Data Sets for Multiple Instance Learning

The multiple-instance learning model is becoming increasingly important in machine learning. Unlike standard supervised learning in which each instance is labeled in the training data, here each example is a set (or bag) of instances which receives a single label equal to the maximum label among the instances in the bag. The individual instances are not given a label. The goal of the learner is to generate a hypothesis to accurately predict the label of previously unseen bags.

Localized Content-Based Image Retrieval
Classic Content-Based Image Retrieval (CBIR) takes a single non-annotated query image, and retrieves similar images from an image repository. Such a search must rely upon a holistic (or global) view of the image. Yet often the desired content of an image is not holistic, but is localized. Specifically, we define Localized Content-Based Image Retrieval as a CBIR task where the user is only interested in a portion of the image, and the rest of the image is irrelevant. ACCIO! is a Localized CBIR System that uses labeled images in conjunction with a multiple-instance learning algorithm to first identify the desired object and re-weight the features, and then to rank images in the database using a similarity measure that is based upon individual regions within the image. We used the following two data sets:
- The SIVAL (Spatially Independent, Variable Area, and Lighting) benchmark
  SIVAL includes 25 different image categories with 60 images per category. This benchmark emphasizes the task of Localized CBIR. The categories consist of images of single objects photographed against highly diverse backgrounds. The objects may occur anywhere spatially in the image and also may be photographed at a wide-angle or close up. We have created this benchmark since most of the Corel object images contain close-ups of an object that is centered in the image and typically occupies a majority of the image.
- We also use a five-category natural scenes image repository with sunset, mountain, waterfall, field and flower images from the Corel suite.

Drug Activity Prediction
The multiple-instance model was motivated by the drug activity prediction problem where each instance is a possible configuration (or shape) for a molecule of interest and each bag (example) contains all low-energy (and hence likely) configurations for the molecule.
Public data sets for learning in the multiple-instance model are for concept learning (i.e. boolean labels). Binding affinity between molecules and receptors is quantitative, borne out in quantities such as the energy released by the molecule-receptor pair upon binding and hence a real-valued label of binding strength is preferable.
The naming convention for the artificial data is LJ-r.f.s where:
- LJ is for Lennard-Jones Potential, which is used as the basis for mimic intermolecular interactions for the generation of artificial data sets.
- r is the number of relevant features.
- f is the number of features (dimensions).
- s is the number of diffent scale factors used for the relevent features.
For more details about the properties of each of the three data sets we are providing a README file.