Week 6 - Possible Readings

[lkcao]

Post a link for a "possibility" reading of your own on the topic of Network & Table Learning [for week 6], accompanied by a 300-400 word reflection that: 1) briefly summarizes the article (e.g., as we do with the first “possibility” reading each week in the syllabus), 2) suggests how its method could be used to extend social science analysis, 3) describes what social data you would use to pilot such a use with enough detail that someone could move forward with implementation.

[JadeBenson]

https://dl.acm.org/doi/pdf/10.1145/3209978.3210077

1) In “Deep learning for Epidemiological Predictions,” Wu et al. propose how deep learning approaches (RNNs and CNNs) could be used for time-series forecasting of predicting new disease outbreaks. They compare these methods to two of the most commonly used techniques in computational epidemiology, autoregressive models and Gaussian process regression, and find meaningful improvements. The datasets they test these models on are relatively simple with only time-series influenza counts from different locations (Japan, US states, and US-HHS districts). They create a full model that combine CNNs, RNNs, and residuals in the final layer (CNNRNN-Res model) which consistently performs the best across data sources. 2) I think this article is interesting because it was one of the first that used deep learning techniques as applied to epidemiology and compared them to classical approaches. Even with basic data sources that are publicly accessible, they were able to create better models that could help predict future outbreaks and, therefore, inform interventions. This deep neural network architecture could be used for future studies using infectious disease data. However, I think it could also be meaningfully expanded by including different types of data. They are only interested in predicting raw number of cases at particular times and locations, which is important, but could be improved. There may be demographics that are at greater risk of severe health consequences if infected, and including counts by demographics could help public health officials target those most in need. I would also be interested to include networks in these models to see not only the numbers of cases, but how they are likely to spread and who they are likely to infect. 3) A simple application of these methods would be scraping the weekly COVID cases reported to the city of Chicago by demographics. This is the same structure of data as the studies explored in this article. Then we could apply the same modeling rounds: CNNs, RNNs, residuals, and then CNNRNN-res to see what the predictive performance was on a hold-out set. I think it would be interesting to see where the model is most inaccurate and if this is because of an unexpected drop in cases and if this corresponded to the timeline of different interventions (masks, lockdowns, etc.).

[linhui1020]

R. G. Guimarães, R. L. Rosa, D. De Gaetano, D. Z. Rodríguez and G. Bressan, "Age Groups Classification in Social Network Using Deep Learning," in IEEE Access, vol. 5, pp. 10805-10816, 2017, doi: 10.1109/ACCESS.2017.2706674.

1. In "Age groups classification in social network using deep learning", the authors construct a build that could successfully predict the age group of users with high prediction accuracy. Despite that social media has a lot of useful data, those data lacks associated demographic information of speakers. When analyzing the sentiment of the sentence, there still exists errors because of the loss of profile information. The accuracy of sentiment analysis would be higher if demographic information is considered. In addition, the authors argue that shared aged groups have shared similarities in terms of writing style, which could be used as a reference for prediction of demographic information. Using a CNN model with around 7000 training samples in the same age group, the authors achieve a 95% accuracy in age group classification. And some interesting features include that adults are more likely to retweet or comment with a url in the tweets.

2. I think this article is interesting because 1) it uses the similarity in writing style of tweets of users to build social network, or more precisely, to classify the age group of users 2) it provides a methodology to infer the age of users, and such demographic information usually is hard for us to get 3) Compared with other machine learning techniques, their deep learning model achieves highest accuracy and the result is consistent with the human subject test. But, I still have some questions, firstly, this paper only identifies teenager and adults, which may be too limited for studies which require more specific ages. Will the prediction accuracy decrease when clustering into more age groups? Can we analyse the friends of that users and the users as way to collect the tweets and construct model so that we not only have tweet data but also has friend relationship data that can be an complementary for the classification tasks? In addition, I am also interested into seeing how age interacts with gender, job and another characters to have impact on the writing behavior on twitter platform.

3. The extension of this study may used for investigating political ideology at different age group. And this also could be used for studying the collective action emergence where shared beliefs may play an important role in. Possibly, when linking the age group with the information diffusion, we also can see whether the information diffusion speed similar or dissimilar among each group, and which group is more focus on which group. It is also interesting if we can link the name of users, the age, the education record and other social media account to build a database for future study on linguistic and human behavior.

[borlasekn]

Berahmand, K., Nasiri, E., Rostami, M., and Forouzandeh, S. (2021). A modifed DeepWalk method for link prediction in attributed social network. Springer Computing.

Link: https://link.springer.com/article/10.1007/s00607-021-00982-2

1. The DeepWalk algorithm is a popular method for embedding graphs that can capture the structure of the network as well as the attributed of the notes (or attributed networks). This paper looks at DeepWalks from a modified lens based on a pure version of random walks in order to solve link prediction in the various attributes and features. It also introduces a new model for random walks which is based on the assumption that two nodes of the network will be linked given they are nearby in a network. This means they would be connected because they share similar attributes. The authors justify this proposal by performing six experiments on real-world attributed networks using network embedding methods. These experiments include datasets such as the Facebook dataset and the BlogCatalog dataset.
2. This method could extend social science analysis because it could simplify the ways that DeepWalks are used for link prediction and increase their accuracy. Rather than going through every possible link, these algorithms should identify two nodes as a potential link that are most similar in terms of the graph structure and their attributes. This should lead to better performance, because the random walks become biased using this attribute and structural information. In the case of predicting links in their experiments, their plan yielded positive results. Enhancing the performance of link prediction can be important in the social sciences because it can suggest links between nodes in a social network that may not have been identified by a weaker algorithm.
3. If I were to pilot my own study using these DeepWalks for link prediction, I would attempt to look at potential collaborators for researchers. In looking at other sources that a given researcher has cited, we can predict potential links between the research of one researcher and another researcher and suggest potential collaborations that previous would not have been considered. In using the algorithm that uses biased random walks and looks at nodes that are close in the network and share similar attributes, we can predict with more accuracy. This would hopefully allow for more fruitful collaborations than a less accurate algorithm.

[isaduan]

Quantifying Spatial Homogeneity of Urban Road Networks via Graph Neural Networks

Link: https://arxiv.org/abs/2101.00307

1. While conventional statistics provide useful information about characteristics of either a single node's direct neighbors or the entire network, such metrics fail to measure the similarities of subnetworks considering local indirect neighborhood relationships. This study uses Graph Neural Networks to quantify the spatial homogeneity of subnetworks. We find that intra-city spatial homogeneity is highly associated with socioeconomic statuses such as GDP and population growth. Moreover, inter-city spatial homogeneity obtained by transferring the model across different cities, reveals the inter-city similarity of urban network structures originating in Europe, passed on to cities in the US and Asia. Socioeconomic development and inter-city similarity revealed using our method can be leveraged to understand and transfer insights across cities.
2. This method could extend social science analysis to measure the similarities of subcommunities (breaking down a large online community e.g. conspiracy theory community into different groups centered on different influencers, or breaking down the academic discipline of sociology into different methods and areas of interests). We can ask: how does the similarity of substructures (almost fractal structures!) of a community affect some end results we care about (e.g radicalization of individuals, rate of intellectual innovation)?
3. I can look at subcommunities of different academic communities (e.g. Chinese vs. America vs. the UK), and see whether the similarity of substructures correlates with higher intellectual diversity or less over time.

[sabinahartnett]

https://www.pnas.org/doi/abs/10.1073/pnas.1618923114 Emotion shapes the diffusion of moralized content in social networks Brady et al., 2017

1. This study uses network analysis to follow the impact of "moral contagion" in online networks and study how social networks transmit moral attitudes and norms. Using social media data, this study reveals that the presence of moral-emotional words in online messages can increase their diffusion by a factor of 20% for each additional word. Within networks of similarly minded members, the presence of moral-emotional words was especially influential in network diffusion (and less influential across in-group networks).
2. This work is grounded in social science theories - extrapolating on Public Health research on health-contagions, this study uses similar framing and language to describe the moral contagion which is also grounded in research on the emotional underpinnings of morality. This research indicates that emotional investment is a significant driver of social contagion in the domain of morality. This work can be further extended beyond moral and emotional messaging to track additional online social phenomena - I would be especially interested in investigating in-group/out-group dynamics and the development and proliferation of 'lingo' online.
3. In order to conduct this analysis, I would take a similar approach to the moral contagion paper, but implement some of the text-analysis tools (especially thinking of this paper from week 4) we've learned to identify 'up and coming' words (based on their tf-idf score in the corpus over time - a spike could indicate a new trendy term) and then track those words as they proliferate across the network. Do the terms arise in direct connections (shared edges) or tight networks (same in-group)?

[thaophuongtran]

On Deep Learning for Trust-Aware Recommendations in Social Networks Shuiguang Deng; Longtao Huang; Guandong Xu; Xindong Wu; Zhaohui Wu Publisher: IEEE https://ieeexplore.ieee.org/abstract/document/7414528

Summary (briefly summarizes the article): In the paper, the authors proposed Deep Learning Based Matrix Factorization (DLMF) approach for trust-aware recommendation in social networks. Given the extensive influx of data and information available for users, the recommender systems become a more valuable and effective mechanism to gain suggestions. However, RSs have three main problems: sparsity problem, cold-start problem, and trustworthiness problem. Even though there is a multitude of emerging solutions, the authors highlighted two shortcomings. The first one concerns initialization of the latent features of users and items relying on rather trivial mechanisms, either random or zero initialization. The second shortcoming is the failure to consider the diversity of user trust networks. With their DLMF approach, they propose a deep learning-based initialization method, where deep autoencoder is utilized to pretrain the initial values of the parameters for our learning model, and a social trust ensemble learning model to consider trusted friends’ recommendation and the community effect. In addition, the authors also provide a community detection algorithm to form community in a trust social network.

Application (suggests how its method could be used to extend social science analysis): As the paper mentioned, RSs can be seen and are utilized every where from a movie recommended by a friend to a referral for a job openings from someone in the company. Thus, this approach can be applied on a wide variety of topics in social science research. For example: using community detection to identify class segregation in social networks by their income, efficiency of RSs in the search and matching process for jobs seekers based on their networks, or how much prediction of PhD admission based on RSs and students' referee networks.

Data (describes what social data you would use to pilot such a use with enough detail that someone could move forward with implementation): I think it would be interesting to retrieve data from applicants' networks (LinkedIn) along with other information from their resumes (previous experience, education, location, skills), job postings information (to compare how well applicants' profile matches with the position), whether the applicant gets the jobs, if not how many/which rounds of interview the applicant went through. The question is to determine the importance of applicants' networks in this prediction model.

[BaotongZh]

DeepInf: Social Influence Prediction with Deep Learning Link: https://doi.org/10.1145/3219819.3220077

1) This paper introduces an end-to-end framework, DeepInf, to learn users’ latent feature representation for predicting social influence.DeepInf takes a user’s local network as the input to a graph neural network for learning her latent social representation. The authors designed strategies to incorporate both network structures and user-specific features into convolutional neural and attention networks. Extensive experiments on Open Academic Graph, Twitter, Weibo, and Digg, representing different types of social and information networks, demonstrate that the proposed end-to-end model, DeepInf, significantly outperforms traditional feature engineering-based approaches, suggesting the effectiveness of representation learning for social applications.

2) The general idea behind the proposed DeepInf can be extended to many network mining tasks. The DeepInf can effectively and efficiently summarize a local area in a network. Such summarized representations can then be fed into various downstream applications, such as link prediction, similarity search, network alignment, etc. Therefore, we would like to explore this promising direction for future work. Another exciting direction is the sampling of near neighbors.

3) This could be really helpful for our group's final project. One part of the dataset of our project is the network. This model may be designed for tackling virtual/online network data instead of the network regarding the real locations and distance between locations. Still, the problems this model and paper trying to predict can be applied to our tasks. We could use DeepInf to predict the action status of a location given the action statuses of its near neighbors and its local structural information. For example, we could input our data and train the model to predict whether a location would be gentrified or not based on its neighbors' status of gentrification. And in the paper, the task the model trying to deal with is a binary classification task, which is the same as our task: to predict whether a place gentrified.

[zihe-yan]

X. Su et al., "A Comprehensive Survey on Community Detection With Deep Learning," in IEEE Transactions on Neural Networks and Learning Systems, doi: 10.1109/TNNLS.2021.3137396.

• Summary: This article surveys methods in the deep learning field in detecting communities in social networks beyond classical spectral clustering and statistic inferences. In this article, the authors classify the techniques of detecting communities in social network analysis into deep neural networks, deep nonnegative matrix factorization, and deep sparse filtering. Besides these state-of-art methodologies, the authors also mention a set of datasets and tools that are available with open-source making the implementation process more straightforward.

• Implementation: Community detection has always been a heated topic in social network analysis. Many fascinating questions stem from this area. For example, anyone who is interested in echo chambers may want to study the formation of each echo chamber, the profile of an echo chamber; how they are different from other echo chambers. By doing a community detection and analyzing the different behaviors between web users in different echo chambers, one may be approaching the core puzzle of how the echo chambers get formed or consolidated.

• Data: To detect communities online, one must first obtain the social network data of people's followers and accounts that the user is following. If possible, some open demographic data can also be obtained such as indicated gender and location or education institutions. Secondly, if one wants to further explore the reasons that echo chambers get consolidated afterward, one should also obtain some text data from the community. By analyzing linguistic features and those texts' dissemination patterns in their own community, some interesting patterns may be detected, hence contributing to our understanding of the formation and development of echo chambers.

[ShiyangLai]

Quantifying spatial homogeneity of urban road networks via graph neural networks https://arxiv.org/pdf/2101.00307.pdf 1) In this research, the authors employed GNN to model the intra-city road network spatial homogeneity. They used 11790 urban road networks across 30 cities worldwide to measure the spatial homogeneity of road networks within each city and across different cities. They find that intra-city spatial homogeneity obtained by transferring the model across different cities, reveals the intra-city similarity of urban network structures originating in Europe, and passed on to cities in the US and Asia. Socioeconomic development and inter-city similarity revealed using this method can be leveraged to understand and transfer insights across cities. It also enables us to address urban policy challenges including network planning in rapidly urbanizing areas and combating regional inequality 2) Local network metrics such as local betweenness and local closeness inherit the same formulas from the global metric and focus on the subnetwork surrounding a specific node. However, they do not directly quantify the similarity of subnetworks due to the lack of subsequent distance functions. The concept of spatial homogeneity used in this study for similarities and the corresponding RNN-based method could be very helpful for capturing the indirect neighborhood relationships. 3) Building my personal research interests, I think this novel subnetwork similarity measurement can be further employed to understand the representation of subculture and subclass in online social media. This method can be used on analyzing direct network data, such as friends networks and follower networks, and generated networks like online comment-based language networks in social media. But other types of network data that have subclusters may also be benefited from this measurement.

[pranathiiyer]

Fake News Detection on Social Media using Geometric Deep Learning Link: https://link.springer.com/article/10.1007/s10586-017-1117-8

1. Social media has made dissemination of information extremely easy and accessible. This paper looks at dissemination of fake news and adopts a novel method to detect it. It uses geometric deep learning which naturally deals with heterogeneous data (suchas user demographic and activity, social network structure, news propagation and content), thus carrying the potential of being a unifying framework for content, social context, and propagation based approaches. The authors use an annotated dataset from twitter from 2013-2018 to identify what is fake news and what is not, and achieve an ROC score of almost 93%. The paper successfully outlines robust behavior in several challenging settings involving large-scale real data, pointing to the great potential of geometric deep learning methods for fake news detection.

2. I think this method could be extremely potent to understand spread and dissemination of information in the realm of social media, which is an extremely important part of the social sciences now. This could have wider implications in attempts to study vitality of news, and spread of information within specific clusters or echo-chambers online. Beyond this, this method could also be used to identify potential players/mediators that play a vital role in dissemination of information throughout networks.

3. I would personally like to use this to understand how information from advertisements or rather advertisements themselves move across networks. Moreover, how these ads and recommendations are influenced by different individuals in the network. While this could offer interesting insights about how some of these social media platforms operate, it might also underscore algorithmic bias.

[min-tae1]

Zhou, H., Guns, R. and Engels, T.C.E. (2021), Measurement of Interdisciplinarity: Quantifying Distance-Based Disparity Using Node2vec. Proceedings of the Association for Information Science and Technology, 58: 563-566. https://doi.org/10.1002/pra2.498

Summary Most scholars include disparity, meaning how different the categories within a system are, within their definition of interdisciplinarity. However, disparity is no easy matter in terms of measurement, as it can involve different combinations of measurement and requires a dissimilarity matrix among all disciplines. This paper aims to tackle this issue by applying Node2vec to the discipline citation network, proposing a distance-based method for the disparity. By employing embeddings vectors to measure the distance between disciplines, deep learning methods have shown less skew and more widespread distribution compared to other methods. It also could be employed to detect variations of disparity among main fields and disciplines.

Implication for Social Science Analysis It is difficult to argue with Aristotle who argues that a definition always involves a genus and a differentia. The difference, therefore, plays a crucial role in what an entity is, and social science is no exception. Institutions, knowledge, and symbols are just a few examples that manifest how difficult it is to define those without involving comparison or contrast with similar, or different, entities. Moreover, there are lots of cases where a definition of an entity could only be grasped by considering its distance from different genera. For instance, understanding an ideology of a politician requires comparison with other politicians as well as the distance from others as well. Hence, the methdology employed in the paper above could be beneficial in understanding concepts or subjects of social science analysis in a much clear manner.

Data Genre is a subject that constantly bothers cinephiles, film critics, and scholars within cinema studies. While it is true that “one can know it when one sees it,” it is difficult to know what makes it belong to a certain genre rather than another. Hence I would like to employ the distance-based method to understand the disparity of movie genres using the IMDB dataset.

[javad-e]

Examining Coivd-19 Forecasting Using Spatio-Temporal Graph Neural Networks (2020)

Kapoor et al. (2020) propose a new model for forecasting Covid-19 cases using graph neural networks and mobility data. There have been popular approaches for epidemiological modeling: 1) a mechanistic approach where parameters such as transmission rate are hardcoded into the mode; or 2) using a time series method such as autoregression. In both these approaches, the prediction only depends on information from that location and observed patterns from other locations. The authors propose taking into account people’s mobility to other regions. In this study, the researchers use Google phone GPS data to create a prediction model at the county level. The created graph has two different edge types. Each node is connected to its neighboring counties in the spatial domain and is also connected to seven prior days of that county. The proposed model outperforms the other tested temporal approaches. The authors compare the results to alternative temporal models such as ARIMA, LSTM, and Seq2Seq all of them with and without mobility data. Although GNN outperforms the tested alternatives overall, the gap is not very significant in this paper.

The proposed approach could be very helpful in answering some social science questions that require both temporal and spatial analyses. In particular, many questions in urban studies are of such a nature. Suppose we want to investigate changes in a city that depend both on geographical factors, such as connection to various parts of the city, but also on their own prior history. We could find many similar examples in other social science fields as well.

As mentioned in the example above, suppose we want to predict urban changes using street-view images, consider Naik et al. (2017) as an example. For each area, we could get time-series data of the location and combine those with the network structure of the city using a stacked model proposed by Kapoor et al.

[Yaweili19]

https://link.springer.com/chapter/10.1007/978-3-319-55696-3_5

The Tangled Program Graph (TPG) is a framework for organizing multiple teams into arbitrarily deep/wide structures through a process of continuous evolution. Benchmarking is conducted using a subset of 20 games from the Arcade Learning Environment (ALE), an Atari 2600 video game emulator. The performance of the proposed approach exceeds that of deep learning in 15 of the 20 games, with 7 of the 15 also exceeding that associated with a human level of competence.

In high-dimensional reinforcement learning challenges, the Tangled Program Graph (TPG) model is proposed for identifying deep combinations of programs that collectively create policies. Although this methodology is currently more often used in video games AI development, especially for Atari titles, similar cases may appear in social science research where TPG could play its part. For example, when deciding how multiple lawmakers create policies, or in-game theories with multiple players, etc.

A graph dataset of policies and team features are required for TPG model to compute its state and action spaces. Moreover, key to TPG is support for emergent modularity: the ability to identify decisions local to different sub-regions of the state–action space and then organize such decisions hierarchically.

[y8script]

T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction https://ieeexplore.ieee.org/document/8809901

1. This study introduced a novel neural network approach that integrates both spatial dependency and temporal dependency in networks by combining the graph convolutional network (GCN) and the gated recurrent unit (GRU). While GCN was used to capture topological patterns, GRU was applied as an additional layer to capture the dynamics of node association changes. These features enabled the model to have the capability of spatio-temporal prediction as well as the possibility of long-term prediction. which beats classical models including the GCN or GRU-only models.

2. Although traffic prediction doesn't seem to be a very social scientific question, this approach may be helpful in many situations where temporal dynamics are important for the identification of network patterns. Most of the network data can be collected and fed into the model in a sequential way rather than being static. With a sufficient amount of data, the models may identify complex temporal interaction patterns that eventually lead to higher accuracy.

3. One possible application of T-GCN may be to identify the temporal dynamics of social network user relationships with time-series data about interpersonal connection. We can sample our snapshot of the user network at a fixed time-lapse(e.g. one week) and obtain a temporally detailed user network dataset. Then we can implement T-GCN on the dataset to see whether we can predict how user interaction changes across time periods.

[yujing-syj]

T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction

https://ieeexplore.ieee.org/abstract/document/8809901?casa_token=zz4CHIjT9lIAAAAA:p-F3PzZ0sfbVeyhcFrwQH3m2nNlC25lxqpEgKgdxjM04Ie3eE9Ur8Iy9RRJvD--9Fn2VgyQX

1. The paper develops a novel neural network based approach for traffic forecasting called T-GCN, which combines the GCN and the GRU. The researchers use a graph network to model the urban road network in which the nodes on the graph represent roads, the edges represent the connection relationships between roads, and the traffic information on the roads is described as the attribute of the nodes on the graph. Eventually the T-GCN model is used to tackle spatio-temporal traffic forecasting tasks. When evaluated on two real-world traffic datasets and compared with the HA model, the ARIMA model, the SVR model, the GCN model, and the GRU model, the T-GCN model achieves a better performance under different prediction horizons.

2. For the T-GCN method introduced in this paper, it could be a very good reference to develop other related models or be used to analyze the city network problems. First, the GCN is used to capture the spatial topological structure of the graph for obtaining the spatial dependence. Second, the GRU model is introduced to capture the dynamic change of node attribute for obtaining the temporal dependence. The combination of these two methods helps the researchers to solve the problems of only analyzing the static data and features. The T-GCN model successfully captures the spatial and temporal features from traffic data so that can be applied to other spatio-temporal tasks.

3. I could also use this model to analyze many urban issues. Besides congestion, the same model could be applied to analyze and predict the crime accidence. With the network data of where the crime is happened and the feature around the spot, we could build a network. Also, since the crime occurs in different time during one day, we will also add the time variation into the model. The model could also be performed to predict the change of some networks (eg: school, hospital, CBD etc.) in the city.

[chentian418]

Shared analyst coverage: Unifying momentum spillover effects https://www.sciencedirect.com/science/article/pii/S0304405X19302533?casa_token=Qu3mzL7mCBQAAAAA:pE1TpVvNN9X9bnQxKgo-soSS2uYrPqeLi7opyGxtIvN2ObzXo-UpJIP8WTelZQKpwp-wmo71

1. The study find that a connected-firm (CF) momentum factor generates a monthly alpha of 1.68% (t = 9.67) by identifying firm connections by shared analyst coverage. In spanning regressions, the alphas of industry, geographic, customer, customer/supplier industry, single- to multi-segment, and technology momentum factors are insignificant/negative after controlling for CF momentum. Similar results hold in cross-sectional regressions and in developed international markets. Sell-side analysts incorporate news about linked firms sluggishly. These effects are stronger for complex and indirect linkages. Consistent with limited investor attention, these results indicate that momentum momentum spillover effects are a unified phenomenon that is captured by shared analyst coverage.
2. We could extend the work by building network between managers, analysts and firms; quantify relationship of managers through analysts as a function of texts and measure impact on stock returns from events of adjacent firms. This relies on neural network to help cluster and classify the nodes and predict future networks.
3. By using conference call data, I can use the analyst coverage data through conference call transcripts with different companies to classify nodes in an identified network, and study the effects of shocks coming from adjacent companies to a company in the network.

[yhchou0904]

Poincaré Embeddings for Learning Hierarchical Representations

1. There are lots of representations for graph data or say network data, and usual embedding representation methods could not account for latent hierarchical structures. In this paper, the writers develop a method of embedding items into a hyperbolic space (or we can say an n-dimensional Poincare ball.) In space, the distance would increase exponentially as the points are closer to the boundary. By embedding into hyperbolic space, the representations could capture hierarchy through the norm of embedding and similarity through the distance between embeddings at the same time, which could not be satisfied by some elder methods. To learn the embedding, this method applies Riemannian optimization, which could be easily parallelizable and scalable.
2. With this Poincare embedding, we can have a better performance than some of the other classical embedding methods. For example, when doing embedding taxonomies, we could not only have a good ability to reconstruct the items from embedding, but also can do well in link prediction tasks. Also, when doing social network embedding through this method, the Poincare embedding could do better than Euclidean embedding. At the same time, the Poincare embedding allows us to make graded assertions about hierarchical relationships, which corresponds well to underlying semantics.
3. The social science data we could apply this data includes any kind of network data that has some hierarchical relationships. For example, we can find this kind of relationship between users in the same network. Also maybe we could build a network by calculating people's interaction amount and then embed it with Poincare embedding.

[ValAlvernUChic]

Link: https://www.cs.cmu.edu/~ylataus/files/ScholandTausczikPennebaker.pdf

1. In social language network analysis, the authors used natural language processing to identify social relationships using language, something that often offers clues to the latent or unrecognized relationships between individuals. To do this, the authors employ three steps of methods. Firstly, they link one or more dyadic pairs in the group. Then, they convert the text associated with particular links to specific psychological, social, or emotional theory. Lastly, they use these metric matrices in a graph-processing algorithm to compute an objective of interest.

2. While this study focuses on a corporate discussion archive, I imagine that this study could extend to other studies that focus on interactions of power between employer and employee over time. For example, can we temporally track the extent to which an immigrant employee assimilates into his/her environment just through the linguistic cues that are given in her text messages or social media posts. This could potentially give insight into how immigrants assimilate into novel and alien environments.

3. I would try to collect text messages between foreign domestic workers and their employers and even between other FDWs and tag them temporally before mapping how their linguistic cues influence one another over time.

[Emily-fyeh]

Wang, S., Tang, J., Aggarwal, C., Chang, Y., & Liu, H. (2017, June). Signed network embedding in social media. In Proceedings of the 2017 SIAM international conference on data mining (pp. 327-335). Society for Industrial and Applied Mathematics.

1. This paper includes negative links in the network embedding application on social media data. The author proposed a deep learning framework SiNE (Signed Network Embedding) to optimize an objective function guided by social theories that provide a fundamental understanding of signed social networks. The structural balance theory proposes that the signed network structures should let the links between friends closer than the links between foes. The article also presents the experiments on two real-world datasets of social media, and the results demonstrate the effectiveness of the proposed SiNE framework–the model outperforms other existing baselines on the signed network prediction task.
2. As the paper mentioned, the properties of real-world social media networks can resonate with the structural balance theory: the density of the positive links is much higher than those of the negative links, and the cost of forming negative relationships is higher than that of the positive ones. The Epinions review platform and Slashdot the tech news site are the subjects of experimentation in the paper, in both of them, the negative connection can be operationalized. Thus I think the SiNE framework can be applied to sites like StackOverflow or even Facebook, where researchers can define upvotes/downvotes and emoji reactions as positive. and negative relationships.
3. I would like to try the framework in my Twitter research, where tagging other users with opposing comments can be operationalized as a negative relationship. Therefore, SiNE can help with the prediction of how and where the dispute and debate will happen within the sphere of a controversial topic.

[hsinkengling]

Representation Learning on Spatial Networks https://papers.nips.cc/paper/2021/file/12e35d9186dd72fe62fd039385890b9c-Paper.pdf

1. The article proposes an algorithm that learns network representations from spatial networks called spatial graph message passing neural network (SGMP). This algorithm learns representations from sampling random trees within spatial networks. One of the main advantages of this algorithm is computational efficiency and immunity to distortion (rotation-invariance, translation-invariance, and information lossless).

2. Given spatial data of a field (eg. relative position of shops, town halls, parks, schools of small towns), this method could be used to study urban and organizational structure. Given enough sample, one could train a neural network to identify network properties that correlate with a number of health and education outcomes.

3. With google maps data, one could harvest network data of small towns/suburban neighborhoods. One could match this with federal data on health, education, crime etc. to train a prediction deep learning algorithm.

[Hongkai040]

Node copying: A random graph model for effective graph sampling https://www.sciencedirect.com/science/article/pii/S0165168421003728

This paper proposed a promising approach of sampling networks while maintaining the attributes of the networks. Unfortunately, Uchicago didn’t buy this paper. So you may need to use other sources.

In this paper, the authors proposed an effective approach of sampling graph using a generative random graph model. Other related works mainly parametric random graph models or learning-based models for sampling. However, parametric random graph models sometimes cannot model some real-word networks attributes like large clus- tering coefficients [21], small world connectivity and exponential degree distributions, and they could be computational expensive for very large graphs.On the other hand, Learning-based models usually can accurately predict, say, links in a network, but tend to fail to reproduce global structural properties. The model proposed by the authors is a learning-based model that could maintain some global structural properties. The model, the node copying model, is based on the assumption that ‘similar nodes should have similar neighborhoods’. So the model randomly replaces the edges of a node with those edges from a randomly chosen similar node then sample from this generated graph. Hence the node copying model preserves the homophily nature of graphs and the cost of sampling scales linearly with the number of nodes.

This approach could be used first used to training models for node classification using smaller graph. The authors say that an ensemble of those graph samples could be integrated into a Bayesian framework for better prediction. The model could also be used to mitigate the negative influence of adversarial attacks on the graph topology by replacing edges. I think this method could also be used to exclude outliers in a graph to make better inferences of the properties of the graph.

Possible application: Using tweets and user profile from twitter to construct user network. ‘Tweet’ and ‘Retweet’ construct a network among users. However, it may be time-consuming if we want to study the dynamics of the network using full dataset, and we may be influenced by the noisy from the data to make some judgements. Hence, we can use the node copying model for sampling and instead studying dynamics of the network on the ensemble of graph samplings to start on. After having some preliminary findings, we can conduct the same analysis on full scale dataset.

[anhdao21]

"Multi-GCN: Graph Convolutional Networks for Multi-View Networks, with Applications to Global Poverty" https://ojs.aaai.org/index.php/AAAI/article/view/3836

1. In this paper, the author develops a GCN on multi-view networks for poverty analyses and shows that it outperforms both shallow (DeepWalk, Node2vec, LINE) and deep (single-view GCN) graph models on three common classification tasks using mobile phone data from developing countries - predicting poverty, product adoption, and gender prediction. The main motivation for building a multi-view GCN is that prior approaches mostly rely on binary relations when constructing social networks and ignore the fact that people are related in myriad ways. The model is constructed in three main steps - merge different views of the same graph using subspace analysis, prune these merged graphs using manifold ranking to highlight the most informative sub-component, and apply a GCN on the pruned graphs for semi-supervised node classification. The results are encouraging: multi-view GCN achieves 10-20% gains in prediction accuracy over shallow models across all three tasks, and 3-5% over single-view GCN for poverty and gender prediction (only small improvements for product adoption).

2. With the rise of smartphone adoption in developing countries, there's vast potential to leverage this approach and build meaningful multi-view networks based on digital trace data for economic development analysis. In the paper, the author relies on phone records to extract views related to phone calls as well as text messages between users while simultaneously building descriptive features for each user (e.g. total call volume and degree of centrality). I can see similar applications using other data sources with the same network-like property (e.g. social network activity, community transaction data), or focusing on different prediction tasks (e.g. predicting creditworthiness among the unbanked population to extend access to financial resources, predicting the vaccination status of children based on parents' communication for targeted vaccination campaign)

3. One interesting application is to predict the creditworthiness of people with no bank accounts. We could use different data sources (e.g. cellphone records, transaction records with local businesses, or social media interactions) to construct multiple views between nodes and weigh more heavily towards views directly related to payments. The major difficulty would lie in collecting and labeling part of the data for a semi-supervised approach. One way to alleviate this is to partner with government enterprises that provide essential goods like utilities or healthcare to obtain anonymous data in a way that protects user privacy.

[sudhamshow]

Graph Convolutional Neural Networks for Web-Scale Recommender Systems

1. The authors propose a data efficient graph convolutional network algorithm for Pinterest that combines efficient random walks and graph convolutions to generate embeddings of nodes that incorporate both graph structure as well as node feature information. The core idea behind GCNs is to learn how to iteratively aggregate feature information from local graph neighbourhoods using neural networks. Unlike purely content-based deep models (RNNs), GCNs leverage both content information as well as graph structure. Although there was much literature about the potential application of GCNs in recommender systems, there were several pragmatic limitations to using it in the real world (computation of the entire graph laplacian, which is infeasible considering the billions of nodes on a social media platform). The authors propose an efficient way of scaled-training (using on the fly convolutions, convolutions using random walks), as well as inference (using map-reduce) of GCN based node embeddings in the paper.
2. Formulated differently, social scientists, advertisers, computer scientists studying recommender systems are all trying to solve similar problems. An interesting phenomenon social scientists would be interested in, is to study how different politicians, businessmen and influencers speak on different platforms (circumstances). Conditioning on ideology and party affiliation (or similar traits for businessmen and influencers), and convolving over then connections (increasing confidence of the observed phenomenon) , one would be able to assess which politicians are similar (albeit being from different geography, speaking different languages) by virtue of choice of words, fervour in the speech made, tone, actions made during speech - in different instances (running for a campaign, addressing the nation in grief etc.). One example of this application would be to study which leaders are becoming more nationalist and which ones are embracing mutual growth by implicit features of the speech (multimodal) that can't be caught without thorough analysis.
3. All data available of a speech has potential for training - audio recording, video recording, photos in newspapers/articles, and even audience reaction. One could collect all information available for different people, model the people database as a graph (using one's own discretion on what constitutes a link) and use the methods described in the paper to model people as vectors using all the multi-modal data collected.