In this blog post I am going to talk about the paper called “Recommendation Systems with Purchase Data” by Anand V. Bodapati (2008).

The majority of firms selling online utilize recommendation systems, which are a decision tool that tries to identify products that the customer is likely to buy, if the product is brought to their attention. They are usually based on analysing the previous purchase behaviour of the user and showing him/her the products that are the most likely to be bought. However, what such a mechanism does not account for is the fact that customers are likely to buy these products even without an explicit recommendation. Therefore, a recommendation may have a higher value if it suggests a product that the customer would not buy unless it is recommended.  Following on that, the main proposition of the author is that recommendation systems should be based not on purchase probabilities but on the sensitivity of the purchase probabilities to the recommendation action.

The methodology of this academic article is highly technical and quantitative. Firstly, the author builds an econometric model that incorporates the role of the recommendation actions of the firm. Secondly, the proposed model is empirically tested using the purchase data of a real life e-commerce company. Then the performance of the proposed model is compared with the one of other benchmark models. The results indicate the superiority of the proposed model.

One of the biggest theoretical contributions of this article is that it suggests the idea that purchasing can be seen as the outcome of two separate factors – awareness and satisfaction. Awareness (A) can be defined as becoming aware of the product and its characteristics. At the same time, Satisfaction (S) involves evaluating the product and buying it, if its utility exceeds a certain threshold level. A consumer buys a product if both A and S occur. The consumer will not buy anything of he/she is unaware of the product, or he/she is aware but decided that the utility is not high enough. Furthermore, the next contribution of the article is showing that these two events can be separated and identified using existing datasets that companies have. These datasets are information about self-initiated purchase data and recommendation response data. Using all of that information, the research paper proposes a decision framework for recommendation system use.

The econometric model that is developed tells firms at what time they should show a product recommendation to the customer in order to optimize the beneficial outcome for the company. The equation takes into consideration the purchasing probability, the timing of the recommendation and the two factors affecting the purchasing decision – awareness (A) and satisfaction (S). The main idea is that recommendations should not increase the purchase probability for products that the customer is already willing to buy, but rather increase the incremental revenue for the company by suggesting items that are otherwise less likely to be bought.

Despite the few limitations of the research, it makes a valuable theoretical and practical contribution.