Recommender systems based on automated collaborative filtering predict new items of interest for a user based on predictive relationships discovered between that user and other participants of a community. Most of the successful research and commercial
systems in collaborative filtering use a nearest-neighbor model for generating predictions. Automated collaborative filtering systems based on the nearest-neighbor method work in three simple phases:
1. Users of an automated collaborative filtering system rate items that they have previously experienced.
2. The automated collaborative filtering system matches the user with other participants of the system who have similar rating patterns (i.e. they have similar opinions on experienced items.) This is usually done through statistical
correlation. The closest matches are selected, becoming known as neighbors of the user, or collectively as the neighborhood.
3. Items that the neighbors have experienced and rated highly, but which the user has not yet experienced, will be recommended to the user, ranked based on the closeness of 数据挖掘研究院
the neighbors to the user and the consistency of opinion within the neighborhood.
Automated collaborative filtering systems are being applied to larger and larger sets of items. With large numbers of items in the prediction domain, we see the occurrence of three significant
negative phenomena:
1. Since users have limited resources to experience items (read articles, see movies, listen to music), the density of user ratings on items decreases. It becomes less likely that any significant number of a user’s neighbors will have experienced the item for which a prediction is being requested 数据挖掘研究院
2. While the density will decrease, the number of items that must be considered in each user to user correlation will still increase, increasing the amount of time necessary to compute the neighborhood.
3. As the number of items in the prediction domain gets large, the diversity of those items will also increase (otherwise, why would you need to recommender system to select between them?). As this diversity increases, it becomes less likely that a user’s opinions on all other items will be relevant to his opinion on a single given item. 数据挖掘研究院
