Topic Title: Tree-based model for next-generation recommender systems


Technical Area:

Recommender systems

Deep learning

Tree-based learning



Recommendation has been widely used by various kinds of content providers.


Personalized recommendation method, based on the intuition that user’s interests can be inferred from their historical behaviors, has been proven to be effective in many applications such as Taobao, YouTube, Netflix, and Amazon.


Collaborative filtering methods are widely deployed in the industry to resolve the recommendation problem. As a representative method in the collaborative filtering family, item-based collaborative filtering method pre-calculates similarity between item pairs and leverages user's historical behaviors as triggers to recall the most similar items. However, there is restriction on the scope of candidate set, i.e., not all items but only items similar to the triggers can be recommended. This nature of the item-based collaborative filtering methods prevents the recommender system from jumping out of historical user behaviors and exploring potential user interests. Also, the recommendation diversity and novelty are often criticized.


In recent years, model-based methods for recommender systems have been studied to provide more precise results. In systems with large candidate set, the amount of calculation to evaluate all user-item pairs is tremendous, which makes the model-based methods difficult to be directly employed. To overcome the calculation barrier, models such as matrix factorization resort to inner product form (i.e., use the inner product of user and item's latent factors) and index such as hashing to perform efficient approximate nearest neighbor searches. However, other more expressive interactions between user and item, e.g., interactions through advanced deep neural networks, are not applicable because of the nature of the inner product form.


We strive to address the challenges above by a novel tree-based deep recommendation model. Tree and tree-based methods were extensively researched in multiclass classification problems and language models. However researchers seldom set foot in the context of recommender systems using tree structures. Actually, hierarchical information structures ubiquitously exist in many domains. For example, in the E-Commerce scenario, “iPhone” is a fine-grained item while “smartphone” is the coarse-grained concept to which “iPhone” belongs. Tree-based methods can leverage this hierarchy of information and turn a large-scale recommendation problem into a series of small ones to overcome the calculation barrier.



We have conducted a preliminary study on tree-based models for recommender systems and the corresponding paper “Learning Tree-based Deep Model for Recommender Systems” is to appear in KDD’18. Tree-based deep models take a new perspective to the large-scale recommendation problem. However, there are many future directions to be explored. For example, the construction and learning approaches of item hierarchy can be studied for more efficient training and more precise prediction. Besides, how to model both long-term and short-term user interests with advanced deep neural networks is also an open question. In this proposal, we hope to investigate more efficient and effective mechanisms for large-scale recommendation with the tree-based model. The research focuses not only on model training but also on online serving. That is, we are interested in deploy the state-of-the-art research results in the online serving systems to enhance the business performance and improve user experience.


Related Research Topics

In order to investigate and employ more efficient and effective mechanisms for large-scale recommendation with tree-based model, the following research topics arises:

1. Construction and learning of tree structure: In tree-based recommendation model, item tree is regarded as an index structure and the quality of item’s hierarchy plays a very important role. How to construct and learn better tree structure is worth studying

2. Modeling user interest evolution: The goal of recommendation is to find user interests and help users explore their favored items. Designing a model to capture user interest evolution from all long-term, short-term and even periodic user behaviors is an open problem.

3. Highly-efficient online serving: Though the inference time complexity is reduced to O(log n) in tree-based model, real-time online serving is still a challenge considering the evolving neural network structure. How to accelerate the inference and strike a balance between efficiency and effectiveness is always an active topic.