Our Project

General Network measures

By plotting on a log-log scale, the review length distribution plots for random and friend network above show that both random and friend network are not Scale-free networks. However, regarding the random network, the regular plot shows that the degree will decrease whilst the number of nodes is decreasing. It is quite surprising that one fifth of the total restaurants have the degree around 500. From personal perspective, I assume that there are mainly two groups of people, and for each group, the people inside have lots of visited restaurants in common, Maybe it is the reason why the degree for these nodes is only around half of the highest degree.

Network plots

In this section we want to explore more differences between random network and friend network,by using node size to its degree, to eigenvector and to betweeness centrality.

The biggest one is node size porportional to its degree, the top right hand side one is the node size proportional to eigenvector centrality whilst the bottom right hand sidre one is node size proportional to betweeness centrality. The color for each node is based on its review count. The higher review count a restaurant has, the darker color it will have. The threshold are 25,50,75,100,150,200,400 and 600, based on the analysis from review count distribution we did secretly. In random network, there are lots of nodes in the centre of the network that have high view count. Obviously, they also have high degree, as we can see from their dark blue colors. A high eigenvector centrality for a given node, means that the node itself is connected to a lot of nodes with high scores. However, since there is not that big difference in terms of the node size for each node in this network, we can say that the nodes are connected without that much preference(since they are in random network and they are not person!). Since betweenness is based on how many times a node appear in a shortest path, the above result probably means that the restaurant nodes connect to each other a lot but is less connected to the rest of the network. From the plot for random network, we can infer that the higher review count a restaurant has, a higher betweenness centrality it will has.

In the left big plot for friend network, node size is proportional to its degree, since the number of restaurants and users are smaller than random network, the size of each node is also smaller than those in random network. It is quite obvious that the nodes are divided into three clusters. Probably it means there are three main groups of restaurants in which people usually visit? We will see a clearer answer in the community detection below. It is even more obvious that the nodes in friend network are divided into three big groups. The plot for eigenvector in friend network differs from the one in random network. The nodes with the biggest size are clustered together, middel size nodes are together and smallest nodes get together. We can infer that the people in friend network have their preference. They will visit the restaurant which are connected to a lot of other restaurants(nodes) with high degrees. Regarding the betweeness plot in friend network, it is quite obvious that some of the nodes in it is much bigger than the nodes in random network. It probably means that there is a group of restaurant nodes connect to each other a lot and also connect to the rest of the network.

Coummnity detection

In our network, each node represents a restaurant, a user can review several different restaurants and a restaurant can be reviewed by multiple users. For a restaurant, the more reviews it has, the higher degree it will have, the size of the node will become bigger.

As we can see from the community detection in random network, since it is called 'random', the community inside is quite trivial. There are lots of communities in this network. One thing that attracts me is the different colors(different communities) are located in 6 directions. I assume that it is based on different areas in Charlotte. People in different area are prone to visit different restaurants in different area

Compared with the community detection in random network above, the communities in friend network is more separate. Based on the analysis before, I can also assume that the different communities are divided by group of people or different area in Charlotte. Anyway, either different group of people or different areas is all related to people's choices, which is also the key proposition in our project.

VADER

The VADER or ‘Valence Aware Dictionary and sentiment Reasoner’ method uses a sentiment and rule-based lexicon. Doing a VADER sentiment analysis on text results in a positive, negative, neutral, and compound score for the text piece. The compound score is the sum of the other three and gives an indication of the general attitude of the text. The score ranges from a negative response = -1 to a positive one = 1. The VADER scores are scaled so a score between 0.5 and 1 is considered positive, a score between -0.5 to 0.5 is considered neutral, and a score below -0.5 is considered negative (Zhou 2019)

We have chosen the top 3 degree highest restaurants, which means this 3 restaurant will have higher chance to be recommended. For higher rating reviews(4 and 5), these reviews are mainly gathering in positive area, but still have some five- star reviews have negative text. For the reviews left by patrons giving 4 or 5 ratings we clearly see that the majority of the reviews have positively loaded sentiment scores as evident by the lower quartile being well above the 0.5 mark.

It is surprising to see that in friendship network top 3 highest degree restaurants, there are no one - star reviews. It can be understood that if a user has been recommended a restaurant by a friend, which means even the restaurant has short length of reviews still have good reputation and high rates, which also means the recommendation list not only recommend most popular restaurant but also recommend less lower rating but from your friends have visited. There is only in total 25 pieces of reviews of top 3 highest degree restaurants, so it is insufficient to prove the correlation between each pair of features.