Zur Hauptnavigation / To main navigation

Zur Sekundärnavigation / To secondary navigation

Zum Inhalt dieser Seite / To the content of this page

Sekundärnavigation / Secondary navigation

Inhaltsbereich / Content

Recommender Systems are understood as systems which propose personalized recommendations of not yet valued goods and which make an important contribution to reduce information overload.

The scientific roots of recommender systems are in the range of Information Retrieval, but actual application fields in the range of Internet Shops. Hybrid Systems combine herby the advantages of Content Based Filtering, which bases on similarities of the objects and their attributes, and on Collaborative Filtering, which generates new objects on the basis of valuation profile similarity of different users.

Selected Publications

Malinowski, J.; Keim, T.; Wendt, O.; Weitzel, T.: Matching People and Jobs: A Bilateral Recommendation Approach, in: 39th Hawaii International Conference on System Sciences 2006; Hawaii.

Malinowski, J.; Keim, T.; Weitzel, T.; Wendt, O.: Decision Support for Team Building: Incorporating Trust into a Recommender-Based Approach, in: The Ninth Pacific Asia Conference on Information Systems 2005; Bangkok, Thailand.

Färber, F.; Keim, T.; Wendt, O.; Weitzel, T.: A Model-based Approach to Recommending Partners, in: Proceedings of the 6. Internationale Tagung Wirtschaftsinformatik 2003 (WI 2003); Dresden.


Problem

Recommender Systems since yet were used nearly almost in the search for objects (i.e. CDS, books) but not in the search for subjects/persons. Application fields for such search systems range from e Recruiting to the search for business or leisure partners in internet.

In spite of the high scope and complexity of such a search the decision support in this field represents a few examined research area until yet.

Classical search methods (i.e. Bool Search by means of key words) require that searcher and wanted person use an identical vocabulary. This is unrealistic for describing persons.

Many providers in Internet image the individual only insufficient only from one partial perspective, f.i. in job agencies by focusing on curriculum vitae disregarding relational attributes.

Certainly, social Networking platforms model this very relations (mostly in private or business field), but allow no further typing of the relation or its attributes (i.e. duration or intensity) beyond the pure existence of the relation.

Thus, classical approaches fail, because

  • they collect the complexity of human and interpersonal attributes only insufficient,
  • the variety of attributes will not be not used for a decision support and
  • the bilateral component of people search will be often neglected.


Method of Resolution

A two-dimensional Problem

Partner matching can be understood as a two-dimensional allocation problem where

  •  persons must be allocated to targets and
  •  persons must be allocated to other persons.

Furthermore the search for subjects is in contrast to the search for objects:

  • a relational problem in which not only those attributes must be considered which directly can be allocated to one person, but relations between these persons can be included in the valuation.
  • a bilateral problem in which not only the preferences of the searcher must be considered but the preferences of the wanted person, too.

Approach

Using different Recommender systems and methods a decision support will be developed in three modules:

Module 1: Implementation of a hybrid Recommender system that recommends candidate profiles which should be considered for a vacant position to the recruiters.

Module 2: Implementation of a Recommender system that recommends personalized relevant vacancies to the job applicants.

Module 3: Conception of a system that – on the basis of existing relations between partners – recommends so far unknown people who the user should get to know.

The planned integration of the modules in a common approach should optimize the matching finally.


Results

Implementation

This model aims at an expansion of the so-called Probabilistic Latent Semantic Analysis (PLSA) for the relational and bilateral case. Without this expansion the valuations of single attributes of single candidates by single evaluators were so far be used to estimate the following stochastic model:

The results of this recommender system (module 1) are embodied in the following figure. As a result, the system recommends already comparitively good the persons suitable for the vacancies: Result

Contact person: Prof. Dr. Oliver Wendt, phone: +49 631 205 3771