If you're new here, you may want to subscribe to my RSS feed. Thanks for visiting!

Viral marketing is a great way to exploit existing social networks for spreading information about your product/service. Although one can understand the positive effects of viral marketing, an exhaustive examination would give us a deeper insight and a greater understanding of how viral marketing works. Fortunately Jure Leskovec, Lada Adamic, and Bernardo Huberman have written a detailed analysis of viral marketing regarding its inner workings. What characteristics of products/services are most effective? What inappropriate uses of viral marketing can be “counterproductive”? How do we measure how influential person-to-person recommendations are over a wide-range of products? Their report gives invaluable information regarding viral marketing and its journey through the intricate structure of social networks with 4 million people, 16 million recommendations, and half a million products.


image credit: www.optimizeandprophesize.com

Case Study Overview

For one of the first times, LAH’s (Leskove, Adamic, and Huberman) study directly observes the movement for hundreds of thousands of products, and the effectiveness of word-of-mouth advertising through it. A total of 15,646,1212 recommendations were made by 3,943,084 users on 548,523 products (99% of these products falling into the following categories book, music, videos or DVDs). Their analysis is centered around a recommendation referal program from a large retailer with the following Criteria:

• Every time a person purchases any item, he/she is given the option of recommending the item to friends through email with a 10% discount.
• The sender also receives 10% credit on their purchase.
• Only the first person to respond to the email by purchasing the following item will receive the discount.

Limitations:

• The observation of the recipient purchasing the product through the same vendor is the only indication of a “successful” recommendation
  • There is no way to know whether the person had decided to purchase the item elsewhere.
• Delivery of recommendation is through email.
  • It’s not quite the same as one person telling another about a product in casual discussions and other similar contexts.
  • Someone reading the email may consider it as spam, or less important than a conversational recommendation.
  • The recipient of the email may consider the friend is just trying to get a discount for themselves than actually recommending it.

While this analysis does have its limitations, it gives us an accurate and detailed view of recommendation dynamics however. The results will reveal to us a greater understanding how influential social networks are in purchasing decisions.

Results

A great amount of information can be obtained from the data of LAH’s results. Their analysis provides several graphs and tables showing different statistics of social network’s influence on product recommendations. I will only highlight the key figures that can give you a broad picture of their analysis. For a deeper understanding, feel free to view the whole report here.

Terms:

• Node: Represents a single person.
• Edge (i, j, p, t): “Indicates that i recommended product p to customer j at time t“[LAH].
• Buy-edge: An edge that points to a node which acted on a recommendation, but did not receive the discount.
• Buy-bit: Flag that indicates a purchase.
• Largest Connect Component: Represents the biggest network of nodes connected by recommendations.

Table 1: Sizes of the Main Product Group Recommendation Networks:

Key Points:

• DVDs are the smallest of the other groups in terms of the number of products, yet it accounts for over half of the total recommendations.
  • The DVD group has about 10 recommendations per node (while books and music have about 2 per node).
• Music recommendations closely followed DVDs in terms of the number of nodes, yet Music only need 1,443,847 recommendations while DVDs used 8,180,393. That’s roughly 5 times fewer than DVDs.
• The number of “unique” edges (recommendations to unique, or different, nodes) is relatively small.
• With total number of recommendations (r) and purchases (the b’s added together), we can do some estimations on the number of recommendations needed to make a new purchase.
  • For books, one out of 69 recommendations resulted in a purchase.
  • DVDs on average needed 108 recommendations for a purchase.
  • For music, we saw 136 recommendations per purchase.
  • Videos needed the most recommendations with 203.

Table 2: Results of the Largest Connected Component for Each Product Group

Interesting data. This is just one of many graphs that details the largest component, but it is the key graph indeed. We can see in this table that all the largest connected components are relatively small compared to their respective networks as a whole. As stated in the analysis, “one would expect that a fraction of the recommendations in the largest component to be proportional to its size”…but this is not the case. This can be shown in the results for DVDs. 84.3% of recommendations are within the largest component for DVDs, yet it only consists of 4.9% of the nodes (a tiny portion of users generated the most of its recommendations). Looking at the bigger picture, the largest component in its entirety contains only 2.54% of the nodes yet has 52.9% of all the recommendations! Key point: The way the largest component behaves is much different from the rest of the network.

Success of Recommendations

Again, this is quite interesting. As one would assume, there is a fine line between recommending too much and too little. If one is too selective and recommends only to few, chances of success will be small. If one recommends too many people and spamming them, success will also be slim. Recommending to the relevant subset of people will be most effective in getting successful buys. The top row of figure 10 shows us when saturation occurs. For books, music, and videos, saturation is apparent when about 10 recommendations are reached. At this point, purchases begin to slow down or decrease. DVDs however, show a different behavior with a steady increase of purchases throughout the graph. LAH’s analysis brings up a good point…that the results become even more interesting when you realize that the receiver of recommendations likely does not know the frequency of other people who have also been recommended the same product. This brings up the question, are there dependencies between the characteristics of the products and the person who recommends the products exhibited through the number of recommendations sent? Some explanations one could consider that LAH’s analysis mentions:

• Widely recommended products are not suitable for viral marketing.
• The recommender didn’t put much thought into the selection of people to recommend.
• People soon start to ignore mass recommenders.

The Timing of Recommendations and Purchases

The mechanism of this referral program encourages people to buy the recommended product as soon as possible in order to maximize the opportunity to receive the discount. In the analysis, we are given several percentages as to the times that certain products are purchased:

• DVDs: 16% of purchases occur more than a week after the last recommendation, 23% of purchases are made within the day of recommendation, and 78% of recommended purchases did not include a discount.
• Books: 10% of purchases occur more than a week after the last recommendation as well within the day of recommendation, and only 21% of recommended purchases did not include a discount.

The last parts of the bullets make sense due to the fact that DVDs have a higher frequency of recommendations and thus, there are much more people who will purchase the product too late and not receive the discount.

When looking at the graph dealing with the purchases and the time of day, we see that recommendations and purchases graph show a similar pattern. Looking at these graphs, we can conclude that recommendations did not greatly effect the amount of purchases. The purchases with discounts, however, show an inverse pattern when one would assume it to look like the other two graphs. What does this mean? Most of the discounted purchases occured when the number of of purchases were small, meaning that these discounted purchases happened in the morning when traffic on the website was relatively low. When you think about it, the graphs actually make sense. Most recommendations happened during the day, therefore if a person wanted to optimize their potential in receiving the discount (since only the first person from the same set of recommendations can obtain it), it would be in their best interest to purchase when traffic on the website was at its lowest.

What Effects do Communities of Interest have on Recommendations?

LAH’s analysis goes into great detail as to what effect specific communities have on recommendations. A community discovery algorithm was applied to detect communities which exchange recommendations amongst themselves and which products they prefer. As one would expect, certain communities would be more effective than others and certain products would be preferred over the other. We can see an example below:

A greater discussion can be found in the actual analysis, but from a broad perspective, LAH states that communities were usually centered around a particular product (books, music, or DVDs). Nearly all, however, shared recommendations for all types.

Observations/Conclusion

I’ve really only scratched the surface of this analysis. From just looking at the analysis in general, we can see that DVDs are purchased in higher frequencies as well as having more recommendations. On the other hand, books are recommended less. A plausible reason could be simply that books take a longer investment in time than a DVD. While a DVD can be viewed in a few hours, a book may take days and even several weeks/months. Another reason that LAH’s analysis discusses is that DVDs are more heavily advertised than books, thus the factor of how informed a customer is on a product can be put examined. Referral websites is an external factor to consider which may have had some effect on the results. These websites allow people who want to purchase a DVD and receive the discount, to ask for recommendations and thus eliminating people to really “know” each other.

LAH’s analysis gives us incredible insight on how viral marketing works which question what we would normally assume in this kind of epidemic propagation (individuals have an equal chance in becoming infected every time an interaction occurs which is not the case in viral marketing…in fact, the probability of infection decreases with repeated interaction when the saturation point is reached). Thus, marketers should be warned that providing too many incentives for customers in order for them to recommend your product could actually cause a backlash and cause the exact opposite as well as losing some credibility. Found within this report are limits to how influential “high-degree” nodes are in the network. This limit is related to the saturation point. When this limit is reached, the success per recommendation decreases, indicating that a person will have influence over a few of their friends, but not everyone they know. This also correlates to the next subject which is that smaller and more tightly knit groups tend to be much more effective in viral marketing.

An interesting idea would be to expand this analysis, but using social media networks, such as StumbleUpon, instead of emails to spread recommendations. Some questions instantly come to my mind:

• Which types of products/services/information would be most “marketable”?
• How difficult would it be in tracking recommendations and “successful results”?
• What would be considered a “successful result”?
• How much of a similarity would there be compared to this analysis?

Although viral marketing is a great technique in advertising/marketing today, the results shown in LAH’s analysis indicate that it’s not as effective as one would expect in general. To effectively grasp viral marketing, one has to understand various key elements, such as the topology and interest of the social network at hand. LAH’s study gives us detailed knowledge about this and a greater understanding of the intricacies of viral marketing.

(All graphs/tables and quotes are from the analysis of Leskovec, Admaic, and Huberman’s paper.)