An improved system and method is provided for learning a weighted index to categorize objects using ranked recall. In an offline embodiment, a learning engine may learn a weighted index for classifying objects using ranked recall by training during an entire initial pass of a training sequence of a collection of objects. In an online embodiment, a learning engine may learn a weighted index for classifying objects using ranked recall by dynamically updating the weighted index as each instance of the collection of objects may be categorized. Advantageously, an instance of a large collection of objects may be accurately and efficiently recalled for many large scale applications with hundreds of thousands of categories by quickly identifying a small set of candidate categories for the given instance of the object.