Twitter is one of today’s largest and most popular social networks. The users of the service generate huge amounts of data each day and rely heavily on the service helping them find interesting tweets in short time. The concept of hashtags aids in this practice but relies on the users choosing to include the correct and commonly used hashtags for the topic of their tweet. Hashtag recommendation has been a target of research before with varying results. This thesis proposes a method taking the location of the users into account when making recommen- dations. The method generated improved results over just using similar tweets as a basis for recommendation. Various factors like the handling of different variations of vocabulary in the tweets, how many tweets the suggestions can be picked from and how the combination of similarity and geographic ranking should function could affect the result. This leads to the conclusion that geographic data can be used to improve hashtag suggestions, but a different approach in handling similarity and alternative combinations of similarity and geographic ranking could cause another result. / Twitter är ett av nutidens största och populäraste sociala nätverk. Tjänstens användare producerar stora mängder data varje dag och förväntar sig att tjänsten ska kunna hjälpa dem att hitta intressanta tweets snabbt. Därmed finns konceptet med hashtags, men detta förutsätter att användare väljer att inkludera vanligt förekommande hashtags som på ett korrekt sätt avspeglar innehållet i tweeten. Automatisk rekommendation av hashtags har därmed varit ett populärt forskningsämne de senaste åren, med varierande resultat. Denna studie undersöker en rekommendationsmetod som väger in användarens geografiska position för att rekommendera så passande hashtags som möjligt. Resultaten visar att denna metod generellt rekommenderar mer passande hashtags än metoder som enbart rekommenderar hashtags genom att analysera likhet mellan tweets. Olika faktorer så som hanterandet av olika varianter av vokabulär, hur många tweets som metoden kan föreslå hashtags från samt hur kombinationen av rekommendation baserat på likhet och geografiskt position ska fungera, kan samtidigt påverka resultaten. Detta leder till slutsatsen att geografisk data kan användas för att förbättra hashtagrekommendation, men att ett annorlunda tillvägagångsätt i att hantera likhet och alternativa kombinationer av likhetsrangordning och geografisk rangordning kan leda till ett annorlunda resultat.

E-commerce has grown at a rapid pace over the last years. At the same time, the return rate of purchased products is high, causing unnecessary transportation of goods to the home of the customer and back. In the clothing industry, most of the returns are related to the size of theproduct. Therefore, a growing demand for digital tools that can assist the customer in finding the correct size before ordering the product, thus avoiding a potential size related return. This thesis applies machine learning to the problem to recommend the correct size of the product to the customer before the purchase is made. Especially, focusing on the stakeholder’s three dimensional size system for trousers by evaluating four different machine learning approaches. Results show increased accuracy compared to the benchmark, yet provide no clear indication ofa specific machine learning approach as favorable using the data sets provided by the stakeholder. Several shortcomings of the data sets with regard to increasing the accuracy are proposed and discussed as potential problems causing noise and confusion into the machine learning models.

The performance of classification algorithms is mainly governed by the hyperparameter configurations deployed. Traditional search-based algorithms tend to require extensive hyperparameter evaluations to select the desirable configurations during the process, and they are often very inefficient for implementations on large-scale tasks. In this dissertation, we resort to solving the problem of hyperparameter selection via meta-learning which provides a mechanism that automatically recommends the promising ones without any inefficient evaluations. In its approach, a meta-learner is constructed on the metadata extracted from historical classification problems which directly determines the success of recommendations. Designing fine meta-learners to recommend effective hyperparameter configurations efficiently is of practical importance. This dissertation divides into six chapters: the first chapter presents the research background and related work, the second to the fifth chapters detail our main work and contributions, and the sixth chapter concludes the dissertation and pictures our possible future work. In the second and third chapters, we propose two (kernel) multivariate sparse-group Lasso (SGLasso) approaches for automatic meta-feature selection. Previously, meta-features were usually picked by researchers manually based on their preferences and experience or by wrapper method, which is either less effective or time-consuming. SGLasso, as an embedded feature selection model, can select the most effective meta-features during the meta-learner training and thus guarantee the optimality of both meta-features and meta-learner which are essential for successful recommendations. In the fourth chapter, we formulate the problem of hyperparameter recommendation as a problem of low-rank tensor completion. The hyperparameter search space was often stretched to a one-dimensional vector, which removes the spatial structure of the search space and ignores the correlations that existed between the adjacent hyperparameters and these characteristics are crucial in meta-learning. Our contributions are to instantiate the search space of hyperparameters as a multi-dimensional tensor and develop a novel kernel tensor completion algorithm that is applied to estimate the performance of hyperparameter configurations. In the fifth chapter, we propose to learn the latent features of performance space via denoising autoencoders. Although the search space is usually high-dimensional, the performance of hyperparameter configurations is usually correlated to each other to a certain degree and its main structure lies in a much lower-dimensional manifold that describes the performance distribution of the search space. Denoising autoencoders are applied to extract the latent features on which two effective recommendation strategies are built. Extensive experiments are conducted to verify the effectiveness of our proposed approaches, and various empirical outcomes have shown that our approaches can recommend promising hyperparameters for real problems and significantly outperform the state-of-the-art meta-learning-based methods as well as search algorithms such as random search, Bayesian optimization, and Hyperband.

No description available.

Broadly, this study focuses on roles of sell-side analysts and examines the determinants and consequences of information discovery and stock timing roles by sell-side analysts. We also re-examine reiterations of prior recommendations by sell-side analysts. In Chapter 1, the contribution is to document that analysts add value by engaging in discovery of private information and this value addition is greater than that due to interpretation of public news or stock timing. The innovation in this Chapter is to read over 3,700 analyst reports from Investext and explicitly identify whether the report contains discovery, interpretation, and/or timing. Analysts discover new information by talking to management sources (personal meetings, investor meetings, and conference calls) or non-management sources (such as channel checks). We find that information discovery is prevalent in 17% of the reports. The cumulative abnormal return (CAR) for reports containing discovery are 6.3% for upgrades and -10.6% for downgrades. The CARs are higher for reports containing discovery relative to those containing interpretation or timing. We find that economic determinants predict whether a report will contain discovery. Discovery from management sources is more likely for reports in the pre-Reg FD period and for reports by optimistic analysts. Discovery from non-management sources is more likely for reports written by All-Star analysts, and for firms that have high information asymmetry and those that are followed by more analysts. In Chapter 2, the contribution is to introduce and document a third role that analysts play that is also valuable to investors, which we term "stock timing." Specifically, we define a timing report as one where the analyst revises his recommendation but does not revise the Price Target or any of the 23 fundamental drivers of stock price (such as EPS, FCF) tracked by I/B/E/S. Because the analyst maintains the same price target as in his prior report but still revises his recommendation, such timing calls are contrarian valuation calls. Analysts issue timing downgrades (upgrades) in response to price increases (declines) since the release of their prior report on the firm. 30% of all revisions are timing reports, indicating the importance of the timing role played by analysts. If analysts have timing ability, then markets should react to the release of the timing report and we should observe that economic determinants explain the cross-sectional variation in timing ability. We find the 3-day announcement return is over 2% in magnitude, 62% of the reports are winners (have announcement returns that have the correct sign), 10% of the reports are large enough to be considered influential, and 37% of the reports are persistent winners. These results suggest that analysts have timing ability. The ability to time is similar is magnitude to information interpretation but smaller compared to information discovery. We find considerable cross-sectional and time-series variation in timing ability. We find that the probability of issuing a timing report is positively related to the opportunities to time the stock provided by potential mispricing. Conditional on issuing a timing report, the probability of issuing a winner, an influential winner, or a persistent winner is positively related to analyst experience and negatively related to the costs associated with issuing a timing report. In Chapter 3, we document that recommendation reiterations are not homogeneous and there is a large subset of reiterations that are as much valued by investors as recommendation revisions. We combine Detail History file containing the measures tracked by I/B/E/S (Price Target, EPS, etc.) and Recommendation file to create the full time series of recommendations (initiations, reiterations, and revisions) made by each analyst for each firm for 14 years from 1999 to 2012. By adopting a modified version of "filling in the holes" method, we find that recommendation reiterations are prevalent, consisting of about 80% of recommendations for our 14-year sample period. Second, market response to recommendation reiterations increases monotonically from Reiteration: Strong Sell to Reiteration: Strong Buy. Third, reiterations coupled with contemporary changes in price targets and/or earning forecasts bring substantial absolute abnormal stock returns to investors. Lastly, when we replicate what Loh and Stulz (2011), we find that the number of reiterations which are influential is more than twice that of recommendation revisions that are influential. / Business Administration/Finance

Spectral reflectance-based, remote sensing technology has been used to adjust in-season nitrogen (N) fertilizer rates for wheat to account for spatial variability in grain yield potential at a sub-meter resolution. The objective of this study was to examine the relationships among spectral reflectance indices, corn tissue N content, chlorophyll measurements, plant size and spacing measurements, and grain yield to develop a similar strategy for variable-rate N management in corn. Irrigated and non-irrigated studies were conducted during the 2002 and 2003 growing seasons in eastern Virginia. Plots were treated with various rates of preplant, starter, and sidedress N fertilizer to establish a wide range of grain yield potential. Spectral measurements, tissue N, chlorophyll measurements, and plant physical measurements were collected at growth stages V6, V8, and V10. At maturity, grain yield was determined and correlated with in-season data and optimum N rate to calibrate in-season, variable-rate N fertilization strategies. Results from these studies indicate that spectral reflectance is well correlated with plant N uptake and chlorophyll meter readings and can also be correlated with final grain yield. These relationships may be used to develop a model to predict in-season, variable N application rates for corn production at a sub-meter resolution. / Master of Science

Online social networks are experiencing an explosive growth in recent years in both the number of users and the amount of information shared. The users join these social networks to connect with each other, share, find content and disseminate information by sending short text messages in near realtime. As a result of the growth of social networks, the users are often experiencing information overload since they interact with many other users and read ever increasing content volume. Thus, finding the "matching" users and content is one of the key challenges for social networks sites. Recommendation systems have been proposed to help users cope with information overload by predicting the items that a user may be interested in. The users' preferences are shaped by personal interests. At the same time, users are affected by their surroundings, as determined by their geographically located communities. Accordingly, our approach takes into account both personal interests and local communities. We first propose a new dynamic recommendation system model that provides better customized content to the user. That is, the model provides the user with the most important tweets according to his individual interests. We then analyze how changes in the surrounding environment can affect the user's experience. Specifically, we study how changes in the geographical community preferences can affect the individual user's interests. These community preferences are generally reflected in the localized trending topics. Consequently, we present TrendFusion, an innovative model that analyzes the trends propagation, predicts the localized diffusion of trends in social networks and recommends the most interesting trends to the user. Our performance evaluation demonstrate the effectiveness of the proposed recommendation system and shows that it improves the precision and recall of identifying important tweets by up to 36% and 80%, respectively. Results also show that TrendFusion accurately predicts places in which a trend will appear, with 98% recall and 80% precision. / Ph. D.

For maximal market penetration, popular mobile applications are typically supported on all major platforms, including Android and iOS. Despite the vast differences in the look-and-feel of major mobile platforms, applications running on these platforms in essence provide the same core functionality. As an application is maintained and evolved, programmers need to replicate the resulting changes on all the supported platforms, a tedious and error-prone programming process. Commercial automated source-to-source translation tools prove inadequate due to the structural and idiomatic differences in how functionalities are expressed across major platforms. In this thesis, we present a new approach---Native-2-Native---that automatically synthesizes code for a mobile application to make use of native resources on one platform, based on the equivalent program transformations performed on another platform. First, the programmer modifies a mobile application's Android version to make use of some native resource, with a plugin capturing code changes. Based on the changes, the system then parameterizes a web search query over popular programming resources (e.g., Google Code, StackOverflow, etc.), to discover equivalent iOS code blocks with the closest similarity to the programmer-written Android code. The discovered iOS code block is then presented to the programmer as an automatically synthesized Swift source file to further fine-tune and subsequently integrate in the mobile application's iOS version. Our evaluation, enhancing mobile applications to make use of common native resources, shows that the presented approach can correctly synthesize more than 86% of Swift code for the subject applications' iOS versions. / Master of Science

The popularity of AI audio applications is growing, it is used in chatbots, automated voice translation, virtual assistants, and text-to-speech translation. Audio classification is crucial in today’s world with a growing need to sort and classify millions of existing audio data with increasing amounts of new data uploaded over time. In the area of classification lies the difficult and lucrative problem of music recommendation. Research in music recommendation has trended over time towards collaborative-based approaches utilizing large amounts of user data. These approaches tend to deal with the cold-start problem of insufficient data and are costly to train. We look to recent advances in music generation to develop a content-based method utilizing a joint embedding space to link text with music audio. This approach has not been previously applied to music recommendation. In this thesis, we will examine the joint embedding methods used by recent AI music generation models and introduce a music recommendation system using joint embeddings. This music recommendation system can avoid cold-start, reduce training costs for music recommendation, and serve as the foundation for a cost-efficient content-based multimedia recommendation system. The current model trained on MusicCaps recommends the correct song per tag input within the top 50%-80% of all songs about 65%-70% of the time and we expect better results after further training.

This study aims to apply user reviews and numerical ratings toward items to create an aspect-aware high-order representation for a recommendation system. We propose a novel aspect-aware knowledge graph recommendation model (AKGR) with the deep learning method to predict users' ratings on non-interacted items, from which more personalized recommendations can be made. First, we create a sequence-to-sequence encoder and decoder model by exploiting contextual and syntactic information in user reviews to extract aspects critical to items. Then we utilize the principal component analysis (PCA) and the K-means clustering to analyze the extracted aspects for category classification. Based on the aspects, we construct a graph structure to connect users and items which share the same aspect-based opinions for mining user preferences and item attributes. Finally, we combine the user and item latent features from the reviews and the user-item rating matrix to complete the rating prediction task by applying the factorization machine model. We conducted experiments on three aspect extraction datasets and five rating prediction datasets. To verify the effectiveness of the proposed aspect extraction model and rating prediction model, comparison experiments were made with some state-of-the-art baseline models, such as double embeddings convolutional neural network (DE-CNN) and dual graph convolutional network (DualGCN). The experiment results revealed that our proposed aspect extraction model had the best performance for the three datasets with an F1 score of 82.41%, 88.57%, and 73.39%. In the experiments of rating prediction, the proposed AKGR model achieved the best MAE and MSE scores on the five datasets, and there was an average improvement of 4.48% against the best baseline.