This paper presents an MDP-based recommender system that models the recommendation process as a sequential decision problem, rather than a static prediction problem. The authors argue that MDPs are more appropriate for this task because they account for the long-term effects of recommendations and their expected value. To build an effective MDP-based recommender system, a strong initial model is required. The authors propose using an n-gram predictive model to generate this initial MDP. This model captures user behavior as a Markov chain, achieving higher predictive accuracy than existing models. The paper describes the predictive model in detail and evaluates its performance on real data. It also shows how the model can be used in an MDP-based recommender system. The authors describe a predictive model based on n-gram techniques, which is used to initialize the MDP. The model is a Markov chain where states represent sequences of user selections. The transition function is based on n-gram models, and the authors describe several improvements to the model, including skipping, clustering, and mixture modeling. These techniques help address data sparsity and computational complexity. The authors then formalize the recommendation problem as an MDP and show how the predictive model can be used to initialize the MDP. They evaluate the performance of their predictive algorithm on real data and compare it with other models, including the Microsoft Commerce Server 2002 Predictor. The results show that the MDP-based approach outperforms traditional models in terms of predictive accuracy. The authors also discuss the challenges of evaluating MDP-based recommender systems, as real-world testing is difficult to conduct. The paper concludes that MDP-based recommender systems offer a more sophisticated approach to recommendation, taking into account the long-term effects of recommendations and their expected value. The authors suggest that future work should focus on improving the predictive model and exploring alternative methods for initializing and updating the MDP. They also note that the model could be enhanced by incorporating user information such as age and gender, as well as item relationships that can be explicitly specified.This paper presents an MDP-based recommender system that models the recommendation process as a sequential decision problem, rather than a static prediction problem. The authors argue that MDPs are more appropriate for this task because they account for the long-term effects of recommendations and their expected value. To build an effective MDP-based recommender system, a strong initial model is required. The authors propose using an n-gram predictive model to generate this initial MDP. This model captures user behavior as a Markov chain, achieving higher predictive accuracy than existing models. The paper describes the predictive model in detail and evaluates its performance on real data. It also shows how the model can be used in an MDP-based recommender system. The authors describe a predictive model based on n-gram techniques, which is used to initialize the MDP. The model is a Markov chain where states represent sequences of user selections. The transition function is based on n-gram models, and the authors describe several improvements to the model, including skipping, clustering, and mixture modeling. These techniques help address data sparsity and computational complexity. The authors then formalize the recommendation problem as an MDP and show how the predictive model can be used to initialize the MDP. They evaluate the performance of their predictive algorithm on real data and compare it with other models, including the Microsoft Commerce Server 2002 Predictor. The results show that the MDP-based approach outperforms traditional models in terms of predictive accuracy. The authors also discuss the challenges of evaluating MDP-based recommender systems, as real-world testing is difficult to conduct. The paper concludes that MDP-based recommender systems offer a more sophisticated approach to recommendation, taking into account the long-term effects of recommendations and their expected value. The authors suggest that future work should focus on improving the predictive model and exploring alternative methods for initializing and updating the MDP. They also note that the model could be enhanced by incorporating user information such as age and gender, as well as item relationships that can be explicitly specified.