on collected data of 14K crowdfunding projects and their 149K reward bundles across a half-year period. Our proposed model is shown to capture the offering and bundling topics, outperform the baselines in predicting reward designs. We also find that the learned offering and bundling topics carry distinguishable meanings and provide insights of key factors on project success.

Emotion analysis of online customer service conservation is important for good user experience and customer satisfaction. However, conventional metrics do not fit this application scenario. In this work, by collecting and labeling online conversations of customer service on Twitter, we identify 8 new metrics, named as tones, to describe emotional information. To better interpret each tone, we extend the Latent Dirichlet Allocation (LDA) model to Tone LDA (T-LDA). In T-LDA, each latent topic is explicitly associated with one of three semantic categories, i.e., tone-related, domain-specific and auxiliary. By integrating tone label into learning, T-LDA can interfere the original unsupervised training process and thus is able to identify representative tone-related words. In evaluation, T-LDA shows better performance than baselines in predicting tone intensity. Also, a case study is conducted to analyze each tone via T-LDA output.

Efficiency of large-scale learning is a hot topic in both academic and industry. The stochastic gradient descent (SGD) algorithm, and its extension mini-batch SGD, allow the model to be updated without scanning the whole data set. However, the use of approximate gradient leads to the uncertainty issue, slowing down the decreasing of objective function. Furthermore, such uncertainty may result in a high frequency of meaningless update on the model, causing a communication issue in parallel learning environment. In this work, we develop a batch-adaptive stochastic gradient descent (BA-SGD) algorithm, which can dynamically choose a proper batch size as learning proceeds. Particularly on the basis of Taylor extension and central limit theorem, it models the decrease of objective value as a Gaussian random walk game with rebound. In this game, a heuristic strategy of determining batch size is adopted to maximize the utility of each incremental sampling. By evaluation on multiple real data sets, we demonstrate that by smartly choosing the batch size, the BA-SGD not only conserves the fast convergence of SGD algorithm but also avoids too frequent model updates.

Distribution channel is a system that partners move products from manufacturer to end users. To increase sales, it is quite common for manufacturers to adjust the product prices to partners according to the product volume per deal. However, the price adjustment is like a double-edged sword. It also spurs some partners to form a cheating alliance, where a cheating seller applies for a falsified big deal with a low price and then re-sells the products to the cheating buyers. Since these cheating behaviors are harmful to a healthy ecosystem of distribution channel, we need the automatic method to guide the tedious audit process.

Understanding social tie development among users is crucial for user engagement in social networking services. In this paper, we analyze the social interactions, both online and offline, of users and investigate the development of their social ties using data trail of “how social ties grow” left in mobile and social networking services. To the best of our knowledge, this is the first research attempt on studying social tie development by considering both online and offline interactions in a heterogeneous yet realistic relationship. In this study, we aim to answer three key questions: 1) is there a correlation between online and offline interactions? 2) how is the social tie developed via heterogeneous interaction channels? 3) would the development of social tie between two users be affected by their common friends? To achieve our goal, we develop a Social-aware Hidden Markov Model (SaHMM) that explicitly takes into account the factor of common friends in measure of the social tie development. Our experiments show that, comparing with results obtained using HMM and other heuristic methods, the social tie development captured by our SaHMM is significantly more consistent to lifetime profiles of users.

Physical properties are obviously essential to study a chaotic system that generates discrete-time signals, but recovering chaotic properties of a signal source from small data is a very troublesome work. Existing chaotic models are weak in dealing with such case in that most of them need big data to exploit those properties. In this paper, geometric theory is considered to solve this problem. We build a smooth trajectory from series to implicitly exhibit the chaotic properties with series-nonuniform rational B-spline (S-NURBS) modeling method, which is presented by our team to model slow-changing chaotic time series. As for the part of validation, we reveal how well our model recovers the properties from both the statistical and the chaotic aspects to confirm the effectiveness of the model. Finally a practical chaotic model is built up to recover the chaotic properties contained in the Musa standard dataset, which is used in analyzing software reliability, thereby further proves the high credibility of this model in practical time series. The effectiveness of the S-NURBS modeling leads us to believe that it is really a feasible and worthy research area to study chaotic systems from geometric perspective. For this reason, we reckon that we have opened up a new horizon for chaotic system research.

Social media is a platform for people to share and vote content. From the analysis of the social media data we found that users are quite inactive in rating/voting. For example, a user on average only votes 2 out of 100 accessed items. Traditional recommendation methods are mostly based on users' votes and thus can not cope with this situation. Based on the observation that the dwell time on an item may reflect the opinion of a user, we aim to enrich the user-vote matrix by converting the dwell time on items into users' ``pseudo votes'' and then help improve recommendation performance. However, it is challenging to correctly interpret the dwell time since many subjective human factors, e.g. user expectation, sensitivity to various item qualities, reading speed, are involved into the casual behavior of online reading. In psychology, it is assumed that people have choice threshold in decision making. The time spent on making decision reflects the decision maker's threshold. This idea inspires us to develop a View-Voting model, which can estimate how much the user likes the viewed item according to her dwell time, and thus make recommendations even if there is no voting data available. Finally, our experimental evaluation shows that the traditional rate-based recommendation's performance is greatly improved with the support of VV model.

Due to the huge and still rapidly growing number of mobile applications (apps), it becomes necessary to provide users an app recommendation service. Different from conventional item recommendation where the user interest is the primary factor, app recommendation also needs to consider factors that invoke a user to replace an old app (if she already has one) with a new app. In this work we propose an Actual- Tempting model that captures such factors in the decision process of mobile app adoption. The model assumes that each owned app has an actual satisfactory value and a new app under consideration has a tempting value. The former stands for the real satisfactory value the owned app brings to the user while the latter represents the estimated value the new app may seemingly have. We argue that the process of app adoption therefore is a contest between the owned apps' actual values and the candidate app's tempting value. Via the extensive experiments we show that the AT model performs significantly better than the conventional recommendation techniques such as collaborative filtering and content-based recommendation. Furthermore, the best recommendation performance is achieved when the AT model is combined with them.

Prediction of popular items in online content sharing systems has recently attracted a lot of attention due to the tremendous need of users and its commercial values. Different from previous works that make prediction by fitting a popularity growth model, we tackle this problem by exploiting the latent conforming and maverick personalities of those who vote to assess the quality of on-line items. We argue that the former personality prompts a user to cast her vote conforming to the majority of the service community while on the contrary the later personality makes her vote different from the community. We thus propose a Conformer-Maverick (CM) model to simulate the voting process and use it to rank top-k potentially popular items based on the early votes they received. Through an extensive experimental evaluation, we validate our ideas and find that our proposed CM model achieves better performance than baseline solutions, especially for smaller k.

In this paper, we develop a semantic annotation technique for location-based social networks to automatically annotate all places with category tags which are a crucial prerequisite for location search, recommendation services, or data cleaning. Our annotation algorithm learns a binary support vector machine (SVM) classifier for each tag in the tag space to support multi-label classification. Based on the check-in behavior of users, we extract features of places from i) explicit patterns (EP) of individual places and ii) implicit relatedness (IR) among similar places. The features extracted from EP are summarized from all check-ins at a specific place. The features from IR are derived by building a novel network of related places (NRP) where similar places are linked by virtual edges. Upon NRP, we determine the probability of a category tag for each place by exploring the relatedness of places. Finally, we conduct a comprehensive experimental study based on a real dataset collected from a location-based social network, Whrrl. The results demonstrate the suitability of our approach and show the strength of taking both EP and IR into account in feature extraction.

In this paper, we aim to provide a point-of-interests (POI) recommendation service for the rapid growing location-based social networks (LBSNs), e.g., Foursquare, Whrrl, etc. Our idea is to explore user preference, social influence and geographical influence for POI recommendations. In addition to deriving user preference based on user-based collaborative filtering and exploring social influence from friends, we put a special emphasis on geographical influence due to the spatial clustering phenomenon exhibited in user check-in activities of LBSNs. We argue that the geographical influence among POIs plays an important role in user check-in behaviors and model it by power law distribution. Accordingly, we develop a collaborative recommendation algorithm based on geographical influence based on naive Bayesian. Furthermore, we propose a unified POI recommendation framework, which fuses user preference to a POI with social influence and geographical influence. Finally, we conduct a comprehensive performance evaluation over two large-scale datasets collected from Foursquare and Whrrl. Experimental results with these real datasets show that the unified collaborative recommendation approach significantly outperforms a wide spectrum of alternative recommendation approaches.

To enhance the quality of document search, recent research studies have started to exploit the social networks of users by considering social influence (SI), measurement of the affinity between a query user and the publisher of a retrieved document, in addition to the commonly used textual relevance (TR). We refer to such document search that considers social networks as social web search. In this paper, we focus on efficient top-k social web search and propose two search strategies: (i) TR-based search and (ii) SI-based search that tailor document examination orders upon TR and SI, respectively. We evaluate the proposed strategies through experimentation.

In this paper, we study the research issues in realizing location recommendation services for large-scale location-based social networks, by exploiting the social and geographical characteristics of users and locations/places. Through our analysis on a dataset collected from Foursquare, a popular location-based social networking system, we observe that there exists strong social and geospatial ties among users and their favorite locations/places in the system. Accordingly, we develop a friend-based collaborative filtering (FCF) approach for location recommendation based on collaborative ratings of places made by social friends. Moreover, we propose a variant of FCF technique, namely Geo-Measured FCF (GM-FCF), based on heuristics derived from observed geospatial characteristics in the Foursquare dataset. Finally, the evaluation results show that the proposed family of FCF techniques holds comparable recommendation effectiveness against the state-of-the-art recommendation algorithms, while incurring significantly lower computational overhead. Meanwhile, the GM-FCF provides additional flexibility in tradeoff between recommendation effectiveness and computational overhead.