Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
A Comparison of How Demographic Data Affects Recommendation                                                               ...
Upcoming SlideShare
Loading in …5

A Comparison of How Demographic Data Affects Recommendation


Published on

Recommender systems attempt to find relevant data for their users. As the amount of data available in the Web becomes larger, this task becomes increasingly harder. In this paper we present a comparison of recommendation results when using different demographic features (age, location, gender, etc.) commonly available in online movie recommendation communities. We assume that demographic information holds implicit information about users' taste and interests, and present results of a simple method that extends standard collaborative filtering algorithms to include one or several of these features. We evaluate the our assumption in movie recommendation scenario and combine different features in order to improve recommendation results.

Published in: Technology
  • Be the first to comment

A Comparison of How Demographic Data Affects Recommendation

  1. 1. A Comparison of How Demographic Data Affects Recommendation Alan Said, Till Plumbaum, and Ernesto W. De Luca ¨ DAI Lab @ Technische Universitat Berlin User Modeling, Adaptation and Personalization (UMAP 2011) Abstract Data characteristics What?’: An assumption that demographic information holds implicit There are very clear differences between the different demographic information about users’ taste and interests. This information can be groups, e.g. the difference between the most “postive” city and the used to improve recommendation results. most “negative” city is 12.3 on a scale of 0-100, women tend to have How?: A comparison of recommendation results when using differ- 100 ratings less than men. Similarly, number of ratings increases by ent demographic features (age, location and gender) which are com- age up to young adulthood and decreases by age afterwards. monly available in online recommendation communities. The com- parison is performed through a simple method that extends stan- dard collaborative filtering algorithms to include the demographic 70.3 features. 59.2 58.0 Results: Results imply that simple demographic data (age, sex, lo- 69.9 cation) can be beneficial for recommendation. 60.8 69.6 Dataset The data comes from German movie recommendation community (a) The top three and bottom (b) Average age, rating and Moviepilota and contains users, movies, the ratings by users on three German cities according number of friends for women movies, user related data such as age, location and gender. to average movie ratings. and men. Feature Number of users City 4, 400 Age 1, 292 Sex 6, 583 Total 10, 000 Table: Number of users with each specific feature. Approach A k-NN algorithm generates neighborhoods of most similar users. Similarity is based on usage-similarity (ratings) as well as group belonging. The similarity of users within the same group is multiplied by a factor proportional to the number of users. Now, users from the same demographic groups are more im- portant when finding the most popular items in the neighbor- hood (i.e. the items to recommend). Demographic relations from Moviepilot In our experiments, we chose to evaluate three basic demographic properties: Location - users from the same cities Results Age - users born in the same decade Gender - users of the same gender Type P@1010K P@10 % MAP10K MAP % City 2.79E − 4 2.66E − 4 4.7% 3.89E − 3 3.81E − 3 2.2% Type # of ratings % Age 2.45E − 4 2.45E − 4 0.0% 3.93E − e 3.93E − 3 0.0% City 991, 845 64% Sex 2.86E − 4 2.33E − 4 22.9% 4.22E − 3 3.82E − 3 10.4% Age 1, 144, 761 74% Table: Results for different demographic features. Gender 1, 398, 732 91% Total 1, 539, 393 100% Our initial tests show that our assumption, “demographic data Table: The number and percentage of total ratings avail- has an impact on CF”, is true. Gender, especially, seems to have a large impact with a resulting increase of 10% for MAP (and able for every demographic property. 22% Precision at 10) are very promising. Age seems, however, a not to have a very high impact. <firstname>.<lastname>