Dienstleistungsinformatik

Permanent URI for this collection

http://hdl.handle.net/2003/27593

Browse

Now showing 1 - 8 of 8

Empirical analysis of session-based recommendation algorithms
(2020-10-20) Ludewig, Malte; Mauro, Noemi; Latifi, Sara; Jannach, Dietmar
Recommender systems are tools that support online users by pointing them to potential items of interest in situations of information overload. In recent years, the class of session-based recommendation algorithms received more attention in the research literature. These algorithms base their recommendations solely on the observed interactions with the user in an ongoing session and do not require the existence of long-term preference profiles. Most recently, a number of deep learning-based (“neural”) approaches to session-based recommendations have been proposed. However, previous research indicates that today’s complex neural recommendation methods are not always better than comparably simple algorithms in terms of prediction accuracy. With this work, our goal is to shed light on the state of the art in the area of session-based recommendation and on the progress that is made with neural approaches. For this purpose, we compare twelve algorithmic approaches, among them six recent neural methods, under identical conditions on various datasets. We find that the progress in terms of prediction accuracy that is achieved with neural methods is still limited. In most cases, our experiments show that simple heuristic methods based on nearest-neighbors schemes are preferable over conceptually and computationally more complex methods. Observations from a user study furthermore indicate that recommendations based on heuristic methods were also well accepted by the study participants. To support future progress and reproducibility in this area, we publicly share the session-rec evaluation framework that was used in our research.
Advances in session-based and session-aware recommendation
(2020) Ludewig, Malte; Jannach, Dietmar; Steffen, Bernhard
As of today, personalized item suggestions provided by an automated recommender system have become a crucial part of many online services, e.g., online shops or media streaming applications, and extensive evidence exists that such systems increase both the user experience as well as the revenue of the providers. In academia, the recommendation problem is often framed as finding suitable items that a user is not yet aware of based on his long-term preference profile. In the real world, however, this problem formulation has a number of problems. Long-term profiles, e.g., are not available for new or anonymous users and recommendations can then only be based on the few most recent interactions in an ongoing usage session. Various approaches to this highly relevant setting of session-based recommendation that recently emerged in the research community were proposed over the recent years. However, in terms of the evaluation procedure, no common standard has been established so far. In this thesis, the author, therefore, proposes a publicly available framework for reproducible research and, furthermore, fairly compares many approaches, of which some were proposed by himself. Extensive experiments and a user study surprisingly showed that comparably simple nearest-neighbor techniques usually outperform recent deep learning models across many domains, datasets, and metrics. Even if long-term preferences are available for the users, recent works indicated that it might still be beneficial to consider the ongoing session, e.g., because a user started the session with a specific intent in mind. The author of this thesis, thus, conducted a systematic statistical analysis to assess what helps recommendations in being effective in such a session-aware scenario. This analysis is based on log data from a fashion retailer and insights were, furthermore, operationalized into novel session-aware recommendation approaches. Matching items of the customer’s ongoing session, reminding him of previously inspected clothes, recommending discounted items, and considering recent trends in the community showed to be particularly effective strategies, not only for item-item recommendation but also in the related scenario of search personalization.
Designing and evaluating recommender systems with the user in the loop
(2019) Jugovac, Michael; Jannach, Dietmar; Teubner, Jens
On many of today's most popular Internet service platforms, users are confronted with a seemingly endless number of options to choose from, such as articles to purchase on online shopping sites, music to listen to on online streaming platforms, or posts to read on social media. As a solution to this choice overload problem, recommender systems have been integrated into more and more websites and applications to help users find items that they might like or that could be useful in their current choice situation. In recent decades, research on recommender systems has mostly been driven by offline performance comparisons, in which each new approach is compared to the state of the art in terms of its ability to retroactively predict user preferences in historical data sets. However, such a purely algorithmic research approach can only capture one of the many factors that contribute to a useful and engaging recommendation experience from a user perspective. In fact, a variety of aspects can influence how recommendations affect users' decision-making processes and how users perceive recommendations, including details regarding the recommender system's user interface or subconscious cognitive effects evoked by the recommendations. In this thesis by publication, selected works of the author are presented that investigate different aspects pertaining to the design and evaluation of recommender systems from a more user-focused perspective. The first part of the thesis outlines each of these publications and positions them within the research context. The presented works investigate (i) how recommender systems interact with their users, (ii) how recommender systems should be evaluated with the user in mind, (iii) possible biases in user studies, (iv) an algorithmic strategy to re-rank recommendation lists according to individual user tendencies, and (v) two phenomena based on which recommendations can subconsciously influence user decision-making processes. The second part of the thesis, the appendix, contains the aforementioned publications in full. The presented studies demonstrate that it is imperative to design and evaluate recommender systems with the user in mind, taking into account the intricacies of interaction details, recommendation list composition, user context, and decision-making processes.
Advances in next-track music recommendation
(2017) Kamehkhosh, Iman; Jannach, Dietmar; Rudolph, Günter
Technological advances in the music industry have dramatically changed how people access and listen to music. Today, online music stores and streaming services offer easy and immediate means to buy or listen to a huge number of songs. One traditional way to find interesting items in such cases when a vast amount of choices are available is to ask others for recommendations. Music providers utilize correspondingly music recommender systems as a software solution to the problem of music overload to provide a better user experience for their customers. At the same time, an enhanced user experience can lead to higher customer retention and higher business value for music providers. Different types of music recommendations can be found on today's music platforms, such as Spotify or Deezer. Providing a list of currently trending music, finding similar tracks to the user's favorite ones, helping users discover new artists, or recommending curated playlists for a certain mood (e.g., romantic) or activity (e.g., driving) are examples of common music recommendation scenarios. "Next-track music recommendation" is a specific form of music recommendation that relies mainly on the user's recently played tracks to create a list of tracks to be played next. Next-track music recommendations are used, for instance, to support users during playlist creation or to provide personalized radio stations. A particular challenge in this context is that the recommended tracks should not only match the general taste of the listener but should also match the characteristics of the most recently played tracks. This thesis by publication focuses on the next-track music recommendation problem and explores some challenges and questions that have not been addressed in previous research. In the first part of this thesis, various next-track music recommendation algorithms as well as approaches to evaluate them from the research literature are reviewed. The recommendation techniques are categorized into the four groups of content-based filtering, collaborative filtering, co-occurrence-based, and sequence-aware algorithms. Moreover, a number of challenges, such as personalizing next-track music recommendations and generating recommendations that are coherent with the user's listening history are discussed. Furthermore, some common approaches in the literature to determine relevant quality criteria for next-track music recommendations and to evaluate the quality of such recommendations are presented. The second part of the thesis contains a selection of the author's publications on next- track music recommendation as follows. 1. The results of comprehensive analyses of the musical characteristics of manually created playlists for music recommendation; 2. the results of a multi-dimensional comparison of different academic and commercial next-track recommending techniques; 3. the results of a multi-faceted comparison of different session-based recommenders, among others, for the next-track music recommendation problem with respect to their accuracy, popularity bias, catalog coverage as well as computational complexity; 4. a two-phase approach to recommend accurate next-track recommendations that also match the characteristics of the most recent listening history; 5. a personalization approach based on multi-dimensional user models that are extracted from the users' long-term preferences; 6. a user study with the aim of determining the quality perception of next-track music recommendations generated by different algorithms.
Session-based item recommendation in e-commerce: on short-term intents, reminders, trends and discounts
(2017-09-01) Jannach, Dietmar; Ludewig, Malte; Lerche, Lukas
In our article Session-based item recommendation in e-commerce: on short-term intents, reminders, trends and discounts accepted for publication in User Modeling and User-Adapted Interaction (UMUAI) we examined a classification-based approach to analyze what makes a recommendation successful. In the process we generated over 95 features for each single recommendation action in our data set provided by the online fashion retailer Zalando. Due to space issues we could only explain some of the most relevant features in the article itself. As an addition, the following table lists all investigated features in detail. Furthermore, in our article we reported the top ten feature weights regarding the label prediction calculated by the methods Gain ratio and Chi-squared to highlight the most important success signals. Here, we additionally reveal the weights for all features and also include the Information gain ratio and the Gini index.
Automated model-based spreadsheet debugging
(2017) Schmitz, Thomas; Jannach, Dietmar; Wotawa, Franz
Spreadsheets are interactive data organization and calculation programs that are developed in spreadsheet environments like Microsoft Excel or LibreOffice Calc. They are probably the most successful example of end-user developed software and are utilized in almost all branches and at all levels of companies. Although spreadsheets often support important decision making processes, they are, like all software, prone to error. In several cases, faults in spreadsheets have caused severe losses of money. Spreadsheet developers are usually not educated in the practices of software development. As they are thus not familiar with quality control methods like systematic testing or debugging, they have to be supported by the spreadsheet environment itself to search for faults in their calculations in order to ensure the correctness and a better overall quality of the developed spreadsheets. This thesis by publication introduces several approaches to locate faults in spreadsheets. The presented approaches are based on the principles of Model-Based Diagnosis (MBD), which is a technique to find the possible reasons why a system does not behave as expected. Several new algorithmic enhancements of the general MBD approach are combined in this thesis to allow spreadsheet users to debug their spreadsheets and to efficiently find the reason of the observed unexpected output values. In order to assure a seamless integration into the environment that is well-known to the spreadsheet developers, the presented approaches are implemented as an extension for Microsoft Excel. The first part of the thesis outlines the different algorithmic approaches that are introduced in this thesis and summarizes the improvements that were achieved over the general MBD approach. In the second part, the appendix, a selection of the author's publications are presented. These publications comprise (a) a survey of the research in the area of spreadsheet quality assurance, (b) a work describing how to adapt the general MBD approach to spreadsheets, (c) two new algorithmic improvements of the general technique to speed up the calculation of the possible reasons of an observed fault, (d) a new concept and algorithm to efficiently determine questions that a user can be asked during debugging in order to reduce the number of possible reasons for the observed unexpected output values, and (e) a new method to find faults in a set of spreadsheets and a new corpus of real-world spreadsheets containing faults that can be used to evaluate the proposed debugging approaches.
Using implicit feedback for recommender systems: characteristics, applications, and challenges
(2016) Lerche, Lukas; Jannach, Dietmar; Zanker, Markus
Recommender systems are software tools to tackle the problem of information overload by helping users to find items that are most relevant for them within an often unmanageable set of choices. To create these personalized recommendations for a user, the algorithmic task of a recommender system is usually to quantify the user's interest in each item by predicting a relevance score, e.g., from the user's current situation or personal preferences in the past. Nowadays, recommender systems are used in various domains to recommend items such as products on e-commerce sites, movies and music on media portals, or people in social networks. To assess the user's preferences, recommender systems proposed in past research often utilized explicit feedback, i.e., deliberately given ratings or like/dislike statements for items. In practice, however, in many of today's application domains of recommender systems this kind of information is not existent. Therefore, recommender systems have to rely on implicit feedback that is derived from the users' behavior and interactions with the system. This information can be extracted from navigation or transaction logs. Using implicit feedback leads to new challenges and open questions regarding, for example, the huge amount of signals to process, the ambiguity of the feedback, and the inevitable noise in the data. This thesis by publication explores some of these challenges and questions that have not been covered in previous research. The thesis is divided into two parts. In the first part, the thesis reviews existing works on implicit feedback and recommender systems that exploit these signals, especially in the Social Information Access domain, which utilizes the "community wisdom" of the social web for recommendations. Common application scenarios for implicit feedback are discussed and a categorization scheme that classifies different types of observable user behavior is established. In addition, state-of-the-art algorithmic approaches for implicit feedback are examined that, e.g., interpret implicit signals directly or convert them to explicit ratings to be able to use "classic" recommendation approaches that were designed for explicit feedback. The second part of the thesis comprises some of the author's publications that deal with selected challenges of implicit feedback based recommendations. These contain (i) a specialized learning-to-rank algorithm that can differentiate different levels of interest indicator strength in implicit signals, (ii) contextualized recommendation techniques for the e-commerce domain that adapt product suggestions to customers' current short-term goals as well as their long-term preferences, and (iii) intelligent reminding approaches that aim at the re-discovery of relevant items in a customer's browsing history. Furthermore, the last paper of the thesis provides an in-depth analysis of different biases of various recommendation algorithms. Especially the popularity bias, the tendency to recommend mostly popular items, can be problematic in practical settings and countermeasures to reduce this bias are proposed.
Leveraging tagging data for recommender systems
(2012-10-09) Gedikli, Fatih; Jannach, Dietmar; Zanker, Markus
The goal of recommender systems is to provide personalized recommendations of products or services to users facing the problem of information overload on the Web. They provide personalized recommendations that best suit a customer's taste, preferences, and individual needs. Especially on large-scale Web sites where millions of items such as books or movies are offered to the users, recommender system technologies play an increasingly important role. One of their main advantages is that they reduce a user's decision-making effort. However, recommender systems are also of high importance from the service provider or system perspective. For instance, they can convince a customer to buy something or develop trust in the system as a whole which ensures customer loyalty and repeat sales gains. With the advent of the Social Web, user generated content has enriched the social dimension of the Web. New types of Web applications have emerged which emphasize content sharing and collaboration. These so-called Social Web platforms turned users from passive recipients of information into active and engaged contributors. As a result, the amount of user contributed information provided by the Social Web poses both new possibilities and challenges for recommender system research. This work deals with the question of how user-provided tagging data can be used to improve the quality of recommender systems. Tag-based recommendations and explanations are the two main areas of contribution in this work. The area of tag-based recommendations deals mainly with the topic of recommending items by exploiting tagging data. A tag recommender algorithm is proposed which can generate highly-accurate tag recommendations in real-time. Furthermore, the concept of user- and item-specific tag preferences is introduced in this work. By attaching feelings to tags users are provided a powerful means to express in detail which features of an item they particularly like or dislike. Additionally, new recommendation schemes are presented that can exploit tag preference data to improve recommendation accuracy. The area of tag-based explanations, on the other hand, deals with questions of how explanations for recommendations should be communicated to the user in the best possible way. New explanation methods based on personalized and non-personalized tag clouds are introduced. The personalized tag cloud interface makes use of the idea of user- and item-specific tag preferences. Furthermore, a first set of possible guidelines for designing or choosing an explanation interface for a recommender system is provided.

Browse

Recent Submissions