LaboUr – Machine Learning for User Modeling - CiteSeerX

13 downloads 0 Views 24KB Size Report
In the following, contributions of machine learning to user modeling will be sketched ... rule construction (like C4.5 [12]), concept formation [18], or inductive logic ...
LaboUr – Machine Learning for User Modeling Wolfgang Pohl GMD – German National Research Center for Information Technology, HCI Research Department, D-53754 St. Augustin, Germany; [email protected] Traditional user modeling systems are often limited, as far as processing of observations about user behavior and handling of user model dynamics are concerned. In this paper, the LaboUr architecture for user modeling systems is discussed. It realizes user modeling as open learning process, thus overcoming the mentioned limitations. 1. Limitations in User Modeling Systems User modeling [15,9] is concerned with acquisition and representation of assumptions about users of technical systems, and with the exploitation of these assumptions for system individualization. In many user modeling systems, the following shortcomings can be observed:

 Assumptions about mental attitudes like knowledge or goals are modeled, while behaviororiented assumptions, e.g. about interaction preferences or behavior patterns, are missing.  Assumptions are acquired with specialized heuristics, which draw conclusions from isolated observations without regarding interaction context.  User behavior, preferences, and mental attitudes are subject to change, which is often not treated adequately.  User models are constructed and exploited mostly within the limits of one application. However, it can be beneficial to share information about users among several applications. These shortcomings can be overcome, if user modeling is realized as open learning process. Such a process is based on machine learning techniques and is implemented as an applicationindependent system. In the following, contributions of machine learning to user modeling will be sketched, and it will be explained how user modeling can become an open learning process. Based on this notion, we developed the LaboUr (Learning about the User) architecture for user modeling systems, which will briefly be discussed. 2. Machine Learning for User Modeling 2.1. Construction of Usage Models User models traditionally contain assumptions about user knowledge, goals, preferences and other mental attitudes [15,9], which are acquired from observed user behavior. Recently, in

several systems user behavior has not only been observed but also modeled as a direct foundation for system individualization. The terms “usage modeling” and “usage profiles” have been coined for systems like Flexcel [6] and Basar [13], which record user actions to obtain information about user behavior. More sophisticated techniques are employed by “interface agents” and “personal assistants” [8,10]. Such a system “becomes gradually more effective as it learns the user’s interests, habits and preferences” [8]. Currently, interface agents mainly learn correlations between situations the user may encounter and the corresponding actions she performs. These data are used to predict user behavior in future situations, to suggest appropriate actions to the user, and perhaps automatically perform actions on the user’s behalf. Several machine learning algorithms have been applied in interface agents: memory-based learning, reinforcement learning and induction of decision trees (ID3) are used in scheduling agents [5,10]; [2] compares different algorithms for classifying feature vectors with respect to their suitability for a WWW advisor. Machine learning has also been used in genuine user modeling systems to construct behaviororiented user models. [16] presents feature-based modeling for learning situation-action correlations (like interface agents) and demonstrates its use in tutoring systems. In the general user ¨ modeling system D OPPELG ANGER [11], statistical and machine learning methods are employed to collect evidence about user behavior patterns. Graph-based induction tries to find such patterns, too, by processing observations and pre-defined domain knowledge [19]. Furthermore, the use of machine learning has recently been investigated for purposes of plan recognition [3] and user classification [4]. 2.2. Construction of User Models So far, the application of machine learning for individualization of software systems has focused on detecting patterns in user behavior. User modeling systems, however, often need assumptions about users’ mental attitudes like beliefs, interests and goals. Machine learning algorithms that can help to acquire such assumptions must be able to produce or lead to the typically symbolic representations of mental attitudes. Promising candidates are induction algorithms for rule construction (like C4.5 [12]), concept formation [18], or inductive logic programming [7]. These techniques can be characterized as “knowledge-based”, in contrast to the mostly statistical and numerical algorithms mentioned in the previous section. Most of them involve domain theories, so that they can be used for acquisition of domain-specific assumptions. In addition to domain theories, user models themselves are an important input for learning techniques. The construction of mentally-oriented assumptions can be influenced by existing behavior-oriented assumptions and vice versa. Integrated acquisition and maintenance of both kinds of data is a central characteristic of the LaboUr architecture. 2.3. Evolution of User Models Keeping user models consistent and up-to-date has proven to be difficult when using deductive reasoning and heuristic rules for user model acquisition. Machine learning offers solutions to this problem. Particularly for logically represented assumptions, knowledge revision techniques may be useful [18]. In general, incremental methods like instance-based learning [1,17] and incremental variants of decision-tree induction [14] will modify previous learning results when new observations are processed. A standard method in user modeling is the use of stereotypes, i.e. pre-defined models of user groups. In order to quickly obtain informative user models, systems with stereotypes try

¨ to determine the user’s membership in the given groups. In D OPPELG ANGER [11], not only group membership is learned, but group models themselves are learned from user models and dynamically revised. Learning group models is an example of discovering information that is implicit in user models. This task is traditionally tackled with deductive reasoning; inductive learning algorithms will be a valuable complement. 2.4. User Modeling as Open Learning Process The benefits that machine learning offers to user modeling will be limited, if the user modeling process remains focussed on single applications. Assumptions about the user can be acquired more reliably based on observations of several applications, and they can be of use for more than one application. Therefore, user modeling should be an open learning process. Open means that a user modeling system constructs and represents all kinds of assumptions (behavior- and mentally-oriented) about many users, and communicates with several providers and consumers of user information. User modeling becomes a learning process, when assumptions about the user are incrementally constructed from observed user behavior and existing user model contents. 3. LaboUr – an Architecture for Learning about the User Based on the notion of user modeling as open learning process, we developed the LaboUr architecture for user modeling systems. The central entities within this architecture are the learning components, each of which incorporates a learning method for user model construction or evolution. The learning method considers its own history or context information and perhaps also domain knowledge and current user model contents when forming assumptions about users. Observations about users are the main input for model construction. They are reported via a communication interface that can also be used between LaboUr components. A learning component possesses a filter mechanism that recognizes appropriate observation data and, if necessary, transforms them into an input format suitable for the learning method. A second filter refrains generated assumptions from being entered into a user model if strength of or confidence in an assumption is not high enough. Model evolution components obtain input from user models and, in return, manipulate user model contents or update group models. Finally, assumptions are communicated to applications, mainly query-driven, but also proactively. At first sight, a LaboUr system is designed as a centralized user modeling server (cf. [11]), which maintains models of a multitude of users, and processes observations and queries (perhaps sent across a network) of a multitude of application systems. However, smaller LaboUr systems can exist on personal, perhaps mobile computers and cooperate with bigger servers. Also, the components of one LaboUr system can be distributed across a network. So, the LaboUr architecture provides enough flexibility for a wide range of user modeling applications. REFERENCES 1. D. W. Aha, D. Kibler, and M. K. Albert. Instance-based learning algorithms. Machine Learning, 6(1):37–66, 1991. 2. R. Armstrong, D. Freitag, Th. Joachims, and T. Mitchell. Webwatcher: A learning apprentice for the world wide web. Proc. of the 1995 AAAI Spring Symposium on Information Gathering from Heterogeneous, Distributed Environments, March 1995.

3. M. Bauer. Machine learning for user modeling and plan recognition. In V. Moustakis J. Herrmann, editor, Proc. ICML’96 Workshop “Machine Learning meets HumanComputer Interaction”, pages 5–16, 1996. 4. J. Finlay. Machine learning: a tool to support improved usability? In V. Moustakis J. Herrmann, editor, Proc. ICML’96 Workshop “Machine Learning meets Human-Computer Interaction”, pages 17–28, 1996. 5. R. Kozierok and P. Maes. A learning interface agent for scheduling meetings. In W. D. Gray, W. E. Hefley, and D. Murray, editors, Proc. of the International Workshop on Intelligent User Interfaces, Orlando FL, pages 81–88, New York, 1993. ACM Press. 6. M. Krogsæter, R. Oppermann, and C. G. Thomas. A user interface integrating adaptability and adaptivity. In R. Oppermann, editor, Adaptive User Support. Lawrence Erlbaum Associates, 1994. 7. N. Lavrac and S. Dzeroski. Inductive Logic Programming – Techniques and Applications. Ellis Horwood, New York, 1994. 8. P. Maes. Agents that reduce work and information overload. Communications of the ACM, 37(7):31–40, July 1994. 9. M. F. McTear. User modelling for adaptive computer systems: a survey. Artificial Intelligence Review, 7(3-4):157–184, August 1993. 10. T. Mitchell, R. Caruana, D. Freitag, J. McDermott, and D. Zabowski. Experience with a learning personal assistant. Communications of the ACM, 37(7):81–91, July 1994. 11. J. Orwant. Heterogeneous learning in the Doppelg¨anger user modeling system. User Modeling and User-Adapted Interaction, 4(2):107–130, 1995. 12. J. R. Quinlan. C4.5: Programs for Machine Learning. The Morgan Kaufmann Series in Machine Learning. Morgan Kaufmann, San Mateo, CA, 1993. 13. C. G. Thomas and G. Fischer. Using agents to improve the usability and the usefulness of the world-wide web. In S. Carberry, D. Chin, and I. Zukerman, editors, Fifth International Conference on User Modeling, pages 5–12. User Modeling, Inc., 1996. 14. P. E. Utgoff. Incremental induction of decision trees. Machine Learning, 4(2):161–186, 1989. 15. W. Wahlster and A. Kobsa. User models in dialog systems. In A. Kobsa and W. Wahlster, editors, User Models in Dialog Systems, pages 4–34. Springer, Berlin, Heidelberg, 1989. 16. Geoffrey I. Webb and Mark Kuzmycz. Feature based modelling: A methodology for producing coherent, consistent, dynamically changing models of agent’s competencies. User Modeling and User-Adapted Interaction, 5(2):117–150, 1996. 17. D. Wettschereck. A Study of Distance-Based Machine Learning Algorithms. PhD thesis, Oregon State University, June 1994. 18. S. Wrobel. Concept Formation and Knowledge Revision. Kluwer Academic Publishers, 1994. 19. K. Yoshida and H. Motoda. Automated user modeling for intelligent interface. International Journal of Human-Computer Interaction, 8(3):237–258, 1996.

Suggest Documents