Recommendation Technology

4 May, 2020 by Jonathan Staniforth
“Artificial Intelligence & AI & Machine Learning” by Mike MacKenzie is licensed under CC BY 2.0.

Introduction & Motivation

Personalized recommendation is ubiquitous, having been at the core of many online services such as E-commerce, social networking, and advertising. At its core is to learn user preference on items based on historical behaviours, such as purchases and views, and then estimate how likely a user would adopt an item. It is becoming more entrenched with businesses. For example, around 65 and 75 percent of video views come from the recommender systems on YouTube and Netflix, respectively. Hence the ability to build effective recommendation models is of crucial significance.

Existing recommender systems have two major limitations that: (1) the data used is limited to the pre-existing features of users and items (e.g., ID or attributes), while high-order relations between users and items are less explored; and (2) the widely-used models belong to the multivariate linear family, failing to capture the nonlinear and complex patterns of user-item interactions. Furthermore, existing approaches lack in-depth consideration for reasons and intentions behind user behaviours, and further fail to explore user interests properly. Such shortcomings would make existing recommenders less intelligent and severely limit their benefits to businesses and our society.

Towards this end, we focus on the vision of intelligent recommendation, aiming to incorporate the advances of AI techniques into the recommender systems, and exploit the rich information, such as (multimodal) knowledge graph and multimedia, to exhibit and reason on user intents more precisely. In particular, this research covers several aspects, including: deep learning-based, explainable, knowledge graph-based, and multimedia recommendation.

Figure 1 Vision of Intelligent Recommendation.

Current Research

For the application of deep learning, we propose two pioneering framework — neural collaborative filtering (NCF) [1] and neural factorization machine (NFM) [2]. In particular, NCF parameterizes ID information of users and items as embeddings, and incorporates non-linear units to automatically reveal nonlinear and complex relations between user and item embeddings. NFM is the first work to subsume factorization machine (FM) under neural networks, which employs a well-designed bi-interaction pooling operation to model high-order and nonlinear feature interactions. Both works have attracted much attention from academia and industry: NCF is selected as one of the TensorFlow official recommendation models, while NFM has been successfully adopted by several Chinese companies for the task of click-through rate (CTR) prediction.

Explainable recommendation attracts increasing attention, as it is important to make users aware of why certain items were recommended to them in order to improve user satisfaction and trust. Towards this end, we proposed cross features of users and items, especially in the form of knowledge proxies, and reason about user-item interactions by synthesizing information within these proxies. In particular, we developed TEM [3], which employs a decision tree model to learn explicit decision rules from user profiles and item attributes, and then uses a neural attention network to select the top rules as a concise explanation for decision. The model has been successfully applied by a Singapore technology company to offer explanations.

Beyond general features, many data sources about users and items (e.g., facts) can be organized in the form of knowledge graph, which are complementary to user-item interactions. Specifically, within a knowledge graph, the connectivity between user and item nodes can be treated as their relations and used as explicit explanations to unseen user-item interactions. We have done several pioneering works. In particular, we proposed KPRN [4] to reason about the qualified paths on why a recommendation is made; KTUP [5] that couples user preferences with knowledge relations, to exhibit the semantics of preference embeddings; and KGAT [6] that generates attention flows from users to items, to model high-order connectivity between users and items; KGPolicy [7] which exploits external knowledge to distil negative signals of user preference. 

Multimedia items such as micro-videos and products in E-commerce are typically associated with multimodal features. For example, a micro-video has visual, acoustic, and textual features. Users usually have inconsistent preferences on different modalities. Hence how to accurately capture users’ modal-specific preferences is of importance. Towards this end, we proposed MMGCN [8] to improve not only the recommendation performance, but also the interpretability of model (e.g., demonstrate the importance of different factors). 

Furthermore, we developed and published a library, NeuRec, which incorporates a variety of state-of-the-arts neural recommendation models. It aims to tackle the general and sequential recommendation tasks. Current version includes 33 neural recommendation models (https://github.com/NExTplusplus/NeuRec). Moreover, we are working on another open source library, AdvRec, which focuses on using adversarial learning to boost recommendation. Current version includes 7 models (https://github.com/NExTplusplus/AdvRec).

Plans for Future Research

We plan to further develop new frameworks on intelligent recommendation, with in-depth consideration for powerful model capability (e.g., establishing expressive representations and modeling deep interactions), efficiency (e.g., capturing dynamic preferences of users in real time), reasoning (e.g., providing concise evidence and understanding of user intents), and diverse data fusion (e.g., providing users’ multiple feedbacks, with different types of features).

Reference

  1. Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, Tat-Seng Chua: Neural Collaborative Filtering. WWW 2017
  2. Xiangnan He, Tat-Seng Chua: Neural Factorization Machines for Sparse Predictive Analytics. SIGIR 2017
  3. Xiang Wang, Xiangnan He, Fuli Feng, Liqiang Nie, Tat-Seng Chua: TEM: Tree-enhanced Embedding Model for Explainable Recommendation. WWW 2018
  4. Xiang Wang, Dingxian Wang, Canran Xu, Xiangnan He, Yixin Cao, Tat-Seng Chua: Explainable Reasoning over Knowledge Graphs for Recommendation. AAAI 2019
  5. Yixin Cao, Xiang Wang, Xiangnan He, Zikun Hu, Tat-Seng Chua: Unifying Knowledge Graph Learning and Recommendation: Towards a Better Understanding of User Preferences. WWW 2019
  6. Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, Tat-Seng Chua: KGAT: Knowledge Graph Attention Network for Recommendation. KDD 2019
  7. Xiang Wang, Yaokun Xu, Xiangnan He, Yixin Cao, Meng Wang and Tat-Seng Chua: Reinforced Negative Sampling over Knowledge Graph for Recommendation. WWW 2020
  8. Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong and Tat-Seng Chua: MMGCN: Multimodal Graph Convolution Network for Personalized Recommendation of Micro-video, MM 2019

Projects

Video Relation Inference and Content Understanding

4 May, 2020

Multimodal and Multilingual Knowledge Graphs

4 May, 2020

Explainable AI

4 May, 2020

Recommendation Technology

4 May, 2020

Multimodal Conversational Search

4 May, 2020

Dialogue and Interactive Systems

4 May, 2020

Heterogeneous Data Mining for Fintech

4 May, 2020

Visually-Aware Fashion Computing

4 May, 2020

Back