ROBIN BURKE feature augmentation and meta-level hy brids therefore all permutations must be considered and these columns do not contain any redundancies. There are 60 non-redundant spaces in the table but some combinations are not possible. Since a knowledge-based technique may take into account any kind of data feature combination does not really represent a possible hybrid. Conversely, the demographic technique is similar to the collaborative in its approach(comparing users against each other), just using different features(demographic data vs. ratings) to do so. Therefore, it does not make sense to distinguish a content-based/ demographic (CN/DM) meta-level hybrid from a content-based/ collaborative (CN/CF)one. The illogical hybrids are marked in black. The white areas of the table enumerate 53 different possible hybrid recommender systems Of the possible hybrids, only 14 seem to have been explored, leaving significant room for further research This chart suggests some interesting types of recommenders that do not yet exist Although collaborative filtering is the most fully explored technique, a number of its hybrids remain unexplored. Content-based/ collaborative feature augmentation hybrid. This possibility was described earlier: a content-based 'filterbot Collaborative/content-based meta-level hybrid, in which collaborative infor- mation is used to generate a representation of overall user ratings for an item and this representation is then used to compare across items Collaborative/demographic augmentation hybrid in which a collaborative technique is used to place the user in a niche of like-minded users, and this information is used as a feature in a demographic rater In addition, four cascade recommenders involving collaborative recommen dation appear untried Other techniques show even fewer examples. Demographic techniques are poorly represented because this kind of data is more difficult to obtain than user ratings Only 4 of the possible 25 such hybrids appear to have been attempted Knowledge and utility-based techniques are also relatively under-explored with 4 of the possible 6 of these combinations researched. Together these techniques account for 36 of the 39 possible hybrid recommenders not yet explored. One reason for this focus on collaborative and content-based techniques is the availability of ratings data bases, such as the popular EachMovie database, which has approximately 45,000 ratings for 250 users. When combined with public data on movies, this data base has ena bled researchers to explore content-based and collaborative tech niques quite thoroughly. While the space remains to be fully explored, research has provided some insight into the question of which hybrid to employ in particular situations. The hybridization strategy must be a function of the characteristics of the recommenders being combined. With demographic, content and colla borative recommenders, this Reproduced with permission of the copyright owner. Further reproduction prohibited without permissionReproduced with permission of the copyright owner. Further reproduction prohibited without permission