Machine Learning for Recommender Systems

Our work in the field of recommendation is focused on researching and developing recommender systems for a diverse range of application fields in e-commerce, chat interactions, clinical diagnostics and publication data. We are interested in sequential recommendation, but have also worked on content-based tag recommendation in the past.

We have cooperated with industrial partners such as adidas to develop personalized product recommendations based on sequences of user interactions. Here, we focus on using additional content to improve recommendations, for example we show how recent transformer models can profit from categorical item information in KeBERT4Rec and propose personalized recommendations based on user attributes in our paper "Personalization through User Attributes for Transformer-based Sequential Recommendation ( Link).

We have also used sequential transformer models for item recommendation in the MOBA game Dota ( Link ) and to predict scientific collaborations in our paper “CoBERT: Scientific Collaboration Prediction via Sequential Recommendation” (Link).

Together with the University Hospital Würzburg we aim to predict patients' next diagnostics requests in the DZ.PTM project. In our recent position paper “Towards Responsible Medical Diagnostics Recommendation Systems ( Link )'' we outline the criteria for working on recommender systems in the medical context while securing fairness, accountability, compliance and safety aspects.

On a more general note, we’ve published “A Case Study on Sampling Strategies for Evaluating Neural Sequential Item Recommendation Models ( Link )” where we examine different sampling strategies for the calculation of metrics and show the influence on the ranking and evaluation of recommendation models.

In KiLiMod, we aim to provide personalized, relevant content recommendations for participants of live-stream chats as available on Twitch for e-sport events.

We have also researched tag recommendation for folksonomies with the help of our social bookmark system BibSonomy and published a number of works. For example, in “Comparison of Content-Based Tag Recommendations in Folksonomy Systems" ( Link) we evaluate several recommendation algorithms in a cold-start scenario on a large real-world dataset. We also have tested and evaluated tag recommendations live in BibSonomy and presented the results in “Testing and Evaluating Tag Recommenders in a Live System" ( Link).

You can find a list of selected publications below.

Projects

Concluded Projects

adidas - Modeling and Personalization for Customer Engagement

Selected Publications

For a complete list of our papers to this topic, see here.

Personalization through User Attributes for Transformer-based Sequential Recommendation. Fischer, Elisabeth; Dallmann, Alexander; Hotho, Andreas. 2022.
- [ BibTeX ]
- [ Download ]
@article{fischer2022personalization, author = {Fischer, Elisabeth and Dallmann, Alexander and Hotho, Andreas}, keywords = {selected}, title = {Personalization through User Attributes for Transformer-based Sequential Recommendation}, year = 2022 }
Towards Responsible Medical Diagnostics Recommendation Systems. Schlör, Daniel; Hotho, Andreas. 2022.
- [ Abstract ]
- [ BibTeX ]
- [ URL ]
The early development and deployment of hospital and healthcare information systems have encouraged the ongoing digitization of processes in hospitals. Many of these processes, which previously required paperwork and telephone arrangements, are now integrated into IT solutions and require physicians and medical staff to interact with appropriate interfaces and tools. Although this shift to digital data management and process support has benefited patient care in many ways, it requires physicians to accurately capture all relevant information digitally for billing and documentation purposes, which takes a lot of time away from actual patient care work. However, systematic collection of healthcare data over a long period of time offers opportunities to improve this process and support medical staff by introducing recommender systems. Based on a practical working example, in this position paper, we will outline the design of a responsible recommender system in the medical context from a technical, application driven perspective and discuss potential design choices and criteria with a specific focus on accountability, safety, and fairness.

@misc{schlor2022towards, abstract = {The early development and deployment of hospital and healthcare information systems have encouraged the ongoing digitization of processes in hospitals. Many of these processes, which previously required paperwork and telephone arrangements, are now integrated into IT solutions and require physicians and medical staff to interact with appropriate interfaces and tools. Although this shift to digital data management and process support has benefited patient care in many ways, it requires physicians to accurately capture all relevant information digitally for billing and documentation purposes, which takes a lot of time away from actual patient care work. However, systematic collection of healthcare data over a long period of time offers opportunities to improve this process and support medical staff by introducing recommender systems. Based on a practical working example, in this position paper, we will outline the design of a responsible recommender system in the medical context from a technical, application driven perspective and discuss potential design choices and criteria with a specific focus on accountability, safety, and fairness.}, author = {Schlör, Daniel and Hotho, Andreas}, keywords = {selected}, note = {cite arxiv:2209.03760Comment: 4th FAccTRec Workshop on Responsible Recommendation}, title = {Towards Responsible Medical Diagnostics Recommendation Systems}, year = 2022 }
Sequential Item Recommendation in the MOBA Game Dota 2. Dallmann, Alexander; Kohlmann, Johannes; Zoller, Daniel; Hotho, Andreas. In CoRR, abs/2201.08724. 2022.
- [ BibTeX ]
- [ URL ]
- [ DOI ]
@article{DBLP:journals/corr/abs-2201-08724, author = {Dallmann, Alexander and Kohlmann, Johannes and Zoller, Daniel and Hotho, Andreas}, journal = {CoRR}, keywords = {selected}, title = {Sequential Item Recommendation in the {MOBA} Game Dota 2}, volume = {abs/2201.08724}, year = 2022 }
CoBERT: Scientific Collaboration Prediction via Sequential Recommendation. Koopmann, Tobias; Kobs, Konstantin; Herud, Konstantin; Hotho, Andreas. In 2021 International Conference on Data Mining Workshops (ICDMW), pp. 45–54. 2021.
- [ Abstract ]
- [ BibTeX ]
- [ DOI ]
- [ Download ]
Collaborations are an Important factor for scientific success, as the joint work leads to results individual scientists cannot easily reach. Recommending collaborations automatically can alleviate the time consuming and tedious search for potential collaborators. Usually, such recommendation systems rely on graph structures modeling co-authorship of papers and content-based relations such as similar paper keywords. Models are then trained to estimate the probability of links between certain authors in these graphs.In this paper, we argue that the order of papers is crucial for reliably predicting future collaborations, which is not considered by graph-based recommendation systems. We thus propose to reformulate the task of collaboration recommendation as a sequential recommendation task. Here, we aim to predict the next co-author in a chronologically sorted sequence of an author’s collaborators. We introduce CoBERT, a BERT4Rec inspired model, that predicts the sequence’s next co-author and thus a potential collaborator. Since the order of co-authors of a single paper is not that important compared to the overall paper order, we leverage positional embeddings encoding paper positions instead of co-author positions in the sequence. Additionally, we inject content features about every paper and their co-authors. We evaluate CoBERT on two datasets consisting of papers from the field of Artificial Intelligence and the journal PlosOne. We show that CoBERT can outperform graph-based methods and BERT4Rec when predicting the co-authors of the next paper. We make our code and data available.

@inproceedings{koopmann2021cobert, abstract = {Collaborations are an Important factor for scientific success, as the joint work leads to results individual scientists cannot easily reach. Recommending collaborations automatically can alleviate the time consuming and tedious search for potential collaborators. Usually, such recommendation systems rely on graph structures modeling co-authorship of papers and content-based relations such as similar paper keywords. Models are then trained to estimate the probability of links between certain authors in these graphs.In this paper, we argue that the order of papers is crucial for reliably predicting future collaborations, which is not considered by graph-based recommendation systems. We thus propose to reformulate the task of collaboration recommendation as a sequential recommendation task. Here, we aim to predict the next co-author in a chronologically sorted sequence of an author’s collaborators. We introduce CoBERT, a BERT4Rec inspired model, that predicts the sequence’s next co-author and thus a potential collaborator. Since the order of co-authors of a single paper is not that important compared to the overall paper order, we leverage positional embeddings encoding paper positions instead of co-author positions in the sequence. Additionally, we inject content features about every paper and their co-authors. We evaluate CoBERT on two datasets consisting of papers from the field of Artificial Intelligence and the journal PlosOne. We show that CoBERT can outperform graph-based methods and BERT4Rec when predicting the co-authors of the next paper. We make our code and data available.}, author = {Koopmann, Tobias and Kobs, Konstantin and Herud, Konstantin and Hotho, Andreas}, booktitle = {2021 International Conference on Data Mining Workshops (ICDMW)}, keywords = {selected}, month = 12, pages = {45-54}, title = {CoBERT: Scientific Collaboration Prediction via Sequential Recommendation}, year = 2021 }
A Case Study on Sampling Strategies for Evaluating Neural Sequential Item Recommendation Models. Dallmann, Alexander; Zoller, Daniel; Hotho, Andreas. In Fifteenth ACM Conference on Recommender Systems. ACM, 2021.
- [ BibTeX ]
- [ DOI ]
- [ Download ]
@inproceedings{dallmann2021study, author = {Dallmann, Alexander and Zoller, Daniel and Hotho, Andreas}, booktitle = {Fifteenth {ACM} Conference on Recommender Systems}, keywords = {selected}, month = {09}, publisher = {{ACM}}, title = {A Case Study on Sampling Strategies for Evaluating Neural Sequential Item Recommendation Models}, year = 2021 }
Leveraging User-Interactions for Time-Aware Tag Recommendations. Zoller, Daniel; Doerfel, Stephan; Pölitz, Christian; Hotho, Andreas. In Proceedings of the Workshop on Temporal Reasoning in Recommender Systems, of CEUR Workshop Proceedings. 2017.
- [ Abstract ]
- [ BibTeX ]
- [ Download ]
For the popular task of tag recommendation, various (complex) approaches have been proposed. Recently however, research has focused on heuristics with low computational effort and particularly, a time-aware heuristic, called BLL, has been shown to compare well to various state-of-the-art methods. Here, we follow up on these results by presenting another time-aware approach leveraging user interaction data in an easily interpretable, on-the-fly computable approach that can successfully be combined with BLL. We investigate the influence of time as a parameter in that approach, and we demonstrate the effectiveness of the proposed method using two datasets from the popular public social tagging system BibSonomy.

@inproceedings{zoller2017leveraging, abstract = {For the popular task of tag recommendation, various (complex) approaches have been proposed. Recently however, research has focused on heuristics with low computational effort and particularly, a time-aware heuristic, called BLL, has been shown to compare well to various state-of-the-art methods. Here, we follow up on these results by presenting another time-aware approach leveraging user interaction data in an easily interpretable, on-the-fly computable approach that can successfully be combined with BLL. We investigate the influence of time as a parameter in that approach, and we demonstrate the effectiveness of the proposed method using two datasets from the popular public social tagging system BibSonomy.}, author = {Zoller, Daniel and Doerfel, Stephan and Pölitz, Christian and Hotho, Andreas}, booktitle = {Proceedings of the Workshop on Temporal Reasoning in Recommender Systems}, keywords = {selected}, series = {{CEUR} Workshop Proceedings}, title = {Leveraging User-Interactions for Time-Aware Tag Recommendations}, year = 2017 }
Improving Session Recommendation with Recurrent Neural Networks by Exploiting Dwell Time. Dallmann, Alexander; Grimm, Alexander; Pölitz, Christian; Zoller, Daniel; Hotho, Andreas. In CoRR, abs/1706.10231. 2017.
- [ BibTeX ]
- [ URL ]
- [ Download ]
@article{journals/corr/DallmannGPZH17, author = {Dallmann, Alexander and Grimm, Alexander and Pölitz, Christian and Zoller, Daniel and Hotho, Andreas}, journal = {CoRR}, keywords = {selected}, title = {Improving Session Recommendation with Recurrent Neural Networks by Exploiting Dwell Time.}, volume = {abs/1706.10231}, year = 2017 }

Machine Learning for Recommender Systems

Projects

KILiMod

HydrAS

BibSonomy

Concluded Projects

Selected Publications

Hinweis zum Datenschutz

Hinweis zum Datenschutz

Bildnachweise