Collaborative Clustering of Parisian Attractions

Yisong Yue <>

This is the project page of the Personalized Collaborative Clustering research project. We studied the problem of Collaborative Clustering, which can be thought of as the clustering analog to Collaborative Filtering. In Collaborative Filtering, each user provides an absolute rating of a subset of items (e.g., rating movies), and the goal is to learn a scoring function that can generalize to items not yet rated by each user. Analogously, in Collaborative Clustering, each user provides a clustering of a subset of items (i.e., a partial clustering), and the goal is to learning a similarity scoring function that can generalize to items not yet clustered by each user.

We collected our clustering data using an clustering interface designed to help users organize attractions in Paris. We asked workers on Amazon Mechanical Turk to use our clustering interface while planning a hypothetical trip to Paris. More information can be found in [1].

The data we collected is available here: This data is free to use for any non-commercial research project. All publications and preprints must cite [1].

  • "attraction_information" -- This is a tab delimited file containing attraction ids, attraction names, and an associated URL. All attractions are from Paris.

  • "attraction_mapping" -- This file contains a a sorted list of attraction ids. This file serves as a mapping from the attraction ids in "attraction_information" to a 0-indexed attraction id used in "clusttering_data". For example, the first attraction d232159 maps to 0.

  • "clustering_data" -- This file contains the clusterings created by users. Each line is a user, and contains clusters delimited by colons. Users only cluster attractions that they rate as itneresting. The last group of attractions are attractions that the user rated as not interesting, and so are not clustered. For example, the fist line contains:
      0 1 2 3 4 5 6 7 8 9:10 11 12:13 14 15 16:17:18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
    This means that the user created 4 clusters. The first cluster has attractions 0-9, the second cluster has attractions 10-12, the third cluster has attractions 13-16, and the final cluster has attraction 17. The user rated attractions 18-47 as not interesting. NOTE that we only use the items clustered (and thus rated interesting by the user) in our experiments.

  • "feature1" -- Each line in this file contains tf-idf features of one attraction. Each line is synchronized with each line in "attraction_mapping". Each line is space delimited, with each feature taking formulation [feature_id]:[feature_value]. All features are text-mined from Wikipedia, so some lines may be empty -- this happens when we cannot find an associated Wikipedia entry for the attraction.

  • "feature1_svd" -- This is a transfomation of "feature1" using dimensionality reduction via the singular value decomposition. This was the actual feature1 dataset used in the experiments in the paper.

  • "feature1_text_tfidf" -- This is the raw tf-idf information of the attractions, as mined from Wikipedia.

  • "feature2" -- Each line in this file contains keyword tagged features of one attraction. Each line is synchronized with each line in "attraction_mapping". Each line is space delimited, taking formulation [feature_id]:[feature_value]. The tagging process is described in Section 5.2 in the paper.

  • "feature2_keywords" -- This is the list of keywords we chose for tagging for "feature2".

  • [1] Yisong Yue, Chong Wang, Khalid El-Arini, and Carlos Guestrin. Personalized Collaborative Clustering, In Proceedings of WWW, 2014 [pdf]

  • [All Content © 2017 Yisong Yue]