Tulmaran Group- The diversity and sheer number of installed news Lens support is provided or forthcoming in Gnus 5. Skip to search form Skip to main content. Similarly, the cost of mistakenly pick- larger set of users and on a larger scale. Maximizing customer satisfaction through an online recommendation system: Remember me on this computer.
|Published (Last):||14 May 2013|
|PDF File Size:||18.23 Mb|
|ePub File Size:||10.61 Mb|
|Price:||Free* [*Free Regsitration Required]|
Mtt netcom. Each line in the body of the article contains a rating of one article by one person. The five fields on each line are the id of the article, the pseudonym of the rater, a rating, the number of seconds the reader spent examining the article before rating it, and the newsgroups the article is in.
The time count is optional. Additional keyword identified fields can also be included at the end of line. Figure 5: a sample matrix of ratings. All the scoring methods we have implemented are based on the heuristic that people who agreed in the past are likely to agree again, at least on articles in the same newsgroup.
This heuristic will mislead on occasion, but preferences for most kinds of articles are likely to be fairly stable over time. To implement this heuristic, our BBBs first correlate ratings on previous articles to determine weights to assign to each of the other people when making predictions for one of them.
Then, they use the weights to combine the ratings that are available for the current article. We have investigated several techniques for correlating past behavior and using the resultant weights, based on reinforcement learning , multivariate regression, and pairwise correlation coefficients that minimize linear error or squared error. First, we compute correlation coefficients , weights between -1 and 1 that indicate how much Ken tended to agree with each of the others on those articles that they both rated.
All the summations and averages in the formula are computed only over those articles that Ken and Lee both rated. We have conveniently arranged for and to be 3 in this example, but that need not be true in practice. That is, Ken tends to disagree with Lee and agree with Meg.
Carrying through similar calculations for Nan yields a lower prediction of 3. The score prediction system is robust with respect to certain differences of interpretation of the rating scale.
If two users are perfectly correlated, but one user gives only scores between 3 and 5 and the other only scores between 1 and 3, a 5 score from the first user will result in a prediction of 3 for the second. If two users would be perfectly correlated, but the first mistakenly thinks 1 is a good score and 5 is bad, the two will be negatively correlated and a 1 score from the first will result in a prediction of 5 for the second. This leads to a clear explanation to the user of how to assign ratings: assign the rating you wish GroupLens had predicted for this article.
It may be, however, that a larger sample of subjects would have yielded some pairs with larger overlaps in their ratings. More importantly, it may be that pairs of people will share interests in some topics but not others. Two people may agree in their evaluations of technical articles, but not jokes. Our BBBs keep separate rating matrices for each newsgroup.
One hopes that the accuracy of the predictions improve as the BBB has more past ratings to use in computing correlations.
Four people at the University of Minnesota participated in a pilot test of an earlier version, using a slightly different scoring function.
While all four participants reported that the predicted scores eventually matched their interests fairly closely, they did observe that there was a start-up interval before the predictions were very useful.
Further experiments and analysis are necessary to determine just how long the start-up interval is likely to be for each new user. It seems likely that better scoring mechanisms can be developed. It may also be helpful to take into account the time people spent reading articles before rating them, information collected but not used by our BBBs. Fortunately, the GroupLens architecture is open: anyone can implement an alternative BBB so long as it posts ratings articles in the format described above and communicates with clients the same way that our BBBs do.
We hope that the development of alternative BBBs will become an active area for future research. As we describe below, our next pilot test should yield rating sets that we will make available to others who wish to evaluate alternative scoring algorithms.
Some may filter out those articles with scores below a threshold. Some may sort the articles based on the scores. Others may simply display the scores, numerically or graphically. One trend in news clients is to display a summary of the unread articles in a newsgroup.
Each line of the summary contains information about one article, typically the author, the subject line and the length. A user browses the summary and requests display of the full text of those articles that seem interesting.
All three of the news clients we modified use this display technique. The three modified clients we implemented make slightly different uses of the scores in the summary display. The modified NN client displays articles in the same order a regular NN client does, namely the order in which the articles arrived at the news server.
It merely adds an additional column containing the predicted scores. In the first version of this client, the scores were displayed numerically. The modified Gnus client uses the predicted scores to alter the order of presentation of articles in the summary.
Gnus clusters articles by thread. The modified Gnus client sorts the threads based on the maximum predicted score over the articles in the thread.
Within each thread, however, articles are still displayed in chronological order, to preserve the flow of discussion. As in the modified NN, the scores are displayed in an additional column in the summary. The Minnesota pilot test included users of both the Gnus and NN clients. As expected, participants tended to believe that the sorting and display mechanisms of their own news reader were best, but all were glad to see the score predictions incorporated into that standard format.
Several users, however, noticed that it was somewhat difficult to visually scan the predictions to find the high ones. A revised version of the NN client Figure 6 rounds off to the nearest integer and reports that as a letter grade A-E , a scale familiar to students at U. The modified NewsWatcher client displays the predicted scores as bar graphs rather than numbers Figure 7 , making it easier to visually scan for articles with high scores longer bars.
Otherwise, it follows the conventions of the original NewsWatcher client. Articles are grouped into threads and the summary display initially shows header lines only for the first article in each thread. Users can twist down the triangle associated with a thread to see the header lines for the rest of the articles. Figure 6: The modified NN client.
The third column displays the number of lines in the article. When no one has evaluated an article, no prediction is made. Figure 7: The modified NewsWatcher client displays predicted scores as bar graphs. Disclaimer: the scores were randomly generated for demonstration purposes. In practice, we would expect articles by Pete Bergstrom one of the authors of this paper to have much higher predicted scores. Scale Issues Further research is needed to understand how performance will change as the scale increases.
In the case of GroupLens, there are several relevant performance measures: prediction quality, user time, Better Bit Bureau compute time and disk storage, and network traffic. The first measure is the quality of score predictions. We expect prediction quality to increase as the number of users increases, since more data will be available to the prediction algorithm. Another measure is how long users have to wait to post ratings and receive predictions.
In an earlier version of GroupLens, the functions of the BBB were incorporated in the news client itself. One major advantage of the separate BBB is that it can pre-fetch ratings and pre-compute predictions rather than computing them when the user starts the news client.
Thus, user time should remain roughly constant as GroupLens grows, even if it takes more CPU time to compute scores. For many possible prediction formulas CPU time will grow even faster than linearly with increases in the number of users.
To reduce CPU time, BBBs could use only a part of the ratings matrix, trading off compute time against quality of predictions. Even though each rating is short, each news article might be read and rated by many raters, so the total volume of ratings could exceed the volume of news. To minimize storage requirements, BBBs may employ algorithms that use and discard ratings as they arrive, rather than storing them.
Three basic techniques could reduce network traffic: reduce the size of the ratings, reduce the number of ratings, and reduce the number of places where each rating is sent. Our BBBs batch several ratings in a single article, a first step toward reducing the amount of storage per rating, but further compression is possible.
The number of ratings could be reduced by limiting the total number of ratings per article or the number of ratings from users with similar profiles.
The separation of the BBBs from the news clients in the GroupLens architecture reduces the number of destinations for each rating: each news client receives only score predictions rather than all the individual ratings that contribute to those predictions. The number of destinations for each rating could be further reduced by sending ratings to some BBBs but not others. For example, BBBs could be clustered, based on geography or interest, and exchange ratings only within clusters.
The size of each cluster must be small enough to limit the amount of ratings information distributed, but large enough to provide an effective peer group. The table below estimates daily network traffic for various cluster sizes assuming each user rates articles per day and each rating requires approximately bytes.
For comparison purposes, the current netnews traffic is around MB per day. The architecture specifies the format of ratings produced in batches by BBBs, the propagation of the ratings by Usenet, and the interface for delivering predictions and ratings between news clients and BBBs.
Otherwise, the architecture is completely open. BBBs and news clients can be freely substituted, providing an environment for experimentation in predicting ratings and in user interfaces for collecting ratings and presenting predictions. These tests led to improvements in both the overall architecture and the user interfaces of news clients, as discussed already. The next step is a larger scale, distributed test, that we plan to carry out this summer. We have established a newsgroup on the news servers at MIT and Minnesota and two slightly different Better Bit Bureaus that communicate ratings through that newsgroup.
The test is not designed to demonstrate that people prefer to read netnews with our collaborative filters than without them. We believe that such an evaluation should wait for at least one more iterative design cycle.
Rather, the goals are to identify any unexpected scaling issues that may arise and to gather a data set that will be useful in evaluating alternative score prediction algorithms. At first glance, it might seem that any large set of ratings would be useful in creating such a benchmark. Upon closer inspection, however, complete ratings matrices are much more valuable than sparse ones. For example, suppose that users read and rate only a small number of articles, based on score predictions they receive from BBB X.
If users read different articles, this generates a sparse matrix of ratings. Now suppose that we wish to compare X to an alternative, Y, that predicts different scores for the users. To allow unbiased comparisons, we are asking each of the participants in the next pilot test to read and rate all the articles in a training set.
GroupLens: applying collaborative filtering to Usenet news
Konstan, Bradley N. Miller, David Maltz, Jonathan L. Herlocker, Lee R. Usenet newsgroups—the individual discussion lists—may carry hundreds of messages each day. While in theory the newsgroup organization allows readers to select the content that most interests them, in practice most newsgroups carry a wide enough spread of messages and advanced language features may be useless to the to make most individuals consider Usenet news to be novice. Furthermore, each The combination of high volume and personal user values a different set of messages. Both taste and taste made Usenet news a promising candidate for prior knowledge are major factors in evaluating news collaborative filtering.
Grouplens: Applying Collaborative Filtering to Usenet News
Kazijinn Thus, Usenet has a high potential Thus, grouppens risk of mistakes is lowest for movies or sci- benefit. Accordingly, many users predictionsusing implicit ratings, and exploring abandon the system before ever receiving benefits the use of filter-bot rating agents. The presence of many high amount of time before the correlations in the rec. One of our test users in Poland approach. The GroupLens Protocol Specification.
GroupLens: An Open Architecture for Collaborative Filtering of Netnews