Michael J. Pazzani

JAAMAS Journal 2026 Journal Article

Adaptive Web Site Agents

• Michael J. Pazzani
• Daniel Billsus

Abstract We discuss the design and evaluation of a class of agents that we call adaptive web site agents. The goal of such an agent is to help a user find additional information at a particular web site, adapting its behavior in response to the actions of the individual user and the actions of other visitors to the web site. The agent recommends related documents to visitors and we show that these recommendations result in increased information read at the site. It integrates and coordinates among different reasons for making recommendations including user preference for subject area, similarity between documents, frequency of citation, frequency of access, and patterns of access by visitors to the web site. We argue that this information is best used not to change the structure or content of the web site but rather to change the behavior of an animated agent that assists the user.

ECAI Conference 2020 Conference Paper

CDeepEx: Contrastive Deep Explanations

• Amir Feghahati
• Christian R. Shelton
• Michael J. Pazzani
• Kevin Tang

We propose a method which can visually explain the classification decision of deep neural networks (DNNs). Many methods have been proposed in machine learning and computer vision seeking to clarify the decision of machine learning black boxes, specifically DNNs. All of these methods try to gain insight into why the network “chose class A” as an answer. Humans search for explanations by asking two types of questions. The first question is, “Why did you choose this answer? ” The second question asks, “Why did you not choose answer B over A? ” The previously proposed methods are not able to provide the latter directly or efficiently. We introduce a method capable of answering the second question both directly and efficiently. In this work, we limit the inputs to be images. In general, the proposed method generates explanations in the input space of any model capable of efficient evaluation and gradient evaluation. It does not require any knowledge of the underlying classifier nor use heuristics in its explanation generation, and it is computationally fast to evaluate. We provide extensive experimental results on three different datasets, showing the robustness of our approach, and its superiority for gaining insight into the inner representations of machine learning models. As an example, we demonstrate our method can detect and explain how a network trained to recognize hair color actually detects eye color, whereas other methods cannot find this bias in the trained classifier.

ICML Conference 2000 Conference Paper

Characterizing Model Erros and Differences

• Stephen D. Bay
• Michael J. Pazzani

ICML Conference 1998 Conference Paper

Learning Collaborative Information Filters

• Daniel Billsus
• Michael J. Pazzani

ICML Conference 1996 Conference Paper

Beyond Independence: Conditions for the Optimality of the Simple Bayesian Classifier

• Pedro M. Domingos
• Michael J. Pazzani

AAAI Conference 1996 Conference Paper

Syskill and Webert: Identifying Interesting Web Sites

• Michael J. Pazzani

We describe Syskill & Webert, a software agent that learns to rate pages on the World Wide Web (WWW), deciding what pages might interest a user. The user rates explored pages on a three point scale, and Syskill & Webert learns a user profile by analyzing the information on each page. The user profile can be used in two ways. First, it can be used to suggest which links a user would be interested in exploring. Second, it can be used to construct a LYCOS query to find pages that would interest a user. We compare six different algorithms from machine learning and information retrieval on this task. We find that the naive Bayesian classifier ofsers several advantages over other learning algorithms on this task. Furthermore, we find that an initial portion of a web page is sufficient for making predictions on its interestingness substantially reducing the amount of network transmission required to make predictions.

ICML Conference 1995 Conference Paper

A Lexical Based Semantic Bias for Theory Revision

• Clifford Brunk
• Michael J. Pazzani

ICML Conference 1995 Conference Paper

Learning Hierarchies from Ambiguous Natural Language Data

• Takefumi Yamazaki
• Michael J. Pazzani
• Christopher J. Merz

LOPSTR Conference 1994 Conference Paper

Avoiding Non-Termination when Learning Logical Programs: A Case Study with FOIL and FOCL

• Giovanni Semeraro
• Floriana Esposito
• Donato Malerba
• Clifford Brunk
• Michael J. Pazzani

Abstract Many systems that learn logic programs from examples adopt θ -subsumption as model of generalization and refer to Plotkin's framework in order to define their search space. However, they seldom take into account the fact that the lattice defined by Plotkin is a set of equivalence classes rather than simple clauses. This may lead to non-terminating learning processes, since the search gets stuck within an equivalence class, which contains an infinite number of clauses. In the paper, we present a task that cannot be solved by two well-known systems that learn logic programs, FOIL and FOCL. The failure is explained on the ground of the previous consideration about the search space. This task can be solved by adopting a weaker, but more mechanizable and manageable, model of generalization, called θ -subsumption under object identity ( θ OI -subsumption). Such a solution has been implemented in a new version of FOCL, called FOCL-OI.

ICML Conference 1994 Conference Paper

Reducing Misclassification Costs

• Michael J. Pazzani
• Christopher J. Merz
• Patrick M. Murphy
• Kamal M. Ali
• Timothy Hume
• Clifford Brunk

ICML Conference 1994 Conference Paper

Revision of Production System Rule-Bases

• Patrick M. Murphy
• Michael J. Pazzani

IJCAI Conference 1989 Conference Paper

Detecting and Correcting Errors of Omission After Explanation-Based Learning

• Michael J. Pazzani

In this paper, we address an issue that arises when the background knowledge used by explanationbased learning is incorrect. In particular, we consider the problems that can be caused by a domain theory that may be overly specific. Under this condition, generalizations formed by explanation-based learning will make errors of omission when they are relied upon to make predictions or explanations. We describe a technique for detecting errors of omission, assigning blame for the error of omission to an inference rule in the domain theory, and revising the domain theory to accommodate new examples.

AAAI Conference 1986 Conference Paper

The Role of Prior Causal Theories in Generalization

• Michael J. Pazzani

OCCAM is a program which organizes memories of events and learns by creating generalizations describing the reasons for the outcomes of the events. OCCAM integrates two sources of information when forming a generalization: l Correlational events. information which reveals perceived regularities in l Prior causal theories which explain regularities in events The former has been extensively studied in machine learning. Recently, there has been interest in explanation-based learning in which the latter source of information is utilized. In OCCAM, prior causal theories are preferred to correlational information when forming generalizations. This strategy is supported by a number of empirical investigations. Generalization rules are used to suggest causal and intentional relational relationships. In familiar domains, these relationships are confirmed or denied by prior causal theories which differentiate the relevant and irrelevant features. In unfamiliar domains, the postulated causal and intentional relationships serve as a basis for the construction of causal theories.

AAAI Conference 1983 Conference Paper

Interactive Script Instantiation

• Michael J. Pazzani

The KNOBS [ENGELMAN 80] planning system is an experimental expert system which assists a user by instantiating a stereotypical solution to his problem. SNUKA, the natural language understanding component of KNOBS, can engage in a dialog with the user to allow him to enter components of a plan or to ask questions about the contents of a database which describes the planning world. User input is processed with respect to several knowledge sources including word definitions, scripts which describe the relationships among the scenes of the problem solution, and four production system rule bases which determine the proper database access for answering questions, infer missing meaning elements, describe how to conduct a conversation, and monitor the topic of the conversation. SNUKA differs from GUS [BOBROW 77], a dialog system similar to SNUKA in its goals, in its use of a script to guide the conversation, interpret indirect answers to questions, determine the referents of nominals, perform inferences to answer the user’s questions, and decide upon the order of asking questions of the user to maintain a coherent conversation. SNUKA differs from other script-based language understanders such as SAM [CULLINGFORD 78] and FRUMP [DEJONG 79] in its role as a conversational participant instead of a story understander.