Title: Recommendation algorithm based on tree structure and deep learning
Technical Area: Recommender System, deep learning, tree-structure
Recommendation algorithms play an important role for large e-commerce platforms. It enables the functions that displays the most suitable items to the interested customers. Matching the products with customers’ interests can be well performed when there are limited number of products and relatively sufficient history data showing the customers’ preference. The problem become more challenging when there are a huge volume of products with a large number of characteristic which can affects customers’ preference. Further, for some of the products the history data can be limited such as durable products, specialized products, new products or luxury products. A network model represents more associations among different characteristics among products and customer interests. Network based prediction algorithms can significantly improve the accuracy because of the nature of multi-linkage, its scope and matching can be well improved. However, it also means the efficiency will be reduced, particularly when the application requires a real time response.
To combine the benefits of the efficiency of tree structured models and accuracy of network structured models, this project aim to develop new mechanisms the corresponding algorithms to better balance between the efficiency and accuracy. An approach is to use multiple levels of structure with hybrid models. The detailed matching involves multiple features where naturally shall be a network structure to achieve high accuracy. At the same time, the span of these close related links is limited. As the subgraph size is limited, the speed can still be reasonably fast. On the other hand, the larger scope of the classes can be generally categorized, which possibly follow a tree structure to improve the efficiency. Establishing a multiple level model combining tree and network structure at different scales can balance achieve high accuracy with sufficient speed to meet real time response requirement.
Related Research Topics
Firstly, we are going to formulate the mathematic model of the problem where the scale and complexity of the classification and queries are essential. Secondly, we will explore a novel tree-network hybrid structure model for deep learning-based recommendation. Thirdly, we will implement the recommendation algorithm into prototype system and test the algorithms with real-world e-commerce datasets to evaluate its performance, comparing with state-of-the-art methods.