Similarity metric relativized to a user's preferences

A method for recommending content based on one or more preferences of a user and a data set comprising a plurality of items, the method comprising, by a computing system comprising one or more computing devices, modifying, supplementing, or replacing a recommendation that is produced by a recommendation system by:

• representing the items by a plurality of item clusters and a plurality of feature clusters, the item clusters and the feature clusters algorithmically derived from the data set, the item clusters each comprising at least one item and the feature clusters each comprising at least one feature;
• creating a probability distribution of the data set by, for each feature cluster in the data set, determining probability distribution data indicative of a proportion of total occurrences of any feature cluster accounted for by that item cluster and feature cluster pair;
• modifying the probability distribution of the data set by mapping user activity data of the user to one or more topics of the data set to produce a user-preference feature distribution;
• wherein the user activity data comprises a subset of the data set;
• wherein the user-preference feature distribution comprises a plurality of weighted probability values;
• wherein a weighted probability value of the plurality of weighted probability values indicates both (i) a likelihood that a particular feature cluster occurs in a particular item cluster of the data set and (ii) a likelihood that the user will select an item from the data set;
• determining a user-specific similarity metric indicative of a similarity between one or more input terms and one or more items of the data set based on the user-preference feature distribution;
• and
• outputting one or more items to the user based on the user-specific similarity metric, where one or more of the item clusters are selected based on the user-specific similarity metric and the one or more items output to the user are chosen from the selected item clusters.

A method for expanding a query produced by a recommendation system, the method comprising, by a computing system comprising one or more computing devices:

• creating a probability distribution of a data set;
• modifying the probability distribution of the data set by mapping user activity data to one or more topics of the data set to produce a user-preference feature distribution;
• wherein the user activity data comprises a subset of the data set;
• wherein the user-preference feature distribution comprises a plurality of weighted probability values;
• wherein a weighted probability value of the plurality of weighted probability values indicates both (i) a likelihood that a particular feature cluster occurs in a particular item cluster of the data set and (ii) a likelihood that the user will select an item from the data set;
• determining one or more query expanders to append to a query of a user, the one or more query expanders being selected from the data set based on the user-preference feature distribution;
• and
• outputting the one or more query expanders to the user.

A method of creating a measure of user preference, the method comprising, with a computing system comprising one or more computing devices:

• accessing a data set, the data set comprising a plurality of items and each item comprising one or more occurrences of a plurality of features,
• clustering the data set into a plurality of item clusters;
• creating a probability distribution of the data set, wherein the probability distribution comprises, for each item cluster and feature combination in the data set, a probability value proportional to the number of occurrences of the feature in the item cluster;
• modifying the probability distribution of the data set by mapping user activity data of a user to one or more topics of the data set to produce a user-preference feature distribution;
• wherein the user activity data comprises a subset of the data set;
• wherein the user-preference feature distribution comprises a plurality of weighted probability values;
• wherein a weighted probability value of the plurality of weighted probability values indicates both (i) a likelihood that a particular feature cluster occurs in a particular item cluster of the data set and (ii) a likelihood that the user will select an item from the data set;
• making available the user-preference feature distribution for use by a recommendation system to modify, supplement, or replace a user-preference feature distribution that is produced by the recommendation system.