Recommender Systems have greatly affected how we consume services, products, and content in recent years. They have various applications in everyday life, such as clothing, restaurants, or song recommendations, rating predictions and businesses to help them understand user choices well. While these systems enhance the user experience, there are concerns about data privacy. Such systems collect private information about users based on their online behavior, cookies, and social interactions, such as user clicks, time, and other data points to improve recommendations. This centralized approach to collecting and storing information is prone to privacy risks and data breaches. The primary goal of this work is to explore the potential of federated learning in addressing these privacy and security issues in recommender systems. We evaluate four algorithms in a federated and non-federated setting across seven diverse datasets to benchmark the performance of federated learning and provide insights into the efficacy of the approach in preserving privacy while maintaining recommender system performance. In this approach, the models are trained on edge devices using the data on user machines. This technique shares each user’s updated parameters using optimal aggregation functions instead of actual data to a shared server. This decentralized way ensures that the data remains local and protects data privacy. We share our code and framework details to enable replication and further expansion of this benchmark work to more datasets and algorithms from the scientific community.

Through lyrics, pitch, and rhythm, music is a natural way of expressing one’s thoughts. As one of the essential music composition elements, lyrics’ composition is complicated as it requires creativity and follows a particular rhythm pattern. In this work we design and train two neural network models for composing lyrics in three genres and propose scoring functions to select and evaluate the generated songs. We treat this problem as a text generation task, and optimize for features particular to song lyrics, including lyrics’ quality, rhyme density, and sentiment ratio, for lyrics in different genres. The neural networks we have experimented with are generative adversarial network (GAN)-based transfer learning for a deep learning model and long short-term memory (LSTM)-based deep learning model. In addition to quantitative evaluation, we also conducted user studies, inviting 25 people to rate the generative songs selected by the scoring functions. Our findings show that the GAN-based models perform better than LSTM-based models and the scoring functions are useful in selecting good songs.

For a long time researchers have focused on the binary classification of domain names sent to DNS servers for resolutions to IP addresses. The objective is to identify malicious domains versus legitimate ones to protect networks from attacks. For legitimate domains, an emerging interest is to classify them into content categories to enable DNS servers deployed in an organization to monitor and potentially block the resolution of irrelevant domains. For example, a financial organization wants to flag gaming related domains, and an elementary school wants to block suspicious adult domains. Classifying a domain by just the domain name is a challenging task. Currently, there exist no publicly available datasets that include an extensive mapping of domains to content categories, since this is usually proprietary information. Our focus is three-fold in this work: a) to develop a data collection methodology and create rich labelled datasets that are appropriate for training such predictive models, b) to share the datasets with the research community by making them publicly available, and c) to evaluate and identify appropriate machine learning and deep learning algorithms for this problem domain. We consider two different datasets. The first is created following a SERP (Search Engine Response Page)-mining approach, having a set of content categories as input. The second is an enhancement of the DMOZ dataset that is publicly available, including both domains and category names as input. In addition to the dataset creation input and methodology, these two datasets differ in the content category number and distribution, yielding different results in our analysis. Overall, we observe that the deep learning-based approach carefully considers the key features of the input data and hence outperforms existing traditional machine learning pipelines, achieving 98.37% and 79.29% accuracy on the respective datasets.

The energy consumption of households has steadily increased over the last couple of decades. Research suggests that user behavior is the most influential factor in the energy waste of a household. Thus, there’s a need for helping consumers change their behavior to make it more energy efficient and environment friendly. In this work we propose a real-time recommender system that assists consumers in improving their household’s energy usage. By monitoring the power demand of each appliance in the household, the system detects the device status (on/off) at any moment, and using pattern mining creates a household profile comprising energy consumption patterns for different periods of the day. An intuitive UI allows users to set energy consumption goals and preferences on the appliances they’d like to save energy from. Based on the household profile, the user’s preferences and the actual power demand the system generates personalized real-time recommendations on which appliances should be turned off at a moment. We employ the UK-DALE (UK Domestic Appliance-Level Electricity) dataset to model and evaluate the entire process, from data preprocessing and transformation of the appliance power demand input to various pattern mining algorithms used to generate appliance usage profiles and recommendations, showing that even small changes in appliance usage behavior can lead to energy savings between 2-17%.

DNS request classification is an area that has received a lot of attention, mostly as part of network security process, in order to classify requests into malicious and non-malicious. However, there exist several categories of web pages that even though not malicious, they belong to “borderline” categories and need to be monitored. For instance, websites selling illegal substances or weapons might be of interest for any public or private organization to monitor as outgoing traffic. In this work, we treat this as a topic classification problem. We present and evaluate a machine learning framework that takes as input a domain name (based on the respective DNS request) and outputs the content category it belongs to. We evaluate several options for feature engineering and classification to find the most appropriate setup for the specific problem domain. We also address the problem of data collection and preprocessing. While there exist several labelled datasets with malicious/non-malicious requests, a similar labelled dataset does not exist for general web content categories. We therefore propose a SERP (Search Engine Response Pages)-mining approach to collect and label an appropriate dataset. Our experimental evaluation uncovers several interesting insights and forms the basis for further work into this interesting domain.

Crime has been prevalent in our society for a very long time and it continues to be so even today. Currently, many cities have released crime-related data as part of an open data initiative. Using this as input, we can apply analytics to be able to predict and hopefully prevent crime in the future. In this work, we applied big data analytics to the San Francisco crime dataset, as collected by the San Francisco Police Department and available through the Open Data initiative. The main focus is to perform an in-depth analysis of the major types of crimes that occurred in the city, observe the trend over the years, and determine how various attributes contribute to specific crimes. Furthermore, we leverage the results of the exploratory data analysis to inform the data preprocessing process, prior to training various machine learning models for crime type prediction. More specifically, the model predicts the type of crime that will occur in each district of the city. We observe that the provided dataset is highly imbalanced, thus metrics used in previous research focus mainly on the majority class, disregarding the performance of the classifiers in minority classes, and propose a methodology to improve this issue. The proposed model finds applications in resource allocation of law enforcement in a Smart City.

Social recommendations have been a very intriguing domain for researchers in the past decade. The main premise is that the social network of a user can be leveraged to enhance the rating-based recommendation process. This has been achieved in various ways, and under different assumptions about the network characteristics, structure, and availability of other information (such as trust, content, etc.) In this work, we create neighborhoods of influence leveraging only the social graph structure. These are in turn introduced in the recommendation process both as a pre-processing step and as a social regularization factor of the matrix factorization algorithm. Our experimental evaluation using real-life datasets demonstrates the effectiveness of the proposed technique.

Social networking is an inevitable behavior of humans living in a society. In recent years, and with the rise of online social networks, personalized recommendations that leverage the social aspect have become a very intriguing domain for researchers. In this work, we explore how influence propagation and the decay in the cascading effect of influence from influential users can be leveraged to generate social graph-based recommendations. Understanding how influence propagates within a social network is itself a challenging problem. Few researchers have considered influence propagation and even fewer have considered decay in the cascading effect of influence in a social network. In this work we model the decay in influence propagation in directed graphs, utilizing the structural properties of the social graph to measure the propagated influence beyond one-hop. We then employ this influence propagation model to form social recommendations, and present our experimental results using real-life datasets.

The advancements in wearable technology, where embedded accelerometers, gyroscopes and other sensors enable the users to actively monitor their activity have made it easier for individuals to pursue a healthy lifestyle. However, most of the existing applications expect continuous commitment from the end users, who need to proactively interact with the application in order to connect with friends and attain their goals. These applications fail to engage and motivate users who have busy schedules, or are not as committed and self-motivated. In this work, we present PRO-Fit, a personalized fitness assistant application that employs machine learning and recommendation algorithms in order to smartly track and identify user's activity, synchronizes with the user's calendar, recommends personalized workout sessions based on the user's preferences, fitness goals, and availability. Moreover, PRO-Fit integrates with the user's social network and recommends “fitness buddies” with similar preferences and availability.

Online advertisements are a major source of profit and customer attraction for web-based businesses. In a successful advertisement campaign, both users and businesses can benefit, as users are expected to respond positively to special offers and recommendations of their liking and businesses are able to reach the most promising potential customers. The extraction of user preferences from content provided in social media and especially in review sites can be a valuable tool both for users and businesses.

Interactive database exploration is a key task in information mining. Relational databases have been long used as a critical infrastructure component to access and analyze large volumes of data in a variety of applications, including ad-hoc analytics over big data, large-scale data warehouses that support business-intelligence tools, and services for scientific-data exploration. To aid the users of such databases, we developed the QueRIE system for personalized query recommendations. Similarly to traditional recommender systems, QueRIE continuously monitors the user's querying behavior and finds matching patterns in the system's query log, identifying "similar" users. Subsequently, these users and their queries are being used to recommend queries that the current user may find useful. We have previously shown that when employing different neighborhood-based collaborative filtering techniques, there exists a trade-off between computational efficiency and accuracy. In this paper we extend our previous work on the QueRIE framework, to address scalability, the most desirable characteristic of applications that rely on the mining of big data. Latent factor collaborative filtering models have been shown to address the scalability problem in traditional rating-based recommender systems, without much compromise to the recommender system's accuracy. In this work, we explore the use of latent factor models when, instead of ratings, the input consists of database-query log data. We show through experimentation that, as in the case of rating-based recommender systems, such techniques offer both scalability and prediction accuracy in the database query recommendations domain, outperforming the neighborhood-based approaches.

As the amount of recorded digital information increases, there is a growing need for flexible recommender systems which can incorporate richly structured data sources to improve recommendations. In this paper, we show how a recently introduced statistical relational learning framework can be used to develop a generic and extensible hybrid recommender system. Our hybrid approach, HyPER (HYbrid Probabilistic Extensible Recommender), incorporates and reasons over a wide range of information sources. Such sources include multiple user-user and item-item similarity measures, content, and social information. HyPER automatically learns to balance these different information signals when making predictions. We build our system using a powerful and intuitive probabilistic programming language called probabilistic soft logic, which enables efficient and accurate prediction by formulating our custom recommender systems with a scalable class of graphical models known as hinge-loss Markov random fields. We experimentally evaluate our approach on two popular recommendation datasets, showing that HyPER can effectively combine multiple information types for improved performance, and can significantly outperform existing state-of-the-art approaches.

The success of a product/service in e-commerce largely depends on the user reviews. A product/service that has a higher average review or rating usually gets picked against a similar product/service with less favorable reviews. Reviews usually have an overall rating, but most of the times there are sub-texts in the review body that describe certain features/aspects of the product. This demonstration presents a system that extracts aspect-specific ratings from reviews and also recommends reviews to users based on their and other users' rating patterns.

Interactive database exploration is a key task in information mining. However, users who lack SQL expertise or familiarity with the database schema face great difficulties in performing this task. To aid these users, we developed the QueRIE system for personalized query recommendations. QueRIE continuously monitors the user's querying behavior and finds matching patterns in the system's query log, in an attempt to identify previous users with similar information needs. Subsequently, QueRIE uses these “similar” users and their queries to recommend queries that the current user may find interesting. In this work we describe an instantiation of the QueRIE framework, where the active user's session is represented by a set of query fragments. The recorded fragments are used to identify similar query fragments in the previously recorded sessions, which are in turn assembled in potentially interesting queries for the active user. We show through experimentation that the proposed method generates meaningful recommendations on real-life traces from the SkyServer database and propose a scalable design that enables the incremental update of similarities, making real-time computations on large amounts of data feasible. Finally, we compare this fragment-based instantiation with our previously proposed tuple-based instantiation discussing the advantages and disadvantages of each approach.

Social network analysis has recently gained a lot of interest because of the advent and the increasing popularity of social media, such as blogs, social networks, micro logging, or customer review sites. Such media often serve as platforms for information dissemination and product placement or promotion. In this environment, influence and trust are becoming essential qualities among user interactions. In this work, we perform an extensive study of various metrics related to the aforementioned elements, and their effect in the process of information propagation in the virtual world. In order to better understand the properties of links and the dynamics of social networks, we distinguish between permanent and transient links and in the latter case, we consider the link freshness. Moreover, we distinguish between local and global influence and compare suggestions provided by locally or globally trusted users.

This demonstration presents QueRIE, a recommender system that supports interactive database exploration. This system aims at assisting non-expert users of scientific databases by tracking their querying behavior and generating personalized query recommendations. The system is supported by two recommendation engines and the underlying recommendation algorithms. The first identifies potentially "interesting" parts of the database related to the corresponding data analysis task by locating those database parts that were accessed by similar users in the past. The second identifies structurally similar queries to the ones posted by the current user. Both approaches result in a recommendation set of SQL queries that is provided to the user to modify, or directly post to the database. The demonstrated system will enable users to query and get real-time recommendations from the SkyServer database, using user traces collected from the SkyServer query log.

Recommendation algorithms aim at proposing "next" pages to a user based on her current visit and the past users' navigational patterns. In the vast majority of related algorithms, only the usage data are used to produce recommendations, whereas the structural properties of the Web graph are ignored. We claim that taking also into account the Web structure and using link analysis algorithms ameliorates the quality of recommendations. In this paper we present UPR, a novel personalization algorithm which combines usage data and link analysis techniques for ranking and recommending Web pages to the end user. Using the Web site's structure and its usage data we produce personalized navigational graph synopsis (prNG) to be used for applying UPR and produce personalized recommendations. Experimental results show that the accuracy of the recommendations is superior to pure usage-based approaches.

Web personalization is the process of customizing a Web site to the needs of each specific user or set of users, taking advantage of the knowledge acquired through the analysis of the user's navigational behavior. Integrating usage data with content, structure or user profile data enhances the results of the personalization process. In this paper, we present SEWeP, a system that makes use of both the usage logs and the semantics of a Web site's content in order to personalize it. Web content is semantically annotated using a conceptual hierarchy (taxonomy). We introduce C-logs, an extended form of Web usage logs that encapsulates knowledge derived from the link semantics. C-logs are used as input to the Web usage mining process, resulting in a broader yet semantically focused set of recommendations.