1 / 49

Prediction Strategies in a TV Recommender System

Outline. IntroductionPrediction techniques and prediction strategiesGeneric modelTechniques versus strategiesExperimentDatasetsMethod for gathering dataMeasuring accuracyValidation processUsed prediction techniques and strategyResultsConclusions and future work. Introduction. For informat

emil
Download Presentation

Prediction Strategies in a TV Recommender System

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


    1. Prediction Strategies in a TV Recommender System Method and Experiments IADIS WWW/Internet 2003, 6 November 2003 Mark van Setten, Mettina Veenstra Telematica Instituut, The Netherlands Anton Nijholt, Betsy van Dijk University of Twente

    2. Outline Introduction Prediction techniques and prediction strategies Generic model Techniques versus strategies Experiment Datasets Method for gathering data Measuring accuracy Validation process Used prediction techniques and strategy Results Conclusions and future work

    3. Introduction For information delivery, personalisation is an important issue Address issue of information overload (necessity) Support people to easily find interesting information (user support) Prediction users interest in information

    4. Prediction Techniques Content-Based Techniques: Structured Querying Information Filtering Case-Based Reasoning (CBR) Content Categories Social-Based Techniques: Social Filtering Item-Item Filtering Social CBR Top-N Demographics

    5. Generic Model

    6. Prediction Strategies Choose one or a combination of prediction techniques at the moment a prediction is required taking into account the most actual knowledge about the current user other users the information for which a prediction is required other information and the system itself

    7. Prediction Strategy Approaches Making decisions about which prediction techniques to use and how to combine them Possible approaches: Hard decision rules (if then else ) Fuzzy rules Artificial neural networks Bayesian networks Case-based reasoning In this experiment, we used hard decision rules created by experts

    8. Techniques vs Strategies: Blackbox

    9. Open Blackbox: Prediction Strategy

    10. Open Blackbox: Prediction Technique

    11. Main Advantage Independent development of prediction techniques Easy reuse of prediction techniques in different domains Creating a library of prediction techniques (toolkit) Easy creation and tuning of predition strategies Example A prediction technique based on CBR can be created without having any domain knowledge In each domain only one function has to be implemented for CBR to have enough domain knowledge: similarity between two items Prediction strategies provide more accurate predictions

    12. Experiments Two datasets MovieLens: Movie Recommendation System TiV: Personalised Electronic TV Guide Results of MovieLens experiment can be found in van Setten, M., Veenstra, M. & Nijholt, A. (2002). Prediction Strategies: Combining Prediction Techniques to Optimize Personalization. Proceedings of workshop Personalization in Future TV02 at Hypermedia 2002. Malaga, Spain, 28 May 2002 Results showed that prediction strategies improved the accuracy of the predictions However, question remained if this would also work in different domains

    13. TiV

    14. Used Prediction Techniques AlreadyRated The rating of an item if the user already rated that item UserAverage The average of all ratings provided by the user TopNDeviation Prediction based on all predictions from other users that already rated the item Social Filtering Prediction based on the idea that people who have rated the same items the same way will probably have similar interests patterns Case-Based Reasoning Prediction based on the idea that if two items are similar and if a rating is known for one of them, the rating for the other will probably be the same

    15. Used Prediction Techniques GenreLMS Prediction is based on the learned interests about the main genres of TV programs SubGenreLMS Prediction is based on the learned interests about the sub genres of TV programs InformationFiltering Similar to GenreLMS and SubGenreLMS, except that it uses all (stemmed) words from the description of TV programs, their frequency and the learned interests in these words to determine a prediction Default The neutral prediction value of zero

    16. Prediction Strategy

    17. TiV Dataset 24 users rated 4 weeks of TV programs containing 40,539 broadcasts distributed over 47 different channels resulting in 31,368 ratings Half-way during the 4 weeks there was the transition of the summer TV season to the winter TV season

    18. Measuring Accuracy Accuracy measure, combination of: Mean Absolute Error (mae) = Coverage = Global Mean Absolute Error (gmae) = mae, except when no prediction can be made the neutral prediction value (zero) is assumed Calculate gmae for each prediction technique and each prediction strategy, including the main strategy Perform paired samples T-tests with 95% confidence interval to determine if differences in gmae are statistically valid (p < 0.05)

    19. Measuring Accuracy Validating throughout system lifecyle: Validation throughout usage lifecycle First 100 ratings of each user Remaining ratings of each user

    20. Validation Process

    21. Validation Process

    22. Validation Process

    23. Validation Process

    24. Validation Process

    25. Validation Process

    26. Validation Process

    27. Validation Process

    28. Validation Process

    29. Validation Process

    30. Validation Process

    31. Validation Process

    32. Validation Process

    33. Validation Process

    34. Validation Process

    35. Validation Process

    36. Validation Process

    37. Validation Process

    38. Validation Process

    39. Validation Process

    40. Validation Process

    41. Validation Process

    42. Validation Process

    43. Validation Process

    44. Accuracy

    45. Used Techniques

    46. First Time versus Established Users

    47. Removing Prediction Techniques

    48. Conclusions Described a new way of looking at prediction engines using a generic method based on prediction strategies Prediction strategies make it possible to quickly create, use and test different strategies using several prediction techniques Showed that prediction strategies provide more accurate predictions because they only decide at the moment a prediction is required which prediction technique(s) to use Be aware of drawback: performance penalty

    49. Future Work Automated prediction strategies Using algorithms that can teach themselves when which prediction techniques can best be used Combining predictions of multiple prediction techniques First results show that combining predictions result in worse performance Looking at motiviations of people to be interested in information Let those (domain dependent) motivations guide the design of prediction strategies

    50. For More Information

More Related