Advanced Machine Learning, Data Mining, and Online Advertising Services
In this blog, I explain how social media marketer can improve their marketing strategies and maximize the number of audiences that they potentially can reach by choosing optimal tweeting time.
One of the most interesting areas in sociology is studying the process by which a new idea or opinion propagates through a social network. Diffusion of an idea is a socioeconomic problem. Viral marketing is an example where economists try to find fast and efficient ways to advertise a new product by using the pre-existing social networks. Spreading an idea through a social network has been studied thoroughly by Everett Rogers in his famous book "Diffusion of Innovations" . Rogers has defined diffusion as the process by which a new idea is communicated through certain channels, e.g. word-of-mouth, among social network members over time. He has also identified four basic elements that are crucial for diffusion process as: an innovation, communication channels, time, and social network.
According to the theory of diffusion of innovations, first of all, an innovation is more likely adopted by people who are seeking new ideas. These people who Rogers calls them Innovators are usually university-level educated, and they tend to try new ideas in advance. If Innovators like the innovation, they tend to adopt it, and they are more likely to introduce the innovation to their friends through communication channels. Second, another group of people who are called Early Adopters are strongly connected to Innovators in the social network and follow Innovators’ opinions. The new opinion then spreads to other less influential groups that Rogers calls the Early Majority, Late Majority, and Laggards, respectively. Laggards who are the last adopters usually hesitate to accept the new idea unless they want to be compatible with their peers.
Based on Rogers’ and other scholars’ works on diffusion of innovation, scientists have found that the cumulative distribution of people who have adopted an innovation has an S-shaped curve over time. The reason that researchers can observe an S-shaped curve for diffusion process, which takes off at some point, is that Laggards, Late Majority, Early Majority, Early Adopters, and Innovators follow one another in a domino-like effect. Rogers and other researchers have also found a strong relationship between diffusion process and social network structure. In other words, the underlying network structure plays an important role on how far an opinion can go and how long it takes for an innovation to be spread out .
Suppose you're a social media marketer promoting a new technology and you want to use twitter to reach your target audience and tell them about the new technology. In this scenario, the innovation is the tweet content that you want to post and the communication channels and social network are governed by Twitter website. However, one interesting question is to find the optimal tweeting time that maximizes the spread of the tweet.
In this article, we ignore the impact of other playing parameters and only focus on tweeting time. To measure the spread of a tweet, we use two metrics. First one is the number of favorites a tweet receives and the second metric is the number of retweets a tweet receives. Readers should note that both favouriting and retweeting a tweet shows that a user agrees to or approves the idea! One can argue that "retweeting" a tweet implies a stronger influence than "favouriting" a tweet.
For finding optimal tweeting time, I crawled my personal twitter account (@kjahanbakhsh) and collected my last 200 posted tweets by August 20th 2014. My tweets are limited to two months: July and August. As a side note, most of my tweets are related to new IT-related technologies. See below for a few tweet examples:
|Tweeting Time||Tweet Content|
|Thu 2014-07-31 19:30:39||Salesforce Buys Big Data Startup RelateIQ For Up To $390M http://t.co/CA9Dhdkuk3 via @techcrunch|
|Fri 2014-07-25 19:31:47||Why Python uses 0-based indexing? http://t.co/qjHKyNppTk #python|
|Fri 2014-07-18 19:45:13||10 Tips for Better Deep Learning Models by @lauradhamilton http://t.co/WDOJ94MFQv|
|Mon 2014-07-14 19:20:54||Defer, Panic, and Recover in #Golang http://t.co/ZazkufOCfk|
We extracted all tweets with their metadata such as time of tweet, number of retweets, and number of favorites each tweet has received. First, we plot the distribution for number of posted tweets vs the tweeting time hour in local timezone (24-hour range). As you see in the following figure, most of my tweeted are posted at nights during 11:00pm - 12:00am interval (22 tweets) and 12:00pm - 13:00pm (21 tweets) time range. This is because I usually get free time to browse web & read interesting articles late at nights and around noons!
Second graph shows the number of retweets my tweets have received as a function of tweeting time (hour). Again, tweets posted in 11:00pm - 12:00am (4 tweets) and 12pm - 13:00pm (3 retweets) time ranges have received the maximum number of retweets.
Finanlly, the third graph shows the number of favorites that my tweets have received as a function of tweeting time (hour). Tweets posted in 11:00pm - 12:00am (7 favourites) and 12:00pm - 13:00pm (11 favourites) time intervals have received the maximum number of favourites.
We have summazrized the numbers for late night and morning tweeting times in the following table for better comparison.
Based on above numbers, we can say that best time intervals for tweeting at nights are 8:00-9:00pm, 11:00-12:00pm, and midnight. It seems for "me" midnight produced the best results. This can be because people see these tweets on the top of their tweet stream when they wake up in the mornings. So, they're more likely to favourite/retweet them if find them interesting. For mornings, it seems 9:00-10:00am, 11:00-12:00pm, and 12:00-13:00pm are effective intervals. Also, for "me" 12:00-13:00pm provided the maximum spread!
We should be careful while generalizing above observations. First, above observations are based on a limited number of tweets (i.e. 200 tweets) for one Twitter tech account. So, statistical significance is a concern. We cannot gaurantee all other playing parameters have uniform distribution in above analysis. In other words, other factors such as day of the week or content of tweet might have had impacts on above observations. To address this issue we need to either control all other parameters and only change tweeting time.