Efficient top-n recommendation for very large scale binary rated datasets

Abstract

We present a simple and scalable algorithm for top-N recommen-\ud dation able to deal with very large datasets and (binary rated) im-\ud plicit feedback. We focus on memory-based collaborative filtering\ud algorithms similar to the well known neighboor based technique\ud for explicit feedback. The major difference, that makes the algo-\ud rithm particularly scalable, is that it uses positive feedback only\ud and no explicit computation of the complete (user-by-user or item-\ud by-item) similarity matrix needs to be performed.\ud The study of the proposed algorithm has been conducted on data\ud from the Million Songs Dataset (MSD) challenge whose task was\ud to suggest a set of songs (out of more than 380k available songs)\ud to more than 100k users given half of the user listening history and\ud complete listening history of other 1 million people.\ud In particular, we investigate on the entire recommendation pipeline,\ud starting from the definition of suitable similarity and scoring func-\ud tions and suggestions on how to aggregate multiple ranking strate-\ud gies to define the overall recommendation. The technique we are\ud proposing extends and improves the one that already won the MSD\ud challenge last year

Towards ‘Human/System Synergistic Development’: How Emergent System Characteristics Change Software Development

“…It can be seen that UserKNN and ARM outperform MF-based methods on all metrics, which is especially prominent in case of D2, where BPRMF has performance comparable to that of the most popular books recommender. This finding conforms to those reported in [4], [3].…”

“…For a given customer, recommendations can be computed either from the neighborhood of similar customers (user-based KNN), or from the neighborhood of products similar to those he or she has bought (item-based KNN). Although the latter approach has been favored in RS literature, we found user-based KNN to be more effective on both our datasets, in line with the results of [4], [16].…”

“…Whereas the main academic interest appears to be the accuracy of predictions, industrial recommenders are meant to be optimized for revenues. As Aiolli [4] pointed out, the choice of an accurate predictor is not all that is required to build a good recommender. Other evaluation criteria that play an important role in e-commerce are coverage of product catalogue, utility and serendipity of recommendations, adaptivity and scalability of an algorithm, etc.…”

Bundle recommendation in ecommerce

“…It has been reported that the average order size for Amazon.com is 1.5 and 2.3 for Walmart.com. 1 Using the transaction data of Walmart.com, we found over 1/3 of orders contain two or more distinct items as shown in Figure 1.…”

Load balancing for partition-based similarity search

“…All Pairs Similarity Search (APSS) [9], which identifies similar objects in a dataset, is used in many applications including collaborative filtering based on user interests or item similarity [1], search query suggestions [22], web mirrors and plagiarism recognition [23], coalition detection for advertisement frauds [20], spam detection [11,17], clustering [7], and near duplicates detection [16]. The complexity of naïve APSS can be quadratic to the dataset size.…”