This article is based on the ECML-PKDD-2020 paper "Social Influence Attentive Neural Network for Friend-Enhanced Recommendation". The authors of the paper are Lu Yuanfu, Xie Ruobing, Ishikawa, Fang Yuan, Wang Wei, Zhang Xuhe from Beijing University of Posts and Telecommunications, Tencent WeChat, Singapore Management University Lin Leyu.
Guide
With the rise of online social networks, a new social-based recommendation method has emerged in many social applications. One of its representative applications is WeChat Look at Friends in the recommendation system. In this article, we refer to WeChat friends watching recommendations as "Friend-Enhanced Recommendation (FER)". In FER, for a product that has never been interacted with before, the user can clearly see the interaction behavior of his friends with the product (these friends are called a friend recommendation circle), such as the article recommendation scenario in the WeChat "take a look" scenario . Different from traditional social recommendation, the unique friend recommendation circle in FER may significantly change the recommendation result due to its importance and interpretability.
In response to this friend-enhanced recommendation scenario, the author proposes a novel social influence focused neural network (SIAN). SIAN models recommendation scenarios from the perspective of heterogeneous information networks, and builds an attention feature aggregator, which aggregates multiple types of heterogeneous information at node-level and type-level features, and learns to obtain vector representations of users and commodities. More importantly, the SIAN model designed a social influence coupler to focus on capturing the influence of the friend recommendation circle.
Based on two public data sets and a WeChat take a look at the data, the author has conducted a wealth of experimental analysis, and the model is significantly improved compared with the comparison method. In particular, after taking a look at the data set, the author conducted some quantitative analysis and discussion, and got some interesting findings and valuable sociological models.
Background introduction
Nowadays, with the rise of online social networks, people are more willing to express their opinions on social platforms and share information with friends. Friends become an important source of information and high-quality information filters. Goods interacted by friends (shared, liked, etc.) will have a significant impact on users, which is likely to become a user's interest in the future. In the industry, there are also many recommendation systems that focus on the social influence of friends. For example, follow the information flow on YouTube and "take a look" article recommendations on WeChat.
In this article, the author formalized a new recommendation scenario based on WeChat "take a look" article recommendation scenario, called Friend-Enhanced Recommendation (FER), which greatly enhanced the friend-to-user recommendation system The influence of behavior. Compared with traditional social recommendation, there are two main differences in friend enhanced recommendation: (1) Since friends can be regarded as high-quality information filters to provide users with high-quality goods, friend enhanced recommendation only recommends users who have interacted with friends. Commodities (read articles); (2) All friends who have interacted with a commodity will be explicitly displayed to the current user, that is, which friends the current user knows have interacted with the current item. This feature emphasizes the impact of explicit social influence on user behavior, which also benefits the interpretability of user behavior.
Figure 1. Friend enhanced recommendation scenario
The left side of the above figure shows the scene of two articles recommended to Jerry, and the following is a friend (such as Tom) who interacts (shares, likes, etc.) with the two articles. The right side shows the formalization of the friend's enhanced recommendation question, in which only articles that friends have interacted with are recommended, and the friends who interacted with the item are explicitly exposed to the user Jerry.
Figure 1 shows a formal example of the WeChat "take a look" scenario. For each user-item pair, the enhanced friend recommendation explicitly displays the set of friends that have interacted with the current item. This set is defined as: for the current product, the current user's Friend Referral Circle (FRC).
For example, for an article about AirPods, Jerry's friend recommendation circle is {Tom, Lily, Jack}. In traditional social recommendation, the friend recommendation circle is invisible to Jerry (will not be shown to Jerry), so Jerry may read an article based on his own interests. However, in the enhanced friend recommendation scenario, the friend recommendation circle is explicitly visible to user Jerry, so he read this article about AirPods, more likely because technology expert Tom has read this article. And Jerry read an article about Disneyland, it is entirely possible that his wife Lily read this article. Furthermore, when the article is related to science and technology, the simultaneous appearance of experts and articles has a greater influence on Jerry's coupling than the wife and science articles. In terms of entertainment, the situation may be completely opposite.
Therefore, in the enhanced friend recommendation scenario, multiple factors lead to user behavior. The reasons why users read/click an article may come from: (1) their own interest in the article (the article itself) (2) expert recommendation (article-friend combination) (3) attention to a friend (friend) . It can be seen that in the enhanced friend recommendation scenario, users tend to pay attention to the content that their friends are watching, rather than just watching the content that they are interested in. It can even be said that traditional social recommendation focuses on combining social information to recommend items, while friend enhanced recommendation aims to recommend a combination of items and friends.
As a key feature of enhanced friend recommendation, the explicit friend recommendation circle brings two challenges to this recommendation scenario: (1) How to extract key information from heterogeneous factors in many aspects? The friend-enhanced recommendation scenario involves multiple heterogeneous targets, such as item content, friend recommendation circle, and user-item interaction behavior. Under different combinations of users, items, and friends' recommendation circles, the influence of these factors is even different. Therefore, the friend-enhanced recommendation scenario is more challenging because it not only needs to learn the user's preference for goods, but also needs to predict the impact of different factors on friends. (2) How to use explicit friend recommendation circle information? Explicit friend recommendation circle recommendation greatly emphasizes the importance of social information in recommendation, which is very important in friend-enhanced recommendation scenarios. However, there is very little work to explore the influence and characteristics of the friend recommendation circle in actual recommendation. Therefore, a well-designed strategy is needed to make full use of the explicit friend recommendation circle information in the friend enhanced recommendation scenario.
In order to solve these problems, the author proposes a novel social influence focused neural network (SIAN). Specifically, SIAN defines enhanced friend recommendation as a user-item interaction prediction task on a heterogeneous social graph, which flexibly integrates rich heterogeneous information into heterogeneous objects and their interaction connections. First, the author designed an attention feature aggregator, which considers both node-level and type-level feature aggregation to learn the representation vectors of users and items. Next, the author implemented a social influence coupler to model the coupling influence spread through the explicit friend recommendation circle. The coupler uses the attention mechanism to couple the influence of multiple factors (for example, friends and items) Together. Overall, the SIAN model captures many valuable factors in the friend-enhanced recommendation scenario, and successfully extracts the user's most basic preferences from heterogeneous information networks and explicit friend recommendation circles.
Model structure
As shown in Figure 2, SIAN uses heterogeneous social networks to model enhanced friend recommendation scenarios. In addition to the vector representation of users and items, SIAN also learns the low-dimensional vector representation of social influence by coupling influential friends and items. By learning the vector representations of users, items, and coupled social influence, SIAN finally predicts the probability of interaction between the user ???? and the item ????.
Figure 2. SIAN model overview
First, considering that different neighbors of the same type may contribute differently to feature aggregation, and different types contain many aspects of information, the author designed hierarchical node-level and type-level attention feature aggregators. At each level, the attention mechanism distinguishes and captures the potential relevance of neighbors and types, so that the SIAN model can encode various heterogeneous information in a more granular manner. Unlike some previous work, SIAN does not need to manually select meta-paths based on any prior knowledge.
Secondly, the author designed a social influence coupler, which is used to capture the coupling influence of an influential friend (such as Tom) and an item (such as Disneyland related articles), so as to quantify the degree of their coupling influence. Finally, it integrates multiple coupling influences from the friend recommendation circle with attention to express the influence of the entire friend recommendation circle on the current user and items.
Finally, based on the influence of users, items, and coupling, the author spliced the three into a two-layer neural network, and predicted the user-item interaction probability through a regression layer. Finally, the optimization model is trained by optimizing the cross-entropy loss function.
experiment analysis
Based on two public data sets (Yelp and Douban) and WeChat Look at Data (FWD), the author conducted a wealth of experimental evaluation and analysis. The statistics of the data set are shown in Table 1 below.
Table 1. Data set statistics
In the main experiment, the author compared four types of representative methods and reported the performance of the model under different representation dimensions. The results can be seen in Table 2. It can be seen that the proposed SIAN model is significantly better than the comparison method on multiple data sets. In particular, compared with traditional social recommendation methods, the SIAN model still performs outstandingly, thanks to the ability of the social influence coupler to encode various social factors for friends to enhance recommendation scenarios, which further confirms the social influence Importance in friend enhancement recommendation.
Table 2. User behavior prediction experiment results
Secondly, the author analyzes the influence of heterogeneous information in many aspects. In the attention feature aggregator, the vector representation of each node is aggregated from various types of heterogeneous neighbors with different weights. The author analyzes and finds that the average attention weight of the friend type learned by the SIAN model The value is significantly larger than the average attention weight of the item type. This shows that the SIAN model pays more attention to users' social relationships, and this attention even exceeds the content itself. This also verifies that the friend recommendation circle is the most important factor in friend enhancement recommendation. The experimental results are shown in Figure 3.
Figure 3. Multi-faceted information impact analysis
Furthermore, taking a look at the data set on WeChat, the author analyzed how different user attributes can affect the user behavior in the friend enhancement recommendation scenario through the influence strength of the social influence coupler, and made some interesting findings. . For example, in Figure 4(a), it can be observed that user behavior is affected by more authoritative friends, and has nothing to do with the user's own authority. At the same time, friends with high authority have greater influence on users with medium authority. Users are usually more susceptible to the influence of authority figures, which is consistent with our common sense. Therefore, in the enhanced recommendation scenario of friends, there may be an interesting phenomenon, that is, sometimes users will pay more attention to the preferences of the authority (or just the authority itself) rather than their actual preferences. We have also observed that other basic user attributes have an impact on user behavior in the friend enhanced recommendation scenario.
Figure 4. Social influence analysis
to sum up
In this work, for the first time, the author formalized the definition of a WeChat friend watching a recommendation scene, that is, friend enhanced recommendation, which explicitly displays friends who have interacted with the current item. In response to this new and widely used recommendation scenario, the author proposes the neural network SIAN for social influence attention. The SIAN model learns the representation of user and item through a two-level attention aggregator, and designs a social influence coupler. Used to extract influence information in the circle of friend recommendation. Experimental results show that on three real data sets, SIAN's performance is significantly better than the latest benchmarks, and rich model analysis reveals interesting sociological patterns.
We believe that in the future, the enhanced recommendation of friends represented by WeChat to see what friends are watching will become an important way for people to obtain information and connect users with the world through recommendations. The research on enhanced friend recommendation is only in its infancy, and we hope to have more new research and analysis in the future.
references:
[1] Chen, Chong, et al. "Efficient Heterogeneous Collaborative Filtering without Negative Sampling for Recommendation." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 01. 2020.
[2] Dong, Yuxiao, Nitesh V. Chawla, and Ananthram Swami. "metapath2vec: Scalable representation learning for heterogeneous networks." Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining. 2017.
[3] Fan, Wenqi, et al. "Graph neural networks for social recommendation." The World Wide Web Conference. 2019.
[4]Grover, Aditya, and Jure Leskovec. "node2vec: Scalable feature learning for networks." Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining. 2016.
[5] Kipf, Thomas N., and Max Welling. "Semi-supervised classification with graph convolutional networks." arXiv preprint arXiv:1609.02907 (2016).
[6] Perozzi, Bryan, Rami Al-Rfou, and Steven Skiena. "Deepwalk: Online learning of social representations." Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 2014.
[7] Shi, Chuan, et al. "Heterogeneous information network embedding for recommendation." IEEE Transactions on Knowledge and Data Engineering 31.2 (2018): 357-370.
[8] Veličković, Petar, et al. "Graph attention networks." arXiv preprint arXiv:1710.10903 (2017).
[9] Wang, Xiao, et al. "Heterogeneous graph attention network." The World Wide Web Conference. 2019
[10] Wu, Le, et al. "A neural influence diffusion model for social recommendation." Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval. 2019.
[11] Xiao, Wenyi, et al. "Beyond personalization: Social content recommendation for creator equality and consumer satisfaction." Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2019.
Live preview
Mini Program Cloud Development Practical Live Class II
Current live time
September 3 (Thursday) 20:00
Live themes of this issue
Kbone+cloud development, quickly help small programs & Web isomorphism
Scan code appointment
Mini Program Cloud Development is an official one-stop backend cloud service for Mini Programs. Using cloud development, developers can develop a small program with front and back ends without having to manage the underlying architecture.
Kbone is a solution dedicated to the homogeneity of the WeChat applet and the web, so that the code on the web can run on the applet without too many changes.