On the Long Tail Products Recommendation using Tripartite Graph

Abstract

The growth of the number of e-commerce users and the items being sold become both opportunities and challenges for e-commerce marketplaces. As the existence of the long-tail phenomenon, the marketplaces need to pay attention to the high number of rarely sold items. The failure to sell these products would be a threat for some B2C e-commerce companies that apply a non-consignment sale system because the products cannot be returned to the manufacturer. Thus, it is important for the marketplace to boost the promotion of long-tail products. The objective of this study is to adapt the graph-based technique to build the recommendation system for long-tail products. The set of products, customers, and categories are represented as nodes in the tripartite graph. The Absorbing Time and Hitting Time algorithms are employed together with the Markov Random Walker to traverse the nodes in the graph. We find that using Absorbing Time achieves better accuracy than the Hitting Time for recommending long-tail products.

“Guess Why I didn't rate it”: A New Preference-based Model for Enhanced TopK Recommendation

“…The latter encompasses two major points that may impact the pre-preference of a user 𝑢 toward an item 𝑖: (1) The popularity or the density of an item 𝐷(𝑖) may influence users' choice of items. By referring to the long-tail distribution of evaluated items in many real-world applications, one may see that a great number of feedback ratings are compressed into a small fraction of popular items [8,9]. (2) A user's prior preference towards an item can also be associated with the average relevancy 𝑅 𝑒 (𝑖) of that item (a user is more likely to watch a movie that the majority of users have positively rated).…”

“…Ignoring these missing interactions International Journal of Intelligent Engineering and Systems, Vol. 16 would bias the model towards higher available rating values, thereby inaccurately ranking tail items [9]. On the other hand, several state-of-the-art imputation approaches consider unobserved items with empty values in the rating matrix as negative examples [10,11] since users are likelier to assign ratings for items they are interested in, which would inaccurately increase the ranking of popular items.…”